Re-introduce the `llama` package (#5034)

* Re-introduce the llama package This PR brings back the llama package, making it possible to call llama.cpp and ggml APIs from Go directly via CGo. This has a few advantages: - C APIs can be called directly from Go without needing to use the previous "server" REST API - On macOS and for CPU builds on Linux and Windows, Ollama can be built without a go generate ./... step, making it easy to get up and running to hack on parts of Ollama that don't require fast inference - Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners takes <5 min on a fast CPU) - No git submodule making it easier to clone and build from source This is a big PR, but much of it is vendor code except for: - llama.go CGo bindings - example/: a simple example of running inference - runner/: a subprocess server designed to replace the llm/ext_server package - Makefile an as minimal as possible Makefile to build the runner package for different targets (cpu, avx, avx2, cuda, rocm) Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Daniel Hiltgen <daniel@ollama.com> * cache: Clear old KV cache entries when evicting a slot When forking a cache entry, if no empty slots are available we evict the least recently used one and copy over the KV entries from the closest match. However, this copy does not overwrite existing values but only adds new ones. Therefore, we need to clear the old slot first. This change fixes two issues: - The KV cache fills up and runs out of space even though we think we are managing it correctly - Performance gets worse over time as we use new cache entries that are not hot in the processor caches * doc: explain golang objc linker warning (#6830) * llama: gather transitive dependencies for rocm for dist packaging (#6848) * Refine go server makefiles to be more DRY (#6924) This breaks up the monolithic Makefile for the Go based runners into a set of utility files as well as recursive Makefiles for the runners. Files starting with the name "Makefile" are buildable, while files that end with ".make" are utilities to include in other Makefiles. This reduces the amount of nearly identical targets and helps set a pattern for future community contributions for new GPU runner architectures. When we are ready to switch over to the Go runners, these files should move to the top of the repo, and we should add targets for the main CLI, as well as a helper "install" (put all the built binaries on the local system in a runnable state) and "dist" target (generate the various tar/zip files for distribution) for local developer use. * llama: don't create extraneous directories (#6988) * llama: Exercise the new build in CI (#6989) Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet. * llama: Refine developer docs for Go server (#6842) This enhances the documentation for development focusing on the new Go server. After we complete the transition further doc refinements can remove the "transition" discussion. * runner.go: Allocate batches for all sequences during init We should tell the model that we could have full batches for all sequences. We already do this when we allocate the batches but it was missed during initialization. * llama.go: Don't return nil from Tokenize on zero length input Potentially receiving nil in a non-error condition is surprising to most callers - it's better to return an empty slice. * runner.go: Remove stop tokens from cache If the last token is EOG then we don't return this and it isn't present in the cache (because it was never submitted to Decode). This works well for extending the cache entry with a new sequence. However, for multi-token stop sequences, we won't return any of the tokens but all but the last one will be in the cache. This means when the conversation continues the cache will contain tokens that don't overlap with the new prompt. This works (we will pick up the portion where there is overlap) but it causes unnecessary cache thrashing because we will fork the original cache entry as it is not a perfect match. By trimming the cache to the tokens that we actually return this issue can be avoided. * runner.go: Simplify flushing of pending tokens * runner.go: Update TODOs * runner.go: Don't panic when processing sequences If there is an error processing a sequence, we should return a clean HTTP error back to Ollama rather than panicing. This will make us more resilient to transient failures. Panics can still occur during startup as there is no way to serve requests if that fails. Co-authored-by: jmorganca <jmorganca@gmail.com> * runner.go: More accurately capture timings Currently prompt processing time doesn't capture the that it takes to tokenize the input, only decoding time. We should capture the full process to more accurately reflect reality. This is especially true once we start processing images where the initial processing can take significant time. This is also more consistent with the existing C++ runner. * runner.go: Support for vision models In addition to bringing feature parity with the C++ runner, this also incorporates several improvements: - Cache prompting works with images, avoiding the need to re-decode embeddings for every message in a conversation - Parallelism is supported, avoiding the need to restrict to one sequence at a time. (Though for now Ollama will not schedule them while we might need to fall back to the old runner.) Co-authored-by: jmorganca <jmorganca@gmail.com> * runner.go: Move Unicode checking code and add tests * runner.go: Export external cache members Runner and cache are in the same package so the change doesn't affect anything but it is more internally consistent. * runner.go: Image embedding cache Generating embeddings from images can take significant time (on my machine between 100ms and 8s depending on the model). Although we already cache the result of decoding these images, the embeddings need to be regenerated every time. This is not necessary if we get the same image over and over again, for example, during a conversation. This currently uses a very small cache with a very simple algorithm but it is easy to improve as is warranted. * llama: catch up on patches Carry forward solar-pro and cli-unicode patches * runner.go: Don't re-allocate memory for every batch We can reuse memory allocated from batch to batch since batch size is fixed. This both saves the cost of reallocation as well keeps the cache lines hot. This results in a roughly 1% performance improvement for token generation with Nvidia GPUs on Linux. * runner.go: Default to classic input cache policy The input cache as part of the go runner implemented a cache policy that aims to maximize hit rate in both single and multi- user scenarios. When there is a cache hit, the response is very fast. However, performance is actually slower when there is an input cache miss due to worse GPU VRAM locality. This means that performance is generally better overall for multi-user scenarios (better input cache hit rate, locality was relatively poor already). But worse for single users (input cache hit rate is about the same, locality is now worse). This defaults the policy back to the old one to avoid a regression but keeps the new one available through an environment variable OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is to improve this in the future to get the best of both worlds without user configuration. For inputs that result in cache misses, on Nvidia/Linux this change improves performance by 31% for prompt processing and 13% for token generation. * runner.go: Increase size of response channel Generally the CPU can easily keep up with handling reponses that are generated but there's no reason not to let generation continue and handle things in larger batches if needed. * llama: Add CI to verify all vendored changes have patches (#7066) Make sure we don't accidentally merge changes in the vendored code that aren't also reflected in the patches. * llama: adjust clip patch for mingw utf-16 (#7065) * llama: adjust clip patch for mingw utf-16 * llama: ensure static linking of runtime libs Avoid runtime dependencies on non-standard libraries * runner.go: Enable llamafile (all platforms) and BLAS (Mac OS) These are two features that are shown on llama.cpp's system info that are currently different between the two runners. On my test systems the performance difference is very small to negligible but it is probably still good to equalize the features. * llm: Don't add BOS/EOS for tokenize requests This is consistent with what server.cpp currently does. It affects things like token processing counts for embedding requests. * runner.go: Don't cache prompts for embeddings Our integration with server.cpp implicitly disables prompt caching because it is not part of the JSON object being parsed, this makes the Go runner behavior similarly. Prompt caching has been seen to affect the results of text completions on certain hardware. The results are not wrong either way but they are non-deterministic. However, embeddings seem to be affected even on hardware that does not show this behavior for completions. For now, it is best to maintain consistency with the existing behavior. * runner.go: Adjust debug log levels Add system info printed at startup and quiet down noisier logging. * llama: fix compiler flag differences (#7082) Adjust the flags for the new Go server to more closely match the generate flow * llama: refine developer docs (#7121) * llama: doc and example clean up (#7122) * llama: doc and example clean up * llama: Move new dockerfile into llama dir Temporary home until we fully transition to the Go server * llama: runner doc cleanup * llama.go: Add description for Tokenize error case --------- Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>

Re-introduce the `llama` package (#5034)
* Re-introduce the llama package This PR brings back the llama package, making it possible to call llama.cpp and ggml APIs from Go directly via CGo. This has a few advantages: - C APIs can be called directly from Go without needing to use the previous "server" REST API - On macOS and for CPU builds on Linux and Windows, Ollama can be built without a go generate ./... step, making it easy to get up and running to hack on parts of Ollama that don't require fast inference - Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners takes <5 min on a fast CPU) - No git submodule making it easier to clone and build from source This is a big PR, but much of it is vendor code except for: - llama.go CGo bindings - example/: a simple example of running inference - runner/: a subprocess server designed to replace the llm/ext_server package - Makefile an as minimal as possible Makefile to build the runner package for different targets (cpu, avx, avx2, cuda, rocm) Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Daniel Hiltgen <daniel@ollama.com> * cache: Clear old KV cache entries when evicting a slot When forking a cache entry, if no empty slots are available we evict the least recently used one and copy over the KV entries from the closest match. However, this copy does not overwrite existing values but only adds new ones. Therefore, we need to clear the old slot first. This change fixes two issues: - The KV cache fills up and runs out of space even though we think we are managing it correctly - Performance gets worse over time as we use new cache entries that are not hot in the processor caches * doc: explain golang objc linker warning (#6830) * llama: gather transitive dependencies for rocm for dist packaging (#6848) * Refine go server makefiles to be more DRY (#6924) This breaks up the monolithic Makefile for the Go based runners into a set of utility files as well as recursive Makefiles for the runners. Files starting with the name "Makefile" are buildable, while files that end with ".make" are utilities to include in other Makefiles. This reduces the amount of nearly identical targets and helps set a pattern for future community contributions for new GPU runner architectures. When we are ready to switch over to the Go runners, these files should move to the top of the repo, and we should add targets for the main CLI, as well as a helper "install" (put all the built binaries on the local system in a runnable state) and "dist" target (generate the various tar/zip files for distribution) for local developer use. * llama: don't create extraneous directories (#6988) * llama: Exercise the new build in CI (#6989) Wire up some basic sanity testing in CI for the Go runner. GPU runners are not covered yet. * llama: Refine developer docs for Go server (#6842) This enhances the documentation for development focusing on the new Go server. After we complete the transition further doc refinements can remove the "transition" discussion. * runner.go: Allocate batches for all sequences during init We should tell the model that we could have full batches for all sequences. We already do this when we allocate the batches but it was missed during initialization. * llama.go: Don't return nil from Tokenize on zero length input Potentially receiving nil in a non-error condition is surprising to most callers - it's better to return an empty slice. * runner.go: Remove stop tokens from cache If the last token is EOG then we don't return this and it isn't present in the cache (because it was never submitted to Decode). This works well for extending the cache entry with a new sequence. However, for multi-token stop sequences, we won't return any of the tokens but all but the last one will be in the cache. This means when the conversation continues the cache will contain tokens that don't overlap with the new prompt. This works (we will pick up the portion where there is overlap) but it causes unnecessary cache thrashing because we will fork the original cache entry as it is not a perfect match. By trimming the cache to the tokens that we actually return this issue can be avoided. * runner.go: Simplify flushing of pending tokens * runner.go: Update TODOs * runner.go: Don't panic when processing sequences If there is an error processing a sequence, we should return a clean HTTP error back to Ollama rather than panicing. This will make us more resilient to transient failures. Panics can still occur during startup as there is no way to serve requests if that fails. Co-authored-by: jmorganca <jmorganca@gmail.com> * runner.go: More accurately capture timings Currently prompt processing time doesn't capture the that it takes to tokenize the input, only decoding time. We should capture the full process to more accurately reflect reality. This is especially true once we start processing images where the initial processing can take significant time. This is also more consistent with the existing C++ runner. * runner.go: Support for vision models In addition to bringing feature parity with the C++ runner, this also incorporates several improvements: - Cache prompting works with images, avoiding the need to re-decode embeddings for every message in a conversation - Parallelism is supported, avoiding the need to restrict to one sequence at a time. (Though for now Ollama will not schedule them while we might need to fall back to the old runner.) Co-authored-by: jmorganca <jmorganca@gmail.com> * runner.go: Move Unicode checking code and add tests * runner.go: Export external cache members Runner and cache are in the same package so the change doesn't affect anything but it is more internally consistent. * runner.go: Image embedding cache Generating embeddings from images can take significant time (on my machine between 100ms and 8s depending on the model). Although we already cache the result of decoding these images, the embeddings need to be regenerated every time. This is not necessary if we get the same image over and over again, for example, during a conversation. This currently uses a very small cache with a very simple algorithm but it is easy to improve as is warranted. * llama: catch up on patches Carry forward solar-pro and cli-unicode patches * runner.go: Don't re-allocate memory for every batch We can reuse memory allocated from batch to batch since batch size is fixed. This both saves the cost of reallocation as well keeps the cache lines hot. This results in a roughly 1% performance improvement for token generation with Nvidia GPUs on Linux. * runner.go: Default to classic input cache policy The input cache as part of the go runner implemented a cache policy that aims to maximize hit rate in both single and multi- user scenarios. When there is a cache hit, the response is very fast. However, performance is actually slower when there is an input cache miss due to worse GPU VRAM locality. This means that performance is generally better overall for multi-user scenarios (better input cache hit rate, locality was relatively poor already). But worse for single users (input cache hit rate is about the same, locality is now worse). This defaults the policy back to the old one to avoid a regression but keeps the new one available through an environment variable OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is to improve this in the future to get the best of both worlds without user configuration. For inputs that result in cache misses, on Nvidia/Linux this change improves performance by 31% for prompt processing and 13% for token generation. * runner.go: Increase size of response channel Generally the CPU can easily keep up with handling reponses that are generated but there's no reason not to let generation continue and handle things in larger batches if needed. * llama: Add CI to verify all vendored changes have patches (#7066) Make sure we don't accidentally merge changes in the vendored code that aren't also reflected in the patches. * llama: adjust clip patch for mingw utf-16 (#7065) * llama: adjust clip patch for mingw utf-16 * llama: ensure static linking of runtime libs Avoid runtime dependencies on non-standard libraries * runner.go: Enable llamafile (all platforms) and BLAS (Mac OS) These are two features that are shown on llama.cpp's system info that are currently different between the two runners. On my test systems the performance difference is very small to negligible but it is probably still good to equalize the features. * llm: Don't add BOS/EOS for tokenize requests This is consistent with what server.cpp currently does. It affects things like token processing counts for embedding requests. * runner.go: Don't cache prompts for embeddings Our integration with server.cpp implicitly disables prompt caching because it is not part of the JSON object being parsed, this makes the Go runner behavior similarly. Prompt caching has been seen to affect the results of text completions on certain hardware. The results are not wrong either way but they are non-deterministic. However, embeddings seem to be affected even on hardware that does not show this behavior for completions. For now, it is best to maintain consistency with the existing behavior. * runner.go: Adjust debug log levels Add system info printed at startup and quiet down noisier logging. * llama: fix compiler flag differences (#7082) Adjust the flags for the new Go server to more closely match the generate flow * llama: refine developer docs (#7121) * llama: doc and example clean up (#7122) * llama: doc and example clean up * llama: Move new dockerfile into llama dir Temporary home until we fully transition to the Go server * llama: runner doc cleanup * llama.go: Add description for Tokenize error case --------- Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
96efd905 · Jeffrey Morgan · GitHub · de982616 · 96efd905 · 96efd905
Unverified Commit 96efd905 authored Oct 08, 2024 by Jeffrey Morgan Committed by GitHub Oct 08, 2024
20 changed files
--- a/.gitattributes
+++ b/.gitattributes
 llm/ext_server/* linguist-vendored
+llama/**/*.{cpp,hpp,h,c,cu,cuh,m} linguist-vendored
+
 * text=auto
 *.go text eol=lf
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -24,6 +24,7 @@ jobs:
      GENERATE: ${{ steps.changes.outputs.GENERATE }}
      GENERATE_CUDA: ${{ steps.changes.outputs.GENERATE_CUDA }}
      GENERATE_ROCM: ${{ steps.changes.outputs.GENERATE_ROCM }}
+      RUNNERS: ${{ steps.changes.outputs.RUNNERS }}
    steps:
      - uses: actions/checkout@v4
        with:
@@ -41,6 +42,7 @@ jobs:
            echo GENERATE=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
            echo GENERATE_CUDA=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
            echo GENERATE_ROCM=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo RUNNERS=$(changed 'llama/**')
          } >>$GITHUB_OUTPUT

  generate:
@@ -213,6 +215,46 @@ jobs:
        env:
          OLLAMA_SKIP_CPU_GENERATE: '1'

+  runners:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-2019]
+        arch: [amd64, arm64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-2019
+            arch: arm64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+      ARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - name: 'Build Windows Go Runners'
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          $gopath=(get-command go).source | split-path -parent
+          $gccpath=(get-command gcc).source | split-path -parent
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$gccpath;$env:PATH"
+          echo $env:PATH
+          make -C llama -j 4      
+      - name: 'Build Unix Go Runners'
+        if: ${{ ! startsWith(matrix.os, 'windows-') }}
+        run: make -C llama -j 4
+      - run: go build .
+
  lint:
    strategy:
      matrix:
@@ -280,3 +322,15 @@ jobs:
      - run: go generate ./...
      - run: go build
      - run: go test -v ./...
+
+  patches:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Verify patches carry all the changes
+        run: |
+          cd llama && ./sync.sh && git diff --compact-summary --exit-code .
\ No newline at end of file
--- a/.gitignore
+++ b/.gitignore
@@ -5,7 +5,6 @@
 .swp
 dist
 ollama
-ggml-metal.metal
 .cache
 *.exe
 .idea

--- a/Dockerfile
+++ b/Dockerfile
@@ -110,9 +110,6 @@ ARG CGO_CFLAGS
 ENV GOARCH=amd64
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate

-FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
 RUN --mount=type=cache,target=/root/.ccache \
    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
@@ -135,9 +132,6 @@ ARG CGO_CFLAGS
 ENV GOARCH=arm64
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate

-FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
 RUN --mount=type=cache,target=/root/.ccache \
    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
@@ -148,7 +142,6 @@ FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
 ENV CGO_ENABLED=1
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
 COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
 COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
 COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
@@ -171,7 +164,6 @@ ENV CGO_ENABLED=1
 ARG GOLANG_VERSION
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
 COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
 COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
 COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
@@ -191,7 +183,7 @@ FROM dist-$TARGETARCH as dist


 # Optimized container images do not cary nested payloads
-FROM --platform=linux/amd64 static-build-amd64 AS container-build-amd64
+FROM --platform=linux/amd64 cpu-builder-amd64 AS container-build-amd64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -199,7 +191,7 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
    go build -trimpath -o dist/linux-amd64/bin/ollama .

-FROM --platform=linux/arm64 static-build-arm64 AS container-build-arm64
+FROM --platform=linux/arm64 cpu-builder-arm64 AS container-build-arm64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS

--- a/docs/development.md
+++ b/docs/development.md
 # Development

+> [!IMPORTANT]
+> The `llm` package that loads and runs models is being updated to use a new [Go runner](#transition-to-go-runner): this should only impact a small set of PRs however it does change how the project is built.
+
 Install required tools:

 - cmake version 3.24 or higher
@@ -166,4 +169,182 @@ Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64
 pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
 ```

-You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
\ No newline at end of file
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
+
+
+## Transition to Go runner
+
+The Ollama team is working on moving to a new Go based runner that loads and runs models in a subprocess to replace the previous code under `ext_server`. During this transition period, this new Go runner is "opt in" at build time, and requires using a different approach to build.
+
+After the transition to use the Go server exclusively, both `make` and `go generate` will build the Go runner.
+
+Install required tools:
+
+- go version 1.22 or higher
+- gcc version 11.4.0 or higher
+
+
+### MacOS
+
+[Download Go](https://go.dev/dl/)
+
+Optionally enable debugging and more verbose logging:
+
+```bash
+# At build time
+export CGO_CFLAGS="-g"
+
+# At runtime
+export OLLAMA_DEBUG=1
+```
+
+Get the required libraries and build the native LLM code:  (Adjust the job count based on your number of processors for a faster build)
+
+```bash
+make -C llama -j 5
+```
+
+Then build ollama:
+
+```bash
+go build .
+```
+
+Now you can run `ollama`:
+
+```bash
+./ollama
+```
+
+#### Xcode 15 warnings
+
+If you are using Xcode newer than version 14, you may see a warning during `go build` about `ld: warning: ignoring duplicate libraries: '-lobjc'` due to Golang issue https://github.com/golang/go/issues/67799 which can be safely ignored.  You can suppress the warning with `export CGO_LDFLAGS="-Wl,-no_warn_duplicate_libraries"`
+
+### Linux
+
+#### Linux CUDA (NVIDIA)
+
+_Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
+
+Install `make`, `gcc` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+development and runtime packages.
+
+Typically the build scripts will auto-detect CUDA, however, if your Linux distro
+or installation approach uses unusual paths, you can specify the location by
+specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
+libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
+a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
+
+Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
+
+```
+make -C llama -j 5
+```
+
+Then build the binary:
+
+```
+go build .
+```
+
+#### Linux ROCm (AMD)
+
+_Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
+
+Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `make`, `gcc`, and `golang`.
+
+Typically the build scripts will auto-detect ROCm, however, if your Linux distro
+or installation approach uses unusual paths, you can specify the location by
+specifying an environment variable `ROCM_PATH` to the location of the ROCm
+install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
+CLBlast install (typically `/usr/lib/cmake/CLBlast`). You can also customize
+the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)
+
+Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
+
+```
+make -C llama -j 5
+```
+
+Then build the binary:
+
+```
+go build .
+```
+
+ROCm requires elevated privileges to access the GPU at runtime. On most distros you can add your user account to the `render` group, or run as root.
+
+#### Advanced CPU Settings
+
+By default, running `make` will compile a few different variations
+of the LLM library based on common CPU families and vector math capabilities,
+including a lowest-common-denominator which should run on almost any 64 bit CPU
+somewhat slowly. At runtime, Ollama will auto-detect the optimal variation to
+load. 
+
+Custom CPU settings are not currently supported in the new Go server build but will be added back after we complete the transition.
+
+#### Containerized Linux Build
+
+If you have Docker available, you can build linux binaries with `OLLAMA_NEW_RUNNERS=1 ./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
+
+### Windows
+
+The following tools are required as a minimal development environment to build CPU inference support.
+
+- Go version 1.22 or higher
+  - https://go.dev/dl/
+- Git
+  - https://git-scm.com/download/win
+- GCC and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
+  - [MSYS2](https://www.msys2.org/)
+    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-ucrt-x86_64-gcc make` to install the required tools
+  - Assuming you used the default install prefix for msys2 above, add `c:\msys64\ucrt64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
+
+Then, build the `ollama` binary:
+
+```powershell
+$env:CGO_ENABLED="1"
+make -C llama -j 8
+go build .
+```
+
+#### GPU Support
+
+The GPU tools require the Microsoft native build tools.  To build either CUDA or ROCm, you must first install MSVC via Visual Studio:
+
+- Make sure to select `Desktop development with C++` as a Workload during the Visual Studio install
+- You must complete the Visual Studio install and run it once **BEFORE** installing CUDA or ROCm for the tools to properly register
+- Add the location of the **64 bit (x64)** compiler (`cl.exe`) to your `PATH`
+- Note: the default Developer Shell may configure the 32 bit (x86) compiler which will lead to build failures.  Ollama requires a 64 bit toolchain.
+
+#### Windows CUDA (NVIDIA)
+
+In addition to the common Windows development tools and MSVC described above:
+
+- [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)
+
+#### Windows ROCm (AMD Radeon)
+
+In addition to the common Windows development tools and MSVC described above:
+
+- [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
+
+#### Windows arm64
+
+The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
+
+```powershell
+import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
+Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
+```
+
+You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`
+
+Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
+
+```
+pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
+```
+
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -160,6 +160,8 @@ var (
 	SchedSpread = Bool("OLLAMA_SCHED_SPREAD")
 	// IntelGPU enables experimental Intel GPU detection.
 	IntelGPU = Bool("OLLAMA_INTEL_GPU")
+	// MultiUserCache optimizes prompt caching for multi-user scenarios
+	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
 )

 func String(s string) func() string {
@@ -245,6 +247,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir(), "Location for temporary files"},
+		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},

 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},

--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@@ -42,7 +42,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 		}
 		resp = [2][]string{
 			{"sunlight"},
-			{"england", "english", "massachusetts", "pilgrims", "british"},
+			{"england", "english", "massachusetts", "pilgrims", "british", "festival"},
 		}
 	)
 	var wg sync.WaitGroup

--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@@ -275,7 +275,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 				break
 			}
 		}
-		require.True(t, atLeastOne, "none of %v found in %s", anyResp, response)
+		require.True(t, atLeastOne, "%s: none of %v found in %s", genReq.Model, anyResp, response)
 		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
 	case <-ctx.Done():
 		t.Error("outer test context done while waiting for generate")

--- a/llama/.gitignore
+++ b/llama/.gitignore
+*.bin
+*.gguf
+build/
\ No newline at end of file
--- a/llama/Dockerfile
+++ b/llama/Dockerfile
+# Note: once we have fully transitioned to the Go server, this will replace the old Dockerfile at the top of the tree
+ARG GOLANG_VERSION=1.22.5
+ARG CMAKE_VERSION=3.22.1
+ARG CUDA_VERSION_11=11.3.1
+ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+ARG CUDA_VERSION_12=12.4.0
+ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
+ARG ROCM_VERSION=6.1.2
+
+### To create a local image for building linux binaries on mac or windows with efficient incremental builds
+#
+# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile.new --target unified-builder-amd64 .
+# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
+#
+### Then incremental builds will be much faster in this container
+#
+# make -C llama -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
+#
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
+COPY ./scripts/rh_linux_deps.sh /
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+# TODO intel oneapi goes here...
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
+
+### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
+# Note: this does not contain jetson variants
+#
+# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile.new --target unified-builder-arm64 .
+# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
+#
+FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
+    dnf config-manager --set-enabled appstream && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
+
+FROM --platform=linux/amd64 unified-builder-amd64 AS runners-amd64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG OLLAMA_SKIP_ROCM_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
+RUN --mount=type=cache,target=/root/.ccache \
+    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
+        make -C llama -j $(expr $(nproc) / 2 ) ; \
+    else \
+        make -C llama -j 5 ; \
+    fi
+
+FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
+RUN --mount=type=cache,target=/root/.ccache \
+    make -C llama -j 8
+
+
+# Intermediate stages used for ./scripts/build_linux.sh
+FROM --platform=linux/amd64 centos:7 AS builder-amd64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH amd64
+WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/amd64 builder-amd64 AS build-amd64
+COPY . .
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
+ARG GOFLAGS
+ARG CGO_CFLAGS
+ARG OLLAMA_SKIP_ROCM_GENERATE
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-amd64/bin/ollama .
+RUN cd dist/linux-$GOARCH && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
+    fi
+
+FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH arm64
+WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/arm64 builder-arm64 AS build-arm64
+COPY . .
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-arm64/bin/ollama .
+RUN cd dist/linux-$GOARCH && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+
+FROM --platform=linux/amd64 scratch AS dist-amd64
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM --platform=linux/arm64 scratch AS dist-arm64
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM dist-$TARGETARCH AS dist
+
+
+# Optimized container images do not cary nested payloads
+FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
+WORKDIR /go/src/github.com/ollama/ollama
+COPY . .
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-amd64/bin/ollama .
+
+FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
+WORKDIR /go/src/github.com/ollama/ollama
+COPY . .
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-arm64/bin/ollama .
+
+# For amd64 container images, filter out cuda/rocm to minimize size
+FROM runners-amd64 AS runners-cuda-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
+    ./dist/linux-amd64/lib/ollama/runners/rocm*
+
+FROM runners-amd64 AS runners-rocm-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
+    ./dist/linux-amd64/lib/ollama/libcu*.so* \
+    ./dist/linux-amd64/lib/ollama/runners/cuda*
+
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+
+# ROCm libraries larger so we keep it distinct from the CPU/CUDA image
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
+# Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
+# across releases
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]
+
+FROM runtime-$TARGETARCH
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+ENV NVIDIA_VISIBLE_DEVICES=all
+
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]
--- a/llama/Makefile
+++ b/llama/Makefile
+# top level makefile for Go server
+include make/common-defs.make
+
+RUNNER_TARGETS := default
+
+# Determine which if any GPU runners we should build
+ifeq ($(OS),windows)
+	CUDA_PATH?=$(shell cygpath -m -s "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\" 2>/dev/null)unknown
+	CUDA_BASE_DIR := $(dir $(shell cygpath -m -s "$(CUDA_PATH)\\.." 2>/dev/null))
+	CUDA_11:=$(shell ls -d $(CUDA_BASE_DIR)/v11.? 2>/dev/null)
+	CUDA_12:=$(shell ls -d $(CUDA_BASE_DIR)/v12.? 2>/dev/null)
+	HIP_PATH_83 := $(shell cygpath -m -s "$(subst \,/,$(HIP_PATH))" 2>/dev/null)
+	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH_83)/lib 2>/dev/null)
+else ifeq ($(OS),linux)
+	HIP_PATH?=/opt/rocm
+	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)
+	CUDA_PATH?=/usr/local/cuda
+	CUDA_11:=$(shell ls -d $(CUDA_PATH)-11 2>/dev/null)
+	CUDA_12:=$(shell ls -d $(CUDA_PATH)-12 2>/dev/null)
+endif
+
+ifeq ($(OLLAMA_SKIP_CUDA_GENERATE),)
+ifneq ($(CUDA_11),)
+	RUNNER_TARGETS += cuda_v11
+endif
+ifneq ($(CUDA_12),)
+	RUNNER_TARGETS += cuda_v12
+endif
+endif
+ifeq ($(OLLAMA_SKIP_ROCM_GENERATE),)
+ifneq ($(HIP_LIB_DIR),)
+	RUNNER_TARGETS += rocm
+endif
+endif
+
+
+all: clean-payload .WAIT runners
+
+runners: $(RUNNER_TARGETS)
+
+$(RUNNER_TARGETS):
+	$(MAKE) -f make/Makefile.$@
+
+clean:
+	rm -rf $(BUILD_DIR) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS) $(RUNNERS_PAYLOAD_DIR)
+
+clean-payload:
+	rm -rf $(addprefix $(RUNNERS_PAYLOAD_DIR)/, $(RUNNER_TARGETS) metal cpu cpu_avx cpu_avx2)
+
+.PHONY: all runners clean clean-payload $(RUNNER_TARGETS) .WAIT
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'
--- a/llama/README.md
+++ b/llama/README.md
+# `llama`
+
+This package integrates the [llama.cpp](https://github.com/ggerganov/llama.cpp) library as a Go package and makes it easy to build it with tags for different CPU and GPU processors.
+
+Supported:
+
+- [x] CPU
+- [x] avx, avx2
+- [x] macOS Metal
+- [x] Windows CUDA
+- [x] Windows ROCm
+- [x] Linux CUDA
+- [x] Linux ROCm
+- [x] Llava
+
+Extra build steps are required for CUDA and ROCm on Windows since `nvcc` and `hipcc` both require using msvc as the host compiler. For these shared libraries are created:
+
+- `ggml_cuda.dll` on Windows or `ggml_cuda.so` on Linux
+- `ggml_hipblas.dll` on Windows or `ggml_hipblas.so` on Linux
+
+> Note: it's important that memory is allocated and freed by the same compiler (e.g. entirely by code compiled with msvc or mingw). Issues from this should be rare, but there are some places where pointers are returned by the CUDA or HIP runtimes and freed elsewhere, causing a a crash. In a future change the same runtime should be used in both cases to avoid crashes.
+
+## Building
+
+```
+go build .
+```
+
+### AVX
+
+```shell
+go build -tags avx .
+```
+
+### AVX2
+
+```shell
+# go doesn't recognize `-mfma` as a valid compiler flag
+# see https://github.com/golang/go/issues/17895
+go env -w "CGO_CFLAGS_ALLOW=-mfma|-mf16c"
+go env -w "CGO_CXXFLAGS_ALLOW=-mfma|-mf16c"
+go build -tags=avx,avx2 .
+```
+
+## Linux
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
+
+```shell
+make ggml_cuda.so
+go build -tags avx,cuda .
+```
+
+### ROCm
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
+
+```shell
+make ggml_hipblas.so
+go build -tags avx,rocm .
+```
+
+## Windows
+
+Download [w64devkit](https://github.com/skeeto/w64devkit/releases/latest) for a simple MinGW development environment.
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive) then build the cuda code:
+
+```shell
+make ggml_cuda.dll
+go build -tags avx,cuda .
+```
+
+### ROCm
+
+Install [ROCm 5.7.1](https://rocm.docs.amd.com/en/docs-5.7.1/).
+
+```shell
+make ggml_hipblas.dll
+go build -tags avx,rocm .
+```
+
+## Building runners
+
+```shell
+# build all runners for this platform
+make -j
+```
+
+## Syncing with llama.cpp
+
+To update this package to the latest llama.cpp code, use the `sync.sh` script:
+
+```
+./sync.sh ../../llama.cpp
+```
--- a/llama/base64.hpp
+++ b/llama/base64.hpp
+/*
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or
+distribute this software, either in source code form or as a compiled
+binary, for any purpose, commercial or non-commercial, and by any
+means.
+
+In jurisdictions that recognize copyright laws, the author or authors
+of this software dedicate any and all copyright interest in the
+software to the public domain. We make this dedication for the benefit
+of the public at large and to the detriment of our heirs and
+successors. We intend this dedication to be an overt act of
+relinquishment in perpetuity of all present and future rights to this
+software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <http://unlicense.org>
+*/
+
+#ifndef PUBLIC_DOMAIN_BASE64_HPP_
+#define PUBLIC_DOMAIN_BASE64_HPP_
+
+#include <cstdint>
+#include <iterator>
+#include <stdexcept>
+#include <string>
+
+class base64_error : public std::runtime_error
+{
+public:
+    using std::runtime_error::runtime_error;
+};
+
+class base64
+{
+public:
+    enum class alphabet
+    {
+        /** the alphabet is detected automatically */
+        auto_,
+        /** the standard base64 alphabet is used */
+        standard,
+        /** like `standard` except that the characters `+` and `/` are replaced by `-` and `_` respectively*/
+        url_filename_safe
+    };
+
+    enum class decoding_behavior
+    {
+        /** if the input is not padded, the remaining bits are ignored */
+        moderate,
+        /** if a padding character is encounter decoding is finished */
+        loose
+    };
+
+    /**
+     Encodes all the elements from `in_begin` to `in_end` to `out`.
+
+     @warning The source and destination cannot overlap. The destination must be able to hold at least
+     `required_encode_size(std::distance(in_begin, in_end))`, otherwise the behavior depends on the output iterator.
+
+     @tparam Input_iterator the source; the returned elements are cast to `std::uint8_t` and should not be greater than
+     8 bits
+     @tparam Output_iterator the destination; the elements written to it are from the type `char`
+     @param in_begin the beginning of the source
+     @param in_end the ending of the source
+     @param out the destination iterator
+     @param alphabet which alphabet should be used
+     @returns the iterator to the next element past the last element copied
+     @throws see `Input_iterator` and `Output_iterator`
+    */
+    template<typename Input_iterator, typename Output_iterator>
+    static Output_iterator encode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
+                                  alphabet alphabet = alphabet::standard)
+    {
+        constexpr auto pad = '=';
+        const char* alpha  = alphabet == alphabet::url_filename_safe
+                                ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
+                                : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+        while (in_begin != in_end) {
+            std::uint8_t i0 = 0, i1 = 0, i2 = 0;
+
+            // first character
+            i0 = static_cast<std::uint8_t>(*in_begin);
+            ++in_begin;
+
+            *out = alpha[i0 >> 2 & 0x3f];
+            ++out;
+
+            // part of first character and second
+            if (in_begin != in_end) {
+                i1 = static_cast<std::uint8_t>(*in_begin);
+                ++in_begin;
+
+                *out = alpha[((i0 & 0x3) << 4) | (i1 >> 4 & 0x0f)];
+                ++out;
+            } else {
+                *out = alpha[(i0 & 0x3) << 4];
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                break;
+            }
+
+            // part of second character and third
+            if (in_begin != in_end) {
+                i2 = static_cast<std::uint8_t>(*in_begin);
+                ++in_begin;
+
+                *out = alpha[((i1 & 0xf) << 2) | (i2 >> 6 & 0x03)];
+                ++out;
+            } else {
+                *out = alpha[(i1 & 0xf) << 2];
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                break;
+            }
+
+            // rest of third
+            *out = alpha[i2 & 0x3f];
+            ++out;
+        }
+
+        return out;
+    }
+    /**
+     Encodes a string.
+
+     @param str the string that should be encoded
+     @param alphabet which alphabet should be used
+     @returns the encoded base64 string
+     @throws see base64::encode()
+    */
+    static std::string encode(const std::string& str, alphabet alphabet = alphabet::standard)
+    {
+        std::string result;
+
+        result.reserve(required_encode_size(str.length()) + 1);
+
+        encode(str.begin(), str.end(), std::back_inserter(result), alphabet);
+
+        return result;
+    }
+    /**
+     Encodes a char array.
+
+     @param buffer the char array
+     @param size the size of the array
+     @param alphabet which alphabet should be used
+     @returns the encoded string
+    */
+    static std::string encode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::standard)
+    {
+        std::string result;
+
+        result.reserve(required_encode_size(size) + 1);
+
+        encode(buffer, buffer + size, std::back_inserter(result), alphabet);
+
+        return result;
+    }
+    /**
+     Decodes all the elements from `in_begin` to `in_end` to `out`. `in_begin` may point to the same location as `out`,
+     in other words: inplace decoding is possible.
+
+     @warning The destination must be able to hold at least `required_decode_size(std::distance(in_begin, in_end))`,
+     otherwise the behavior depends on the output iterator.
+
+     @tparam Input_iterator the source; the returned elements are cast to `char`
+     @tparam Output_iterator the destination; the elements written to it are from the type `std::uint8_t`
+     @param in_begin the beginning of the source
+     @param in_end the ending of the source
+     @param out the destination iterator
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the iterator to the next element past the last element copied
+     @throws base64_error depending on the set behavior
+     @throws see `Input_iterator` and `Output_iterator`
+    */
+    template<typename Input_iterator, typename Output_iterator>
+    static Output_iterator decode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
+                                  alphabet alphabet          = alphabet::auto_,
+                                  decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        //constexpr auto pad = '=';
+        std::uint8_t last  = 0;
+        auto bits          = 0;
+
+        while (in_begin != in_end) {
+            auto c = *in_begin;
+            ++in_begin;
+
+            if (c == '=') {
+                break;
+            }
+
+            auto part = _base64_value(alphabet, c);
+
+            // enough bits for one byte
+            if (bits + 6 >= 8) {
+                *out = (last << (8 - bits)) | (part >> (bits - 2));
+                ++out;
+
+                bits -= 2;
+            } else {
+                bits += 6;
+            }
+
+            last = part;
+        }
+
+        // check padding
+        if (behavior != decoding_behavior::loose) {
+            while (in_begin != in_end) {
+                auto c = *in_begin;
+                ++in_begin;
+
+                if (c != '=') {
+                    throw base64_error("invalid base64 character.");
+                }
+            }
+        }
+
+        return out;
+    }
+    /**
+     Decodes a string.
+
+     @param str the base64 encoded string
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the decoded string
+     @throws see base64::decode()
+    */
+    static std::string decode(const std::string& str, alphabet alphabet = alphabet::auto_,
+                              decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        std::string result;
+
+        result.reserve(max_decode_size(str.length()));
+
+        decode(str.begin(), str.end(), std::back_inserter(result), alphabet, behavior);
+
+        return result;
+    }
+    /**
+     Decodes a string.
+
+     @param buffer the base64 encoded buffer
+     @param size the size of the buffer
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the decoded string
+     @throws see base64::decode()
+    */
+    static std::string decode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::auto_,
+                              decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        std::string result;
+
+        result.reserve(max_decode_size(size));
+
+        decode(buffer, buffer + size, std::back_inserter(result), alphabet, behavior);
+
+        return result;
+    }
+    /**
+     Decodes a string inplace.
+
+     @param[in,out] str the base64 encoded string
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @throws base64::decode_inplace()
+    */
+    static void decode_inplace(std::string& str, alphabet alphabet = alphabet::auto_,
+                               decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        str.resize(decode(str.begin(), str.end(), str.begin(), alphabet, behavior) - str.begin());
+    }
+    /**
+     Decodes a char array inplace.
+
+     @param[in,out] str the string array
+     @param size the length of the array
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the pointer to the next element past the last element decoded
+     @throws base64::decode_inplace()
+    */
+    static char* decode_inplace(char* str, std::size_t size, alphabet alphabet = alphabet::auto_,
+                                decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        return decode(str, str + size, str, alphabet, behavior);
+    }
+    /**
+     Returns the required decoding size for a given size. The value is calculated with the following formula:
+
+     $$
+     \lceil \frac{size}{4} \rceil \cdot 3
+     $$
+
+     @param size the size of the encoded input
+     @returns the size of the resulting decoded buffer; this the absolute maximum
+    */
+    static std::size_t max_decode_size(std::size_t size) noexcept
+    {
+        return (size / 4 + (size % 4 ? 1 : 0)) * 3;
+    }
+    /**
+     Returns the required encoding size for a given size. The value is calculated with the following formula:
+
+     $$
+     \lceil \frac{size}{3} \rceil \cdot 4
+     $$
+
+     @param size the size of the decoded input
+     @returns the size of the resulting encoded buffer
+    */
+    static std::size_t required_encode_size(std::size_t size) noexcept
+    {
+        return (size / 3 + (size % 3 ? 1 : 0)) * 4;
+    }
+
+private:
+    static std::uint8_t _base64_value(alphabet& alphabet, char c)
+    {
+        if (c >= 'A' && c <= 'Z') {
+            return c - 'A';
+        } else if (c >= 'a' && c <= 'z') {
+            return c - 'a' + 26;
+        } else if (c >= '0' && c <= '9') {
+            return c - '0' + 52;
+        }
+
+        // comes down to alphabet
+        if (alphabet == alphabet::standard) {
+            if (c == '+') {
+                return 62;
+            } else if (c == '/') {
+                return 63;
+            }
+        } else if (alphabet == alphabet::url_filename_safe) {
+            if (c == '-') {
+                return 62;
+            } else if (c == '_') {
+                return 63;
+            }
+        } // auto detect
+        else {
+            if (c == '+') {
+                alphabet = alphabet::standard;
+
+                return 62;
+            } else if (c == '/') {
+                alphabet = alphabet::standard;
+
+                return 63;
+            } else if (c == '-') {
+                alphabet = alphabet::url_filename_safe;
+
+                return 62;
+            } else if (c == '_') {
+                alphabet = alphabet::url_filename_safe;
+
+                return 63;
+            }
+        }
+
+        throw base64_error("invalid base64 character.");
+    }
+};
+
+#endif // !PUBLIC_DOMAIN_BASE64_HPP_
--- a/llama/build-info.cpp
+++ b/llama/build-info.cpp
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+int LLAMA_BUILD_NUMBER = 0;
+char const *LLAMA_COMMIT = "";
+char const *LLAMA_COMPILER = "";
+char const *LLAMA_BUILD_TARGET = "";
--- a/llama/clip.cpp
+++ b/llama/clip.cpp
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// NOTE: This is modified from clip.cpp only for LLaVA,
+// so there might be still unnecessary artifacts hanging around
+// I'll gradually clean and extend it
+// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
+#include "clip.h"
+#include "log.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb_image.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <stdexcept>
+#include <vector>
+#include <sstream>
+#include <cinttypes>
+#include <limits>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#if __GLIBCXX__
+#include <cstdio>
+#include <ext/stdio_filebuf.h>
+#include <fcntl.h>
+#endif
+#endif
+
+//#define CLIP_DEBUG_FUNCTIONS
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+static std::string format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), buf.size());
+}
+
+//
+// key constants
+//
+
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+
+#define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
+#define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
+#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+
+
+//
+// tensor name constants
+//
+
+#define TN_TOKEN_EMBD      "%s.token_embd.weight"
+#define TN_POS_EMBD        "%s.position_embd.weight"
+#define TN_CLASS_EMBD      "v.class_embd"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
+#define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
+#define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
+#define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s"
+#define TN_LN_PRE          "%s.pre_ln.%s"
+#define TN_LN_POST         "%s.post_ln.%s"
+#define TN_TEXT_PROJ       "text_projection.weight"
+#define TN_VIS_PROJ        "visual_projection.weight"
+#define TN_LLAVA_PROJ      "mm.%d.%s"
+#define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
+#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
+#define TN_IMAGE_NEWLINE   "model.image_newline"
+
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY "resampler.query"
+#define TN_MINICPMV_PROJ "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ "resampler.kv.weight"
+#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN "resampler.ln_%s.%s"
+
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_UNKNOWN,
+};
+
+static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
+    { PROJECTOR_TYPE_MLP, "mlp" },
+    { PROJECTOR_TYPE_LDP, "ldp" },
+    { PROJECTOR_TYPE_LDPV2, "ldpv2"},
+    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+};
+
+
+//
+// utilities to get data from a gguf file
+//
+
+static int get_key_idx(const gguf_context * ctx, const char * key) {
+    int i = gguf_find_key(ctx, key);
+    if (i == -1) {
+        LOG_TEE("key %s not found in file\n", key);
+        throw std::runtime_error(format("Missing required key: %s", key));
+    }
+
+    return i;
+}
+
+static uint32_t get_u32(const gguf_context * ctx, const std::string & key) {
+    const int i = get_key_idx(ctx, key.c_str());
+
+    return gguf_get_val_u32(ctx, i);
+}
+
+static float get_f32(const gguf_context * ctx, const std::string & key) {
+    const int i = get_key_idx(ctx, key.c_str());
+
+    return gguf_get_val_f32(ctx, i);
+}
+
+static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::string & name) {
+    struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
+    if (!cur) {
+        throw std::runtime_error(format("%s: unable to find tensor %s\n", __func__, name.c_str()));
+    }
+
+    return cur;
+}
+
+static std::string get_ftype(int ftype) {
+    return ggml_type_name(static_cast<ggml_type>(ftype));
+}
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}
+
+static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
+    size_t tensor_size = ggml_nbytes(tensor);
+    LOG_TEE("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
+            prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
+            tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
+}
+
+static projector_type clip_projector_type_from_string(const std::string & name) {
+    for (const auto & kv : PROJECTOR_TYPE_NAMES) { // NOLINT
+        if (kv.second == name) {
+            return kv.first;
+        }
+    }
+    return PROJECTOR_TYPE_UNKNOWN;
+}
+
+#ifdef CLIP_DEBUG_FUNCTIONS
+static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
+    std::ofstream file(filename, std::ios::binary);
+    if (!file.is_open()) {
+        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        return;
+    }
+
+    // PPM header: P6 format, width, height, and max color value
+    file << "P6\n" << img.nx << " " << img.ny << "\n255\n";
+
+    // Write pixel data
+    for (size_t i = 0; i < img.buf.size(); i += 3) {
+        // PPM expects binary data in RGB format, which matches our image buffer
+        file.write(reinterpret_cast<const char*>(&img.buf[i]), 3);
+    }
+
+    file.close();
+}
+
+static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
+    std::ofstream file(filename, std::ios::binary);
+    if (!file.is_open()) {
+        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        return;
+    }
+
+    int fileSize = 54 + 3 * img.nx * img.ny; // File header + info header + pixel data
+    int bytesPerPixel = 3;
+    int widthInBytes = img.nx * bytesPerPixel;
+    int paddingAmount = (4 - (widthInBytes % 4)) % 4;
+    int stride = widthInBytes + paddingAmount;
+
+    // Bitmap file header
+    unsigned char fileHeader[14] = {
+        'B','M',     // Signature
+        0,0,0,0,    // Image file size in bytes
+        0,0,0,0,    // Reserved
+        54,0,0,0    // Start of pixel array
+    };
+
+    // Total file size
+    fileSize = 54 + (stride * img.ny);
+    fileHeader[2] = (unsigned char)(fileSize);
+    fileHeader[3] = (unsigned char)(fileSize >> 8);
+    fileHeader[4] = (unsigned char)(fileSize >> 16);
+    fileHeader[5] = (unsigned char)(fileSize >> 24);
+
+    // Bitmap information header (BITMAPINFOHEADER)
+    unsigned char infoHeader[40] = {
+        40,0,0,0,   // Size of this header (40 bytes)
+        0,0,0,0,    // Image width
+        0,0,0,0,    // Image height
+        1,0,        // Number of color planes
+        24,0,       // Bits per pixel
+        0,0,0,0,    // No compression
+        0,0,0,0,    // Image size (can be 0 for no compression)
+        0,0,0,0,    // X pixels per meter (not specified)
+        0,0,0,0,    // Y pixels per meter (not specified)
+        0,0,0,0,    // Total colors (color table not used)
+        0,0,0,0     // Important colors (all are important)
+    };
+
+    // Width and height in the information header
+    infoHeader[4] = (unsigned char)(img.nx);
+    infoHeader[5] = (unsigned char)(img.nx >> 8);
+    infoHeader[6] = (unsigned char)(img.nx >> 16);
+    infoHeader[7] = (unsigned char)(img.nx >> 24);
+    infoHeader[8] = (unsigned char)(img.ny);
+    infoHeader[9] = (unsigned char)(img.ny >> 8);
+    infoHeader[10] = (unsigned char)(img.ny >> 16);
+    infoHeader[11] = (unsigned char)(img.ny >> 24);
+
+    // Write file headers
+    file.write(reinterpret_cast<char*>(fileHeader), sizeof(fileHeader));
+    file.write(reinterpret_cast<char*>(infoHeader), sizeof(infoHeader));
+
+    // Pixel data
+    std::vector<unsigned char> padding(3, 0); // Max padding size to be added to each row
+    for (int y = img.ny - 1; y >= 0; --y) { // BMP files are stored bottom-to-top
+        for (int x = 0; x < img.nx; ++x) {
+            // Each pixel
+            size_t pixelIndex = (y * img.nx + x) * 3;
+            unsigned char pixel[3] = {
+                img.buf[pixelIndex + 2], // BMP stores pixels in BGR format
+                img.buf[pixelIndex + 1],
+                img.buf[pixelIndex]
+            };
+            file.write(reinterpret_cast<char*>(pixel), 3);
+        }
+        // Write padding for the row
+        file.write(reinterpret_cast<char*>(padding.data()), paddingAmount);
+    }
+
+    file.close();
+}
+
+// debug function to convert f32 to u8
+static void clip_image_convert_f32_to_u8(const clip_image_f32& src, clip_image_u8& dst) {
+    dst.nx = src.nx;
+    dst.ny = src.ny;
+    dst.buf.resize(3 * src.nx * src.ny);
+    for (size_t i = 0; i < src.buf.size(); ++i) {
+        dst.buf[i] = static_cast<uint8_t>(std::min(std::max(int(src.buf[i] * 255.0f), 0), 255));
+    }
+}
+#endif
+
+
+//
+// clip layers
+//
+
+struct clip_hparams {
+    int32_t image_size;
+    int32_t patch_size;
+    int32_t hidden_size;
+    int32_t n_intermediate;
+    int32_t projection_dim;
+    int32_t n_head;
+    int32_t n_layer;
+
+    float eps;
+
+    char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default)
+
+    int32_t image_grid_pinpoints[32];
+    int32_t image_crop_resolution;
+};
+
+struct clip_layer {
+    // attention
+    struct ggml_tensor * k_w;
+    struct ggml_tensor * k_b;
+    struct ggml_tensor * q_w;
+    struct ggml_tensor * q_b;
+    struct ggml_tensor * v_w;
+    struct ggml_tensor * v_b;
+
+    struct ggml_tensor * o_w;
+    struct ggml_tensor * o_b;
+
+    // layernorm 1
+    struct ggml_tensor * ln_1_w;
+    struct ggml_tensor * ln_1_b;
+
+    // ff
+    struct ggml_tensor * ff_i_w;
+    struct ggml_tensor * ff_i_b;
+
+    struct ggml_tensor * ff_o_w;
+    struct ggml_tensor * ff_o_b;
+
+    // layernorm 2
+    struct ggml_tensor * ln_2_w;
+    struct ggml_tensor * ln_2_b;
+};
+
+struct clip_vision_model {
+    struct clip_hparams hparams;
+
+    // embeddings
+    struct ggml_tensor * class_embedding;
+    struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_bias;
+    struct ggml_tensor * position_embeddings;
+
+    struct ggml_tensor * pre_ln_w;
+    struct ggml_tensor * pre_ln_b;
+
+    std::vector<clip_layer> layers;
+
+    struct ggml_tensor * post_ln_w;
+    struct ggml_tensor * post_ln_b;
+
+    struct ggml_tensor * projection;
+
+    // LLaVA projection
+    struct ggml_tensor * mm_0_w = NULL;
+    struct ggml_tensor * mm_0_b = NULL;
+    struct ggml_tensor * mm_2_w = NULL;
+    struct ggml_tensor * mm_2_b = NULL;
+
+    struct ggml_tensor * image_newline = NULL;
+
+    // Yi type models with mlp+normalization projection
+    struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4
+    struct ggml_tensor * mm_1_b = NULL;
+    struct ggml_tensor * mm_3_w = NULL;
+    struct ggml_tensor * mm_3_b = NULL;
+    struct ggml_tensor * mm_4_w = NULL;
+    struct ggml_tensor * mm_4_b = NULL;
+
+    // MobileVLM projection
+    struct ggml_tensor * mm_model_mlp_1_w;
+    struct ggml_tensor * mm_model_mlp_1_b;
+    struct ggml_tensor * mm_model_mlp_3_w;
+    struct ggml_tensor * mm_model_mlp_3_b;
+    struct ggml_tensor * mm_model_block_1_block_0_0_w;
+    struct ggml_tensor * mm_model_block_1_block_0_1_w;
+    struct ggml_tensor * mm_model_block_1_block_0_1_b;
+    struct ggml_tensor * mm_model_block_1_block_1_fc1_w;
+    struct ggml_tensor * mm_model_block_1_block_1_fc1_b;
+    struct ggml_tensor * mm_model_block_1_block_1_fc2_w;
+    struct ggml_tensor * mm_model_block_1_block_1_fc2_b;
+    struct ggml_tensor * mm_model_block_1_block_2_0_w;
+    struct ggml_tensor * mm_model_block_1_block_2_1_w;
+    struct ggml_tensor * mm_model_block_1_block_2_1_b;
+    struct ggml_tensor * mm_model_block_2_block_0_0_w;
+    struct ggml_tensor * mm_model_block_2_block_0_1_w;
+    struct ggml_tensor * mm_model_block_2_block_0_1_b;
+    struct ggml_tensor * mm_model_block_2_block_1_fc1_w;
+    struct ggml_tensor * mm_model_block_2_block_1_fc1_b;
+    struct ggml_tensor * mm_model_block_2_block_1_fc2_w;
+    struct ggml_tensor * mm_model_block_2_block_1_fc2_b;
+    struct ggml_tensor * mm_model_block_2_block_2_0_w;
+    struct ggml_tensor * mm_model_block_2_block_2_1_w;
+    struct ggml_tensor * mm_model_block_2_block_2_1_b;
+
+    // MobileVLM_V2 projection
+    struct ggml_tensor * mm_model_mlp_0_w;
+    struct ggml_tensor * mm_model_mlp_0_b;
+    struct ggml_tensor * mm_model_mlp_2_w;
+    struct ggml_tensor * mm_model_mlp_2_b;
+    struct ggml_tensor * mm_model_peg_0_w;
+    struct ggml_tensor * mm_model_peg_0_b;
+
+    // MINICPMV projection
+    struct ggml_tensor * mm_model_pos_embed_k;
+    struct ggml_tensor * mm_model_query;
+    struct ggml_tensor * mm_model_proj;
+    struct ggml_tensor * mm_model_kv_proj;
+    struct ggml_tensor * mm_model_attn_q_w;
+    struct ggml_tensor * mm_model_attn_q_b;
+    struct ggml_tensor * mm_model_attn_k_w;
+    struct ggml_tensor * mm_model_attn_k_b;
+    struct ggml_tensor * mm_model_attn_v_w;
+    struct ggml_tensor * mm_model_attn_v_b;
+    struct ggml_tensor * mm_model_attn_o_w;
+    struct ggml_tensor * mm_model_attn_o_b;
+    struct ggml_tensor * mm_model_ln_q_w;
+    struct ggml_tensor * mm_model_ln_q_b;
+    struct ggml_tensor * mm_model_ln_kv_w;
+    struct ggml_tensor * mm_model_ln_kv_b;
+    struct ggml_tensor * mm_model_ln_post_w;
+    struct ggml_tensor * mm_model_ln_post_b;
+};
+
+struct clip_ctx {
+    bool has_text_encoder    = false;
+    bool has_vision_encoder  = false;
+    bool has_llava_projector = false;
+    bool has_minicpmv_projector = false;
+    int minicpmv_version = 2;
+
+    struct clip_vision_model vision_model;
+    projector_type proj_type = PROJECTOR_TYPE_MLP;
+
+    float image_mean[3];
+    float image_std[3];
+    bool use_gelu = false;
+    int32_t ftype = 1;
+
+    bool has_class_embedding = true;
+    bool has_pre_norm = true;
+    bool has_post_norm = false;
+    bool has_patch_bias = false;
+
+    struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
+
+    // memory buffers to evaluate the model
+    ggml_backend_buffer_t params_buffer  = NULL;
+
+    ggml_backend_t backend       = NULL;
+    ggml_gallocr_t compute_alloc = NULL;
+
+    struct clip_image_size * load_image_size;
+};
+
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return nullptr;
+    }
+
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        if (load_image_size == nullptr) {
+            load_image_size = clip_image_size_init();
+        }
+        LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        image_size_width  = load_image_size->width;
+        image_size_height = load_image_size->height;
+        if (is_inf) {
+            image_size_width  = imgs->data->nx;
+            image_size_height = imgs->data->ny;
+        }
+    }
+    const int patch_size           = hparams.patch_size;
+    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    const int hidden_size          = hparams.hidden_size;
+    const int n_head               = hparams.n_head;
+    const int d_head               = hidden_size / n_head;
+    int n_layer                    = hparams.n_layer;
+    const float eps                = hparams.eps;
+
+    const int batch_size = imgs->size;
+
+    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
+        GGML_ASSERT(batch_size == 1);
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
+        /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+
+    if (ctx->has_patch_bias) {
+        // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+    struct ggml_tensor * embeddings = inp;
+    struct ggml_tensor * pos_embed = nullptr;
+
+    if (ctx->has_llava_projector) {
+        // concat class_embeddings and patch_embeddings
+        if (ctx->has_class_embedding) {
+            embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+            ggml_set_name(embeddings, "embeddings");
+            ggml_set_input(embeddings);
+            embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+            embeddings = ggml_acc(ctx0, embeddings, inp,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+        }
+    }
+
+    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    embeddings =
+        ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
+
+    if (ctx->has_minicpmv_projector) {
+        int pos_w = image_size_width/patch_size;
+        int pos_h = image_size_height/patch_size;
+        if (ctx->minicpmv_version == 2) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        }
+        else if (ctx->minicpmv_version == 3) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
+        }
+        ggml_set_name(pos_embed, "pos_embed");
+        ggml_set_input(pos_embed);
+    }
+
+    // pre-layernorm
+    if (ctx->has_pre_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "pre_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
+    }
+
+    // loop over layers
+    if (ctx->has_minicpmv_projector) {
+        n_layer += 1;
+    }
+    for (int il = 0; il < n_layer - 1; il++) {
+        struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
+
+        //const size_t nb_q_w = model.layers[il].q_w->nb[0];
+
+        // layernorm1
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w),
+                           model.layers[il].ln_1_b);
+        }
+
+        // self-attention
+        {
+
+            struct ggml_tensor * Q =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
+
+            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+            Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
+            Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+            Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
+
+            struct ggml_tensor * K =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
+
+            K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+            K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+            K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+
+            struct ggml_tensor * V =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b);
+
+            V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+            V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+            V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            KQ = ggml_soft_max_inplace(ctx0, KQ);
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+            KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size);
+            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
+        }
+
+        // attention output
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b);
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, embeddings);
+
+        embeddings = cur; // embeddings = residual, cur = hidden_states
+
+        // layernorm2
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b);
+        }
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b);
+
+        if (ctx->use_gelu) {
+            cur = ggml_gelu_inplace(ctx0, cur);
+        } else {
+            cur = ggml_gelu_quick_inplace(ctx0, cur);
+        }
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b);
+
+        // residual 2
+        cur = ggml_add(ctx0, embeddings, cur);
+
+        embeddings = cur;
+    }
+
+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
+    // llava projector
+    if (ctx->has_llava_projector) {
+        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
+
+        struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
+        ggml_set_name(patches, "patches");
+        ggml_set_input(patches);
+
+        // shape [1, 576, 1024]
+        // ne is whcn, ne = [1024, 576, 1, 1]
+        embeddings = ggml_get_rows(ctx0, embeddings, patches);
+
+        // print_tensor_info(embeddings, "embeddings");
+
+        // llava projector
+        if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+            embeddings = ggml_gelu(ctx0, embeddings);
+            embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projector
+            int n_patch = 24;
+            struct ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
+            mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
+            mlp_1 = ggml_gelu(ctx0, mlp_1);
+            struct ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
+            mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
+            // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]
+
+            // block 1
+            struct ggml_tensor * block_1 = nullptr;
+            {
+                // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
+                mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
+                mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                // stride = 1, padding = 1, bias is nullptr
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+
+                // layer norm
+                // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+
+                // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // hardswish
+                struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+                // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // residual
+                block_1 = ggml_add(ctx0, mlp_3, block_1);
+            }
+
+            // block_2
+            {
+                // stride = 2
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // layer norm
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // hardswish
+                struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                // not sure the parameters is right for globalAvgPooling
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+
+                // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
+                block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
+                // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
+            }
+            embeddings = block_1;
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
+        else {
+            GGML_ABORT("fatal error");
+        }
+    }
+    // minicpmv projector
+    else if (ctx->has_minicpmv_projector)
+    {
+        if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            struct ggml_tensor * q = model.mm_model_query;
+            { // layernorm
+                q = ggml_norm(ctx0, q, eps);
+                q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            }
+            struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+            { // layernorm
+                v = ggml_norm(ctx0, v, eps);
+                v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b);
+            }
+            struct ggml_tensor * k;
+            { // position
+                // q = ggml_add(ctx0, q, model.mm_model_pos_embed);
+                k = ggml_add(ctx0, v, pos_embed);
+            }
+
+            { // attention
+                int hidden_size = 4096;
+                const int d_head = 128;
+                int n_head = hidden_size/d_head;
+                int num_query = 96;
+                if (ctx->minicpmv_version == 2) {
+                    hidden_size = 4096;
+                    n_head = hidden_size/d_head;
+                    num_query = 96;
+                }
+                else if (ctx->minicpmv_version == 3) {
+                    hidden_size = 3584;
+                    n_head = hidden_size/d_head;
+                    num_query = 64;
+                }
+
+                struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
+                Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+                struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b);
+                struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b);
+                // permute
+                Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size);
+                Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+                Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size);
+                K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+                K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+                K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+                V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+                V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+                V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+                KQ = ggml_soft_max_inplace(ctx0, KQ);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+                KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size);
+                KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size);
+
+                embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b);
+            }
+            { // layernorm
+                embeddings = ggml_norm(ctx0, embeddings, eps);
+                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b);
+            }
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+        }
+        else {
+            GGML_ASSERT(false);
+        }
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+// read and create ggml_context containing the tensors and their data
+struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+    struct ggml_context * meta = NULL;
+
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ &meta,
+    };
+
+    struct gguf_context * ctx = gguf_init_from_file(fname, params);
+    if (!ctx) {
+        throw std::runtime_error(format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname));
+    }
+
+    if (verbosity >= 1) {
+        const int n_tensors = gguf_get_n_tensors(ctx);
+        const int n_kv = gguf_get_n_kv(ctx);
+        const int ftype = get_u32(ctx, KEY_FTYPE);
+        const std::string ftype_str = get_ftype(ftype);
+        const int idx_desc = get_key_idx(ctx, KEY_DESCRIPTION);
+        const std::string description = gguf_get_val_str(ctx, idx_desc);
+        const int idx_name = gguf_find_key(ctx, KEY_NAME);
+        if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
+            const std::string name = gguf_get_val_str(ctx, idx_name);
+            LOG_TEE("%s: model name:   %s\n", __func__, name.c_str());
+        }
+        LOG_TEE("%s: description:  %s\n", __func__, description.c_str());
+        LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
+        LOG_TEE("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
+        LOG_TEE("%s: n_tensors:    %d\n", __func__, n_tensors);
+        LOG_TEE("%s: n_kv:         %d\n", __func__, n_kv);
+        LOG_TEE("%s: ftype:        %s\n", __func__, ftype_str.c_str());
+        LOG_TEE("\n");
+    }
+    const int n_tensors = gguf_get_n_tensors(ctx);
+
+    // kv
+    const int n_kv = gguf_get_n_kv(ctx);
+    LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+        __func__, n_kv, n_tensors, fname);
+    {
+        std::map<enum ggml_type, uint32_t> n_type;
+
+        for (int i = 0; i < n_tensors; i++) {
+            enum ggml_type type = gguf_get_tensor_type(ctx, i);
+
+            n_type[type]++;
+        }
+
+        LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+        for (int i = 0; i < n_kv; i++) {
+            const char * name           = gguf_get_key(ctx, i);
+            const enum gguf_type type   = gguf_get_kv_type(ctx, i);
+            const std::string type_name =
+                type == GGUF_TYPE_ARRAY
+                ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx, i)), gguf_get_arr_n(ctx, i))
+                : gguf_type_name(type);
+
+            std::string value          = gguf_kv_to_str(ctx, i);
+            const size_t MAX_VALUE_LEN = 40;
+            if (value.size() > MAX_VALUE_LEN) {
+                value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str());
+            }
+            replace_all(value, "\n", "\\n");
+
+            LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+        }
+
+        // print type counts
+        for (auto & kv : n_type) {
+            if (kv.second == 0) {
+                continue;
+            }
+
+            LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
+        }
+    }
+
+    // data
+    size_t model_size = 0;
+    {
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            const size_t offset = gguf_get_tensor_offset(ctx, i);
+            enum ggml_type type = gguf_get_tensor_type(ctx, i);
+            struct ggml_tensor * cur = ggml_get_tensor(meta, name);
+            size_t tensor_size = ggml_nbytes(cur);
+            model_size += tensor_size;
+            if (verbosity >= 3) {
+                LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
+                       __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
+            }
+        }
+    }
+
+    clip_ctx * new_clip = new clip_ctx{};
+
+    // update projector type
+    {
+        int idx = gguf_find_key(ctx, KEY_PROJ_TYPE);
+        if (idx != -1) {
+            const std::string proj_type = gguf_get_val_str(ctx, idx);
+            new_clip->proj_type = clip_projector_type_from_string(proj_type);
+        } else {
+            new_clip->proj_type = PROJECTOR_TYPE_MLP;
+        }
+
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+            if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) {
+                new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM;
+            }
+        }
+    }
+
+#ifdef GGML_USE_CUDA
+    new_clip->backend = ggml_backend_cuda_init(0);
+    LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+    new_clip->backend = ggml_backend_metal_init();
+    LOG_TEE("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_CANN
+    new_clip->backend = ggml_backend_cann_init(0);
+    LOG_TEE("%s: CLIP using CANN backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_VULKAN
+    new_clip->backend = ggml_backend_vk_init(0);
+    LOG_TEE("%s: CLIP using Vulkan backend\n", __func__);
+#endif
+
+    if (!new_clip->backend) {
+        new_clip->backend = ggml_backend_cpu_init();
+        LOG_TEE("%s: CLIP using CPU backend\n", __func__);
+    }
+
+    // model size and capabilities
+    {
+        int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC);
+        new_clip->has_text_encoder = gguf_get_val_bool(ctx, idx);
+
+        idx = get_key_idx(ctx, KEY_HAS_VIS_ENC);
+        new_clip->has_vision_encoder = gguf_get_val_bool(ctx, idx);
+
+        idx = gguf_find_key(ctx, KEY_HAS_LLAVA_PROJ);
+        if (idx != -1) {
+            new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
+        }
+
+        idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ);
+        if (idx != -1) {
+            new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
+        }
+
+        idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION);
+        if (idx != -1) {
+            new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
+        }
+
+        // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+
+        GGML_ASSERT(new_clip->has_vision_encoder);
+        GGML_ASSERT(!new_clip->has_text_encoder);
+
+        idx = get_key_idx(ctx, KEY_USE_GELU);
+        new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
+
+        if (verbosity >= 1) {
+            LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
+            LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
+            LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_TEE("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
+            LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
+            LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
+        }
+    }
+
+    LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
+
+    // load tensors
+    {
+        std::vector<uint8_t> read_buf;
+        struct ggml_init_params params = {
+            /*.mem_size =*/ (n_tensors + 1) * ggml_tensor_overhead(),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc =*/ true,
+        };
+
+        new_clip->ctx_data = ggml_init(params);
+        if (!new_clip->ctx_data) {
+            LOG_TEE("%s: ggml_init() failed\n", __func__);
+            clip_free(new_clip);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+#ifdef _WIN32
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+        if (!wlen) {
+            return NULL;
+        }
+        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
+        if (!wlen) {
+            free(wbuf);
+            return NULL;
+        }
+#if __GLIBCXX__
+        int fd = _wopen(wbuf, _O_RDONLY | _O_BINARY);
+        __gnu_cxx::stdio_filebuf<char> buffer(fd, std::ios_base::in);
+        std::istream fin(&buffer);
+#else // MSVC
+        // unused in our current build
+        auto fin = std::ifstream(wbuf, std::ios::binary);
+#endif
+        free(wbuf);
+#else
+        auto fin = std::ifstream(fname, std::ios::binary);
+#endif
+        if (!fin) {
+            LOG_TEE("cannot open model file for loading tensors\n");
+            clip_free(new_clip);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+        // add tensors to context
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor * t = ggml_get_tensor(meta, name);
+            struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx_data, t);
+            ggml_set_name(cur, name);
+        }
+
+        // alloc memory and offload data
+        new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
+            const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
+            fin.seekg(offset, std::ios::beg);
+            if (!fin) {
+                LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
+                clip_free(new_clip);
+                gguf_free(ctx);
+                return nullptr;
+            }
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_clip->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
+        }
+#if defined(_WIN32) && defined(__GLIBCXX__)
+        close(fd);
+#else
+        fin.close();
+#endif
+    }
+
+    // vision model
+    if (new_clip->has_vision_encoder) {
+        // load vision model
+        auto & vision_model = new_clip->vision_model;
+        auto & hparams = vision_model.hparams;
+        hparams.hidden_size    = get_u32(ctx, format(KEY_N_EMBD, "vision"));
+        hparams.n_head         = get_u32(ctx, format(KEY_N_HEAD, "vision"));
+        hparams.n_intermediate = get_u32(ctx, format(KEY_N_FF, "vision"));
+        hparams.n_layer        = get_u32(ctx, format(KEY_N_BLOCK, "vision"));
+        hparams.image_size     = get_u32(ctx, KEY_IMAGE_SIZE);
+        hparams.patch_size     = get_u32(ctx, KEY_PATCH_SIZE);
+        hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision"));
+        hparams.eps            = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
+
+        try {
+            int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS);
+            int n = gguf_get_arr_n(ctx, idx);
+            const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx);
+            for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) {
+                hparams.image_grid_pinpoints[i] = pinpoints[i];
+            }
+            if (n < 32)
+                hparams.image_grid_pinpoints[n] = 0;
+        } catch (std::runtime_error & /*e*/) {
+            hparams.image_grid_pinpoints[0]=0;
+        }
+
+        try {
+            int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
+            strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx));
+        } catch (std::runtime_error & /*e*/) {
+            strcpy(hparams.mm_patch_merge_type, "flat");
+        }
+
+        try {
+            hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6
+        } catch(const std::exception& /*e*/) {
+            hparams.image_crop_resolution = hparams.image_size;
+        }
+
+        int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
+        int idx_std  = get_key_idx(ctx, KEY_IMAGE_STD);
+
+        const float * mean_data = (const float *)gguf_get_arr_data(ctx, idx_mean);
+        const float * std_data  = (const float *)gguf_get_arr_data(ctx, idx_std);
+
+        for (int i = 0; i < 3; ++i) {
+            new_clip->image_mean[i] = mean_data[i];
+            new_clip->image_std[i]  = std_data[i];
+        }
+
+        if (verbosity >= 2) {
+            LOG_TEE("\n%s: vision model hparams\n", __func__);
+            LOG_TEE("image_size         %d\n", hparams.image_size);
+            LOG_TEE("patch_size         %d\n", hparams.patch_size);
+            LOG_TEE("v_hidden_size      %d\n", hparams.hidden_size);
+            LOG_TEE("v_n_intermediate   %d\n", hparams.n_intermediate);
+            LOG_TEE("v_projection_dim   %d\n", hparams.projection_dim);
+            LOG_TEE("v_n_head           %d\n", hparams.n_head);
+            LOG_TEE("v_n_layer          %d\n", hparams.n_layer);
+            LOG_TEE("v_eps              %f\n", hparams.eps);
+            LOG_TEE("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
+            LOG_TEE("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
+            LOG_TEE("v_image_grid_pinpoints: ");
+            for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
+                LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
+            }
+            LOG_TEE("\n");
+            LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
+
+        }
+
+        try {
+            vision_model.class_embedding  = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
+            new_clip->has_class_embedding = true;
+        } catch (const std::exception& /*e*/) {
+            new_clip->has_class_embedding = false;
+        }
+
+        try {
+            vision_model.pre_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
+            vision_model.pre_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+            new_clip->has_pre_norm = true;
+        } catch (std::exception & /*e*/) {
+            new_clip->has_pre_norm = false;
+        }
+
+        try {
+            vision_model.post_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight"));
+            vision_model.post_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias"));
+            new_clip->has_post_norm = true;
+        } catch (std::exception & /*e*/) {
+            new_clip->has_post_norm = false;
+        }
+
+        try {
+            vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS);
+            new_clip->has_patch_bias = true;
+        } catch (std::exception & /*e*/) {
+            new_clip->has_patch_bias = false;
+        }
+
+        try {
+            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
+        } catch(const std::exception& /*e*/) {
+            LOG_TEE("%s: failed to load vision model tensors\n", __func__);
+        }
+
+        // LLaVA projection
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
+            vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+            try {
+                // Yi-type llava
+                vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight"));
+                vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias"));
+            } catch (std::runtime_error & /*e*/) { }
+            try {
+                // missing in Yi-type llava
+                vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+                vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            } catch (std::runtime_error & /*e*/) { }
+            try {
+                // Yi-type llava
+                vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight"));
+                vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias"));
+            } catch (std::runtime_error & /*e*/) { }
+            try {
+                // Yi-type llava
+                vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
+                vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
+            } catch (std::runtime_error & /*e*/) { }
+            try {
+                vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
+                // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
+            } catch (std::runtime_error & /*e*/) { }
+        } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projection
+            vision_model.mm_model_mlp_1_w               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight"));
+            vision_model.mm_model_mlp_1_b               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias"));
+            vision_model.mm_model_mlp_3_w               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight"));
+            vision_model.mm_model_mlp_3_b               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias"));
+            vision_model.mm_model_block_1_block_0_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight"));
+            vision_model.mm_model_block_1_block_0_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight"));
+            vision_model.mm_model_block_1_block_0_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias"));
+            vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight"));
+            vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias"));
+            vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight"));
+            vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias"));
+            vision_model.mm_model_block_1_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight"));
+            vision_model.mm_model_block_1_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight"));
+            vision_model.mm_model_block_1_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias"));
+            vision_model.mm_model_block_2_block_0_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight"));
+            vision_model.mm_model_block_2_block_0_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight"));
+            vision_model.mm_model_block_2_block_0_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias"));
+            vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight"));
+            vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias"));
+            vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight"));
+            vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias"));
+            vision_model.mm_model_block_2_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
+            vision_model.mm_model_block_2_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
+            vision_model.mm_model_block_2_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            // MobilVLM_V2 projection
+            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight"));
+            vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias"));
+            vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight"));
+            vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias"));
+            vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
+            vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD);
+            vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K);
+            vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY);
+            vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ);
+            vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ);
+            vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight"));
+            vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight"));
+            vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight"));
+            vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias"));
+            vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias"));
+            vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias"));
+            vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight"));
+            vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias"));
+            vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight"));
+            vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias"));
+            vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight"));
+            vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias"));
+            vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
+            vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
+        }
+        else {
+            std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
+            throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
+        }
+
+        vision_model.layers.resize(hparams.n_layer);
+
+        for (int il = 0; il < hparams.n_layer; ++il) {
+            auto & layer = vision_model.layers[il];
+            layer.k_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "weight"));
+            layer.q_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "weight"));
+            layer.v_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "weight"));
+            layer.o_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "weight"));
+            layer.ln_1_w = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "weight"));
+            layer.ln_2_w = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "weight"));
+            layer.ff_i_w = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "weight"));
+            layer.ff_o_w = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "weight"));
+            layer.k_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "bias"));
+            layer.q_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "bias"));
+            layer.v_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "bias"));
+            layer.o_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "bias"));
+            layer.ln_1_b = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "bias"));
+            layer.ln_2_b = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "bias"));
+            layer.ff_i_b = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "bias"));
+            layer.ff_o_b = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "bias"));
+        }
+    }
+
+    ggml_free(meta);
+
+    new_clip->ctx_gguf = ctx;
+
+    // measure mem requirement and allocate
+    {
+        new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
+        clip_image_f32_batch batch;
+        batch.size = 1;
+        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
+        ggml_gallocr_reserve(new_clip->compute_alloc, gf);
+        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
+        LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+    }
+
+    return new_clip;
+}
+
+void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
+    ctx_clip->load_image_size = load_image_size;
+}
+
+struct clip_image_size * clip_image_size_init() {
+    struct clip_image_size * load_image_size = new struct clip_image_size();
+    load_image_size->width = 448;
+    load_image_size->height = 448;
+    return load_image_size;
+}
+
+struct clip_image_u8 * clip_image_u8_init() {
+    return new clip_image_u8();
+}
+
+struct clip_image_f32 * clip_image_f32_init() {
+    return new clip_image_f32();
+}
+
+void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
+void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
+void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+
+static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
+    img->nx = nx;
+    img->ny = ny;
+    img->buf.resize(3 * nx * ny);
+    memcpy(img->buf.data(), data, img->buf.size());
+}
+
+bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
+    if (!data) {
+        LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
+    stbi_image_free(data);
+    return true;
+}
+
+bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
+    if (!data) {
+        LOG_TEE("%s: failed to decode image bytes\n", __func__);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
+    stbi_image_free(data);
+    return true;
+}
+
+// Linear interpolation between two points
+inline float clip_lerp(float s, float e, float t) {
+    return s + (e - s) * t;
+}
+// Bilinear resize function
+static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int target_width, int target_height) {
+    dst.nx = target_width;
+    dst.ny = target_height;
+    dst.buf.resize(3 * target_width * target_height);
+
+    float x_ratio = static_cast<float>(src.nx - 1) / target_width;
+    float y_ratio = static_cast<float>(src.ny - 1) / target_height;
+
+    for (int y = 0; y < target_height; y++) {
+        for (int x = 0; x < target_width; x++) {
+            float px = x_ratio * x;
+            float py = y_ratio * y;
+            int x_floor = static_cast<int>(px);
+            int y_floor = static_cast<int>(py);
+            float x_lerp = px - x_floor;
+            float y_lerp = py - y_floor;
+
+            for (int c = 0; c < 3; c++) {
+                float top = clip_lerp(
+                    static_cast<float>(src.buf[3 * (y_floor * src.nx + x_floor) + c]),
+                    static_cast<float>(src.buf[3 * (y_floor * src.nx + (x_floor + 1)) + c]),
+                    x_lerp
+                );
+                float bottom = clip_lerp(
+                    static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + x_floor) + c]),
+                    static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + (x_floor + 1)) + c]),
+                    x_lerp
+                );
+                dst.buf[3 * (y * target_width + x) + c] = static_cast<uint8_t>(clip_lerp(top, bottom, y_lerp));
+            }
+        }
+    }
+}
+
+// Normalize image to float32 - careful with pytorch .to(model.device, dtype=torch.float16) - this sometimes reduces precision (32>16>32), sometimes not
+static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* dst, const float mean[3], const float std[3]) {
+    dst->nx = src->nx;
+    dst->ny = src->ny;
+    dst->buf.resize(src->buf.size());
+
+    for (size_t i = 0; i < src->buf.size(); ++i) {
+        int c = i % 3; // rgb
+        dst->buf[i] = (static_cast<float>(src->buf[i]) / 255.0f - mean[c]) / std[c];
+    }
+}
+
+inline int clip(int x, int lower, int upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int target_width, int target_height) {
+    const int nx = img.nx;
+    const int ny = img.ny;
+
+    dst.nx = target_width;
+    dst.ny = target_height;
+    dst.buf.resize(3 * target_width * target_height);
+
+    float Cc;
+    float C[5];
+    float d0, d2, d3, a0, a1, a2, a3;
+    int i, j, k, jj;
+    int x, y;
+    float dx, dy;
+    float tx, ty;
+
+    tx = (float)nx / (float)target_width;
+    ty = (float)ny / (float)target_height;
+
+    // Bicubic interpolation; adapted from ViT.cpp, inspired from :
+    //    -> https://github.com/yglukhov/bicubic-interpolation-image-processing/blob/master/libimage.c#L36
+    //    -> https://en.wikipedia.org/wiki/Bicubic_interpolation
+
+    for (i = 0; i < target_height; i++) {
+        for (j = 0; j < target_width; j++) {
+            x = (int)(tx * j);
+            y = (int)(ty * i);
+
+            dx = tx * j - x;
+            dy = ty * i - y;
+
+            for (k = 0; k < 3; k++) {
+                for (jj = 0; jj <= 3; jj++) {
+                    d0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x - 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+                    d2 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+                    d3 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 2, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+                    a0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
+
+                    a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
+                    a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
+                    a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
+
+                    C[jj] = a0 + a1 * dx + a2 * dx * dx + a3 * dx * dx * dx;
+
+                    d0 = C[0] - C[1];
+                    d2 = C[2] - C[1];
+                    d3 = C[3] - C[1];
+                    a0 = C[1];
+                    a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
+                    a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
+                    a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
+                    Cc = a0 + a1 * dy + a2 * dy * dy + a3 * dy * dy * dy;
+
+                    const uint8_t Cc2 = std::min(std::max(std::round(Cc), 0.0f), 255.0f);
+                    dst.buf[(i * target_width + j) * 3 + k] = float(Cc2);
+                }
+            }
+        }
+    }
+
+    return true;
+}
+
+// llava-1.6 type of resize_and_pad (black)
+static void resize_and_pad_image(const clip_image_u8& image, clip_image_u8 &image_output, const std::pair<int, int>& target_resolution) {
+    int target_width = target_resolution.first;
+    int target_height = target_resolution.second;
+
+    float scale_w = static_cast<float>(target_width) / image.nx;
+    float scale_h = static_cast<float>(target_height) / image.ny;
+
+    int new_width, new_height;
+
+    if (scale_w < scale_h) {
+        new_width = target_width;
+        new_height = std::min(static_cast<int>(std::ceil(image.ny * scale_w)), target_height);
+    } else {
+        new_height = target_height;
+        new_width = std::min(static_cast<int>(std::ceil(image.nx * scale_h)), target_width);
+    }
+
+    clip_image_u8 resized_image;
+    // bilinear_resize(image, resized_image, new_width, new_height);
+    bicubic_resize(image, resized_image, new_width, new_height);
+
+    clip_image_u8 padded_image;
+    padded_image.nx = target_width;
+    padded_image.ny = target_height;
+    padded_image.buf.resize(3 * target_width * target_height, 0); // Initialize with black
+
+    // Calculate padding offsets
+    int pad_x = (target_width - new_width) / 2;
+    int pad_y = (target_height - new_height) / 2;
+
+    // Copy the resized image into the center of the padded buffer
+    for (int y = 0; y < new_height; ++y) {
+        for (int x = 0; x < new_width; ++x) {
+            for (int c = 0; c < 3; ++c) {
+                padded_image.buf[3 * ((y + pad_y) * target_width + (x + pad_x)) + c] = resized_image.buf[3 * (y * new_width + x) + c];
+            }
+        }
+    }
+    image_output = std::move(padded_image);
+}
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int> & original_size, const std::vector<std::pair<int, int>> & possible_resolutions) {
+    int original_width = original_size.first;
+    int original_height = original_size.second;
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & image, int patch_size) {
+    std::vector<clip_image_u8*> patches;
+    int width = image.nx;
+    int height = image.ny;
+    for (int i = 0; i < height; i += patch_size) {
+        for (int j = 0; j < width; j += patch_size) {
+            clip_image_u8 *patch = clip_image_u8_init();
+            patch->nx = std::min(patch_size, width - j);
+            patch->ny = std::min(patch_size, height - i);
+            patch->buf.resize(3 * patch->nx * patch->ny);
+            for (int y = 0; y < patch->ny; ++y) {
+                for (int x = 0; x < patch->nx; ++x) {
+                    for (int c = 0; c < 3; ++c) {
+                        patch->buf[3 * (y * patch->nx + x) + c] = image.buf[3 * ((i + y) * width + (j + x)) + c];
+                    }
+                }
+            }
+            patches.push_back(patch);
+        }
+    }
+    return patches;
+}
+
+static int ensure_divide(int length, int patch_size) {
+    return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
+}
+
+static std::pair<int, int> uhd_find_best_resize(std::pair<int, int> original_size, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width = original_size.first;
+    int height = original_size.second;
+    if ((width * height > scale_resolution * scale_resolution) || allow_upscale) {
+        float r = static_cast<float>(width) / height;
+        height = static_cast<int>(scale_resolution / std::sqrt(r));
+        width = static_cast<int>(height * r);
+    }
+    int best_width = ensure_divide(width, patch_size);
+    int best_height = ensure_divide(height, patch_size);
+    return std::make_pair(best_width, best_height);
+}
+
+static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size, std::pair<int, int> grid, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width, height;
+    std::tie(width, height) = original_size;
+    int grid_x, grid_y;
+    std::tie(grid_x, grid_y) = grid;
+
+    int refine_width = ensure_divide(width, grid_x);
+    int refine_height = ensure_divide(height, grid_y);
+
+    int grid_width = refine_width / grid_x;
+    int grid_height = refine_height / grid_y;
+
+   // auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line)
+    auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair
+    int best_grid_width, best_grid_height;
+    std::tie(best_grid_width, best_grid_height) = best_grid_size;
+
+  //  std::pair<int, int> refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line)
+    std::pair<int, int> refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line)
+    return refine_size;
+}
+
+static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
+    std::vector<int> candidate_split_grids_nums;
+    for (int i : {multiple - 1, multiple, multiple + 1}) {
+        if (i == 1 || i > max_slice_nums) {
+            continue;
+        }
+        candidate_split_grids_nums.push_back(i);
+    }
+
+    std::vector<std::pair<int, int>> candidate_grids;
+    for (int split_grids_nums : candidate_split_grids_nums) {
+        int m = 1;
+        while (m <= split_grids_nums) {
+            if (split_grids_nums % m == 0) {
+                candidate_grids.emplace_back(m, split_grids_nums / m);
+            }
+            ++m;
+        }
+    }
+
+    std::pair<int, int> best_grid{1, 1};
+    float min_error = std::numeric_limits<float>::infinity();
+    for (const auto& grid : candidate_grids) {
+        float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second));
+        if (error < min_error) {
+            best_grid = grid;
+            min_error = error;
+        }
+    }
+    return best_grid;
+}
+
+// inspired from LLaVA-UHD:
+//    -> https://arxiv.org/pdf/2403.11703
+//    -> https://github.com/thunlp/LLaVA-UHD
+//    -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118
+static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
+    const std::pair<int, int> original_size={img->nx,img->ny};
+    const int original_width = img->nx;
+    const int original_height = img->ny;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+
+    std::vector<std::vector<clip_image_u8 *>> images;
+    LOG_TEE("%s: multiple %d\n", __func__, multiple);
+    images.push_back(std::vector<clip_image_u8 *>());
+
+    if (multiple <= 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.resize(best_size, Image.Resampling.BICUBIC)
+        images[images.size()-1].push_back(source_image);
+    }
+    else if (multiple > 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
+        LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
+        images[images.size()-1].push_back(source_image);
+
+        std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+        LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
+
+        auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
+        clip_image_u8 * refine_image = clip_image_u8_init();
+        bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
+
+        LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
+
+        // split_to_patches
+        int width = refine_image->nx;
+        int height = refine_image->ny;
+        int grid_x = int(width / best_grid.first);
+        int grid_y = int(height / best_grid.second);
+        for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
+            images.push_back(std::vector<clip_image_u8 *>());
+            for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
+                clip_image_u8 * patch = clip_image_u8_init();
+                patch->nx = grid_x;
+                patch->ny = grid_y;
+                patch->buf.resize(3 * patch->nx * patch->ny);
+                for (int y = patches_i; y < patches_i + grid_y; ++y) {
+                    for (int x = patches_j; x < patches_j + grid_x; ++x) {
+                        const int i = 3 * (y * refine_image->nx + x);
+                        const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j));
+                        patch->buf[j]   = refine_image->buf[i];
+                        patch->buf[j+1] = refine_image->buf[i+1];
+                        patch->buf[j+2] = refine_image->buf[i+2];
+                    }
+                }
+                images[images.size()-1].push_back(patch);
+            }
+        }
+    }
+    return images;
+}
+
+int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
+    const int max_slice_nums=9;
+    const int scale_resolution=448;
+    const int original_width = ctx_clip->load_image_size->width;
+    const int original_height = ctx_clip->load_image_size->height;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+    std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+    return best_grid.first;
+}
+
+// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
+// res_imgs memory is being allocated here, previous allocations will be freed if found
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
+
+    if(clip_is_minicpmv(ctx)){
+        int max_slice_nums = 9;
+        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img, max_slice_nums);
+        res_imgs->size = 0;
+        for (size_t i = 0; i < imgs.size(); ++i){
+            res_imgs->size += imgs[i].size();
+        }
+        res_imgs->data = new clip_image_f32[res_imgs->size];
+        int idx = 0;
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            for (size_t j = 0; j < imgs[i].size(); ++j) {
+                LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
+                clip_image_f32 * res = clip_image_f32_init();
+                normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
+                res_imgs->data[idx++] = *res;
+                clip_image_f32_free(res);
+            }
+        }
+        return true;
+    }
+
+    bool pad_to_square = true;
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return false;
+    }
+    auto & params = ctx->vision_model.hparams;
+    // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing
+    if (strcmp(params.mm_patch_merge_type, "spatial_unpad") == 0) {
+        pad_to_square = false;
+    }
+    // free the previous res_imgs if any set
+    if (res_imgs->size > 0) {
+        clip_image_f32_batch_free(res_imgs);
+    }
+    res_imgs->data = nullptr;
+    res_imgs->size = 0;
+
+    // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
+    // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
+
+    clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily
+    if (pad_to_square && img->nx != img->ny) {
+        int longer_side = std::max(img->nx, img->ny);
+        temp->nx = longer_side;
+        temp->ny = longer_side;
+        temp->buf.resize(3 * longer_side * longer_side);
+        const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA (this is the mean rgb color * 255)
+
+        // fill with background color
+        for (size_t i = 0; i < temp->buf.size(); i++) {
+            temp->buf[i] = bc[i % 3];
+        }
+
+        // copy from the input image
+        for (int y = 0; y < img->ny; y++) {
+            for (int x = 0; x < img->nx; x++) {
+                const int i = 3 * (y * img->nx + x);
+                const int j = 3 * (y * temp->nx + x);
+                temp->buf[j]   = img->buf[i];
+                temp->buf[j+1] = img->buf[i+1];
+                temp->buf[j+2] = img->buf[i+2];
+            }
+        }
+    } else {
+        if (params.image_grid_pinpoints[0] != 0) {
+            // "spatial_unpad" with "anyres" processing for llava-1.6
+            std::vector<std::pair<int, int>> possible_resolutions;
+            for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+                possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+            }
+            std::pair<int, int> best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions);
+            // clip_image_save_to_bmp(*img, "input.bmp");
+            resize_and_pad_image(*img, *temp, best_resolution);  // we do not pad with mean-bg color anymore in llava-1.6
+            // clip_image_save_to_bmp(*temp, "resized.bmp");
+            // visually verify normalized image:
+            // normalize_image_u8_to_f32(*temp, *res, ctx->image_mean, ctx->image_std);
+            // {
+            //     clip_image_u8 * temp2 = clip_image_u8_init();
+            //     clip_image_convert_f32_to_u8(*res, *temp2);
+            //     clip_image_save_to_bmp(*temp2, "resized_normalized_f32.bmp");
+            //     clip_image_u8_free(temp2);
+            // }
+
+            std::vector<clip_image_u8 *> patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6)
+
+            clip_image_u8 *image_original_resize = clip_image_u8_init();
+            // bilinear_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
+            bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
+            patches.insert(patches.begin(), image_original_resize);
+            // clip_image_f32_batch_init(patches.size());
+            res_imgs->size = patches.size();
+            res_imgs->data = new clip_image_f32[res_imgs->size];
+            int num=0;
+            for (auto& patch : patches) {
+                normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std);
+                num++;
+            }
+
+            for (size_t i = 0; i < patches.size(); i++) {
+                // LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
+                clip_image_u8_free(patches[i]);
+            }
+
+            clip_image_u8_free(temp);
+
+            return true;
+        } else {
+            temp->nx = img->nx;
+            temp->ny = img->ny;
+            temp->buf.resize(img->buf.size());
+            memcpy(temp->buf.data(), img->buf.data(), temp->buf.size());
+        }
+    }
+
+    const int nx = temp->nx;
+    const int ny = temp->ny;
+    // clip_image_save_to_bmp(*temp, "resized_vanilla.bmp");
+
+    const int nx2 = ctx->vision_model.hparams.image_size;
+    const int ny2 = ctx->vision_model.hparams.image_size;
+    clip_image_f32 * res = clip_image_f32_init();
+    res->nx = nx2;
+    res->ny = ny2;
+    res->buf.resize(3 * nx2 * ny2);
+
+    const float scale = std::max(nx, ny) / (float)ctx->vision_model.hparams.image_size;
+
+    const int nx3 = int(nx / scale + 0.5f);
+    const int ny3 = int(ny / scale + 0.5f);
+
+    const auto & m3 = ctx->image_mean; // {0.48145466f, 0.4578275f, 0.40821073f};
+    const auto & s3 = ctx->image_std;  // {0.26862954f, 0.26130258f, 0.27577711f};
+
+    for (int y = 0; y < ny3; y++) {
+        for (int x = 0; x < nx3; x++) {
+            for (int c = 0; c < 3; c++) {
+                // linear interpolation
+                const float sx = (x + 0.5f) * scale - 0.5f;
+                const float sy = (y + 0.5f) * scale - 0.5f;
+
+                const int x0 = std::max(0, (int)std::floor(sx));
+                const int y0 = std::max(0, (int)std::floor(sy));
+
+                const int x1 = std::min(x0 + 1, nx - 1);
+                const int y1 = std::min(y0 + 1, ny - 1);
+
+                const float dx = sx - x0;
+                const float dy = sy - y0;
+
+                const int j00 = 3 * (y0 * nx + x0) + c;
+                const int j01 = 3 * (y0 * nx + x1) + c;
+                const int j10 = 3 * (y1 * nx + x0) + c;
+                const int j11 = 3 * (y1 * nx + x1) + c;
+
+                const float v00 = temp->buf[j00];
+                const float v01 = temp->buf[j01];
+                const float v10 = temp->buf[j10];
+                const float v11 = temp->buf[j11];
+
+                const float v0 = v00 * (1.0f - dx) + v01 * dx;
+                const float v1 = v10 * (1.0f - dx) + v11 * dx;
+
+                const float v = v0 * (1.0f - dy) + v1 * dy;
+
+                const uint8_t v2 = std::min(std::max(std::round(v), 0.0f), 255.0f);
+
+                const int i = 3 * (y * nx3 + x) + c;
+
+                res->buf[i] = ((float(v2) / 255.0f) - m3[c]) / s3[c];
+            }
+        }
+    }
+    clip_image_u8_free(temp);
+
+    // {
+    //     clip_image_u8 * temp2 = clip_image_u8_init();
+    //     clip_image_convert_f32_to_u8(*res, *temp2);
+    //     clip_image_save_to_bmp(*temp2, "resized_normalized_f32_vanilla.bmp");
+    //     clip_image_u8_free(temp2);
+    // }
+    // res_imgs.push_back(res);
+
+    res_imgs->size = 1;
+    res_imgs->data = new clip_image_f32[res_imgs->size];
+    res_imgs->data[0] = *res;
+    clip_image_f32_free(res);
+
+    return true;
+}
+
+ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
+    return ctx->vision_model.image_newline;
+}
+
+void clip_free(clip_ctx * ctx) {
+    ggml_free(ctx->ctx_data);
+    gguf_free(ctx->ctx_gguf);
+
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
+    delete ctx;
+}
+
+size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
+    return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
+}
+
+int32_t clip_image_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.image_size;
+}
+
+int32_t clip_patch_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.patch_size;
+}
+
+int32_t clip_hidden_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.hidden_size;
+}
+
+const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.mm_patch_merge_type;
+}
+
+const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.image_grid_pinpoints;
+}
+
+int clip_n_patches(const struct clip_ctx * ctx) {
+    const auto & params = ctx->vision_model.hparams;
+
+    int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
+
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        n_patches /= 4;
+    } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        if (ctx->minicpmv_version == 2) {
+            n_patches = 96;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            n_patches = 64;
+        }
+    }
+
+    return n_patches;
+}
+
+static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
+    assert(embed_dim % 2 == 0);
+    int H = pos.size();
+    int W = pos[0].size();
+
+    std::vector<float> omega(embed_dim / 2);
+    for (int i = 0; i < embed_dim / 2; ++i) {
+        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
+    }
+
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                float out_value = pos[h][w] * omega[d];
+                emb[h][w][d] = sin(out_value);
+                emb[h][w][d + embed_dim / 2] = cos(out_value);
+            }
+        }
+    }
+
+    return emb;
+}
+
+static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
+    assert(embed_dim % 2 == 0);
+    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
+    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
+
+    int H = emb_h.size();
+    int W = emb_h[0].size();
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                emb[h][w][d] = emb_h[h][w][d];
+                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
+            }
+        }
+    }
+    return emb;
+}
+
+static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
+    int grid_h_size = image_size.first;
+    int grid_w_size = image_size.second;
+
+    std::vector<float> grid_h(grid_h_size);
+    std::vector<float> grid_w(grid_w_size);
+
+    for (int i = 0; i < grid_h_size; ++i) {
+        grid_h[i] = static_cast<float>(i);
+    }
+    for (int i = 0; i < grid_w_size; ++i) {
+        grid_w[i] = static_cast<float>(i);
+    }
+
+    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid[h][w] = grid_w[w];
+        }
+    }
+    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid_2d[0][h][w] = grid_h[h];
+            grid_2d[1][h][w] = grid_w[w];
+        }
+    }
+
+    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
+
+    int H = image_size.first;
+    int W = image_size.second;
+    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
+        }
+    }
+
+    return pos_embed_2d;
+}
+
+bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return false;
+    }
+
+    clip_image_f32_batch imgs{};
+    imgs.size = 1;
+    imgs.data = img;
+    return clip_image_batch_encode(ctx, n_threads, &imgs, vec);
+}
+
+bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
+    if (!ctx->has_vision_encoder) {
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        return false;
+    }
+
+    int batch_size = imgs->size;
+    if (ctx->has_llava_projector) {
+        GGML_ASSERT(batch_size == 1); // TODO: support multiple images
+    }
+    if (ctx->has_minicpmv_projector) {
+        GGML_ASSERT(batch_size == 1);
+    }
+
+    // build the inference graph
+    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
+    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+
+    // set inputs
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        image_size_width  = imgs->data[0].nx;
+        image_size_height = imgs->data[0].ny;
+    }
+    const int patch_size    = hparams.patch_size;
+    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    if(ctx->load_image_size==nullptr){
+        ctx->load_image_size= clip_image_size_init();
+    }
+    const int pos_w = ctx->load_image_size->width/patch_size;
+    const int pos_h = ctx->load_image_size->height/patch_size;
+
+    {
+        struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
+        float * data = (float *)malloc(ggml_nbytes(inp_raw));
+
+        for (size_t i = 0; i < imgs->size; i++) {
+            const int nx = imgs->data[i].nx;
+            const int ny = imgs->data[i].ny;
+            if (!ctx->has_minicpmv_projector) {
+                GGML_ASSERT(nx == image_size && ny == image_size);
+            }
+
+            const int n = nx * ny;
+
+            for (int b = 0; b < batch_size; b++) {
+                for (int k = 0; k < 3; k++) {
+                    for (int y = 0; y < ny; y++) {
+                        for (int x = 0; x < nx; x++) {
+                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
+                        }
+                    }
+                }
+            }
+        }
+        ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
+        free(data);
+    }
+    if (ctx->has_minicpmv_projector) {
+        {
+            // inspired from siglip:
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            int bucket_coords_h[70];
+            int bucket_coords_w[70];
+            for (int i = 0; i < pos_h; i++){
+                bucket_coords_h[i] = std::floor(70.0*i/pos_h);
+            }
+            for (int i = 0; i < pos_w; i++){
+                bucket_coords_w[i] = std::floor(70.0*i/pos_w);
+            }
+            for (int i = 0, id = 0; i < pos_h; i++){
+                for (int j = 0; j < pos_w; j++){
+                    positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
+                }
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+
+        {
+            // inspired from resampler of Qwen-VL:
+            //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
+            //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
+            struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
+            int embed_dim = 4096;
+            if (ctx->minicpmv_version == 2) {
+                embed_dim = 4096;
+            }
+            else if (ctx->minicpmv_version == 3) {
+                embed_dim = 3584;
+            }
+            auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
+
+            float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
+            for(int i=0;i<pos_w * pos_h;++i){
+                for(int j=0;j<embed_dim;++j){
+                    pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
+                }
+            }
+
+            ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
+            free(pos_embed_data);
+        }
+    }
+    else{
+        {
+            if (ctx->has_class_embedding) {
+                struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
+
+                void* zero_mem = malloc(ggml_nbytes(embeddings));
+                memset(zero_mem, 0, ggml_nbytes(embeddings));
+                ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+                free(zero_mem);
+            }
+        }
+
+        {
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+
+        {
+            struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+            int* patches_data = (int*)malloc(ggml_nbytes(patches));
+            for (int i = 0; i < num_patches; i++) {
+                patches_data[i] = i + 1;
+            }
+            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+            free(patches_data);
+        }
+    }
+
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
+    }
+
+#ifdef GGML_USE_METAL
+    if (ggml_backend_is_metal(ctx->backend)) {
+        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
+    }
+#endif
+
+    ggml_backend_graph_compute(ctx->backend, gf);
+
+    // the last node is the embedding tensor
+    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+
+    // copy the embeddings to the location passed by the user
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
+
+    return true;
+}
+
+bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    assert(itype < GGML_TYPE_COUNT);
+    type = static_cast<ggml_type>(itype);
+
+    auto * ctx_clip = clip_model_load(fname_inp, 2);
+
+    const auto & ctx_src = ctx_clip->ctx_gguf;
+    const auto & ctx_data = ctx_clip->ctx_data;
+
+    auto * ctx_out = gguf_init_empty();
+    gguf_set_kv(ctx_out, ctx_src);
+    gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
+    gguf_set_val_u32(ctx_out, "general.file_type", itype);
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+
+    const int n_tensors = gguf_get_n_tensors(ctx_src);
+
+    for (int i = 0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(ctx_src, i);
+        struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
+        gguf_add_tensor(ctx_out, cur);
+    }
+
+    const size_t meta_size = gguf_get_meta_size(ctx_out);
+    for (size_t i = 0; i < meta_size; ++i) {
+        fout.put(0);
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> k_names = {
+        ".*weight",
+    };
+
+    std::vector<uint8_t> work(512);
+    std::vector<float> conv_buf(512);
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    for (int i = 0; i < n_tensors; ++i) {
+        const std::string name = gguf_get_tensor_name(ctx_src, i);
+        struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name.c_str());
+
+        enum ggml_type new_type;
+        void * new_data;
+        size_t new_size;
+
+        bool quantize = false;
+        for (const auto & s : k_names) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = true;
+                break;
+            }
+        }
+
+        // quantize only 2D tensors
+        quantize &= (ggml_n_dims(cur) == 2);
+
+        if (quantize) {
+            new_type = type;
+            if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
+                new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
+                // LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+            }
+            const size_t n_elms = ggml_nelements(cur);
+            float * f32_data;
+
+            switch (cur->type) {
+            case GGML_TYPE_F32:
+                f32_data = (float *)cur->data;
+                break;
+            case GGML_TYPE_F16:
+                if (conv_buf.size() < n_elms) {
+                    conv_buf.resize(n_elms);
+                }
+                for (size_t j = 0; j < n_elms; ++j) {
+                    conv_buf[j] = ggml_fp16_to_fp32(((ggml_fp16_t *)cur->data)[j]);
+                }
+                f32_data = (float *)conv_buf.data();
+                break;
+            default:
+                LOG_TEE("Please use an input file in f32 or f16\n");
+                gguf_free(ctx_out);
+                return false;
+            }
+
+            if (work.size() < n_elms * 4) {
+                work.resize(n_elms * 4);
+            }
+            new_data = work.data();
+
+            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr);
+        } else {
+            new_type = cur->type;
+            new_data = cur->data;
+            new_size = ggml_nbytes(cur);
+        }
+        const size_t orig_size = ggml_nbytes(cur);
+        total_size_org += orig_size;
+        total_size_new += new_size;
+        gguf_set_tensor_type(ctx_out, name.c_str(), new_type);
+        gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size);
+        fout.write((const char *)new_data, new_size);
+        size_t pad = GGML_PAD(new_size, gguf_get_alignment(ctx_out)) - new_size;
+        for (size_t j = 0; j < pad; ++j) {
+            fout.put(0);
+        }
+
+        LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+               orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+    }
+
+    // go back to beginning of file and write the updated metadata
+    fout.seekp(0, std::ios::beg);
+    std::vector<uint8_t> meta(meta_size);
+    gguf_get_meta_data(ctx_out, meta.data());
+    fout.write((const char *)meta.data(), meta_size);
+
+    fout.close();
+
+    clip_free(ctx_clip);
+    gguf_free(ctx_out);
+
+    {
+        LOG_TEE("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+    }
+
+    return true;
+}
+
+int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+        return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        return ctx->vision_model.mm_model_peg_0_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
+        return ctx->vision_model.mm_2_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        return ctx->vision_model.mm_3_b->ne[0];
+    }
+    if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        if (ctx->minicpmv_version == 2) {
+            return 4096;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            return 3584;
+        }
+    }
+
+    std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
+    throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
+}
+
+int clip_is_minicpmv(const struct clip_ctx * ctx) {
+    if (ctx->has_minicpmv_projector) {
+        return ctx->minicpmv_version;
+    }
+    return 0;
+}
--- a/llama/clip.h
+++ b/llama/clip.h
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef CLIP_H
+#define CLIP_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define CLIP_API __declspec(dllexport)
+#        else
+#            define CLIP_API __declspec(dllimport)
+#        endif
+#    else
+#        define CLIP_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define CLIP_API
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct clip_ctx;
+
+struct clip_image_size {
+    int width;
+    int height;
+};
+
+struct clip_image_u8_batch {
+    struct clip_image_u8 * data;
+    size_t size;
+};
+
+struct clip_image_f32_batch {
+    struct clip_image_f32 * data;
+    size_t size;
+};
+
+CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
+
+CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
+CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
+CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+
+CLIP_API struct clip_image_size * clip_image_size_init();
+CLIP_API struct clip_image_u8  * clip_image_u8_init ();
+CLIP_API struct clip_image_f32 * clip_image_f32_init();
+
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
+CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
+
+CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+
+/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
+CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
+
+/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
+CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
+
+CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
+
+CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // CLIP_H
--- a/llama/common.cpp
+++ b/llama/common.cpp
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
+#include "common.h"
+// Change JSON_ASSERT from assert() to GGML_ASSERT:
+#define JSON_ASSERT GGML_ASSERT
+#include "json.hpp"
+#include "json-schema-to-grammar.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <cmath>
+#include <codecvt>
+#include <cstdarg>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <iterator>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#if defined(__APPLE__) && defined(__MACH__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <locale>
+#include <windows.h>
+#include <fcntl.h>
+#include <io.h>
+#else
+#include <sys/ioctl.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#endif
+#if defined(LLAMA_USE_CURL)
+#include <curl/curl.h>
+#include <curl/easy.h>
+#include <thread>
+#include <future>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#if (defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL))
+#define GGML_USE_CUDA_SYCL
+#endif
+
+#if (defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)) || defined(GGML_USE_VULKAN)
+#define GGML_USE_CUDA_SYCL_VULKAN
+#endif
+
+#if defined(LLAMA_USE_CURL)
+#ifdef __linux__
+#include <linux/limits.h>
+#elif defined(_WIN32)
+#define PATH_MAX MAX_PATH
+#else
+#include <sys/syslimits.h>
+#endif
+#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
+#endif // LLAMA_USE_CURL
+
+using json = nlohmann::ordered_json;
+
+//
+// Environment variable utils
+//
+
+template<typename T>
+static typename std::enable_if<std::is_same<T, std::string>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::string(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<!std::is_same<T, bool>::value && std::is_integral<T>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::stoi(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<std::is_floating_point<T>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::stof(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<std::is_same<T, bool>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    if (value) {
+        std::string val(value);
+        target = val == "1" || val == "true";
+    }
+}
+
+//
+// CPU utils
+//
+
+int32_t cpu_get_num_physical_cores() {
+#ifdef __linux__
+    // enumerate the set of thread siblings, num entries is num cores
+    std::unordered_set<std::string> siblings;
+    for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
+        std::ifstream thread_siblings("/sys/devices/system/cpu/cpu"
+            + std::to_string(cpu) + "/topology/thread_siblings");
+        if (!thread_siblings.is_open()) {
+            break; // no more cpus
+        }
+        std::string line;
+        if (std::getline(thread_siblings, line)) {
+            siblings.insert(line);
+        }
+    }
+    if (!siblings.empty()) {
+        return static_cast<int32_t>(siblings.size());
+    }
+#elif defined(__APPLE__) && defined(__MACH__)
+    int32_t num_physical_cores;
+    size_t len = sizeof(num_physical_cores);
+    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    unsigned int n_threads_win = std::thread::hardware_concurrency();
+    unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4;
+
+    DWORD buffer_size = 0;
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
+        if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
+            return default_threads;
+        }
+    }
+
+    std::vector<char> buffer(buffer_size);
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
+        return default_threads;
+    }
+
+    int32_t num_physical_cores = 0;
+    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
+    while (buffer_size > 0) {
+        if (info->Relationship == RelationProcessorCore) {
+            num_physical_cores += info->Processor.GroupCount;
+        }
+        buffer_size -= info->Size;
+        info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
+    }
+
+    return num_physical_cores > 0 ? num_physical_cores : default_threads;
+#endif
+    unsigned int n_threads = std::thread::hardware_concurrency();
+    return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
+}
+
+#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
+#include <pthread.h>
+
+static void cpuid(unsigned leaf, unsigned subleaf,
+                  unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx) {
+    __asm__("movq\t%%rbx,%%rsi\n\t"
+            "cpuid\n\t"
+            "xchgq\t%%rbx,%%rsi"
+            : "=a"(*eax), "=S"(*ebx), "=c"(*ecx), "=d"(*edx)
+            : "0"(leaf), "2"(subleaf));
+}
+
+static int pin_cpu(int cpu) {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    CPU_SET(cpu, &mask);
+    return pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
+}
+
+static bool is_hybrid_cpu(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(7, 0, &eax, &ebx, &ecx, &edx);
+    return !!(edx & (1u << 15));
+}
+
+static bool is_running_on_efficiency_core(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(0x1a, 0, &eax, &ebx, &ecx, &edx);
+    int intel_atom = 0x20;
+    int core_type = (eax & 0xff000000u) >> 24;
+    return core_type == intel_atom;
+}
+
+static int cpu_count_math_cpus(int n_cpu) {
+    int result = 0;
+    for (int cpu = 0; cpu < n_cpu; ++cpu) {
+        if (pin_cpu(cpu)) {
+            return -1;
+        }
+        if (is_running_on_efficiency_core()) {
+            continue; // efficiency cores harm lockstep threading
+        }
+        ++cpu; // hyperthreading isn't useful for linear algebra
+        ++result;
+    }
+    return result;
+}
+
+#endif // __x86_64__ && __linux__
+
+/**
+ * Returns number of CPUs on system that are useful for math.
+ */
+int32_t cpu_get_num_math() {
+#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
+    int n_cpu = sysconf(_SC_NPROCESSORS_ONLN);
+    if (n_cpu < 1) {
+        return cpu_get_num_physical_cores();
+    }
+    if (is_hybrid_cpu()) {
+        cpu_set_t affinity;
+        if (!pthread_getaffinity_np(pthread_self(), sizeof(affinity), &affinity)) {
+            int result = cpu_count_math_cpus(n_cpu);
+            pthread_setaffinity_np(pthread_self(), sizeof(affinity), &affinity);
+            if (result > 0) {
+                return result;
+            }
+        }
+    }
+#endif
+    return cpu_get_num_physical_cores();
+}
+
+// Helper for setting process priority
+
+#if defined(_WIN32)
+
+bool set_process_priority(enum ggml_sched_priority prio) {
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        return true;
+    }
+
+    DWORD p = NORMAL_PRIORITY_CLASS;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
+        case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
+        case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS;     break;
+    }
+
+    if (!SetPriorityClass(GetCurrentProcess(), p)) {
+        fprintf(stderr, "warn: failed to set process priority class %d : (%d)\n", prio, (int) GetLastError());
+        return false;
+    }
+
+    return true;
+}
+
+#else // MacOS and POSIX
+#include <sys/types.h>
+#include <sys/resource.h>
+
+bool set_process_priority(enum ggml_sched_priority prio) {
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        return true;
+    }
+
+    int p = 0;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
+        case GGML_SCHED_PRIO_HIGH:     p = -10; break;
+        case GGML_SCHED_PRIO_REALTIME: p = -20; break;
+    }
+
+    if (!setpriority(PRIO_PROCESS, 0, p)) {
+        fprintf(stderr, "warn: failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
+        return false;
+    }
+    return true;
+}
+
+#endif
+
+//
+// CLI argument parsing
+//
+
+void gpt_params_handle_model_default(gpt_params & params) {
+    if (!params.hf_repo.empty()) {
+        // short-hand to avoid specifying --hf-file -> default it to --model
+        if (params.hf_file.empty()) {
+            if (params.model.empty()) {
+                throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
+            }
+            params.hf_file = params.model;
+        } else if (params.model.empty()) {
+            params.model = fs_get_cache_file(string_split(params.hf_file, '/').back());
+        }
+    } else if (!params.model_url.empty()) {
+        if (params.model.empty()) {
+            auto f = string_split(params.model_url, '#').front();
+            f = string_split(f, '?').front();
+            params.model = fs_get_cache_file(string_split(f, '/').back());
+        }
+    } else if (params.model.empty()) {
+        params.model = DEFAULT_MODEL_PATH;
+    }
+}
+
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
+    int32_t n_set = 0;
+
+    if (cpuparams.n_threads < 0) {
+        // Assuming everything about cpuparams is invalid
+        if (role_model != nullptr) {
+            cpuparams = *role_model;
+        } else {
+            cpuparams.n_threads = cpu_get_num_math();
+        }
+    }
+
+    for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
+        if (cpuparams.cpumask[i]) {
+            n_set++;
+        }
+    }
+
+    if (n_set && n_set < cpuparams.n_threads) {
+        // Not enough set bits, may experience performance issues.
+        fprintf(stderr, "warn: Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads);
+    }
+}
+
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
+    bool invalid_param = false;
+    std::string arg;
+    const std::string arg_prefix = "--";
+    llama_sampling_params & sparams = params.sparams;
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+        if (!gpt_params_find_arg(argc, argv, arg, params, i, invalid_param)) {
+            throw std::invalid_argument("error: unknown argument: " + arg);
+        }
+        if (invalid_param) {
+            throw std::invalid_argument("error: invalid parameter for argument: " + arg);
+        }
+    }
+
+    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
+    if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
+        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
+    }
+
+    gpt_params_handle_model_default(params);
+
+    if (params.hf_token.empty()) {
+        get_env("HF_TOKEN", params.hf_token);
+    }
+
+    if (params.escape) {
+        string_process_escapes(params.prompt);
+        string_process_escapes(params.input_prefix);
+        string_process_escapes(params.input_suffix);
+        string_process_escapes(sparams.cfg_negative_prompt);
+        for (auto & antiprompt : params.antiprompt) {
+            string_process_escapes(antiprompt);
+        }
+    }
+
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back();
+        params.kv_overrides.back().key[0] = 0;
+    }
+
+    return true;
+}
+
+void gpt_params_parse_from_env(gpt_params & params) {
+    // we only care about server-related params for now
+    get_env("LLAMA_ARG_MODEL",            params.model);
+    get_env("LLAMA_ARG_MODEL_URL",        params.model_url);
+    get_env("LLAMA_ARG_MODEL_ALIAS",      params.model_alias);
+    get_env("LLAMA_ARG_HF_REPO",          params.hf_repo);
+    get_env("LLAMA_ARG_HF_FILE",          params.hf_file);
+    get_env("LLAMA_ARG_THREADS",          params.cpuparams.n_threads);
+    get_env("LLAMA_ARG_CTX_SIZE",         params.n_ctx);
+    get_env("LLAMA_ARG_N_PARALLEL",       params.n_parallel);
+    get_env("LLAMA_ARG_BATCH",            params.n_batch);
+    get_env("LLAMA_ARG_UBATCH",           params.n_ubatch);
+    get_env("LLAMA_ARG_N_GPU_LAYERS",     params.n_gpu_layers);
+    get_env("LLAMA_ARG_THREADS_HTTP",     params.n_threads_http);
+    get_env("LLAMA_ARG_CHAT_TEMPLATE",    params.chat_template);
+    get_env("LLAMA_ARG_N_PREDICT",        params.n_predict);
+    get_env("LLAMA_ARG_ENDPOINT_METRICS", params.endpoint_metrics);
+    get_env("LLAMA_ARG_ENDPOINT_SLOTS",   params.endpoint_slots);
+    get_env("LLAMA_ARG_EMBEDDINGS",       params.embedding);
+    get_env("LLAMA_ARG_FLASH_ATTN",       params.flash_attn);
+    get_env("LLAMA_ARG_DEFRAG_THOLD",     params.defrag_thold);
+    get_env("LLAMA_ARG_CONT_BATCHING",    params.cont_batching);
+    get_env("LLAMA_ARG_HOST",             params.hostname);
+    get_env("LLAMA_ARG_PORT",             params.port);
+}
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    const auto params_org = params; // the example can modify the default params
+
+    try {
+        if (!gpt_params_parse_ex(argc, argv, params) || params.usage) {
+            params = params_org;
+            params.usage = true;
+            return false;
+        }
+    } catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        params = params_org;
+        return false;
+    }
+
+    return true;
+}
+
+bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) {
+    size_t dash_loc = range.find('-');
+    if (dash_loc == std::string::npos) {
+        fprintf(stderr, "Format of CPU range is invalid! Expected [<start>]-[<end>].\n");
+        return false;
+    }
+
+    size_t start_i;
+    size_t end_i;
+
+    if (dash_loc == 0) {
+        start_i = 0;
+    } else {
+        start_i = std::stoull(range.substr(0, dash_loc));
+        if (start_i >= GGML_MAX_N_THREADS) {
+            fprintf(stderr, "Start index out of bounds!\n");
+            return false;
+        }
+    }
+
+    if (dash_loc == range.length() - 1) {
+        end_i = GGML_MAX_N_THREADS - 1;
+    } else {
+        end_i = std::stoull(range.substr(dash_loc + 1));
+        if (end_i >= GGML_MAX_N_THREADS) {
+            fprintf(stderr, "End index out of bounds!\n");
+            return false;
+        }
+    }
+
+    for (size_t i = start_i; i <= end_i; i++) {
+        boolmask[i] = true;
+    }
+
+    return true;
+}
+
+bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) {
+    // Discard potential 0x prefix
+    size_t start_i = 0;
+    if (mask.length() >= 2 && mask.substr(0, 2) == "0x") {
+        start_i = 2;
+    }
+
+    size_t num_digits = mask.length() - start_i;
+    if (num_digits > 128) num_digits = 128;
+
+    size_t end_i = num_digits + start_i;
+
+    for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) {
+        char c = mask.at(i);
+        int8_t id = c;
+
+        if ((c >= '0' && c <= '9')) {
+            id -= '0';
+        } else if (c >= 'a' && c <= 'f') {
+            id -= 'a' - 10;
+        } else if (c >= 'A' && c <= 'F') {
+            id -= 'A' - 10;
+        } else {
+            fprintf(stderr, "Invalid hex character '%c' at position %d\n", c, int32_t(i));
+            return false;
+        }
+
+        boolmask[  n  ] = boolmask[  n  ] || ((id & 8) != 0);
+        boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0);
+        boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0);
+        boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0);
+    }
+
+    return true;
+}
+
+#define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; }
+
+bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) {
+    const char split_delim = ',';
+
+    llama_sampling_params & sparams = params.sparams;
+
+    if (arg == "-s" || arg == "--seed") {
+        CHECK_ARG
+        // TODO: this is temporary, in the future the sampling state will be moved fully to llama_sampling_context.
+        params.seed = std::stoul(argv[i]);
+        sparams.seed = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "-t" || arg == "--threads") {
+        CHECK_ARG
+        params.cpuparams.n_threads = std::stoi(argv[i]);
+        if (params.cpuparams.n_threads <= 0) {
+            params.cpuparams.n_threads = std::thread::hardware_concurrency();
+        }
+        return true;
+    }
+    if (arg == "-C" || arg == "--cpu-mask") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "-Cr" || arg == "--cpu-range") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "--prio") {
+        CHECK_ARG
+        params.cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict") {
+        CHECK_ARG
+        params.cpuparams.strict_cpu = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--poll") {
+        CHECK_ARG
+        params.cpuparams.poll = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "-tb" || arg == "--threads-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.n_threads = std::stoi(argv[i]);
+        if (params.cpuparams_batch.n_threads <= 0) {
+            params.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
+        }
+        return true;
+    }
+    if (arg == "-Cb" || arg == "--cpu-mask-batch") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "-Crb" || arg == "--cpu-range_batch") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "--prio-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-batch") {
+        params.cpuparams_batch.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.poll = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "-td" || arg == "--threads-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.n_threads = std::stoi(argv[i]);
+        if (params.draft_cpuparams.n_threads <= 0) {
+            params.draft_cpuparams.n_threads = std::thread::hardware_concurrency();
+        }
+        return true;
+    }
+        if (arg == "-Cd" || arg == "--cpu-mask-draft") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.draft_cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.draft_cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "-Crd" || arg == "--cpu-range-draft") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.draft_cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.draft_cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "--prio-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-draft") {
+        params.draft_cpuparams.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.poll = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "-tbd" || arg == "--threads-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]);
+        if (params.draft_cpuparams_batch.n_threads <= 0) {
+            params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency();
+        }
+        return true;
+    }
+    if (arg == "-Crbd" || arg == "--cpu-range-batch-draft") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.draft_cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.draft_cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "--prio-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-batch-draft") {
+        params.draft_cpuparams_batch.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.poll = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "-p" || arg == "--prompt") {
+        CHECK_ARG
+        params.prompt = argv[i];
+        return true;
+    }
+    if (arg == "-e" || arg == "--escape") {
+        params.escape = true;
+        return true;
+    }
+    if (arg == "--no-escape") {
+        params.escape = false;
+        return true;
+    }
+    if (arg == "--prompt-cache") {
+        CHECK_ARG
+        params.path_prompt_cache = argv[i];
+        return true;
+    }
+    if (arg == "--prompt-cache-all") {
+        params.prompt_cache_all = true;
+        return true;
+    }
+    if (arg == "--prompt-cache-ro") {
+        params.prompt_cache_ro = true;
+        return true;
+    }
+    if (arg == "-bf" || arg == "--binary-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i], std::ios::binary);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        // store the external file name in params
+        params.prompt_file = argv[i];
+        std::ostringstream ss;
+        ss << file.rdbuf();
+        params.prompt = ss.str();
+        fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), argv[i]);
+        return true;
+    }
+    if (arg == "-f" || arg == "--file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        // store the external file name in params
+        params.prompt_file = argv[i];
+        std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+        if (!params.prompt.empty() && params.prompt.back() == '\n') {
+            params.prompt.pop_back();
+        }
+        return true;
+    }
+    if (arg == "--in-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        params.in_files.push_back(argv[i]);
+        return true;
+    }
+    if (arg == "-n" || arg == "--predict" || arg == "--n-predict") {
+        CHECK_ARG
+        params.n_predict = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--top-k") {
+        CHECK_ARG
+        sparams.top_k = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-c" || arg == "--ctx-size") {
+        CHECK_ARG
+        params.n_ctx = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--grp-attn-n" || arg == "-gan") {
+        CHECK_ARG
+        params.grp_attn_n = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--grp-attn-w" || arg == "-gaw") {
+        CHECK_ARG
+        params.grp_attn_w = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--rope-freq-base") {
+        CHECK_ARG
+        params.rope_freq_base = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--rope-freq-scale") {
+        CHECK_ARG
+        params.rope_freq_scale = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--rope-scaling") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
+        else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
+        else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
+        else { invalid_param = true; }
+        return true;
+    }
+    if (arg == "--rope-scale") {
+        CHECK_ARG
+        params.rope_freq_scale = 1.0f / std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--yarn-orig-ctx") {
+        CHECK_ARG
+        params.yarn_orig_ctx = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--yarn-ext-factor") {
+        CHECK_ARG
+        params.yarn_ext_factor = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--yarn-attn-factor") {
+        CHECK_ARG
+        params.yarn_attn_factor = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--yarn-beta-fast") {
+        CHECK_ARG
+        params.yarn_beta_fast = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--yarn-beta-slow") {
+        CHECK_ARG
+        params.yarn_beta_slow = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--pooling") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; }
+        else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; }
+        else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; }
+        else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; }
+        else { invalid_param = true; }
+        return true;
+    }
+    if (arg == "--attention") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; }
+        else if (value == "non-causal") { params.attention_type = LLAMA_ATTENTION_TYPE_NON_CAUSAL; }
+        else { invalid_param = true; }
+        return true;
+    }
+    if (arg == "--defrag-thold" || arg == "-dt") {
+        CHECK_ARG
+        params.defrag_thold = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--samplers") {
+        CHECK_ARG
+        const auto sampler_names = string_split(argv[i], ';');
+        sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, true);
+        return true;
+    }
+    if (arg == "--sampling-seq") {
+        CHECK_ARG
+        sparams.samplers_sequence = llama_sampling_types_from_chars(argv[i]);
+        return true;
+    }
+    if (arg == "--top-p") {
+        CHECK_ARG
+        sparams.top_p = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--min-p") {
+        CHECK_ARG
+        sparams.min_p = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--temp") {
+        CHECK_ARG
+        sparams.temp = std::stof(argv[i]);
+        sparams.temp = std::max(sparams.temp, 0.0f);
+        return true;
+    }
+    if (arg == "--tfs") {
+        CHECK_ARG
+        sparams.tfs_z = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--typical") {
+        CHECK_ARG
+        sparams.typical_p = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--repeat-last-n") {
+        CHECK_ARG
+        sparams.penalty_last_n = std::stoi(argv[i]);
+        sparams.n_prev = std::max(sparams.n_prev, sparams.penalty_last_n);
+        return true;
+    }
+    if (arg == "--repeat-penalty") {
+        CHECK_ARG
+        sparams.penalty_repeat = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--frequency-penalty") {
+        CHECK_ARG
+        sparams.penalty_freq = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--presence-penalty") {
+        CHECK_ARG
+        sparams.penalty_present = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--dynatemp-range") {
+        CHECK_ARG
+        sparams.dynatemp_range = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--dynatemp-exp") {
+        CHECK_ARG
+        sparams.dynatemp_exponent = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--mirostat") {
+        CHECK_ARG
+        sparams.mirostat = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--mirostat-lr") {
+        CHECK_ARG
+        sparams.mirostat_eta = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--mirostat-ent") {
+        CHECK_ARG
+        sparams.mirostat_tau = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "--cfg-negative-prompt") {
+        CHECK_ARG
+        sparams.cfg_negative_prompt = argv[i];
+        return true;
+    }
+    if (arg == "--cfg-negative-prompt-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(sparams.cfg_negative_prompt));
+        if (!sparams.cfg_negative_prompt.empty() && sparams.cfg_negative_prompt.back() == '\n') {
+            sparams.cfg_negative_prompt.pop_back();
+        }
+        return true;
+    }
+    if (arg == "--cfg-scale") {
+        CHECK_ARG
+        sparams.cfg_scale = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "-b" || arg == "--batch-size") {
+        CHECK_ARG
+        params.n_batch = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-ub" || arg == "--ubatch-size") {
+        CHECK_ARG
+        params.n_ubatch = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--keep") {
+        CHECK_ARG
+        params.n_keep = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--draft") {
+        CHECK_ARG
+        params.n_draft = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--chunks") {
+        CHECK_ARG
+        params.n_chunks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-np" || arg == "--parallel") {
+        CHECK_ARG
+        params.n_parallel = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-ns" || arg == "--sequences") {
+        CHECK_ARG
+        params.n_sequences = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--p-split" || arg == "-ps") {
+        CHECK_ARG
+        params.p_split = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "-m" || arg == "--model") {
+        CHECK_ARG
+        params.model = argv[i];
+        return true;
+    }
+    if (arg == "-md" || arg == "--model-draft") {
+        CHECK_ARG
+        params.model_draft = argv[i];
+        return true;
+    }
+    if (arg == "-a" || arg == "--alias") {
+        CHECK_ARG
+        params.model_alias = argv[i];
+        return true;
+    }
+    if (arg == "-mu" || arg == "--model-url") {
+        CHECK_ARG
+        params.model_url = argv[i];
+        return true;
+    }
+    if (arg == "-hft" || arg == "--hf-token") {
+        if (++i >= argc) {
+          invalid_param = true;
+          return true;
+        }
+        params.hf_token = argv[i];
+        return true;
+    }
+    if (arg == "-hfr" || arg == "--hf-repo") {
+        CHECK_ARG
+        params.hf_repo = argv[i];
+        return true;
+    }
+    if (arg == "-hff" || arg == "--hf-file") {
+        CHECK_ARG
+        params.hf_file = argv[i];
+        return true;
+    }
+    if (arg == "--lora") {
+        CHECK_ARG
+        params.lora_adapters.push_back({
+            std::string(argv[i]),
+            1.0,
+        });
+        return true;
+    }
+    if (arg == "--lora-scaled") {
+        CHECK_ARG
+        std::string lora_adapter = argv[i];
+        CHECK_ARG
+        params.lora_adapters.push_back({
+            lora_adapter,
+            std::stof(argv[i]),
+        });
+        return true;
+    }
+    if (arg == "--lora-init-without-apply") {
+        params.lora_init_without_apply = true;
+        return true;
+    }
+    if (arg == "--control-vector") {
+        CHECK_ARG
+        params.control_vectors.push_back({ 1.0f, argv[i], });
+        return true;
+    }
+    if (arg == "--control-vector-scaled") {
+        CHECK_ARG
+        const char* fname = argv[i];
+        CHECK_ARG
+        params.control_vectors.push_back({ std::stof(argv[i]), fname, });
+        return true;
+    }
+    if (arg == "--control-vector-layer-range") {
+        CHECK_ARG
+        params.control_vector_layer_start = std::stoi(argv[i]);
+        CHECK_ARG
+        params.control_vector_layer_end = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--mmproj") {
+        CHECK_ARG
+        params.mmproj = argv[i];
+        return true;
+    }
+    if (arg == "--image") {
+        CHECK_ARG
+        params.image.emplace_back(argv[i]);
+        return true;
+    }
+    if (arg == "-i" || arg == "--interactive") {
+        params.interactive = true;
+        return true;
+    }
+    if (arg == "-sp" || arg == "--special") {
+        params.special = true;
+        return true;
+    }
+    if (arg == "--embedding" || arg == "--embeddings") {
+        params.embedding = true;
+        return true;
+    }
+    if (arg == "--embd-normalize") {
+        CHECK_ARG
+        params.embd_normalize = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--embd-output-format") {
+        CHECK_ARG
+        params.embd_out = argv[i];
+        return true;
+    }
+    if (arg == "--embd-separator") {
+        CHECK_ARG
+        params.embd_sep = argv[i];
+        return true;
+    }
+    if (arg == "-if" || arg == "--interactive-first") {
+        params.interactive_first = true;
+        return true;
+    }
+    if (arg == "-cnv" || arg == "--conversation") {
+        params.conversation = true;
+        return true;
+    }
+    if (arg == "--infill") {
+        params.infill = true;
+        return true;
+    }
+    if (arg == "-dkvc" || arg == "--dump-kv-cache") {
+        params.dump_kv_cache = true;
+        return true;
+    }
+    if (arg == "-nkvo" || arg == "--no-kv-offload") {
+        params.no_kv_offload = true;
+        return true;
+    }
+    if (arg == "-ctk" || arg == "--cache-type-k") {
+        params.cache_type_k = argv[++i];
+        return true;
+    }
+    if (arg == "-ctv" || arg == "--cache-type-v") {
+        params.cache_type_v = argv[++i];
+        return true;
+    }
+    if (arg == "-mli" || arg == "--multiline-input") {
+        params.multiline_input = true;
+        return true;
+    }
+    if (arg == "--simple-io") {
+        params.simple_io = true;
+        return true;
+    }
+    if (arg == "-cb" || arg == "--cont-batching") {
+        params.cont_batching = true;
+        return true;
+    }
+    if (arg == "-nocb" || arg == "--no-cont-batching") {
+        params.cont_batching = false;
+        return true;
+    }
+    if (arg == "-fa" || arg == "--flash-attn") {
+        params.flash_attn = true;
+        return true;
+    }
+    if (arg == "-co" || arg == "--color") {
+        params.use_color = true;
+        return true;
+    }
+    if (arg == "--mlock") {
+        params.use_mlock = true;
+        return true;
+    }
+    if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+        CHECK_ARG
+        params.n_gpu_layers = std::stoi(argv[i]);
+        if (!llama_supports_gpu_offload()) {
+            fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+        }
+        return true;
+    }
+    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--n-gpu-layers-draft") {
+        CHECK_ARG
+        params.n_gpu_layers_draft = std::stoi(argv[i]);
+        if (!llama_supports_gpu_offload()) {
+            fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+        }
+        return true;
+    }
+    if (arg == "--main-gpu" || arg == "-mg") {
+        CHECK_ARG
+        params.main_gpu = std::stoi(argv[i]);
+#ifndef GGML_USE_CUDA_SYCL_VULKAN
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL_VULKAN
+        return true;
+    }
+    if (arg == "--split-mode" || arg == "-sm") {
+        CHECK_ARG
+        std::string arg_next = argv[i];
+        if (arg_next == "none") {
+            params.split_mode = LLAMA_SPLIT_MODE_NONE;
+        }
+        else if (arg_next == "layer") {
+            params.split_mode = LLAMA_SPLIT_MODE_LAYER;
+        }
+        else if (arg_next == "row") {
+#ifdef GGML_USE_SYCL
+            fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n");
+            exit(1);
+#endif // GGML_USE_SYCL
+            params.split_mode = LLAMA_SPLIT_MODE_ROW;
+        }
+        else {
+            invalid_param = true;
+            return true;
+        }
+#ifndef GGML_USE_CUDA_SYCL_VULKAN
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL_VULKAN
+        return true;
+    }
+    if (arg == "--tensor-split" || arg == "-ts") {
+        CHECK_ARG
+        std::string arg_next = argv[i];
+
+        // split string by , and /
+        const std::regex regex{ R"([,/]+)" };
+        std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
+        std::vector<std::string> split_arg{ it, {} };
+        if (split_arg.size() >= llama_max_devices()) {
+            invalid_param = true;
+            return true;
+        }
+        for (size_t i = 0; i < llama_max_devices(); ++i) {
+            if (i < split_arg.size()) {
+                params.tensor_split[i] = std::stof(split_arg[i]);
+            }
+            else {
+                params.tensor_split[i] = 0.0f;
+            }
+        }
+#ifndef GGML_USE_CUDA_SYCL_VULKAN
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting a tensor split has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL_VULKAN
+        return true;
+    }
+    if (arg == "--rpc") {
+        CHECK_ARG
+        params.rpc_servers = argv[i];
+        return true;
+    }
+    if (arg == "--no-mmap") {
+        params.use_mmap = false;
+        return true;
+    }
+    if (arg == "--numa") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+        else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+        else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+        else { invalid_param = true; }
+        return true;
+    }
+    if (arg == "-v" || arg == "--verbose") {
+        params.verbosity = 1;
+        return true;
+    }
+    if (arg == "--verbosity") {
+        CHECK_ARG
+        params.verbosity = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--verbose-prompt") {
+        params.verbose_prompt = true;
+        return true;
+    }
+    if (arg == "--no-display-prompt") {
+        params.display_prompt = false;
+        return true;
+    }
+    if (arg == "-r" || arg == "--reverse-prompt") {
+        CHECK_ARG
+        params.antiprompt.emplace_back(argv[i]);
+        return true;
+    }
+    if (arg == "-ld" || arg == "--logdir") {
+        CHECK_ARG
+        params.logdir = argv[i];
+
+        if (params.logdir.back() != DIRECTORY_SEPARATOR) {
+            params.logdir += DIRECTORY_SEPARATOR;
+        }
+        return true;
+    }
+    if (arg == "-lcs" || arg == "--lookup-cache-static") {
+        CHECK_ARG
+        params.lookup_cache_static = argv[i];
+        return true;
+    }
+    if (arg == "-lcd" || arg == "--lookup-cache-dynamic") {
+        CHECK_ARG
+        params.lookup_cache_dynamic = argv[i];
+        return true;
+    }
+    if (arg == "--save-all-logits" || arg == "--kl-divergence-base") {
+        CHECK_ARG
+        params.logits_file = argv[i];
+        return true;
+    }
+    if (arg == "--perplexity" || arg == "--all-logits") {
+        params.logits_all = true;
+        return true;
+    }
+    if (arg == "--ppl-stride") {
+        CHECK_ARG
+        params.ppl_stride = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--ppl-output-type") {
+        CHECK_ARG
+        params.ppl_output_type = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-ptc" || arg == "--print-token-count") {
+        CHECK_ARG
+        params.n_print = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--check-tensors") {
+        params.check_tensors = true;
+        return true;
+    }
+    if (arg == "--hellaswag") {
+        params.hellaswag = true;
+        return true;
+    }
+    if (arg == "--hellaswag-tasks") {
+        CHECK_ARG
+        params.hellaswag_tasks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--winogrande") {
+        params.winogrande = true;
+        return true;
+    }
+    if (arg == "--winogrande-tasks") {
+        CHECK_ARG
+        params.winogrande_tasks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--multiple-choice") {
+        params.multiple_choice = true;
+        return true;
+    }
+    if (arg == "--multiple-choice-tasks") {
+        CHECK_ARG
+        params.multiple_choice_tasks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--kl-divergence") {
+        params.kl_divergence = true;
+        return true;
+    }
+    if (arg == "--ignore-eos") {
+        params.ignore_eos = true;
+        return true;
+    }
+    if (arg == "--penalize-nl") {
+        sparams.penalize_nl = true;
+        return true;
+    }
+    if (arg == "-l" || arg == "--logit-bias") {
+        CHECK_ARG
+        std::stringstream ss(argv[i]);
+        llama_token key;
+        char sign;
+        std::string value_str;
+        try {
+            if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
+                sparams.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
+            }
+            else {
+                throw std::exception();
+            }
+        }
+        catch (const std::exception&) {
+            invalid_param = true;
+            return true;
+        }
+        return true;
+    }
+    if (arg == "-h" || arg == "--help" || arg == "--usage"  ) {
+        params.usage = true;
+        return true;
+    }
+    if (arg == "--version") {
+        fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+        fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+        exit(0);
+    }
+    if (arg == "--in-prefix-bos") {
+        params.input_prefix_bos = true;
+        params.enable_chat_template = false;
+        return true;
+    }
+    if (arg == "--in-prefix") {
+        CHECK_ARG
+        params.input_prefix = argv[i];
+        params.enable_chat_template = false;
+        return true;
+    }
+    if (arg == "--in-suffix") {
+        CHECK_ARG
+        params.input_suffix = argv[i];
+        params.enable_chat_template = false;
+        return true;
+    }
+    if (arg == "--spm-infill") {
+        params.spm_infill = true;
+        return true;
+    }
+    if (arg == "--grammar") {
+        CHECK_ARG
+        sparams.grammar = argv[i];
+        return true;
+    }
+    if (arg == "--grammar-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        std::copy(
+            std::istreambuf_iterator<char>(file),
+            std::istreambuf_iterator<char>(),
+            std::back_inserter(sparams.grammar)
+        );
+        return true;
+    }
+    if (arg == "-j" || arg == "--json-schema") {
+        CHECK_ARG
+        sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
+        return true;
+    }
+    if (arg == "--override-kv") {
+        CHECK_ARG
+        if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
+            fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        return true;
+    }
+    if (arg == "--host") {
+        CHECK_ARG
+        params.hostname = argv[i];
+        return true;
+    }
+    if (arg == "--port") {
+        CHECK_ARG
+        params.port = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--path") {
+        CHECK_ARG
+        params.public_path = argv[i];
+        return true;
+    }
+    if (arg == "--api-key") {
+        CHECK_ARG
+        params.api_keys.push_back(argv[i]);
+        return true;
+    }
+    if (arg == "--api-key-file") {
+        CHECK_ARG
+        std::ifstream key_file(argv[i]);
+        if (!key_file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        std::string key;
+        while (std::getline(key_file, key)) {
+            if (!key.empty()) {
+                params.api_keys.push_back(key);
+            }
+        }
+        key_file.close();
+        return true;
+    }
+    if (arg == "--ssl-key-file") {
+        CHECK_ARG
+        params.ssl_file_key = argv[i];
+        return true;
+    }
+    if (arg == "--ssl-cert-file") {
+        CHECK_ARG
+        params.ssl_file_cert = argv[i];
+        return true;
+    }
+    if (arg == "--timeout" || arg == "-to") {
+        CHECK_ARG
+        params.timeout_read  = std::stoi(argv[i]);
+        params.timeout_write = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--threads-http") {
+        CHECK_ARG
+        params.n_threads_http = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-spf" || arg == "--system-prompt-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        std::string system_prompt;
+        std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(system_prompt)
+                );
+        params.system_prompt = system_prompt;
+        return true;
+    }
+    if (arg == "--log-format") {
+        CHECK_ARG
+        if (std::strcmp(argv[i], "json") == 0) {
+            params.log_json = true;
+        } else if (std::strcmp(argv[i], "text") == 0) {
+            params.log_json = false;
+        } else {
+            invalid_param = true;
+            return true;
+        }
+        return true;
+    }
+    if (arg == "--no-slots") {
+        params.endpoint_slots = false;
+        return true;
+    }
+    if (arg == "--metrics") {
+        params.endpoint_metrics = true;
+        return true;
+    }
+    if (arg == "--slot-save-path") {
+        CHECK_ARG
+        params.slot_save_path = argv[i];
+        // if doesn't end with DIRECTORY_SEPARATOR, add it
+        if (!params.slot_save_path.empty() && params.slot_save_path[params.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) {
+            params.slot_save_path += DIRECTORY_SEPARATOR;
+        }
+        return true;
+    }
+    if (arg == "--chat-template") {
+        CHECK_ARG
+        if (!llama_chat_verify_template(argv[i])) {
+            fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
+            fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
+            invalid_param = true;
+            return true;
+        }
+        params.chat_template = argv[i];
+        return true;
+    }
+    if (arg == "--slot-prompt-similarity" || arg == "-sps") {
+        CHECK_ARG
+        params.slot_prompt_similarity = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "-pps") {
+        params.is_pp_shared = true;
+        return true;
+    }
+    if (arg == "-npp") {
+        CHECK_ARG
+        auto p = string_split<int>(argv[i], split_delim);
+        params.n_pp.insert(params.n_pp.end(), p.begin(), p.end());
+        return true;
+    }
+    if (arg == "-ntg") {
+        CHECK_ARG
+        auto p = string_split<int>(argv[i], split_delim);
+        params.n_tg.insert(params.n_tg.end(), p.begin(), p.end());
+        return true;
+    }
+    if (arg == "-npl") {
+        CHECK_ARG
+        auto p = string_split<int>(argv[i], split_delim);
+        params.n_pl.insert(params.n_pl.end(), p.begin(), p.end());
+        return true;
+    }
+    if (arg == "--context-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i], std::ios::binary);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        params.context_files.push_back(argv[i]);
+        return true;
+    }
+    if (arg == "--chunk-size") {
+        CHECK_ARG
+        params.chunk_size = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--chunk-separator") {
+        CHECK_ARG
+        params.chunk_separator = argv[i];
+        return true;
+    }
+    if (arg == "--junk") {
+        CHECK_ARG
+        params.n_junk = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--pos") {
+        CHECK_ARG
+        params.i_pos = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-o" || arg == "--output" || arg == "--output-file") {
+        CHECK_ARG
+        params.out_file = argv[i];
+        params.cvector_outfile = argv[i];
+        params.lora_outfile = argv[i];
+        return true;
+    }
+    if (arg == "-ofreq" || arg == "--output-frequency") {
+        CHECK_ARG
+        params.n_out_freq = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--save-frequency") {
+        CHECK_ARG
+        params.n_save_freq = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--process-output") {
+        params.process_output = true;
+        return true;
+    }
+    if (arg == "--no-ppl") {
+        params.compute_ppl = false;
+        return true;
+    }
+    if (arg == "--chunk" || arg == "--from-chunk") {
+        CHECK_ARG
+        params.i_chunk = std::stoi(argv[i]);
+        return true;
+    }
+    // cvector params
+    if (arg == "--positive-file") {
+        CHECK_ARG
+        params.cvector_positive_file = argv[i];
+        return true;
+    }
+    if (arg == "--negative-file") {
+        CHECK_ARG
+        params.cvector_negative_file = argv[i];
+        return true;
+    }
+    if (arg == "--pca-batch") {
+        CHECK_ARG
+        params.n_pca_batch = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--pca-iter") {
+        CHECK_ARG
+        params.n_pca_iterations = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--method") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "pca") { params.cvector_dimre_method = DIMRE_METHOD_PCA; }
+        else if (value == "mean") { params.cvector_dimre_method = DIMRE_METHOD_MEAN; }
+        else { invalid_param = true; }
+        return true;
+    }
+    if (arg == "--no-warmup") {
+        params.warmup = false;
+        return true;
+    }
+#ifndef LOG_DISABLE_LOGS
+    // Parse args for logging parameters
+    if (log_param_single_parse(argv[i])) {
+        // Do nothing, log_param_single_parse automatically does it's thing
+        //  and returns if a match was found and parsed.
+        return true;
+    }
+    if (log_param_pair_parse( /*check_but_dont_parse*/ true, argv[i])) {
+        // We have a matching known parameter requiring an argument,
+        //  now we need to check if there is anything after this argv
+        //  and flag invalid_param or parse it.
+        CHECK_ARG
+        if (!log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i - 1], argv[i])) {
+            invalid_param = true;
+            return true;
+        }
+        return true;
+    }
+    // End of Parse args for logging parameters
+#endif // LOG_DISABLE_LOGS
+
+    return false;
+}
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
+#endif
+
+void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
+    const llama_sampling_params & sparams = params.sparams;
+
+    std::string sampler_type_chars;
+    std::string sampler_type_names;
+    for (const auto sampler_type : sparams.samplers_sequence) {
+        sampler_type_chars += static_cast<char>(sampler_type);
+        sampler_type_names += llama_sampling_type_to_str(sampler_type) + ";";
+    }
+    sampler_type_names.pop_back();
+
+    struct option_info {
+        LLAMA_COMMON_ATTRIBUTE_FORMAT(4, 5)
+        option_info(const std::string & tags, const char * args, const char * desc, ...) : tags(tags), args(args), desc(desc) {
+            va_list args_list;
+            va_start(args_list, desc);
+            char buffer[1024];
+            vsnprintf(buffer, sizeof(buffer), desc, args_list);
+            va_end(args_list);
+            this->desc = buffer;
+        }
+
+        option_info(const std::string & grp) : grp(grp) {}
+
+        std::string tags;
+        std::string args;
+        std::string desc;
+        std::string grp;
+    };
+
+    std::vector<option_info> options;
+
+    // TODO: filter by tags
+
+    options.push_back({ "general" });
+    options.push_back({ "*",           "-h,    --help, --usage",        "print usage and exit" });
+    options.push_back({ "*",           "       --version",              "show version and build info" });
+    options.push_back({ "*",           "-v,    --verbose",              "print verbose information" });
+    options.push_back({ "*",           "       --verbosity N",          "set specific verbosity level (default: %d)", params.verbosity });
+    options.push_back({ "*",           "       --verbose-prompt",       "print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false" });
+    options.push_back({ "*",           "       --no-display-prompt",    "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" });
+    options.push_back({ "*",           "-co,   --color",                "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" });
+    options.push_back({ "*",           "-s,    --seed SEED",            "RNG seed (default: %d, use random seed for < 0)", params.seed });
+    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during generation (default: %d)", params.cpuparams.n_threads });
+    options.push_back({ "*",           "-tb,   --threads-batch N",      "number of threads to use during batch and prompt processing (default: same as --threads)" });
+    options.push_back({ "speculative", "-td,   --threads-draft N",      "number of threads to use during generation (default: same as --threads)" });
+    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N","number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
+
+#ifndef GGML_USE_OPENMP
+    // these options are available only with the internal threadpool
+    options.push_back({ "*",           "-C,    --cpu-mask M",            "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")"});
+    options.push_back({ "*",           "-Cr,   --cpu-range lo-hi",       "range of CPUs for affinity. Complements --cpu-mask"});
+    options.push_back({ "*",           "       --cpu-strict <0|1>",      "use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu});
+    options.push_back({ "*",           "       --priority N",            "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority});
+    options.push_back({ "*",           "       --poll <0...100>",        "use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll});
+
+    options.push_back({ "*",           "-Cb,   --cpu-mask-batch M",      "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)"});
+    options.push_back({ "*",           "-Crb,  --cpu-range-batch lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch"});
+    options.push_back({ "*",           "       --cpu-strict-batch <0|1>","use strict CPU placement (default: same as --cpu-strict)"});
+    options.push_back({ "*",           "       --priority-batch N",      "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority)"});
+    options.push_back({ "*",           "       --poll-batch <0|1>",      "use polling to wait for work (default: same as --poll"});
+
+    options.push_back({ "speculative", "-Cd,   --cpu-mask-draft M",      "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)"});
+    options.push_back({ "speculative", "-Crd,  --cpu-range-draft lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft"});
+    options.push_back({ "speculative", "       --cpu-strict-draft <0|1>","Use strict CPU placement for draft model (default: same as --cpu-strict)"});
+    options.push_back({ "speculative", "       --priority-draft N",      "Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: same as --priority)"});
+    options.push_back({ "speculative", "       --poll-draft <0|1>",      "Use polling to wait for draft model work (default: same as --poll])"});
+
+    options.push_back({ "speculative", "-Cbd,  --cpu-mask-batch-draft M","Draft model CPU affinity mask. Complements cpu-range-draft-batch (default: same as --cpu-mask-draft)"});
+    options.push_back({ "speculative", "-Crbd, --cpu-range-batch-draft lo-hi",
+                                                                         "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)"});
+    options.push_back({ "speculative", "       --cpu-strict-batch-draft <0|1>",
+                                                                         "Use strict CPU placement for draft model (default: --cpu-strict-draft)"});
+    options.push_back({ "speculative", "       --priority-batch-draft N","Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority-draft)"});
+    options.push_back({ "speculative", "       --poll-batch-draft <0|1>","Use polling to wait for draft model work (default: --poll-draft)"});
+#endif // GGML_USE_OPENMP
+
+    options.push_back({ "speculative", "       --draft N",              "number of tokens to draft for speculative decoding (default: %d)", params.n_draft });
+    options.push_back({ "speculative", "-ps,   --p-split N",            "speculative decoding split probability (default: %.1f)", (double)params.p_split });
+    options.push_back({ "*",           "-lcs,  --lookup-cache-static FNAME",
+                                                                        "path to static lookup cache to use for lookup decoding (not updated by generation)" });
+    options.push_back({ "*",           "-lcd,  --lookup-cache-dynamic FNAME",
+                                                                        "path to dynamic lookup cache to use for lookup decoding (updated by generation)" });
+
+    options.push_back({ "*",           "-c,    --ctx-size N",           "size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx });
+    options.push_back({ "*",           "-n,    --predict N",            "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict });
+    options.push_back({ "*",           "-b,    --batch-size N",         "logical maximum batch size (default: %d)", params.n_batch });
+    options.push_back({ "*",           "-ub,   --ubatch-size N",        "physical maximum batch size (default: %d)", params.n_ubatch });
+    options.push_back({ "*",           "       --keep N",               "number of tokens to keep from the initial prompt (default: %d, -1 = all)", params.n_keep });
+    options.push_back({ "*",           "       --chunks N",             "max number of chunks to process (default: %d, -1 = all)", params.n_chunks });
+    options.push_back({ "*",           "-fa,   --flash-attn",           "enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled" });
+    options.push_back({ "*",           "-p,    --prompt PROMPT",        "prompt to start generation with\n"
+                                                                        "in conversation mode, this will be used as system prompt\n"
+                                                                        "(default: '%s')", params.prompt.c_str() });
+    options.push_back({ "*",           "-f,    --file FNAME",           "a file containing the prompt (default: none)" });
+    options.push_back({ "*",           "       --in-file FNAME",        "an input file (repeat to specify multiple files)" });
+    options.push_back({ "*",           "-bf,   --binary-file FNAME",    "binary file containing the prompt (default: none)" });
+    options.push_back({ "*",           "-e,    --escape",               "process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false" });
+    options.push_back({ "*",           "       --no-escape",            "do not process escape sequences" });
+    options.push_back({ "main",        "-ptc,  --print-token-count N",  "print token count every N tokens (default: %d)", params.n_print });
+    options.push_back({ "main",        "       --prompt-cache FNAME",   "file to cache prompt state for faster startup (default: none)" });
+    options.push_back({ "main",        "       --prompt-cache-all",     "if specified, saves user input and generations to cache as well\n"
+                                                                        "not supported with --interactive or other interactive options" });
+    options.push_back({ "main",        "       --prompt-cache-ro",      "if specified, uses the prompt cache but does not update it" });
+    options.push_back({ "main",        "-r,    --reverse-prompt PROMPT",
+                                                                        "halt generation at PROMPT, return control in interactive mode\n"
+                                                                        "can be specified more than once for multiple prompts" });
+    options.push_back({ "main",        "-sp,   --special",              "special tokens output enabled (default: %s)", params.special ? "true" : "false" });
+    options.push_back({ "main",        "-cnv,  --conversation",         "run in conversation mode, does not print special tokens and suffix/prefix\n"
+                                                                        "if suffix/prefix are not specified, default chat template will be used\n"
+                                                                        "(default: %s)", params.conversation ? "true" : "false" });
+    options.push_back({ "main infill", "-i,    --interactive",          "run in interactive mode (default: %s)", params.interactive ? "true" : "false" });
+    options.push_back({ "main infill", "-if,   --interactive-first",    "run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false" });
+    options.push_back({ "main infill", "-mli,  --multiline-input",      "allows you to write or paste multiple lines without ending each in '\\'" });
+    options.push_back({ "main infill", "       --in-prefix-bos",        "prefix BOS to user inputs, preceding the `--in-prefix` string" });
+    options.push_back({ "main infill", "       --in-prefix STRING",     "string to prefix user inputs with (default: empty)" });
+    options.push_back({ "main infill", "       --in-suffix STRING",     "string to suffix after user inputs with (default: empty)" });
+    options.push_back({ "main",        "       --no-warmup",            "skip warming up the model with an empty run" });
+    options.push_back({ "server infill",
+                                       "       --spm-infill",           "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", params.spm_infill ? "enabled" : "disabled" });
+
+    options.push_back({ "sampling" });
+    options.push_back({ "*",           "       --samplers SAMPLERS",    "samplers that will be used for generation in the order, separated by \';\'\n"
+                                                                        "(default: %s)", sampler_type_names.c_str() });
+    options.push_back({ "*",           "       --sampling-seq SEQUENCE",
+                                                                        "simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str() });
+    options.push_back({ "*",           "       --ignore-eos",           "ignore end of stream token and continue generating (implies --logit-bias EOS-inf)" });
+    options.push_back({ "*",           "       --penalize-nl",          "penalize newline tokens (default: %s)", sparams.penalize_nl ? "true" : "false" });
+    options.push_back({ "*",           "       --temp N",               "temperature (default: %.1f)", (double)sparams.temp });
+    options.push_back({ "*",           "       --top-k N",              "top-k sampling (default: %d, 0 = disabled)", sparams.top_k });
+    options.push_back({ "*",           "       --top-p N",              "top-p sampling (default: %.1f, 1.0 = disabled)", (double)sparams.top_p });
+    options.push_back({ "*",           "       --min-p N",              "min-p sampling (default: %.1f, 0.0 = disabled)", (double)sparams.min_p });
+    options.push_back({ "*",           "       --tfs N",                "tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)sparams.tfs_z });
+    options.push_back({ "*",           "       --typical N",            "locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)sparams.typical_p });
+    options.push_back({ "*",           "       --repeat-last-n N",      "last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", sparams.penalty_last_n });
+    options.push_back({ "*",           "       --repeat-penalty N",     "penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)sparams.penalty_repeat });
+    options.push_back({ "*",           "       --presence-penalty N",   "repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)sparams.penalty_present });
+    options.push_back({ "*",           "       --frequency-penalty N",  "repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)sparams.penalty_freq });
+    options.push_back({ "*",           "       --dynatemp-range N",     "dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)sparams.dynatemp_range });
+    options.push_back({ "*",           "       --dynatemp-exp N",       "dynamic temperature exponent (default: %.1f)", (double)sparams.dynatemp_exponent });
+    options.push_back({ "*",           "       --mirostat N",           "use Mirostat sampling.\n"
+                                                                        "Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n"
+                                                                        "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", sparams.mirostat });
+    options.push_back({ "*",           "       --mirostat-lr N",        "Mirostat learning rate, parameter eta (default: %.1f)", (double)sparams.mirostat_eta });
+    options.push_back({ "*",           "       --mirostat-ent N",       "Mirostat target entropy, parameter tau (default: %.1f)", (double)sparams.mirostat_tau });
+    options.push_back({ "*",           "       -l TOKEN_ID(+/-)BIAS",   "modifies the likelihood of token appearing in the completion,\n"
+                                                                        "i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n"
+                                                                        "or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'" });
+    options.push_back({ "main",        "       --cfg-negative-prompt PROMPT",
+                                                                        "negative prompt to use for guidance (default: '%s')", sparams.cfg_negative_prompt.c_str() });
+    options.push_back({ "main",        "       --cfg-negative-prompt-file FNAME",
+                                                                        "negative prompt file to use for guidance" });
+    options.push_back({ "main",        "       --cfg-scale N",          "strength of guidance (default: %.1f, 1.0 = disable)", (double)sparams.cfg_scale });
+    options.push_back({ "main",        "       --chat-template JINJA_TEMPLATE",
+                                                                        "set custom jinja chat template (default: template taken from model's metadata)\n"
+                                                                        "if suffix/prefix are specified, template will be disabled\n"
+                                                                        "only commonly used templates are accepted:\n"
+                                                                        "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
+    options.push_back({ "grammar" });
+    options.push_back({ "*",           "       --grammar GRAMMAR",      "BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", sparams.grammar.c_str() });
+    options.push_back({ "*",           "       --grammar-file FNAME",   "file to read grammar from" });
+    options.push_back({ "*",           "-j,    --json-schema SCHEMA",
+                                                                        "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\n"
+                                                                        "For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead" });
+
+    options.push_back({ "embedding" });
+    options.push_back({ "embedding",   "       --pooling {none,mean,cls,last}",
+                                                                        "pooling type for embeddings, use model default if unspecified" });
+    options.push_back({ "embedding",   "       --attention {causal,non-causal}",
+                                                                        "attention type for embeddings, use model default if unspecified" });
+
+    options.push_back({ "context hacking" });
+    options.push_back({ "*",           "       --rope-scaling {none,linear,yarn}",
+                                                                        "RoPE frequency scaling method, defaults to linear unless specified by the model" });
+    options.push_back({ "*",           "       --rope-scale N",         "RoPE context scaling factor, expands context by a factor of N" });
+    options.push_back({ "*",           "       --rope-freq-base N",     "RoPE base frequency, used by NTK-aware scaling (default: loaded from model)" });
+    options.push_back({ "*",           "       --rope-freq-scale N",    "RoPE frequency scaling factor, expands context by a factor of 1/N" });
+    options.push_back({ "*",           "       --yarn-orig-ctx N",      "YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx });
+    options.push_back({ "*",           "       --yarn-ext-factor N",    "YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor });
+    options.push_back({ "*",           "       --yarn-attn-factor N",   "YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor });
+    options.push_back({ "*",           "       --yarn-beta-slow N",     "YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow });
+    options.push_back({ "*",           "       --yarn-beta-fast N",     "YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast });
+    options.push_back({ "*",           "-gan,  --grp-attn-n N",         "group-attention factor (default: %d)", params.grp_attn_n });
+    options.push_back({ "*",           "-gaw,  --grp-attn-w N",         "group-attention width (default: %.1f)", (double)params.grp_attn_w });
+    options.push_back({ "*",           "-dkvc, --dump-kv-cache",        "verbose print of the KV cache" });
+    options.push_back({ "*",           "-nkvo, --no-kv-offload",        "disable KV offload" });
+    options.push_back({ "*",           "-ctk,  --cache-type-k TYPE",    "KV cache data type for K (default: %s)", params.cache_type_k.c_str() });
+    options.push_back({ "*",           "-ctv,  --cache-type-v TYPE",    "KV cache data type for V (default: %s)", params.cache_type_v.c_str() });
+
+    options.push_back({ "perplexity" });
+    options.push_back({ "perplexity",  "       --all-logits",           "return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false" });
+    options.push_back({ "perplexity",  "       --hellaswag",            "compute HellaSwag score over random tasks from datafile supplied with -f" });
+    options.push_back({ "perplexity",  "       --hellaswag-tasks N",    "number of tasks to use when computing the HellaSwag score (default: %zu)", params.hellaswag_tasks });
+    options.push_back({ "perplexity",  "       --winogrande",           "compute Winogrande score over random tasks from datafile supplied with -f" });
+    options.push_back({ "perplexity",  "       --winogrande-tasks N",   "number of tasks to use when computing the Winogrande score (default: %zu)", params.winogrande_tasks });
+    options.push_back({ "perplexity",  "       --multiple-choice",      "compute multiple choice score over random tasks from datafile supplied with -f" });
+    options.push_back({ "perplexity",  "       --multiple-choice-tasks N",
+                                                                        "number of tasks to use when computing the multiple choice score (default: %zu)", params.multiple_choice_tasks });
+    options.push_back({ "perplexity",  "       --kl-divergence",        "computes KL-divergence to logits provided via --kl-divergence-base" });
+    options.push_back({ "perplexity",  "       --ppl-stride N",         "stride for perplexity calculation (default: %d)", params.ppl_stride });
+    options.push_back({ "perplexity",  "       --ppl-output-type {0,1}",
+                                                                        "output type for perplexity calculation (default: %d)", params.ppl_output_type });
+
+    options.push_back({ "parallel" });
+    options.push_back({ "*",           "-dt,   --defrag-thold N",       "KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold });
+    options.push_back({ "*",           "-np,   --parallel N",           "number of parallel sequences to decode (default: %d)", params.n_parallel });
+    options.push_back({ "*",           "-ns,   --sequences N",          "number of sequences to decode (default: %d)", params.n_sequences });
+    options.push_back({ "*",           "-cb,   --cont-batching",        "enable continuous batching (a.k.a dynamic batching) (default: %s)", params.cont_batching ? "enabled" : "disabled" });
+    options.push_back({ "*",           "-nocb, --no-cont-batching",     "disable continuous batching" });
+
+    options.push_back({ "multi-modality" });
+    options.push_back({ "*",           "       --mmproj FILE",          "path to a multimodal projector file for LLaVA. see examples/llava/README.md" });
+    options.push_back({ "*",           "       --image FILE",           "path to an image file. use with multimodal models. Specify multiple times for batching" });
+
+    options.push_back({ "backend" });
+    options.push_back({ "*",           "       --rpc SERVERS",          "comma separated list of RPC servers" });
+
+    if (llama_supports_mlock()) {
+        options.push_back({ "*",           "       --mlock",                "force system to keep model in RAM rather than swapping or compressing" });
+    }
+    if (llama_supports_mmap()) {
+        options.push_back({ "*",           "       --no-mmap",              "do not memory-map model (slower load but may reduce pageouts if not using mlock)" });
+    }
+    options.push_back({ "*",           "       --numa TYPE",            "attempt optimizations that help on some NUMA systems\n"
+                                                                        "  - distribute: spread execution evenly over all nodes\n"
+                                                                        "  - isolate: only spawn threads on CPUs on the node that execution started on\n"
+                                                                        "  - numactl: use the CPU map provided by numactl\n"
+                                                                        "if run without this previously, it is recommended to drop the system page cache before using this\n"
+                                                                        "see https://github.com/ggerganov/llama.cpp/issues/1437" });
+
+    if (llama_supports_gpu_offload()) {
+        options.push_back({ "*",           "-ngl,  --gpu-layers N",
+                                                                        "number of layers to store in VRAM" });
+        options.push_back({ "*",           "-ngld, --gpu-layers-draft N",
+                                                                        "number of layers to store in VRAM for the draft model" });
+        options.push_back({ "*",           "-sm,   --split-mode SPLIT_MODE",
+                                                                        "how to split the model across multiple GPUs, one of:\n"
+                                                                        "  - none: use one GPU only\n"
+                                                                        "  - layer (default): split layers and KV across GPUs\n"
+                                                                        "  - row: split rows across GPUs" });
+        options.push_back({ "*",           "-ts,   --tensor-split SPLIT",
+                                                                        "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1" });
+        options.push_back({ "*",           "-mg,   --main-gpu i",       "the GPU to use for the model (with split-mode = none),\n"
+                                                                        "or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu });
+    }
+
+    options.push_back({ "model" });
+    options.push_back({ "*",           "       --check-tensors",        "check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false" });
+    options.push_back({ "*",           "       --override-kv KEY=TYPE:VALUE",
+                                                                        "advanced option to override model metadata by key. may be specified multiple times.\n"
+                                                                        "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false" });
+    options.push_back({ "*",           "       --lora FNAME",           "apply LoRA adapter (can be repeated to use multiple adapters)" });
+    options.push_back({ "*",           "       --lora-scaled FNAME S",  "apply LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" });
+    options.push_back({ "*",           "       --control-vector FNAME", "add a control vector\n"
+                                                                        "note: this argument can be repeated to add multiple control vectors" });
+    options.push_back({ "*",           "       --control-vector-scaled FNAME SCALE",
+                                                                        "add a control vector with user defined scaling SCALE\n"
+                                                                        "note: this argument can be repeated to add multiple scaled control vectors" });
+    options.push_back({ "*",           "       --control-vector-layer-range START END",
+                                                                        "layer range to apply the control vector(s) to, start and end inclusive" });
+    options.push_back({ "*",           "-m,    --model FNAME",          "model path (default: models/$filename with filename from --hf-file\n"
+                                                                        "or --model-url if set, otherwise %s)", DEFAULT_MODEL_PATH });
+    options.push_back({ "*",           "-md,   --model-draft FNAME",    "draft model for speculative decoding (default: unused)" });
+    options.push_back({ "*",           "-mu,   --model-url MODEL_URL",  "model download url (default: unused)" });
+    options.push_back({ "*",           "-hfr,  --hf-repo REPO",         "Hugging Face model repository (default: unused)" });
+    options.push_back({ "*",           "-hff,  --hf-file FILE",         "Hugging Face model file (default: unused)" });
+    options.push_back({ "*",           "-hft,  --hf-token TOKEN",       "Hugging Face access token (default: value from HF_TOKEN environment variable)" });
+
+    options.push_back({ "retrieval" });
+    options.push_back({ "retrieval",   "       --context-file FNAME",   "file to load context from (repeat to specify multiple files)" });
+    options.push_back({ "retrieval",   "       --chunk-size N",         "minimum length of embedded text chunks (default: %d)", params.chunk_size });
+    options.push_back({ "retrieval",   "       --chunk-separator STRING",
+                                                                        "separator between chunks (default: '%s')", params.chunk_separator.c_str() });
+
+    options.push_back({ "passkey" });
+    options.push_back({ "passkey",     "       --junk N",               "number of times to repeat the junk text (default: %d)", params.n_junk });
+    options.push_back({ "passkey",     "       --pos N",                "position of the passkey in the junk text (default: %d)", params.i_pos });
+
+    options.push_back({ "imatrix" });
+    options.push_back({ "imatrix",     "-o,    --output FNAME",         "output file (default: '%s')", params.out_file.c_str() });
+    options.push_back({ "imatrix",     "       --output-frequency N",   "output the imatrix every N iterations (default: %d)", params.n_out_freq });
+    options.push_back({ "imatrix",     "       --save-frequency N",     "save an imatrix copy every N iterations (default: %d)", params.n_save_freq });
+    options.push_back({ "imatrix",     "       --process-output",       "collect data for the output tensor (default: %s)", params.process_output ? "true" : "false" });
+    options.push_back({ "imatrix",     "       --no-ppl",               "do not compute perplexity (default: %s)", params.compute_ppl ? "true" : "false" });
+    options.push_back({ "imatrix",     "       --chunk N",              "start processing the input from chunk N (default: %d)", params.i_chunk });
+
+    options.push_back({ "bench" });
+    options.push_back({ "bench",       "-pps",                          "is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false" });
+    options.push_back({ "bench",       "-npp n0,n1,...",                "number of prompt tokens" });
+    options.push_back({ "bench",       "-ntg n0,n1,...",                "number of text generation tokens" });
+    options.push_back({ "bench",       "-npl n0,n1,...",                "number of parallel prompts" });
+
+    options.push_back({ "embedding" });
+    options.push_back({ "embedding",   "       --embd-normalize",       "normalisation for embendings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize });
+    options.push_back({ "embedding",   "       --embd-output-format",   "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix" });
+    options.push_back({ "embedding",   "       --embd-separator",       "separator of embendings (default \\n) for example \"<#sep#>\"" });
+
+    options.push_back({ "server" });
+    options.push_back({ "server",      "       --host HOST",            "ip address to listen (default: %s)", params.hostname.c_str() });
+    options.push_back({ "server",      "       --port PORT",            "port to listen (default: %d)", params.port });
+    options.push_back({ "server",      "       --path PATH",            "path to serve static files from (default: %s)", params.public_path.c_str() });
+    options.push_back({ "server",      "       --embedding(s)",         "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --api-key KEY",          "API key to use for authentication (default: none)" });
+    options.push_back({ "server",      "       --api-key-file FNAME",   "path to file containing API keys (default: none)" });
+    options.push_back({ "server",      "       --ssl-key-file FNAME",   "path to file a PEM-encoded SSL private key" });
+    options.push_back({ "server",      "       --ssl-cert-file FNAME",  "path to file a PEM-encoded SSL certificate" });
+    options.push_back({ "server",      "       --timeout N",            "server read/write timeout in seconds (default: %d)", params.timeout_read });
+    options.push_back({ "server",      "       --threads-http N",       "number of threads used to process HTTP requests (default: %d)", params.n_threads_http });
+    options.push_back({ "server",      "       --system-prompt-file FNAME",
+                                                                        "set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications" });
+    options.push_back({ "server",      "       --log-format {text,json}",
+                                                                        "log output format: json or text (default: json)" });
+    options.push_back({ "server",      "       --metrics",              "enable prometheus compatible metrics endpoint (default: %s)", params.endpoint_metrics ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --no-slots",             "disables slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --slot-save-path PATH",  "path to save slot kv cache (default: disabled)" });
+    options.push_back({ "server",      "       --chat-template JINJA_TEMPLATE",
+                                                                        "set custom jinja chat template (default: template taken from model's metadata)\n"
+                                                                        "only commonly used templates are accepted:\n"
+                                                                        "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
+    options.push_back({ "server",      "-sps,  --slot-prompt-similarity SIMILARITY",
+                                                                        "how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity });
+    options.push_back({ "server",      "       --lora-init-without-apply",     "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"});
+
+#ifndef LOG_DISABLE_LOGS
+    options.push_back({ "logging" });
+    options.push_back({ "*",           "       --simple-io",            "use basic IO for better compatibility in subprocesses and limited consoles" });
+    options.push_back({ "*",           "-ld,   --logdir LOGDIR",        "path under which to save YAML logs (no logging if unset)" });
+    options.push_back({ "logging",     "       --log-test",             "Run simple logging test" });
+    options.push_back({ "logging",     "       --log-disable",          "Disable trace logs" });
+    options.push_back({ "logging",     "       --log-enable",           "Enable trace logs" });
+    options.push_back({ "logging",     "       --log-file FNAME",       "Specify a log filename (without extension)" });
+    options.push_back({ "logging",     "       --log-new",              "Create a separate new log file on start. "
+                                                                        "Each log file will have unique name: \"<name>.<ID>.log\"" });
+    options.push_back({ "logging",     "       --log-append",           "Don't truncate the old log file." });
+#endif // LOG_DISABLE_LOGS
+
+    options.push_back({ "cvector" });
+    options.push_back({ "cvector",     "-o,    --output FNAME",         "output file (default: '%s')", params.cvector_outfile.c_str() });
+    options.push_back({ "cvector",     "       --positive-file FNAME",  "positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str() });
+    options.push_back({ "cvector",     "       --negative-file FNAME",  "negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str() });
+    options.push_back({ "cvector",     "       --pca-batch N",          "batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch });
+    options.push_back({ "cvector",     "       --pca-iter N",           "number of iterations used for PCA (default: %d)", params.n_pca_iterations });
+    options.push_back({ "cvector",     "       --method {pca,mean}",    "dimensionality reduction method to be used (default: pca)" });
+
+    options.push_back({ "export-lora" });
+    options.push_back({ "export-lora", "-m,    --model",                "model path from which to load base model (default '%s')", params.model.c_str() });
+    options.push_back({ "export-lora", "       --lora FNAME",           "path to LoRA adapter  (can be repeated to use multiple adapters)" });
+    options.push_back({ "export-lora", "       --lora-scaled FNAME S",  "path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)" });
+    options.push_back({ "export-lora", "-o,    --output FNAME",         "output file (default: '%s')", params.lora_outfile.c_str() });
+
+    printf("usage: %s [options]\n", argv[0]);
+
+    for (const auto & o : options) {
+        if (!o.grp.empty()) {
+            printf("\n%s:\n\n", o.grp.c_str());
+            continue;
+        }
+        printf("  %-32s", o.args.c_str());
+        if (o.args.length() > 30) {
+            printf("\n%34s", "");
+        }
+
+        const auto desc = o.desc;
+        size_t start = 0;
+        size_t end = desc.find('\n');
+        while (end != std::string::npos) {
+            printf("%s\n%34s", desc.substr(start, end - start).c_str(), "");
+            start = end + 1;
+            end = desc.find('\n', start);
+        }
+
+        printf("%s\n", desc.substr(start).c_str());
+    }
+    printf("\n");
+}
+
+std::string gpt_params_get_system_info(const gpt_params & params) {
+    std::ostringstream os;
+
+    os << "system_info: n_threads = " << params.cpuparams.n_threads;
+    if (params.cpuparams_batch.n_threads != -1) {
+        os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")";
+    }
+#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
+    os << " / " << logicalProcessorCount << " | " << llama_print_system_info();
+#else
+    os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+#endif
+
+    return os.str();
+}
+
+//
+// String utils
+//
+
+std::vector<std::string> string_split(std::string input, char separator) {
+    std::vector<std::string> parts;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(0, separator_pos);
+        parts.emplace_back(part);
+        input = input.substr(separator_pos + 1);
+        separator_pos = input.find(separator);
+    }
+    parts.emplace_back(input);
+    return parts;
+}
+
+std::string string_strip(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+    while (start < end && std::isspace(str[start])) {
+        start++;
+    }
+    while (end > start && std::isspace(str[end - 1])) {
+        end--;
+    }
+    return str.substr(start, end - start);
+}
+
+std::string string_get_sortable_timestamp() {
+    using clock = std::chrono::system_clock;
+
+    const clock::time_point current_time = clock::now();
+    const time_t as_time_t = clock::to_time_t(current_time);
+    char timestamp_no_ns[100];
+    std::strftime(timestamp_no_ns, 100, "%Y_%m_%d-%H_%M_%S", std::localtime(&as_time_t));
+
+    const int64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
+        current_time.time_since_epoch() % 1000000000).count();
+    char timestamp_ns[11];
+    snprintf(timestamp_ns, 11, "%09" PRId64, ns);
+
+    return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
+}
+
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+void string_process_escapes(std::string & input) {
+    std::size_t input_len = input.length();
+    std::size_t output_idx = 0;
+
+    for (std::size_t input_idx = 0; input_idx < input_len; ++input_idx) {
+        if (input[input_idx] == '\\' && input_idx + 1 < input_len) {
+            switch (input[++input_idx]) {
+                case 'n':  input[output_idx++] = '\n'; break;
+                case 'r':  input[output_idx++] = '\r'; break;
+                case 't':  input[output_idx++] = '\t'; break;
+                case '\'': input[output_idx++] = '\''; break;
+                case '\"': input[output_idx++] = '\"'; break;
+                case '\\': input[output_idx++] = '\\'; break;
+                case 'x':
+                    // Handle \x12, etc
+                    if (input_idx + 2 < input_len) {
+                        const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 };
+                        char *err_p = nullptr;
+                        const long val = std::strtol(x, &err_p, 16);
+                        if (err_p == x + 2) {
+                            input_idx += 2;
+                            input[output_idx++] = char(val);
+                            break;
+                        }
+                    }
+                    // fall through
+                default:   input[output_idx++] = '\\';
+                           input[output_idx++] = input[input_idx]; break;
+            }
+        } else {
+            input[output_idx++] = input[input_idx];
+        }
+    }
+
+    input.resize(output_idx);
+}
+
+bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides) {
+    const char * sep = strchr(data, '=');
+    if (sep == nullptr || sep - data >= 128) {
+        fprintf(stderr, "%s: malformed KV override '%s'\n", __func__, data);
+        return false;
+    }
+    llama_model_kv_override kvo;
+    std::strncpy(kvo.key, data, sep - data);
+    kvo.key[sep - data] = 0;
+    sep++;
+    if (strncmp(sep, "int:", 4) == 0) {
+        sep += 4;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+        kvo.val_i64 = std::atol(sep);
+    } else if (strncmp(sep, "float:", 6) == 0) {
+        sep += 6;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
+        kvo.val_f64 = std::atof(sep);
+    } else if (strncmp(sep, "bool:", 5) == 0) {
+        sep += 5;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
+        if (std::strcmp(sep, "true") == 0) {
+            kvo.val_bool = true;
+        } else if (std::strcmp(sep, "false") == 0) {
+            kvo.val_bool = false;
+        } else {
+            fprintf(stderr, "%s: invalid boolean value for KV override '%s'\n", __func__, data);
+            return false;
+        }
+    } else if (strncmp(sep, "str:", 4) == 0) {
+        sep += 4;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+        if (strlen(sep) > 127) {
+            fprintf(stderr, "%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data);
+            return false;
+        }
+        strncpy(kvo.val_str, sep, 127);
+        kvo.val_str[127] = '\0';
+    } else {
+        fprintf(stderr, "%s: invalid type for KV override '%s'\n", __func__, data);
+        return false;
+    }
+    overrides.emplace_back(std::move(kvo));
+    return true;
+}
+
+//
+// Filesystem utils
+//
+
+// Validate if a filename is safe to use
+// To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
+bool fs_validate_filename(const std::string & filename) {
+    if (!filename.length()) {
+        // Empty filename invalid
+        return false;
+    }
+    if (filename.length() > 255) {
+        // Limit at common largest possible filename on Linux filesystems
+        // to avoid unnecessary further validation
+        // (On systems with smaller limits it will be caught by the OS)
+        return false;
+    }
+
+    std::u32string filename_utf32;
+    try {
+        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+        filename_utf32 = converter.from_bytes(filename);
+
+        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
+        // or invalid encodings were encountered. Reject such attempts
+        std::string filename_reencoded = converter.to_bytes(filename_utf32);
+        if (filename_reencoded != filename) {
+            return false;
+        }
+    } catch (const std::exception &) {
+        return false;
+    }
+
+    // Check for forbidden codepoints:
+    // - Control characters
+    // - Unicode equivalents of illegal characters
+    // - UTF-16 surrogate pairs
+    // - UTF-8 replacement character
+    // - Byte order mark (BOM)
+    // - Illegal characters: / \ : * ? " < > |
+    for (char32_t c : filename_utf32) {
+        if (c <= 0x1F // Control characters (C0)
+            || c == 0x7F // Control characters (DEL)
+            || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
+            || c == 0xFF0E // Fullwidth Full Stop (period equivalent)
+            || c == 0x2215 // Division Slash (forward slash equivalent)
+            || c == 0x2216 // Set Minus (backslash equivalent)
+            || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
+            || c == 0xFFFD // Replacement Character (UTF-8)
+            || c == 0xFEFF // Byte Order Mark (BOM)
+            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
+            || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
+            return false;
+        }
+    }
+
+    // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
+    // Unicode and other whitespace is not affected, only 0x20 space
+    if (filename.front() == ' ' || filename.back() == ' ' || filename.back() == '.') {
+        return false;
+    }
+
+    // Reject any ".." (currently stricter than necessary, it should be fine to just check for == ".." instead)
+    if (filename.find("..") != std::string::npos) {
+        return false;
+    }
+
+    // Reject "."
+    if (filename == ".") {
+        return false;
+    }
+
+    return true;
+}
+
+// returns true if successful, false otherwise
+bool fs_create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+std::string fs_get_cache_directory() {
+    std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
+#ifdef __linux__
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        }
+#elif defined(__APPLE__)
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
+#elif defined(_WIN32)
+        cache_directory = std::getenv("LOCALAPPDATA");
+#endif // __linux__
+        cache_directory = ensure_trailing_slash(cache_directory);
+        cache_directory += "llama.cpp";
+    }
+    return ensure_trailing_slash(cache_directory);
+}
+
+std::string fs_get_cache_file(const std::string & filename) {
+    GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
+    std::string cache_directory = fs_get_cache_directory();
+    const bool success = fs_create_directory_with_parents(cache_directory);
+    if (!success) {
+        throw std::runtime_error("failed to create cache directory: " + cache_directory);
+    }
+    return cache_directory + filename;
+}
+
+
+//
+// Model utils
+//
+struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+    llama_init_result iparams;
+    auto mparams = llama_model_params_from_gpt_params(params);
+
+    llama_model * model = nullptr;
+
+    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
+        model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
+    } else if (!params.model_url.empty()) {
+        model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
+    } else {
+        model = llama_load_model_from_file(params.model.c_str(), mparams);
+    }
+
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+        return iparams;
+    }
+
+    auto cparams = llama_context_params_from_gpt_params(params);
+
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
+    if (lctx == NULL) {
+        fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
+        llama_free_model(model);
+        return iparams;
+    }
+
+    if (!params.control_vectors.empty()) {
+        if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
+        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
+
+        const auto cvec = llama_control_vector_load(params.control_vectors);
+        if (cvec.n_embd == -1) {
+            llama_free(lctx);
+            llama_free_model(model);
+            return iparams;
+        }
+
+        int err = llama_control_vector_apply(lctx,
+                                             cvec.data.data(),
+                                             cvec.data.size(),
+                                             cvec.n_embd,
+                                             params.control_vector_layer_start,
+                                             params.control_vector_layer_end);
+        if (err) {
+            llama_free(lctx);
+            llama_free_model(model);
+            return iparams;
+        }
+    }
+
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
+            llama_free(lctx);
+            llama_free_model(model);
+            return iparams;
+        }
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+    }
+    if (!params.lora_init_without_apply) {
+        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
+    }
+
+    if (params.ignore_eos) {
+        params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
+    }
+
+    if (params.warmup) {
+        LOG("warming up the model with an empty run\n");
+
+        std::vector<llama_token> tmp;
+        llama_token bos = llama_token_bos(model);
+        llama_token eos = llama_token_eos(model);
+        // some models (e.g. T5) don't have a BOS token
+        if (bos != -1) {
+            tmp.push_back(bos);
+        }
+        tmp.push_back(eos);
+
+        if (llama_model_has_encoder(model)) {
+            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
+            llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+            if (decoder_start_token_id == -1) {
+                decoder_start_token_id = bos;
+            }
+            tmp.clear();
+            tmp.push_back(decoder_start_token_id);
+        }
+        if (llama_model_has_decoder(model)) {
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        }
+        llama_kv_cache_clear(lctx);
+        llama_synchronize(lctx);
+        llama_reset_timings(lctx);
+    }
+
+    iparams.model   = model;
+    iparams.context = lctx;
+    return iparams;
+}
+
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
+        if (la.scale != 0.0f) {
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+        }
+    }
+}
+
+struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
+    auto mparams = llama_model_default_params();
+
+    if (params.n_gpu_layers != -1) {
+        mparams.n_gpu_layers = params.n_gpu_layers;
+    }
+    mparams.rpc_servers     = params.rpc_servers.c_str();
+    mparams.main_gpu        = params.main_gpu;
+    mparams.split_mode      = params.split_mode;
+    mparams.tensor_split    = params.tensor_split;
+    mparams.use_mmap        = params.use_mmap;
+    mparams.use_mlock       = params.use_mlock;
+    mparams.check_tensors   = params.check_tensors;
+    if (params.kv_overrides.empty()) {
+        mparams.kv_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
+        mparams.kv_overrides = params.kv_overrides.data();
+    }
+
+    return mparams;
+}
+
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f32") {
+        return GGML_TYPE_F32;
+    }
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "iq4_nl") {
+        return GGML_TYPE_IQ4_NL;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    throw std::runtime_error("Invalid cache type: " + s);
+}
+
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
+    auto cparams = llama_context_default_params();
+
+    cparams.n_ctx             = params.n_ctx;
+    cparams.n_seq_max         = params.n_parallel;
+    cparams.n_batch           = params.n_batch;
+    cparams.n_ubatch          = params.n_ubatch;
+    cparams.n_threads         = params.cpuparams.n_threads;
+    cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
+                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+    cparams.seed              = params.seed;
+    cparams.logits_all        = params.logits_all;
+    cparams.embeddings        = params.embedding;
+    cparams.rope_scaling_type = params.rope_scaling_type;
+    cparams.rope_freq_base    = params.rope_freq_base;
+    cparams.rope_freq_scale   = params.rope_freq_scale;
+    cparams.yarn_ext_factor   = params.yarn_ext_factor;
+    cparams.yarn_attn_factor  = params.yarn_attn_factor;
+    cparams.yarn_beta_fast    = params.yarn_beta_fast;
+    cparams.yarn_beta_slow    = params.yarn_beta_slow;
+    cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
+    cparams.pooling_type      = params.pooling_type;
+    cparams.attention_type    = params.attention_type;
+    cparams.defrag_thold      = params.defrag_thold;
+    cparams.cb_eval           = params.cb_eval;
+    cparams.cb_eval_user_data = params.cb_eval_user_data;
+    cparams.offload_kqv       = !params.no_kv_offload;
+    cparams.flash_attn        = params.flash_attn;
+
+    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
+    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
+
+    return cparams;
+}
+
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) {
+    struct ggml_threadpool_params tpp;
+
+    ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults
+
+    if (params.mask_valid) {
+        std::memcpy(&tpp.cpumask, &params.cpumask, GGML_MAX_N_THREADS);
+    }
+
+    tpp.prio       = params.priority;
+    tpp.poll       = params.poll;
+    tpp.strict_cpu = params.strict_cpu;
+
+    return tpp;
+}
+
+#ifdef LLAMA_USE_CURL
+
+static bool starts_with(const std::string & str, const std::string & prefix) {
+    // While we wait for C++20's std::string::starts_with...
+    return str.rfind(prefix, 0) == 0;
+}
+
+static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
+
+    // Initialize libcurl
+    std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
+    if (!curl) {
+        fprintf(stderr, "%s: error initializing libcurl\n", __func__);
+        return false;
+    }
+
+    bool force_download = false;
+
+    // Set the URL, allow to follow http redirection
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+
+    // Check if hf-token or bearer-token was specified
+    if (!hf_token.empty()) {
+      std::string auth_header = "Authorization: Bearer ";
+      auth_header += hf_token.c_str();
+      struct curl_slist *http_headers = NULL;
+      http_headers = curl_slist_append(http_headers, auth_header.c_str());
+      curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
+    }
+
+#if defined(_WIN32)
+    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
+    //   operating system. Currently implemented under MS-Windows.
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+
+    // Check if the file already exists locally
+    struct stat model_file_info;
+    auto file_exists = (stat(path.c_str(), &model_file_info) == 0);
+
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
+    nlohmann::json metadata;
+    std::string etag;
+    std::string last_modified;
+
+    if (file_exists) {
+        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
+        std::ifstream metadata_in(metadata_path);
+        if (metadata_in.good()) {
+            try {
+                metadata_in >> metadata;
+                fprintf(stderr, "%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
+                if (metadata.contains("url") && metadata.at("url").is_string()) {
+                    auto previous_url = metadata.at("url").get<std::string>();
+                    if (previous_url != url) {
+                        fprintf(stderr, "%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
+                        return false;
+                    }
+                }
+                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
+                    etag = metadata.at("etag");
+                }
+                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
+                    last_modified = metadata.at("lastModified");
+                }
+            } catch (const nlohmann::json::exception & e) {
+                fprintf(stderr, "%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+                return false;
+            }
+        }
+    } else {
+        fprintf(stderr, "%s: no previous model file found %s\n", __func__, path.c_str());
+    }
+
+    // Send a HEAD request to retrieve the etag and last-modified headers
+    struct llama_load_model_from_url_headers {
+        std::string etag;
+        std::string last_modified;
+    };
+    llama_load_model_from_url_headers headers;
+    {
+        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
+        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
+            llama_load_model_from_url_headers *headers = (llama_load_model_from_url_headers *) userdata;
+
+            static std::regex header_regex("([^:]+): (.*)\r\n");
+            static std::regex etag_regex("ETag", std::regex_constants::icase);
+            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
+
+            std::string header(buffer, n_items);
+            std::smatch match;
+            if (std::regex_match(header, match, header_regex)) {
+                const std::string & key = match[1];
+                const std::string & value = match[2];
+                if (std::regex_match(key, match, etag_regex)) {
+                    headers->etag = value;
+                } else if (std::regex_match(key, match, last_modified_regex)) {
+                    headers->last_modified = value;
+                }
+            }
+            return n_items;
+        };
+
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
+
+        CURLcode res = curl_easy_perform(curl.get());
+        if (res != CURLE_OK) {
+            fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code != 200) {
+            // HEAD not supported, we don't know if the file has changed
+            // force trigger downloading
+            force_download = true;
+            fprintf(stderr, "%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        }
+    }
+
+    bool should_download = !file_exists || force_download;
+    if (!should_download) {
+        if (!etag.empty() && etag != headers.etag) {
+            fprintf(stderr, "%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
+            should_download = true;
+        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
+            fprintf(stderr, "%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
+            should_download = true;
+        }
+    }
+    if (should_download) {
+        std::string path_temporary = path + ".downloadInProgress";
+        if (file_exists) {
+            fprintf(stderr, "%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
+            if (remove(path.c_str()) != 0) {
+                fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path.c_str());
+                return false;
+            }
+        }
+
+        // Set the output file
+
+        struct FILE_deleter {
+            void operator()(FILE * f) const {
+                fclose(f);
+            }
+        };
+
+        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
+        if (!outfile) {
+            fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path.c_str());
+            return false;
+        }
+
+        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
+        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
+            return fwrite(data, size, nmemb, (FILE *)fd);
+        };
+        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
+
+        //  display download progress
+        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
+
+        // helper function to hide password in URL
+        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
+            std::size_t protocol_pos = url.find("://");
+            if (protocol_pos == std::string::npos) {
+                return url;  // Malformed URL
+            }
+
+            std::size_t at_pos = url.find('@', protocol_pos + 3);
+            if (at_pos == std::string::npos) {
+                return url;  // No password in URL
+            }
+
+            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
+        };
+
+        // start the download
+        fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+                llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
+        auto res = curl_easy_perform(curl.get());
+        if (res != CURLE_OK) {
+            fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
+            return false;
+        }
+
+        long http_code = 0;
+        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code < 200 || http_code >= 400) {
+            fprintf(stderr, "%s: invalid http status code received: %ld\n", __func__, http_code);
+            return false;
+        }
+
+        // Causes file to be closed explicitly here before we rename it.
+        outfile.reset();
+
+        // Write the updated JSON metadata file.
+        metadata.update({
+            {"url", url},
+            {"etag", headers.etag},
+            {"lastModified", headers.last_modified}
+        });
+        std::ofstream(metadata_path) << metadata.dump(4);
+        fprintf(stderr, "%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+
+        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
+            fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+            return false;
+        }
+    }
+
+    return true;
+}
+
+struct llama_model * llama_load_model_from_url(
+        const char * model_url,
+        const char * path_model,
+        const char * hf_token,
+        const struct llama_model_params & params) {
+    // Basic validation of the model_url
+    if (!model_url || strlen(model_url) == 0) {
+        fprintf(stderr, "%s: invalid model_url\n", __func__);
+        return NULL;
+    }
+
+    if (!llama_download_file(model_url, path_model, hf_token)) {
+        return NULL;
+    }
+
+    // check for additional GGUFs split to download
+    int n_split = 0;
+    {
+        struct gguf_init_params gguf_params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ NULL,
+        };
+        auto * ctx_gguf = gguf_init_from_file(path_model, gguf_params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, path_model);
+            return NULL;
+        }
+
+        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+        if (key_n_split >= 0) {
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+        }
+
+        gguf_free(ctx_gguf);
+    }
+
+    if (n_split > 1) {
+        char split_prefix[PATH_MAX] = {0};
+        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+
+        // Verify the first split file format
+        // and extract split URL and PATH prefixes
+        {
+            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), path_model, 0, n_split)) {
+                fprintf(stderr, "\n%s: unexpected model file name: %s"
+                                " n_split=%d\n", __func__, path_model, n_split);
+                return NULL;
+            }
+
+            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url, 0, n_split)) {
+                fprintf(stderr, "\n%s: unexpected model url: %s"
+                                " n_split=%d\n", __func__, model_url, n_split);
+                return NULL;
+            }
+        }
+
+        // Prepare download in parallel
+        std::vector<std::future<bool>> futures_download;
+        for (int idx = 1; idx < n_split; idx++) {
+            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
+                char split_path[PATH_MAX] = {0};
+                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
+
+                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
+
+                return llama_download_file(split_url, split_path, hf_token);
+            }, idx));
+        }
+
+        // Wait for all downloads to complete
+        for (auto & f : futures_download) {
+            if (!f.get()) {
+                return NULL;
+            }
+        }
+    }
+
+    return llama_load_model_from_file(path_model, params);
+}
+
+struct llama_model * llama_load_model_from_hf(
+        const char * repo,
+        const char * model,
+        const char * path_model,
+        const char * hf_token,
+        const struct llama_model_params & params) {
+    // construct hugging face model url:
+    //
+    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
+    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
+    //
+    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
+    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
+    //
+
+    std::string model_url = "https://huggingface.co/";
+    model_url += repo;
+    model_url += "/resolve/main/";
+    model_url += model;
+
+    return llama_load_model_from_url(model_url.c_str(), path_model, hf_token, params);
+}
+
+#else
+
+struct llama_model * llama_load_model_from_url(
+        const char * /*model_url*/,
+        const char * /*path_model*/,
+        const char * /*hf_token*/,
+        const struct llama_model_params & /*params*/) {
+    fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
+    return nullptr;
+}
+
+struct llama_model * llama_load_model_from_hf(
+        const char * /*repo*/,
+        const char * /*model*/,
+        const char * /*path_model*/,
+        const char * /*hf_token*/,
+        const struct llama_model_params & /*params*/) {
+    fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
+    return nullptr;
+}
+
+#endif // LLAMA_USE_CURL
+
+//
+// Batch utils
+//
+
+void llama_batch_clear(struct llama_batch & batch) {
+    batch.n_tokens = 0;
+}
+
+void llama_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits) {
+    batch.token   [batch.n_tokens] = id;
+    batch.pos     [batch.n_tokens] = pos;
+    batch.n_seq_id[batch.n_tokens] = seq_ids.size();
+    for (size_t i = 0; i < seq_ids.size(); ++i) {
+        batch.seq_id[batch.n_tokens][i] = seq_ids[i];
+    }
+    batch.logits  [batch.n_tokens] = logits;
+
+    batch.n_tokens++;
+}
+
+//
+// Vocab utils
+//
+
+std::vector<llama_token> llama_tokenize(
+  const struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_special,
+                        bool   parse_special) {
+    return llama_tokenize(llama_get_model(ctx), text, add_special, parse_special);
+}
+
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
+           const std::string & text,
+                        bool   add_special,
+                        bool   parse_special) {
+    // upper limit for the number of tokens
+    int n_tokens = text.length() + 2 * add_special;
+    std::vector<llama_token> result(n_tokens);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+    return result;
+}
+
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
+    std::string piece;
+    piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
+    const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+    if (n_chars < 0) {
+        piece.resize(-n_chars);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+        GGML_ASSERT(check == -n_chars);
+    }
+    else {
+        piece.resize(n_chars);
+    }
+
+    return piece;
+}
+
+std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
+    std::string text;
+    text.resize(std::max(text.capacity(), tokens.size()));
+    int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    if (n_chars < 0) {
+        text.resize(-n_chars);
+        n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
+    }
+
+    text.resize(n_chars);
+
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
+    return text;
+}
+
+//
+// Chat template utils
+//
+
+bool llama_chat_verify_template(const std::string & tmpl) {
+    llama_chat_message chat[] = {{"user", "test"}};
+    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
+    return res >= 0;
+}
+
+std::string llama_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & msgs,
+        bool add_ass) {
+    int alloc_size = 0;
+    bool fallback = false; // indicate if we must fallback to default chatml
+    std::vector<llama_chat_message> chat;
+    for (auto & msg : msgs) {
+        chat.push_back({msg.role.c_str(), msg.content.c_str()});
+        alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
+    }
+
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    std::vector<char> buf(alloc_size);
+
+    // run the first time to get the total output length
+    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+
+    // error: chat template is not supported
+    if (res < 0) {
+        if (ptr_tmpl != nullptr) {
+            // if the custom "tmpl" is not supported, we throw an error
+            // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
+            throw std::runtime_error("this custom template is not supported");
+        } else {
+            // If the built-in template is not supported, we default to chatml
+            res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+            fallback = true;
+        }
+    }
+
+    // if it turns out that our buffer is too small, we resize it
+    if ((size_t) res > buf.size()) {
+        buf.resize(res);
+        res = llama_chat_apply_template(
+            fallback ? nullptr : model,
+            fallback ? "chatml" : ptr_tmpl,
+            chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+    }
+
+    std::string formatted_chat(buf.data(), res);
+    return formatted_chat;
+}
+
+std::string llama_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & past_msg,
+        const llama_chat_msg & new_msg,
+        bool add_ass) {
+    std::ostringstream ss;
+    auto fmt_past_msg = past_msg.empty() ? "" : llama_chat_apply_template(model, tmpl, past_msg, false);
+    std::vector<llama_chat_msg> chat_new(past_msg);
+    // if the past_msg ends with a newline, we must preserve it in the formatted version
+    if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
+        ss << "\n";
+    };
+    // format chat with new_msg
+    chat_new.push_back(new_msg);
+    auto fmt_new_msg = llama_chat_apply_template(model, tmpl, chat_new, add_ass);
+    // get the diff part
+    ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
+    return ss.str();
+}
+
+std::string llama_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl) {
+    std::vector<llama_chat_msg> msgs = {
+        {"system",    "You are a helpful assistant"},
+        {"user",      "Hello"},
+        {"assistant", "Hi there"},
+        {"user",      "How are you?"},
+    };
+    return llama_chat_apply_template(model, tmpl, msgs, true);
+}
+
+//
+// KV cache utils
+//
+
+void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        int seq_count = 0;
+        for (int j = 0; j < view.n_seq_max; j++) {
+            if (cs_curr[j] >= 0) { seq_count++; }
+        }
+        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
+void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    std::unordered_map<llama_seq_id, size_t> seqs;
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        for (int j = 0; j < view.n_seq_max; j++) {
+            if (cs_curr[j] < 0) { continue; }
+            if (seqs.find(cs_curr[j]) == seqs.end()) {
+                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+                const size_t sz = seqs.size();
+                seqs[cs_curr[j]] = sz;
+            }
+        }
+        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+    }
+
+    printf("=== Sequence legend: ");
+    for (const auto & it : seqs) {
+        printf("%zu=%d, ", it.second, it.first);
+    }
+    printf("'+'=other sequence ids");
+
+    c_curr = view.cells;
+    cs_curr = view.cells_sequences;
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        for (int j = 0; j < view.n_seq_max; j++) {
+            if (cs_curr[j] >= 0) {
+                const auto & it = seqs.find(cs_curr[j]);
+                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
+            } else {
+                putchar('.');
+            }
+        }
+        putchar(' ');
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
+//
+// Embedding utils
+//
+
+void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm) {
+    double sum = 0.0;
+
+    switch (embd_norm) {
+        case -1: // no normalisation
+            sum = 1.0;
+            break;
+        case 0: // max absolute
+            for (int i = 0; i < n; i++) {
+                if (sum < std::abs(inp[i])) sum = std::abs(inp[i]);
+            }
+            sum /= 32760.0; // make an int16 range
+            break;
+        case 2: // euclidean
+            for (int i = 0; i < n; i++) {
+                sum += inp[i] * inp[i];
+            }
+            sum = std::sqrt(sum);
+            break;
+        default: // p-norm (euclidean is p-norm p=2)
+            for (int i = 0; i < n; i++) {
+                sum += std::pow(std::abs(inp[i]), embd_norm);
+            }
+            sum = std::pow(sum, 1.0 / embd_norm);
+            break;
+    }
+
+    const float norm = sum > 0.0 ? 1.0 / sum : 0.0f;
+
+    for (int i = 0; i < n; i++) {
+        out[i] = inp[i] * norm;
+    }
+}
+
+float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){
+    double sum  = 0.0;
+    double sum1 = 0.0;
+    double sum2 = 0.0;
+
+    for (int i = 0; i < n; i++) {
+        sum  += embd1[i] * embd2[i];
+        sum1 += embd1[i] * embd1[i];
+        sum2 += embd2[i] * embd2[i];
+    }
+
+    // Handle the case where one or both vectors are zero vectors
+    if (sum1 == 0.0 || sum2 == 0.0) {
+        if (sum1 == 0.0 && sum2 == 0.0) {
+            return 1.0f; // two zero vectors are similar
+        }
+        return 0.0f;
+    }
+
+    return sum / (sqrt(sum1) * sqrt(sum2));
+}
+
+//
+// Control vector utils
+//
+
+static llama_control_vector_data llama_control_vector_load_one(const llama_control_vector_load_info & load_info) {
+    llama_control_vector_data result = { -1, {} };
+
+    ggml_context * ctx = nullptr;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ false,
+        /* .ctx      = */ &ctx,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s: failed to load control vector file from %s\n", __func__, load_info.fname.c_str());
+        return result;
+    }
+
+    int32_t n_tensors = gguf_get_n_tensors(ctx_gguf);
+    if (n_tensors == 0) {
+        fprintf(stderr, "%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str());
+    }
+
+    for (int i = 0; i < n_tensors; i++) {
+        std::string name = gguf_get_tensor_name(ctx_gguf, i);
+
+        int layer_idx = -1;
+
+        // split on '.'
+        size_t dotpos = name.find('.');
+        if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") {
+            try {
+                layer_idx = std::stoi(name.substr(dotpos + 1));
+            } catch (...) {
+                layer_idx = -1;
+            }
+        }
+        if (layer_idx < 0) {
+            fprintf(stderr, "%s: invalid/unparsable direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        } else if (layer_idx == 0) {
+            fprintf(stderr, "%s: invalid (zero) direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
+
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str());
+        if (tensor->type != GGML_TYPE_F32) {
+            fprintf(stderr, "%s: invalid (non-F32) direction tensor type in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
+        if (ggml_n_dims(tensor) != 1) {
+            fprintf(stderr, "%s: invalid (non-1D) direction tensor shape in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
+
+        if (result.n_embd == -1) {
+            result.n_embd = ggml_nelements(tensor);
+        } else if (ggml_nelements(tensor) != result.n_embd) {
+            fprintf(stderr, "%s: direction tensor in %s does not match previous dimensions\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
+
+        // extend if necessary - do not store data for layer 0 (it's not used)
+        result.data.resize(std::max(result.data.size(), static_cast<size_t>(result.n_embd * layer_idx)), 0.0f);
+
+        const float * src = (const float *) tensor->data;
+        float * dst = result.data.data() + result.n_embd * (layer_idx - 1);  // layer 1 at [0]
+        for (int j = 0; j < result.n_embd; j++) {
+            dst[j] += src[j] * load_info.strength;  // allows multiple directions for same layer in same file
+        }
+
+    }
+
+    if (result.n_embd == -1) {
+        fprintf(stderr, "%s: skipping %s due to invalid direction tensors\n", __func__, load_info.fname.c_str());
+        result.data.clear();
+    }
+
+    gguf_free(ctx_gguf);
+    ggml_free(ctx);
+
+    return result;
+}
+
+llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos) {
+    llama_control_vector_data result = { -1, {} };
+
+    for (const auto & info : load_infos) {
+        auto cur = llama_control_vector_load_one(info);
+
+        if (cur.n_embd == -1) {
+            result.n_embd = -1;
+            break;
+        }
+        if (result.n_embd != -1 && result.n_embd != cur.n_embd) {
+            fprintf(stderr, "%s: control vectors in %s does not match previous dimensions\n", __func__, info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
+
+        if (result.n_embd == -1) {
+            result = std::move(cur);
+        } else {
+            result.data.resize(std::max(result.data.size(), cur.data.size()), 0.0f);  // extend if necessary
+            for (size_t i = 0; i < cur.data.size(); i++) {
+                result.data[i] += cur.data[i];
+            }
+        }
+    }
+
+    if (result.n_embd == -1) {
+        fprintf(stderr, "%s: no valid control vector files passed\n", __func__);
+        result.data.clear();
+    }
+
+    return result;
+}
+
+//
+// YAML utils
+//
+
+void yaml_dump_vector_float(FILE * stream, const char * prop_name, const std::vector<float> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%e, ", data[i]);
+    }
+    fprintf(stream, "%e]\n", data.back());
+}
+
+void yaml_dump_vector_int(FILE * stream, const char * prop_name, const std::vector<int> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%d, ", data[i]);
+    }
+    fprintf(stream, "%d]\n", data.back());
+}
+
+void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data) {
+    std::string data_str(data == NULL ? "" : data);
+
+    if (data_str.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    size_t pos_start = 0;
+    size_t pos_found = 0;
+
+    if (std::isspace(data_str[0]) || std::isspace(data_str.back())) {
+        data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
+        data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
+        data_str = std::regex_replace(data_str, std::regex(R"(\\[^n"])"), R"(\$&)");
+        data_str = "\"" + data_str + "\"";
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    if (data_str.find('\n') == std::string::npos) {
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    fprintf(stream, "%s: |\n", prop_name);
+    while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
+        fprintf(stream, "  %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
+        pos_start = pos_found + 1;
+    }
+}
+
+void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const llama_context * lctx,
+                               const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    const llama_sampling_params & sparams = params.sparams;
+
+    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
+    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
+    fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
+    fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_avx_vnni: %s\n",    ggml_cpu_has_avx_vnni()    ? "true" : "false");
+    fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
+    fprintf(stream, "cpu_has_cuda: %s\n",        ggml_cpu_has_cuda()        ? "true" : "false");
+    fprintf(stream, "cpu_has_vulkan: %s\n",      ggml_cpu_has_vulkan()      ? "true" : "false");
+    fprintf(stream, "cpu_has_kompute: %s\n",     ggml_cpu_has_kompute()     ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sve: %s\n",         ggml_cpu_has_sve()         ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");
+
+#ifdef NDEBUG
+    fprintf(stream, "debug: false\n");
+#else
+    fprintf(stream, "debug: true\n");
+#endif // NDEBUG
+
+    fprintf(stream, "model_desc: %s\n", model_desc);
+    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
+
+#ifdef __OPTIMIZE__
+    fprintf(stream, "optimize: true\n");
+#else
+    fprintf(stream, "optimize: false\n");
+#endif // __OPTIMIZE__
+
+    fprintf(stream, "time: %s\n", timestamp.c_str());
+
+    fprintf(stream, "\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "# User Inputs #\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "\n");
+
+    fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
+    fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
+    yaml_dump_string_multiline(stream, "cfg_negative_prompt", sparams.cfg_negative_prompt.c_str());
+    fprintf(stream, "cfg_scale: %f # default: 1.0\n", sparams.cfg_scale);
+    fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
+    fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
+    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
+    fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
+    yaml_dump_string_multiline(stream, "grammar", sparams.grammar.c_str());
+    fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
+    fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
+    fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
+
+    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx)));
+    const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
+    fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
+
+    yaml_dump_string_multiline(stream, "in_prefix", params.input_prefix.c_str());
+    fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
+    yaml_dump_string_multiline(stream, "in_suffix", params.input_prefix.c_str());
+    fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
+    fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
+    fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
+    fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
+
+    fprintf(stream, "logit_bias:\n");
+    for (std::pair<llama_token, float> lb : sparams.logit_bias) {
+        if (ignore_eos && lb.first == logit_bias_eos->first) {
+            continue;
+        }
+        fprintf(stream, "  %d: %f", lb.first, lb.second);
+    }
+
+    fprintf(stream, "lora:\n");
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
+        }
+    }
+    fprintf(stream, "lora_scaled:\n");
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
+        }
+    }
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
+    fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
+    fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
+    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
+    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
+    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
+    fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
+    fprintf(stream, "model: %s # default: %s\n", params.model.c_str(), DEFAULT_MODEL_PATH);
+    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
+    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
+    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
+    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
+    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
+    fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
+    fprintf(stream, "penalize_nl: %s # default: false\n", sparams.penalize_nl ? "true" : "false");
+    fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
+    fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
+    yaml_dump_string_multiline(stream, "prompt", params.prompt.c_str());
+    fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
+    fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
+    fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
+    yaml_dump_vector_int(stream, "prompt_tokens", prompt_tokens);
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
+
+    fprintf(stream, "reverse_prompt:\n");
+    for (std::string ap : params.antiprompt) {
+        size_t pos = 0;
+        while ((pos = ap.find('\n', pos)) != std::string::npos) {
+            ap.replace(pos, 1, "\\n");
+            pos += 1;
+        }
+
+        fprintf(stream, "  - %s\n", ap.c_str());
+    }
+
+    fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
+    fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
+    fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
+    fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
+    fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false");
+    fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
+
+    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
+    yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
+
+    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
+    fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
+    fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
+    fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
+    fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
+    fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
+    fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
+}
--- a/llama/common.h
+++ b/llama/common.h
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// Various helper functions and utilities
+
+#pragma once
+
+#include "llama.h"
+
+#include "sampling.h"
+
+#define LOG_NO_FILE_LINE_FUNCTION
+#include "log.h"
+
+#include <cmath>
+#include <string>
+#include <vector>
+#include <random>
+#include <thread>
+#include <unordered_map>
+#include <tuple>
+
+#ifdef _WIN32
+#define DIRECTORY_SEPARATOR '\\'
+#else
+#define DIRECTORY_SEPARATOR '/'
+#endif // _WIN32
+
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define print_build_info() do {                                                                     \
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
+} while(0)
+
+#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
+
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
+// build info
+extern int LLAMA_BUILD_NUMBER;
+extern char const * LLAMA_COMMIT;
+extern char const * LLAMA_COMPILER;
+extern char const * LLAMA_BUILD_TARGET;
+
+struct llama_control_vector_load_info;
+
+//
+// CPU utils
+//
+
+int32_t cpu_get_num_physical_cores();
+int32_t cpu_get_num_math();
+
+//
+// CLI argument parsing
+//
+
+// dimensionality reduction methods, used by cvector-generator
+enum dimre_method {
+    DIMRE_METHOD_PCA,
+    DIMRE_METHOD_MEAN,
+};
+
+struct cpu_params {
+    int      n_threads                   = -1;
+    bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
+    bool     mask_valid                  = false;   // Default: any CPU
+    enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
+    bool     strict_cpu                  = false;   // Use strict CPU placement
+    uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
+};
+
+struct gpt_params {
+    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+
+    int32_t n_predict             =    -1; // new tokens to predict
+    int32_t n_ctx                 =     0; // context size
+    int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                =     0; // number of tokens to keep from initial prompt
+    int32_t n_draft               =     5; // number of tokens to draft during speculative decoding
+    int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel            =     1; // number of parallel sequences to decode
+    int32_t n_sequences           =     1; // number of sequences to decode
+    float   p_split               =  0.1f; // speculative decoding split probability
+    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
+    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
+    int32_t grp_attn_n            =     1; // group-attention factor
+    int32_t grp_attn_w            =   512; // group-attention width
+    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
+    float   rope_freq_base        =  0.0f; // RoPE base frequency
+    float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
+    float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
+    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
+    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
+    int32_t yarn_orig_ctx         =     0; // YaRN original context length
+    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
+
+    struct cpu_params cpuparams;
+    struct cpu_params cpuparams_batch;
+    struct cpu_params draft_cpuparams;
+    struct cpu_params draft_cpuparams_batch;
+
+    ggml_backend_sched_eval_callback cb_eval = nullptr;
+    void * cb_eval_user_data                 = nullptr;
+
+    ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
+
+    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
+    enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
+
+    // // sampling parameters
+    struct llama_sampling_params sparams;
+
+    std::string model                = ""; // model path
+    std::string model_draft          = ""; // draft model for speculative decoding
+    std::string model_alias          = "unknown"; // model alias
+    std::string model_url            = ""; // model url to download
+    std::string hf_token             = ""; // HF token
+    std::string hf_repo              = ""; // HF repo
+    std::string hf_file              = ""; // HF file
+    std::string prompt               = "";
+    std::string prompt_file          = ""; // store the external prompt file name
+    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state
+    std::string input_prefix         = ""; // string to prefix user inputs with
+    std::string input_suffix         = ""; // string to suffix user inputs with
+    std::string logdir               = ""; // directory in which to save YAML log files
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
+    std::string logits_file          = ""; // file for saving *all* logits
+    std::string rpc_servers          = ""; // comma separated list of RPC servers
+
+    std::vector<std::string> in_files;   // all input files
+    std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
+    std::vector<llama_model_kv_override> kv_overrides;
+
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
+
+    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
+
+    int32_t verbosity                  = 0;
+    int32_t control_vector_layer_start = -1; // layer range for control vector
+    int32_t control_vector_layer_end   = -1; // layer range for control vector
+
+    int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
+    int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
+                                     //                                       (which is more convenient to use for plotting)
+                                     //
+    bool   hellaswag        = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    size_t hellaswag_tasks  = 400;   // number of tasks to use when computing the HellaSwag score
+
+    bool   winogrande       = false; // compute Winogrande score over random tasks from datafile supplied in prompt
+    size_t winogrande_tasks = 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
+
+    bool   multiple_choice  = false;  // compute TruthfulQA score over random tasks from datafile supplied in prompt
+    size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
+
+    bool   kl_divergence    = false; // compute KL divergence
+
+    bool usage             = false; // print usage
+    bool use_color         = false; // use color to distinguish generations and inputs
+    bool special           = false; // enable special token output
+    bool interactive       = false; // interactive mode
+    bool interactive_first = false; // wait for user input immediately
+    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
+    bool prompt_cache_all  = false; // save user input and generations to prompt cache
+    bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
+
+    bool escape            = true;  // escape "\n", "\r", "\t", "\'", "\"", and "\\"
+    bool multiline_input   = false; // reverse the usage of `\`
+    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
+    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
+    bool flash_attn        = false; // flash attention
+
+    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
+    bool ignore_eos        = false; // ignore generated EOS tokens
+    bool logits_all        = false; // return logits for all tokens in the batch
+    bool use_mmap          = true;  // use mmap for faster loads
+    bool use_mlock         = false; // use mlock to keep model in memory
+    bool verbose_prompt    = false; // print prompt tokens before generation
+    bool display_prompt    = true;  // print prompt before generation
+    bool infill            = false; // use infill mode
+    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
+    bool no_kv_offload     = false; // disable KV offloading
+    bool warmup            = true;  // warmup run
+    bool check_tensors     = false; // validate tensor data
+
+    std::string cache_type_k = "f16"; // KV cache data type for the K
+    std::string cache_type_v = "f16"; // KV cache data type for the V
+
+    // multimodal models (see examples/llava)
+    std::string mmproj = "";        // path to multimodal projector
+    std::vector<std::string> image; // path to image file(s)
+
+    // embedding
+    bool embedding         = false; // get only sentence embedding
+    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+    std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
+    std::string embd_sep   = "\n";  // separator of embendings
+
+    // server params
+    int32_t port           = 8080;         // server listens on this network port
+    int32_t timeout_read   = 600;          // http read timeout in seconds
+    int32_t timeout_write  = timeout_read; // http write timeout in seconds
+    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+
+    std::string hostname      = "127.0.0.1";
+    std::string public_path   = "";
+    std::string chat_template = "";
+    std::string system_prompt = "";
+    bool enable_chat_template = true;
+
+    std::vector<std::string> api_keys;
+
+    std::string ssl_file_key  = "";
+    std::string ssl_file_cert = "";
+
+    bool endpoint_slots   = true;
+    bool endpoint_metrics = false;
+
+    bool log_json = false;
+
+    std::string slot_save_path;
+
+    float slot_prompt_similarity = 0.5f;
+
+    // batched-bench params
+    bool is_pp_shared = false;
+
+    std::vector<int32_t> n_pp;
+    std::vector<int32_t> n_tg;
+    std::vector<int32_t> n_pl;
+
+    // retrieval params
+    std::vector<std::string> context_files; // context files to embed
+
+    int32_t chunk_size = 64; // chunk size for context embedding
+
+    std::string chunk_separator = "\n"; // chunk separator for context embedding
+
+    // passkey params
+    int32_t n_junk = 250; // number of times to repeat the junk text
+    int32_t i_pos  = -1;  // position of the passkey in the junk text
+
+    // imatrix params
+    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
+
+    int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
+    int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
+    int32_t i_chunk     =  0; // start processing from this chunk
+
+    bool process_output = false; // collect data for the output tensor
+    bool compute_ppl    = true;  // whether to compute perplexity
+
+    // cvector-generator params
+    int n_pca_batch = 100;
+    int n_pca_iterations = 1000;
+    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
+    std::string cvector_outfile       = "control_vector.gguf";
+    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
+
+    bool spm_infill = false; // suffix/prefix/middle pattern for infill
+
+    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
+};
+
+void gpt_params_parse_from_env(gpt_params & params);
+void gpt_params_handle_model_default(gpt_params & params);
+
+bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
+bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
+bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_params_get_system_info(const gpt_params & params);
+
+bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
+bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
+bool set_process_priority(enum ggml_sched_priority prio);
+
+//
+// String utils
+//
+
+std::vector<std::string> string_split(std::string input, char separator);
+
+std::string string_strip(const std::string & str);
+std::string string_get_sortable_timestamp();
+
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
+template<class T>
+static std::vector<T> string_split(const std::string & str, char delim) {
+    std::vector<T> values;
+    std::istringstream str_stream(str);
+    std::string token;
+    while (std::getline(str_stream, token, delim)) {
+        T value;
+        std::istringstream token_stream(token);
+        token_stream >> value;
+        values.push_back(value);
+    }
+    return values;
+}
+
+bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+void string_process_escapes(std::string & input);
+
+//
+// Filesystem utils
+//
+
+bool fs_validate_filename(const std::string & filename);
+bool fs_create_directory_with_parents(const std::string & path);
+
+std::string fs_get_cache_directory();
+std::string fs_get_cache_file(const std::string & filename);
+
+//
+// Model utils
+//
+
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
+};
+
+struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
+
+struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
+struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
+
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+
+// Batch utils
+
+void llama_batch_clear(struct llama_batch & batch);
+
+void llama_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits);
+
+//
+// Vocab utils
+//
+
+// tokenizes a string into a vector of tokens
+// should work similar to Python's `tokenizer.encode`
+std::vector<llama_token> llama_tokenize(
+  const struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_special,
+                        bool   parse_special = false);
+
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
+           const std::string & text,
+                        bool   add_special,
+                        bool   parse_special = false);
+
+// tokenizes a token into a piece, optionally renders special/control tokens
+// should work similar to Python's `tokenizer.id_to_piece`
+std::string llama_token_to_piece(
+        const struct llama_context * ctx,
+                       llama_token   token,
+                       bool          special = true);
+
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+// optionally renders special/control tokens
+std::string llama_detokenize(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens,
+                                  bool   special = true);
+
+//
+// Chat template utils
+//
+
+// same with llama_chat_message, but uses std::string
+struct llama_chat_msg {
+    std::string role;
+    std::string content;
+};
+
+// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
+bool llama_chat_verify_template(const std::string & tmpl);
+
+// CPP wrapper for llama_chat_apply_template
+// If the built-in template is not supported, we default to chatml
+// If the custom "tmpl" is not supported, we throw an error
+std::string llama_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & chat,
+        bool add_ass);
+
+// Format single message, while taking into account the position of that message in chat history
+std::string llama_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<llama_chat_msg> & past_msg,
+        const llama_chat_msg & new_msg,
+        bool add_ass);
+
+// Returns an example of formatted chat
+std::string llama_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl);
+
+//
+// KV cache utils
+//
+
+// Dump the KV cache view with the number of sequences per cell.
+void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
+
+// Dump the KV cache view showing individual sequences in each cell (long output).
+void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+
+//
+// Embedding utils
+//
+
+void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
+
+float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
+
+//
+// Control vector utils
+//
+
+struct llama_control_vector_data {
+    int n_embd;
+
+    // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
+    std::vector<float> data;
+};
+
+struct llama_control_vector_load_info {
+    float strength;
+
+    std::string fname;
+};
+
+// Load control vectors, scale each by strength, and add them together.
+// On error, returns {-1, empty}
+llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
+
+//
+// Split utils
+//
+
+static const char * const LLM_KV_SPLIT_NO            = "split.no";
+static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+
+//
+// YAML utils
+//
+
+void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
+void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
+void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
+
+void yaml_dump_non_result_info(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
--- a/llama/ggml-aarch64.c
+++ b/llama/ggml-aarch64.c
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#include "ggml-aarch64.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#elif defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define UNUSED GGML_UNUSED
+
+// Functions to create the interleaved data layout formats
+
+// interleave 4 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x4
+// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
+// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
+//
+// - in                  : an array of block_q4_0 pointers
+// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
+//                         blck_size_interleave bytes
+// - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
+//                         from bias offset form to pure sign form (this saves subtract
+//                         operations durin unpacking)
+//
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+    block_q4_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    for (int i = 0; i < QK4_0 * 2; i++) {
+        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
+
+        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    }
+
+    return out;
+}
+
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x8
+// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+    block_q4_0x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    for (int i = 0; i < QK4_0 * 4; i++) {
+        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
+
+        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    }
+
+    return out;
+}
+
+void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 8; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 4; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][2 * j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][2 * j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][2 * j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    if (blck_size_interleave == 4) {
+        quantize_q8_0_4x4(x, vy, n_per_row);
+    } else if (blck_size_interleave == 8) {
+        quantize_q8_0_4x8(x, vy, n_per_row);
+    } else {
+        assert(false);
+    }
+}
+
+static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
+    assert(n_per_row % QK4_0 == 0);
+    const int nb = n_per_row / QK4_0;
+
+    void * out_ptr = NULL;
+    if (nrows_interleaved == 8) {
+        out_ptr = (block_q4_0x8 *) dst;
+    }
+    else if (nrows_interleaved == 4) {
+        out_ptr = (block_q4_0x4 *) dst;
+    }
+    assert(nrows_interleaved <= 8);
+    block_q4_0 dst_tmp[8];
+
+    for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
+
+        for (int64_t x = 0; x < nb; x++) {
+
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
+            }
+
+            if (nrows_interleaved == 8) {
+                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
+                out_ptr = (block_q4_0x8 *) out_ptr + 1;
+            }
+            else if (nrows_interleaved == 4) {
+                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
+                out_ptr = (block_q4_0x4 *) out_ptr + 1;
+            }
+        }
+    }
+
+    return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
+}
+
+size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
+}
+
+size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
+}
+
+size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
+}
+
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
+                "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+
+    __asm__ __volatile__(
+        "movi v31.16b, #0x4\n"
+        "movi v30.16b, #0xf0\n"
+        "add %x[b_ptr], %x[b_ptr], #0x8\n"
+        "1:"  // Column loop
+        "add x22, %x[a_ptr], #0x2\n"
+        "movi v29.16b, #0x0\n"
+        "mov x21, %x[nb]\n"
+        "2:"  // Block loop
+        "ldr q28, [%x[b_ptr], #0x0]\n"
+        "ldr q27, [x22, #0x0]\n"
+        "movi v26.4s, #0x0\n"
+        "sub x20, x22, #0x2\n"
+        "ldr q25, [x22, #0x10]\n"
+        "ldr q24, [%x[b_ptr], #0x10]\n"
+        "sub x21, x21, #0x1\n"
+        "add x22, x22, #0x22\n"
+        "ldr q23, [%x[b_ptr], #0x20]\n"
+        "ldr q22, [%x[b_ptr], #0x30]\n"
+        "ld1r { v21.8h }, [x20]\n"
+        "ldr q20, [%x[b_ptr], #-0x8]\n"
+        "sshl v16.16b, v28.16b, v31.16b\n"
+        "and v28.16b, v28.16b, v30.16b\n"
+        "sshl v19.16b, v24.16b, v31.16b\n"
+        "and v24.16b, v24.16b, v30.16b\n"
+        "add %x[b_ptr], %x[b_ptr], #0x48\n"
+        "sshl v18.16b, v23.16b, v31.16b\n"
+        "and v23.16b, v23.16b, v30.16b\n"
+        ".inst 0x4f9be21a  // sdot v26.4s, v16.16b, v27.4b[0]\n"
+        "sshl v17.16b, v22.16b, v31.16b\n"
+        "and v22.16b, v22.16b, v30.16b\n"
+        "fcvtl v21.4s, v21.4h\n"
+        "fcvtl v16.4s, v20.4h\n"
+        ".inst 0x4f99e39a  // sdot v26.4s, v28.16b, v25.4b[0]\n"
+        "fmul v16.4s, v16.4s, v21.4s\n"
+        ".inst 0x4fbbe27a  // sdot v26.4s, v19.16b, v27.4b[1]\n"
+        ".inst 0x4fb9e31a  // sdot v26.4s, v24.16b, v25.4b[1]\n"
+        ".inst 0x4f9bea5a  // sdot v26.4s, v18.16b, v27.4b[2]\n"
+        ".inst 0x4f99eafa  // sdot v26.4s, v23.16b, v25.4b[2]\n"
+        ".inst 0x4fbbea3a  // sdot v26.4s, v17.16b, v27.4b[3]\n"
+        ".inst 0x4fb9eada  // sdot v26.4s, v22.16b, v25.4b[3]\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "fmla v29.4s, v26.4s, v16.4s\n"
+        "cbnz x21, 2b\n"
+        "sub %x[nc], %x[nc], #0x4\n"
+        "str q29, [%x[res_ptr], #0x0]\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "cbnz %x[nc], 1b\n"
+        : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+        : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+        : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22"
+    );
+#else
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+#endif
+}
+
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+
+    __asm__ __volatile__(
+        "movi v2.16b, #0x4\n"
+        "movi v1.16b, #0xf0\n"
+        "add %x[b_ptr], %x[b_ptr], #0x8\n"
+        "1:"  // Column loop
+        "add x23, %x[a_ptr], #0x2\n"
+        "movi v0.16b, #0x0\n"
+        "mov x22, %x[nb]\n"
+        "2:"  // Block loop
+        "ldr q31, [%x[b_ptr], #0x0]\n"
+        "ldr q30, [%x[b_ptr], #0x10]\n"
+        "mov x21, x23\n"
+        "movi v29.4s, #0x0\n"
+        "ldr q28, [%x[b_ptr], #0x20]\n"
+        "ldr q27, [%x[b_ptr], #0x30]\n"
+        "movi v26.4s, #0x0\n"
+        "sub x20, x23, #0x2\n"
+        "ld1r { v25.8h }, [x20]\n"
+        "ldr q24, [%x[b_ptr], #-0x8]\n"
+        "sub x22, x22, #0x1\n"
+        "add x23, x23, #0x22\n"
+        "ld1r { v23.2d }, [x21], #0x8\n"
+        "sshl v22.16b, v31.16b, v2.16b\n"
+        "sshl v16.16b, v30.16b, v2.16b\n"
+        "add %x[b_ptr], %x[b_ptr], #0x48\n"
+        "ld1r { v21.2d }, [x21], #0x8\n"
+        "sshl v20.16b, v28.16b, v2.16b\n"
+        "sshl v19.16b, v27.16b, v2.16b\n"
+        "ld1r { v18.2d }, [x21], #0x8\n"
+        "ld1r { v17.2d }, [x21], #0x8\n"
+        "and v31.16b, v31.16b, v1.16b\n"
+        "and v30.16b, v30.16b, v1.16b\n"
+        ".inst 0x4e9796dd  // sdot v29.4s, v22.16b, v23.16b\n"
+        ".inst 0x4e97961a  // sdot v26.4s, v16.16b, v23.16b\n"
+        "and v28.16b, v28.16b, v1.16b\n"
+        "and v27.16b, v27.16b, v1.16b\n"
+        "fcvtl v25.4s, v25.4h\n"
+        "fcvtl v16.4s, v24.4h\n"
+        ".inst 0x4e95969d  // sdot v29.4s, v20.16b, v21.16b\n"
+        ".inst 0x4e95967a  // sdot v26.4s, v19.16b, v21.16b\n"
+        "fmul v16.4s, v16.4s, v25.4s\n"
+        ".inst 0x4e9297fd  // sdot v29.4s, v31.16b, v18.16b\n"
+        ".inst 0x4e9297da  // sdot v26.4s, v30.16b, v18.16b\n"
+        ".inst 0x4e91979d  // sdot v29.4s, v28.16b, v17.16b\n"
+        ".inst 0x4e91977a  // sdot v26.4s, v27.16b, v17.16b\n"
+        "addp v29.4s, v29.4s, v26.4s\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "fmla v0.4s, v29.4s, v16.4s\n"
+        "cbnz x22, 2b\n"
+        "sub %x[nc], %x[nc], #0x4\n"
+        "str q0, [%x[res_ptr], #0x0]\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "cbnz %x[nc], 1b\n"
+        : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+        : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+        : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
+    );
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+#endif
+}
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    if (ggml_sve_cnt_b == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "ptrue p0.b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x10\n"
+            "1:"  // Column loop
+            "add x22, %x[a_ptr], #0x2\n"
+            "mov z31.b, #0x0\n"
+            "mov x21, %x[nb]\n"
+            "2:"  // Block loop
+            "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
+            "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
+            "mov z28.s, #0x0\n"
+            "mov z27.s, #0x0\n"
+            "ld1rd { z26.d }, p0/Z, [x22]\n"
+            "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
+            "sub x20, x22, #0x2\n"
+            "sub x21, x21, #0x1\n"
+            "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
+            "ld1rd { z23.d }, p0/Z, [x22, #8]\n"
+            "lsl z22.b, z30.b, #0x4\n"
+            "lsl z16.b, z29.b, #0x4\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1rd { z21.d }, p0/Z, [x22, #16]\n"
+            "ld1rd { z20.d }, p0/Z, [x22, #24]\n"
+            "lsl z19.b, z25.b, #0x4\n"
+            "and z25.b, z25.b, #0xf0\n"
+            "ld1rh { z17.h }, p0/Z, [x20]\n"
+            "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
+            "sdot z28.s, z22.b, z26.b\n"
+            "sdot z27.s, z16.b, z26.b\n"
+            "lsl z16.b, z24.b, #0x4\n"
+            "add x22, x22, #0x22\n"
+            "and z24.b, z24.b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x90\n"
+            "fcvt z17.s, p0/m, z17.h\n"
+            "fcvt z18.s, p0/m, z18.h\n"
+            "sdot z28.s, z19.b, z23.b\n"
+            "sdot z27.s, z16.b, z23.b\n"
+            "fmul z18.s, z18.s, z17.s\n"
+            "sdot z28.s, z30.b, z21.b\n"
+            "sdot z27.s, z29.b, z21.b\n"
+            "sdot z28.s, z25.b, z20.b\n"
+            "sdot z27.s, z24.b, z20.b\n"
+            "uzp1 z17.s, z28.s, z27.s\n"
+            "uzp2 z16.s, z28.s, z27.s\n"
+            "add z17.s, z17.s, z16.s\n"
+            "asr z17.s, z17.s, #0x4\n"
+            "scvtf z17.s, p0/m, z17.s\n"
+            "fmla z31.s, p0/M, z17.s, z18.s\n"
+            "cbnz x21, 2b\n"
+            "sub %x[nc], %x[nc], #0x8\n"
+            "st1w { z31.s }, p0, [%x[res_ptr]]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+    else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
+                    "performance");
+    }
+    else if (ggml_cpu_has_neon()) {
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
+                    "quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(ggml_cpu_has_sve() &&
+                "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+#endif
+}
+
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
+                "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+    size_t res_stride = bs * sizeof(float);
+
+    __asm__ __volatile__(
+        "mov x10, %x[nr]\n"
+        "mov x9, #0x88\n"
+        "cmp x10, #0x10\n"
+        "mul x9, %x[nb], x9\n"
+        "blt 4f\n"
+        "1:"  // Row loop
+        "add x28, %x[b_ptr], #0x8\n"
+        "mov x27, %x[nc]\n"
+        "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+        "2:"  // Column loop
+        "add x25, %x[a_ptr], #0x8\n"
+        "movi v15.16b, #0x0\n"
+        "movi v19.16b, #0x0\n"
+        "mov x24, %x[nb]\n"
+        "add x23, x25, x9\n"
+        "movi v18.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "add x22, x23, x9\n"
+        "movi v11.16b, #0x0\n"
+        "movi v13.16b, #0x0\n"
+        "add x21, x22, x9\n"
+        "movi v23.16b, #0x0\n"
+        "movi v16.16b, #0x0\n"
+        "movi v25.16b, #0x0\n"
+        "movi v7.16b, #0x0\n"
+        "movi v0.16b, #0x0\n"
+        "movi v4.16b, #0x0\n"
+        "movi v5.16b, #0x0\n"
+        "movi v21.16b, #0x0\n"
+        "movi v8.16b, #0x0\n"
+        "movi v1.16b, #0x0\n"
+        "3:"  // Block loop
+        "ldr q3, [x28, #0x0]\n"
+        "ldr q31, [x25, #0x0]\n"
+        "movi v28.16b, #0x4\n"
+        "movi v10.4s, #0x0\n"
+        "ldr q22, [x28, #0x10]\n"
+        "ldr q6, [x25, #0x10]\n"
+        "movi v29.4s, #0x0\n"
+        "movi v9.4s, #0x0\n"
+        "ldr q27, [x28, #0x20]\n"
+        "ldr q30, [x28, #0x30]\n"
+        "movi v20.4s, #0x0\n"
+        "movi v24.16b, #0xf0\n"
+        "ldr d2, [x25, #-0x8]\n"
+        "ldr d26, [x23, #-0x8]\n"
+        "sshl v12.16b, v3.16b, v28.16b\n"
+        "sub x20, x28, #0x8\n"
+        "ldr d17, [x20, #0x0]\n"
+        "and v3.16b, v3.16b, v24.16b\n"
+        "subs x24, x24, #0x1\n"
+        "add x28, x28, #0x48\n"
+        ".inst 0x4f9fe18a  // sdot v10.4s, v12.16b, v31.4b[0]\n"
+        ".inst 0x4fbfe19d  // sdot v29.4s, v12.16b, v31.4b[1]\n"
+        ".inst 0x4f9fe989  // sdot v9.4s, v12.16b, v31.4b[2]\n"
+        ".inst 0x4fbfe994  // sdot v20.4s, v12.16b, v31.4b[3]\n"
+        "sshl v31.16b, v22.16b, v28.16b\n"
+        "and v22.16b, v22.16b, v24.16b\n"
+        "fcvtl v17.4s, v17.4h\n"
+        "fcvtl v2.4s, v2.4h\n"
+        "fcvtl v26.4s, v26.4h\n"
+        ".inst 0x4f86e3ea  // sdot v10.4s, v31.16b, v6.4b[0]\n"
+        ".inst 0x4fa6e3fd  // sdot v29.4s, v31.16b, v6.4b[1]\n"
+        ".inst 0x4f86ebe9  // sdot v9.4s, v31.16b, v6.4b[2]\n"
+        ".inst 0x4fa6ebf4  // sdot v20.4s, v31.16b, v6.4b[3]\n"
+        "sshl v6.16b, v27.16b, v28.16b\n"
+        "sshl v28.16b, v30.16b, v28.16b\n"
+        "and v27.16b, v27.16b, v24.16b\n"
+        "and v30.16b, v30.16b, v24.16b\n"
+        "ldr q24, [x25, #0x20]\n"
+        ".inst 0x4f98e0ca  // sdot v10.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8c9  // sdot v9.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8d4  // sdot v20.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x30]\n"
+        ".inst 0x4f98e38a  // sdot v10.4s, v28.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e39d  // sdot v29.4s, v28.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb89  // sdot v9.4s, v28.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb94  // sdot v20.4s, v28.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x40]\n"
+        ".inst 0x4f98e06a  // sdot v10.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e869  // sdot v9.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e874  // sdot v20.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x50]\n"
+        ".inst 0x4f98e2ca  // sdot v10.4s, v22.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e2dd  // sdot v29.4s, v22.16b, v24.4b[1]\n"
+        ".inst 0x4f98eac9  // sdot v9.4s, v22.16b, v24.4b[2]\n"
+        ".inst 0x4fb8ead4  // sdot v20.4s, v22.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x60]\n"
+        ".inst 0x4f98e36a  // sdot v10.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb69  // sdot v9.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb74  // sdot v20.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x70]\n"
+        "add x25, x25, #0x88\n"
+        ".inst 0x4f98e3ca  // sdot v10.4s, v30.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e3dd  // sdot v29.4s, v30.16b, v24.4b[1]\n"
+        ".inst 0x4f98ebc9  // sdot v9.4s, v30.16b, v24.4b[2]\n"
+        ".inst 0x4fb8ebd4  // sdot v20.4s, v30.16b, v24.4b[3]\n"
+        "fmul v24.4s, v17.4s, v2.s[0]\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v15.4s, v10.4s, v24.4s\n"
+        "ldr q24, [x23, #0x0]\n"
+        "fmul v10.4s, v17.4s, v2.s[1]\n"
+        "fmla v19.4s, v29.4s, v10.4s\n"
+        "ldr q10, [x23, #0x10]\n"
+        "fmul v29.4s, v17.4s, v2.s[2]\n"
+        "fmul v2.4s, v17.4s, v2.s[3]\n"
+        "fmla v18.4s, v9.4s, v29.4s\n"
+        "movi v9.4s, #0x0\n"
+        "movi v29.4s, #0x0\n"
+        ".inst 0x4f98e189  // sdot v9.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e19d  // sdot v29.4s, v12.16b, v24.4b[1]\n"
+        "fmla v14.4s, v20.4s, v2.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v2.4s, #0x0\n"
+        ".inst 0x4f98e994  // sdot v20.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x20]\n"
+        ".inst 0x4f8ae3e9  // sdot v9.4s, v31.16b, v10.4b[0]\n"
+        ".inst 0x4faae3fd  // sdot v29.4s, v31.16b, v10.4b[1]\n"
+        ".inst 0x4f8aebf4  // sdot v20.4s, v31.16b, v10.4b[2]\n"
+        ".inst 0x4faaebe2  // sdot v2.4s, v31.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x30]\n"
+        ".inst 0x4f98e0c9  // sdot v9.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8d4  // sdot v20.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x40]\n"
+        ".inst 0x4f8ae389  // sdot v9.4s, v28.16b, v10.4b[0]\n"
+        ".inst 0x4faae39d  // sdot v29.4s, v28.16b, v10.4b[1]\n"
+        ".inst 0x4f8aeb94  // sdot v20.4s, v28.16b, v10.4b[2]\n"
+        ".inst 0x4faaeb82  // sdot v2.4s, v28.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x50]\n"
+        ".inst 0x4f98e069  // sdot v9.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e874  // sdot v20.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x60]\n"
+        ".inst 0x4f8ae2c9  // sdot v9.4s, v22.16b, v10.4b[0]\n"
+        ".inst 0x4faae2dd  // sdot v29.4s, v22.16b, v10.4b[1]\n"
+        ".inst 0x4f8aead4  // sdot v20.4s, v22.16b, v10.4b[2]\n"
+        ".inst 0x4faaeac2  // sdot v2.4s, v22.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x70]\n"
+        "add x23, x23, #0x88\n"
+        ".inst 0x4f98e369  // sdot v9.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb74  // sdot v20.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x0]\n"
+        ".inst 0x4f8ae3c9  // sdot v9.4s, v30.16b, v10.4b[0]\n"
+        ".inst 0x4faae3dd  // sdot v29.4s, v30.16b, v10.4b[1]\n"
+        ".inst 0x4f8aebd4  // sdot v20.4s, v30.16b, v10.4b[2]\n"
+        ".inst 0x4faaebc2  // sdot v2.4s, v30.16b, v10.4b[3]\n"
+        "fmul v10.4s, v17.4s, v26.s[0]\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "fmla v11.4s, v9.4s, v10.4s\n"
+        "ldr q9, [x22, #0x10]\n"
+        "fmul v10.4s, v17.4s, v26.s[1]\n"
+        "fmla v13.4s, v29.4s, v10.4s\n"
+        "ldr d29, [x22, #-0x8]\n"
+        "fmul v10.4s, v17.4s, v26.s[2]\n"
+        "fmul v26.4s, v17.4s, v26.s[3]\n"
+        "fcvtl v29.4s, v29.4h\n"
+        "fmla v23.4s, v20.4s, v10.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v10.4s, #0x0\n"
+        "fmla v16.4s, v2.4s, v26.4s\n"
+        "movi v26.4s, #0x0\n"
+        "movi v2.4s, #0x0\n"
+        ".inst 0x4f98e194  // sdot v20.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+        ".inst 0x4f98e99a  // sdot v26.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x20]\n"
+        ".inst 0x4f89e3f4  // sdot v20.4s, v31.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebfa  // sdot v26.4s, v31.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebe2  // sdot v2.4s, v31.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x30]\n"
+        ".inst 0x4f98e0d4  // sdot v20.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0ca  // sdot v10.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8da  // sdot v26.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x40]\n"
+        ".inst 0x4f89e394  // sdot v20.4s, v28.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+        ".inst 0x4f89eb9a  // sdot v26.4s, v28.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eb82  // sdot v2.4s, v28.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x50]\n"
+        ".inst 0x4f98e074  // sdot v20.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e06a  // sdot v10.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e87a  // sdot v26.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x60]\n"
+        ".inst 0x4f89e2d4  // sdot v20.4s, v22.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+        ".inst 0x4f89eada  // sdot v26.4s, v22.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eac2  // sdot v2.4s, v22.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x70]\n"
+        "add x22, x22, #0x88\n"
+        ".inst 0x4f98e374  // sdot v20.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e36a  // sdot v10.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb7a  // sdot v26.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x21, #0x0]\n"
+        ".inst 0x4f89e3d4  // sdot v20.4s, v30.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ca  // sdot v10.4s, v30.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebda  // sdot v26.4s, v30.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebc2  // sdot v2.4s, v30.16b, v9.4b[3]\n"
+        "fmul v9.4s, v17.4s, v29.s[0]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "fmla v25.4s, v20.4s, v9.4s\n"
+        "ldr q9, [x21, #0x10]\n"
+        "fmul v20.4s, v17.4s, v29.s[1]\n"
+        "fmla v7.4s, v10.4s, v20.4s\n"
+        "ldr d20, [x21, #-0x8]\n"
+        "fmul v10.4s, v17.4s, v29.s[2]\n"
+        "fmul v29.4s, v17.4s, v29.s[3]\n"
+        "fcvtl v20.4s, v20.4h\n"
+        "fmla v0.4s, v26.4s, v10.4s\n"
+        "movi v26.4s, #0x0\n"
+        "movi v10.4s, #0x0\n"
+        "fmla v4.4s, v2.4s, v29.4s\n"
+        "movi v2.4s, #0x0\n"
+        "movi v29.4s, #0x0\n"
+        ".inst 0x4f98e19a  // sdot v26.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+        ".inst 0x4f98e982  // sdot v2.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e99d  // sdot v29.4s, v12.16b, v24.4b[3]\n"
+        "ldr q12, [x21, #0x20]\n"
+        "fmul v24.4s, v17.4s, v20.s[0]\n"
+        ".inst 0x4f89e3fa  // sdot v26.4s, v31.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebe2  // sdot v2.4s, v31.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebfd  // sdot v29.4s, v31.16b, v9.4b[3]\n"
+        "ldr q9, [x21, #0x30]\n"
+        "fmul v31.4s, v17.4s, v20.s[1]\n"
+        ".inst 0x4f8ce0da  // sdot v26.4s, v6.16b, v12.4b[0]\n"
+        ".inst 0x4face0ca  // sdot v10.4s, v6.16b, v12.4b[1]\n"
+        ".inst 0x4f8ce8c2  // sdot v2.4s, v6.16b, v12.4b[2]\n"
+        ".inst 0x4face8dd  // sdot v29.4s, v6.16b, v12.4b[3]\n"
+        "ldr q12, [x21, #0x40]\n"
+        "fmul v6.4s, v17.4s, v20.s[2]\n"
+        "fmul v20.4s, v17.4s, v20.s[3]\n"
+        ".inst 0x4f89e39a  // sdot v26.4s, v28.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+        ".inst 0x4f89eb82  // sdot v2.4s, v28.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eb9d  // sdot v29.4s, v28.16b, v9.4b[3]\n"
+        "ldr q9, [x21, #0x50]\n"
+        ".inst 0x4f8ce07a  // sdot v26.4s, v3.16b, v12.4b[0]\n"
+        ".inst 0x4face06a  // sdot v10.4s, v3.16b, v12.4b[1]\n"
+        ".inst 0x4f8ce862  // sdot v2.4s, v3.16b, v12.4b[2]\n"
+        ".inst 0x4face87d  // sdot v29.4s, v3.16b, v12.4b[3]\n"
+        "ldr q12, [x21, #0x60]\n"
+        ".inst 0x4f89e2da  // sdot v26.4s, v22.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+        ".inst 0x4f89eac2  // sdot v2.4s, v22.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eadd  // sdot v29.4s, v22.16b, v9.4b[3]\n"
+        "ldr q17, [x21, #0x70]\n"
+        "add x21, x21, #0x88\n"
+        ".inst 0x4f8ce37a  // sdot v26.4s, v27.16b, v12.4b[0]\n"
+        ".inst 0x4face36a  // sdot v10.4s, v27.16b, v12.4b[1]\n"
+        ".inst 0x4f8ceb62  // sdot v2.4s, v27.16b, v12.4b[2]\n"
+        ".inst 0x4faceb7d  // sdot v29.4s, v27.16b, v12.4b[3]\n"
+        ".inst 0x4f91e3da  // sdot v26.4s, v30.16b, v17.4b[0]\n"
+        ".inst 0x4fb1e3ca  // sdot v10.4s, v30.16b, v17.4b[1]\n"
+        ".inst 0x4f91ebc2  // sdot v2.4s, v30.16b, v17.4b[2]\n"
+        ".inst 0x4fb1ebdd  // sdot v29.4s, v30.16b, v17.4b[3]\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "fmla v5.4s, v26.4s, v24.4s\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "fmla v21.4s, v10.4s, v31.4s\n"
+        "fmla v8.4s, v2.4s, v6.4s\n"
+        "fmla v1.4s, v29.4s, v20.4s\n"
+        "bgt 3b\n"
+        "mov x20, %x[res_ptr]\n"
+        "subs x27, x27, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "str q15, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q19, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q18, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q14, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q11, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q13, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q23, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q16, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q25, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q7, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q0, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q4, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q5, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q21, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q8, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q1, [x20, #0x0]\n"
+        "bne 2b\n"
+        "mov x20, #0x4\n"
+        "sub x10, x10, #0x10\n"
+        "cmp x10, #0x10\n"
+        "mov %x[res_ptr], x26\n"
+        "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+        "bge 1b\n"
+        "4:"  // Row loop skip
+        "cbz x10, 9f\n"
+        "5:"  // Row tail: Row loop
+        "add x24, %x[b_ptr], #0x8\n"
+        "mov x23, %x[nc]\n"
+        "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+        "6:"  // Row tail: Column loop
+        "movi v15.16b, #0x0\n"
+        "movi v19.16b, #0x0\n"
+        "add x25, %x[a_ptr], #0x8\n"
+        "mov x21, %x[nb]\n"
+        "movi v18.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "7:"  // Row tail: Block loop
+        "ldr q7, [x24, #0x0]\n"
+        "ldr q5, [x25, #0x0]\n"
+        "movi v9.16b, #0x4\n"
+        "movi v4.4s, #0x0\n"
+        "ldr q3, [x24, #0x10]\n"
+        "ldr q2, [x25, #0x10]\n"
+        "movi v1.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        "ldr q13, [x24, #0x20]\n"
+        "ldr q31, [x25, #0x20]\n"
+        "movi v30.4s, #0x0\n"
+        "movi v29.16b, #0xf0\n"
+        "ldr q28, [x24, #0x30]\n"
+        "ldr q27, [x25, #0x30]\n"
+        "sshl v20.16b, v7.16b, v9.16b\n"
+        "sub x20, x24, #0x8\n"
+        "ldr q26, [x25, #0x40]\n"
+        "ldr q25, [x25, #0x50]\n"
+        "sshl v17.16b, v3.16b, v9.16b\n"
+        "and v7.16b, v7.16b, v29.16b\n"
+        "ldr q24, [x25, #0x60]\n"
+        "ldr q16, [x25, #0x70]\n"
+        "sshl v22.16b, v13.16b, v9.16b\n"
+        "and v3.16b, v3.16b, v29.16b\n"
+        "ldr d21, [x20, #0x0]\n"
+        "ldr d12, [x25, #-0x8]\n"
+        ".inst 0x4f85e284  // sdot v4.4s, v20.16b, v5.4b[0]\n"
+        ".inst 0x4fa5e281  // sdot v1.4s, v20.16b, v5.4b[1]\n"
+        ".inst 0x4f85ea80  // sdot v0.4s, v20.16b, v5.4b[2]\n"
+        ".inst 0x4fa5ea9e  // sdot v30.4s, v20.16b, v5.4b[3]\n"
+        "sshl v9.16b, v28.16b, v9.16b\n"
+        "subs x21, x21, #0x1\n"
+        "and v13.16b, v13.16b, v29.16b\n"
+        "and v28.16b, v28.16b, v29.16b\n"
+        "add x25, x25, #0x88\n"
+        "add x24, x24, #0x48\n"
+        "fcvtl v21.4s, v21.4h\n"
+        "fcvtl v12.4s, v12.4h\n"
+        ".inst 0x4f82e224  // sdot v4.4s, v17.16b, v2.4b[0]\n"
+        ".inst 0x4fa2e221  // sdot v1.4s, v17.16b, v2.4b[1]\n"
+        ".inst 0x4f82ea20  // sdot v0.4s, v17.16b, v2.4b[2]\n"
+        ".inst 0x4fa2ea3e  // sdot v30.4s, v17.16b, v2.4b[3]\n"
+        "fmul v11.4s, v21.4s, v12.s[0]\n"
+        "fmul v23.4s, v21.4s, v12.s[1]\n"
+        "fmul v17.4s, v21.4s, v12.s[2]\n"
+        ".inst 0x4f9fe2c4  // sdot v4.4s, v22.16b, v31.4b[0]\n"
+        "fmul v6.4s, v21.4s, v12.s[3]\n"
+        ".inst 0x4fbfe2c1  // sdot v1.4s, v22.16b, v31.4b[1]\n"
+        ".inst 0x4f9feac0  // sdot v0.4s, v22.16b, v31.4b[2]\n"
+        ".inst 0x4fbfeade  // sdot v30.4s, v22.16b, v31.4b[3]\n"
+        ".inst 0x4f9be124  // sdot v4.4s, v9.16b, v27.4b[0]\n"
+        ".inst 0x4fbbe121  // sdot v1.4s, v9.16b, v27.4b[1]\n"
+        ".inst 0x4f9be920  // sdot v0.4s, v9.16b, v27.4b[2]\n"
+        ".inst 0x4fbbe93e  // sdot v30.4s, v9.16b, v27.4b[3]\n"
+        ".inst 0x4f9ae0e4  // sdot v4.4s, v7.16b, v26.4b[0]\n"
+        ".inst 0x4fbae0e1  // sdot v1.4s, v7.16b, v26.4b[1]\n"
+        ".inst 0x4f9ae8e0  // sdot v0.4s, v7.16b, v26.4b[2]\n"
+        ".inst 0x4fbae8fe  // sdot v30.4s, v7.16b, v26.4b[3]\n"
+        ".inst 0x4f99e064  // sdot v4.4s, v3.16b, v25.4b[0]\n"
+        ".inst 0x4fb9e061  // sdot v1.4s, v3.16b, v25.4b[1]\n"
+        ".inst 0x4f99e860  // sdot v0.4s, v3.16b, v25.4b[2]\n"
+        ".inst 0x4fb9e87e  // sdot v30.4s, v3.16b, v25.4b[3]\n"
+        ".inst 0x4f98e1a4  // sdot v4.4s, v13.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e1a1  // sdot v1.4s, v13.16b, v24.4b[1]\n"
+        ".inst 0x4f98e9a0  // sdot v0.4s, v13.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e9be  // sdot v30.4s, v13.16b, v24.4b[3]\n"
+        ".inst 0x4f90e384  // sdot v4.4s, v28.16b, v16.4b[0]\n"
+        ".inst 0x4fb0e381  // sdot v1.4s, v28.16b, v16.4b[1]\n"
+        ".inst 0x4f90eb80  // sdot v0.4s, v28.16b, v16.4b[2]\n"
+        ".inst 0x4fb0eb9e  // sdot v30.4s, v28.16b, v16.4b[3]\n"
+        "scvtf v4.4s, v4.4s, #0x4\n"
+        "scvtf v1.4s, v1.4s, #0x4\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "fmla v15.4s, v4.4s, v11.4s\n"
+        "scvtf v30.4s, v30.4s, #0x4\n"
+        "fmla v19.4s, v1.4s, v23.4s\n"
+        "fmla v18.4s, v0.4s, v17.4s\n"
+        "fmla v14.4s, v30.4s, v6.4s\n"
+        "bgt 7b\n"
+        "mov x20, %x[res_ptr]\n"
+        "cmp x10, #0x1\n"
+        "str q15, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x2\n"
+        "str q19, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x3\n"
+        "str q18, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "str q14, [x20, #0x0]\n"
+        "8:"  // Row tail: Accumulator store skip
+        "subs x23, x23, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "bne 6b\n"
+        "subs x10, x10, #0x4\n"
+        "add %x[a_ptr], %x[a_ptr], x9\n"
+        "mov %x[res_ptr], x22\n"
+        "bgt 5b\n"
+        "9:"  // Row tail: Row loop skip
+        : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+        : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+#else
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+#endif
+}
+
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+    size_t res_stride = bs * sizeof(float);
+
+    __asm__ __volatile__(
+        "mov x10, %x[nr]\n"
+        "mov x9, #0x88\n"
+        "cmp x10, #0x10\n"
+        "mul x9, %x[nb], x9\n"
+        "blt 4f\n"
+        "1:"  // Row loop
+        "add x28, %x[b_ptr], #0x8\n"
+        "mov x27, %x[nc]\n"
+        "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+        "2:"  // Column loop
+        "add x25, %x[a_ptr], #0x8\n"
+        "movi v2.16b, #0x0\n"
+        "movi v10.16b, #0x0\n"
+        "mov x24, %x[nb]\n"
+        "add x23, x25, x9\n"
+        "movi v12.16b, #0x0\n"
+        "movi v28.16b, #0x0\n"
+        "add x22, x23, x9\n"
+        "movi v11.16b, #0x0\n"
+        "movi v13.16b, #0x0\n"
+        "add x21, x22, x9\n"
+        "movi v22.16b, #0x0\n"
+        "movi v23.16b, #0x0\n"
+        "movi v25.16b, #0x0\n"
+        "movi v5.16b, #0x0\n"
+        "movi v7.16b, #0x0\n"
+        "movi v4.16b, #0x0\n"
+        "movi v6.16b, #0x0\n"
+        "movi v30.16b, #0x0\n"
+        "movi v24.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "3:"  // Block loop
+        "ldr q21, [x28, #0x0]\n"
+        "ldr q16, [x28, #0x10]\n"
+        "movi v1.16b, #0x4\n"
+        "movi v19.4s, #0x0\n"
+        "ldr q27, [x25, #0x0]\n"
+        "ldr q15, [x25, #0x10]\n"
+        "movi v26.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        "ldr q29, [x28, #0x20]\n"
+        "ldr q3, [x28, #0x30]\n"
+        "movi v17.4s, #0x0\n"
+        "movi v0.16b, #0xf0\n"
+        "ldr d20, [x25, #-0x8]\n"
+        "ldr d9, [x23, #-0x8]\n"
+        "sshl v8.16b, v21.16b, v1.16b\n"
+        "sshl v31.16b, v16.16b, v1.16b\n"
+        "and v21.16b, v21.16b, v0.16b\n"
+        "and v16.16b, v16.16b, v0.16b\n"
+        "sub x20, x28, #0x8\n"
+        "subs x24, x24, #0x1\n"
+        "add x28, x28, #0x48\n"
+        ".inst 0x4e88a773  // smmla v19.4s, v27.16b, v8.16b\n"
+        ".inst 0x4e9fa77a  // smmla v26.4s, v27.16b, v31.16b\n"
+        "ldr q27, [x25, #0x20]\n"
+        ".inst 0x4e88a5f2  // smmla v18.4s, v15.16b, v8.16b\n"
+        ".inst 0x4e9fa5f1  // smmla v17.4s, v15.16b, v31.16b\n"
+        "sshl v15.16b, v29.16b, v1.16b\n"
+        "sshl v1.16b, v3.16b, v1.16b\n"
+        "and v29.16b, v29.16b, v0.16b\n"
+        "and v3.16b, v3.16b, v0.16b\n"
+        "ldr q0, [x25, #0x30]\n"
+        "fcvtl v20.4s, v20.4h\n"
+        ".inst 0x4e8fa773  // smmla v19.4s, v27.16b, v15.16b\n"
+        "fcvtl v9.4s, v9.4h\n"
+        ".inst 0x4e81a77a  // smmla v26.4s, v27.16b, v1.16b\n"
+        "ldr q27, [x25, #0x40]\n"
+        ".inst 0x4e8fa412  // smmla v18.4s, v0.16b, v15.16b\n"
+        ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+        "ldr q0, [x25, #0x50]\n"
+        ".inst 0x4e95a773  // smmla v19.4s, v27.16b, v21.16b\n"
+        ".inst 0x4e90a77a  // smmla v26.4s, v27.16b, v16.16b\n"
+        "ldr q27, [x25, #0x60]\n"
+        ".inst 0x4e95a412  // smmla v18.4s, v0.16b, v21.16b\n"
+        ".inst 0x4e90a411  // smmla v17.4s, v0.16b, v16.16b\n"
+        "ldr q0, [x25, #0x70]\n"
+        "add x25, x25, #0x88\n"
+        ".inst 0x4e9da773  // smmla v19.4s, v27.16b, v29.16b\n"
+        ".inst 0x4e83a77a  // smmla v26.4s, v27.16b, v3.16b\n"
+        "ldr d27, [x20, #0x0]\n"
+        ".inst 0x4e9da412  // smmla v18.4s, v0.16b, v29.16b\n"
+        ".inst 0x4e83a411  // smmla v17.4s, v0.16b, v3.16b\n"
+        "fcvtl v27.4s, v27.4h\n"
+        "uzp1 v0.2d, v19.2d, v26.2d\n"
+        "uzp2 v26.2d, v19.2d, v26.2d\n"
+        "fmul v19.4s, v27.4s, v20.s[0]\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "fmla v2.4s, v0.4s, v19.4s\n"
+        "ldr q19, [x23, #0x0]\n"
+        "uzp1 v0.2d, v18.2d, v17.2d\n"
+        "uzp2 v18.2d, v18.2d, v17.2d\n"
+        "fmul v17.4s, v27.4s, v20.s[1]\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "fmla v10.4s, v26.4s, v17.4s\n"
+        "ldr q17, [x23, #0x10]\n"
+        "fmul v26.4s, v27.4s, v20.s[2]\n"
+        "fmul v20.4s, v27.4s, v20.s[3]\n"
+        "fmla v12.4s, v0.4s, v26.4s\n"
+        "ldr d0, [x22, #-0x8]\n"
+        "ldr d26, [x21, #-0x8]\n"
+        "fcvtl v0.4s, v0.4h\n"
+        "fmla v28.4s, v18.4s, v20.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+        ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+        "ldr q19, [x23, #0x20]\n"
+        "fcvtl v26.4s, v26.4h\n"
+        ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+        ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+        "ldr q19, [x23, #0x40]\n"
+        ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+        ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+        "ldr q19, [x23, #0x60]\n"
+        ".inst 0x4e9da674  // smmla v20.4s, v19.16b, v29.16b\n"
+        ".inst 0x4e83a672  // smmla v18.4s, v19.16b, v3.16b\n"
+        "uzp1 v19.2d, v20.2d, v18.2d\n"
+        "scvtf v19.4s, v19.4s, #0x4\n"
+        "uzp2 v20.2d, v20.2d, v18.2d\n"
+        "fmul v18.4s, v27.4s, v9.s[0]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v11.4s, v19.4s, v18.4s\n"
+        "ldr q18, [x22, #0x0]\n"
+        "fmul v19.4s, v27.4s, v9.s[1]\n"
+        "fmla v13.4s, v20.4s, v19.4s\n"
+        "movi v19.4s, #0x0\n"
+        "movi v20.4s, #0x0\n"
+        ".inst 0x4e88a633  // smmla v19.4s, v17.16b, v8.16b\n"
+        ".inst 0x4e9fa634  // smmla v20.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x23, #0x30]\n"
+        ".inst 0x4e8fa633  // smmla v19.4s, v17.16b, v15.16b\n"
+        ".inst 0x4e81a634  // smmla v20.4s, v17.16b, v1.16b\n"
+        "ldr q17, [x23, #0x50]\n"
+        ".inst 0x4e95a633  // smmla v19.4s, v17.16b, v21.16b\n"
+        ".inst 0x4e90a634  // smmla v20.4s, v17.16b, v16.16b\n"
+        "ldr q17, [x23, #0x70]\n"
+        "add x23, x23, #0x88\n"
+        ".inst 0x4e9da633  // smmla v19.4s, v17.16b, v29.16b\n"
+        ".inst 0x4e83a634  // smmla v20.4s, v17.16b, v3.16b\n"
+        "uzp1 v17.2d, v19.2d, v20.2d\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "uzp2 v20.2d, v19.2d, v20.2d\n"
+        "fmul v19.4s, v27.4s, v9.s[2]\n"
+        "fmul v9.4s, v27.4s, v9.s[3]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v22.4s, v17.4s, v19.4s\n"
+        "ldr q17, [x22, #0x10]\n"
+        "movi v19.4s, #0x0\n"
+        ".inst 0x4e88a653  // smmla v19.4s, v18.16b, v8.16b\n"
+        "fmla v23.4s, v20.4s, v9.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v9.4s, #0x0\n"
+        ".inst 0x4e9fa654  // smmla v20.4s, v18.16b, v31.16b\n"
+        "ldr q18, [x22, #0x20]\n"
+        ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+        ".inst 0x4e8fa653  // smmla v19.4s, v18.16b, v15.16b\n"
+        ".inst 0x4e81a654  // smmla v20.4s, v18.16b, v1.16b\n"
+        "ldr q18, [x22, #0x40]\n"
+        ".inst 0x4e95a653  // smmla v19.4s, v18.16b, v21.16b\n"
+        ".inst 0x4e90a654  // smmla v20.4s, v18.16b, v16.16b\n"
+        "ldr q18, [x22, #0x60]\n"
+        ".inst 0x4e9da653  // smmla v19.4s, v18.16b, v29.16b\n"
+        ".inst 0x4e83a654  // smmla v20.4s, v18.16b, v3.16b\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e9fa632  // smmla v18.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x22, #0x30]\n"
+        ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+        ".inst 0x4e81a632  // smmla v18.4s, v17.16b, v1.16b\n"
+        "ldr q17, [x22, #0x50]\n"
+        ".inst 0x4e95a629  // smmla v9.4s, v17.16b, v21.16b\n"
+        ".inst 0x4e90a632  // smmla v18.4s, v17.16b, v16.16b\n"
+        "ldr q17, [x22, #0x70]\n"
+        "add x22, x22, #0x88\n"
+        ".inst 0x4e9da629  // smmla v9.4s, v17.16b, v29.16b\n"
+        ".inst 0x4e83a632  // smmla v18.4s, v17.16b, v3.16b\n"
+        "uzp1 v17.2d, v19.2d, v20.2d\n"
+        "uzp2 v20.2d, v19.2d, v20.2d\n"
+        "fmul v19.4s, v27.4s, v0.s[0]\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v25.4s, v17.4s, v19.4s\n"
+        "ldr q19, [x21, #0x0]\n"
+        "fmul v17.4s, v27.4s, v0.s[1]\n"
+        "fmla v5.4s, v20.4s, v17.4s\n"
+        "ldr q17, [x21, #0x10]\n"
+        "uzp1 v20.2d, v9.2d, v18.2d\n"
+        "uzp2 v9.2d, v9.2d, v18.2d\n"
+        "fmul v18.4s, v27.4s, v0.s[2]\n"
+        "fmul v0.4s, v27.4s, v0.s[3]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "fmla v7.4s, v20.4s, v18.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+        ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+        "ldr q19, [x21, #0x20]\n"
+        "fmla v4.4s, v9.4s, v0.4s\n"
+        "movi v9.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+        "fmul v8.4s, v27.4s, v26.s[0]\n"
+        ".inst 0x4e9fa620  // smmla v0.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x21, #0x30]\n"
+        ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+        "fmul v31.4s, v27.4s, v26.s[1]\n"
+        ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+        "ldr q19, [x21, #0x40]\n"
+        ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+        "fmul v15.4s, v27.4s, v26.s[2]\n"
+        "fmul v27.4s, v27.4s, v26.s[3]\n"
+        ".inst 0x4e81a620  // smmla v0.4s, v17.16b, v1.16b\n"
+        "ldr q1, [x21, #0x50]\n"
+        ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+        ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+        "ldr q26, [x21, #0x60]\n"
+        ".inst 0x4e95a429  // smmla v9.4s, v1.16b, v21.16b\n"
+        ".inst 0x4e90a420  // smmla v0.4s, v1.16b, v16.16b\n"
+        "ldr q21, [x21, #0x70]\n"
+        "add x21, x21, #0x88\n"
+        ".inst 0x4e9da754  // smmla v20.4s, v26.16b, v29.16b\n"
+        ".inst 0x4e83a752  // smmla v18.4s, v26.16b, v3.16b\n"
+        ".inst 0x4e9da6a9  // smmla v9.4s, v21.16b, v29.16b\n"
+        ".inst 0x4e83a6a0  // smmla v0.4s, v21.16b, v3.16b\n"
+        "uzp1 v29.2d, v20.2d, v18.2d\n"
+        "uzp2 v21.2d, v20.2d, v18.2d\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "uzp1 v18.2d, v9.2d, v0.2d\n"
+        "uzp2 v16.2d, v9.2d, v0.2d\n"
+        "scvtf v21.4s, v21.4s, #0x4\n"
+        "fmla v6.4s, v29.4s, v8.4s\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "scvtf v16.4s, v16.4s, #0x4\n"
+        "fmla v30.4s, v21.4s, v31.4s\n"
+        "fmla v24.4s, v18.4s, v15.4s\n"
+        "fmla v14.4s, v16.4s, v27.4s\n"
+        "bgt 3b\n"
+        "mov x20, %x[res_ptr]\n"
+        "subs x27, x27, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "str q2, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q10, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q12, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q28, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q11, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q13, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q22, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q23, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q25, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q5, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q7, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q4, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q6, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q30, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q24, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q14, [x20, #0x0]\n"
+        "bne 2b\n"
+        "mov x20, #0x4\n"
+        "sub x10, x10, #0x10\n"
+        "cmp x10, #0x10\n"
+        "mov %x[res_ptr], x26\n"
+        "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+        "bge 1b\n"
+        "4:"  // Row loop skip
+        "cbz x10, 9f\n"
+        "5:"  // Row tail: Row loop
+        "add x24, %x[b_ptr], #0x8\n"
+        "mov x23, %x[nc]\n"
+        "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+        "6:"  // Row tail: Column loop
+        "movi v2.16b, #0x0\n"
+        "movi v10.16b, #0x0\n"
+        "add x25, %x[a_ptr], #0x8\n"
+        "mov x21, %x[nb]\n"
+        "movi v12.16b, #0x0\n"
+        "movi v28.16b, #0x0\n"
+        "7:"  // Row tail: Block loop
+        "ldr q6, [x24, #0x0]\n"
+        "ldr q5, [x24, #0x10]\n"
+        "movi v17.16b, #0x4\n"
+        "movi v8.4s, #0x0\n"
+        "ldr q4, [x25, #0x0]\n"
+        "ldr q13, [x25, #0x10]\n"
+        "movi v27.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        "ldr q31, [x24, #0x20]\n"
+        "ldr q14, [x24, #0x30]\n"
+        "movi v29.4s, #0x0\n"
+        "movi v22.16b, #0xf0\n"
+        "ldr q11, [x25, #0x20]\n"
+        "ldr q23, [x25, #0x30]\n"
+        "sshl v21.16b, v6.16b, v17.16b\n"
+        "sshl v16.16b, v5.16b, v17.16b\n"
+        "ldr q20, [x25, #0x40]\n"
+        "ldr q26, [x25, #0x50]\n"
+        "and v6.16b, v6.16b, v22.16b\n"
+        "and v5.16b, v5.16b, v22.16b\n"
+        "ldr q25, [x25, #0x60]\n"
+        "ldr q3, [x25, #0x70]\n"
+        "sshl v19.16b, v31.16b, v17.16b\n"
+        "sshl v18.16b, v14.16b, v17.16b\n"
+        "ldr d17, [x25, #-0x8]\n"
+        ".inst 0x4e95a488  // smmla v8.4s, v4.16b, v21.16b\n"
+        ".inst 0x4e90a49b  // smmla v27.4s, v4.16b, v16.16b\n"
+        "and v31.16b, v31.16b, v22.16b\n"
+        ".inst 0x4e95a5a0  // smmla v0.4s, v13.16b, v21.16b\n"
+        ".inst 0x4e90a5bd  // smmla v29.4s, v13.16b, v16.16b\n"
+        "and v14.16b, v14.16b, v22.16b\n"
+        "sub x20, x24, #0x8\n"
+        "ldr d16, [x20, #0x0]\n"
+        "subs x21, x21, #0x1\n"
+        "add x25, x25, #0x88\n"
+        "fcvtl v17.4s, v17.4h\n"
+        "add x24, x24, #0x48\n"
+        ".inst 0x4e93a568  // smmla v8.4s, v11.16b, v19.16b\n"
+        ".inst 0x4e92a57b  // smmla v27.4s, v11.16b, v18.16b\n"
+        ".inst 0x4e93a6e0  // smmla v0.4s, v23.16b, v19.16b\n"
+        ".inst 0x4e92a6fd  // smmla v29.4s, v23.16b, v18.16b\n"
+        "fcvtl v16.4s, v16.4h\n"
+        ".inst 0x4e86a688  // smmla v8.4s, v20.16b, v6.16b\n"
+        ".inst 0x4e85a69b  // smmla v27.4s, v20.16b, v5.16b\n"
+        "fmul v23.4s, v16.4s, v17.s[0]\n"
+        "fmul v21.4s, v16.4s, v17.s[1]\n"
+        "fmul v1.4s, v16.4s, v17.s[2]\n"
+        "fmul v20.4s, v16.4s, v17.s[3]\n"
+        ".inst 0x4e86a740  // smmla v0.4s, v26.16b, v6.16b\n"
+        ".inst 0x4e85a75d  // smmla v29.4s, v26.16b, v5.16b\n"
+        ".inst 0x4e9fa728  // smmla v8.4s, v25.16b, v31.16b\n"
+        ".inst 0x4e8ea73b  // smmla v27.4s, v25.16b, v14.16b\n"
+        ".inst 0x4e9fa460  // smmla v0.4s, v3.16b, v31.16b\n"
+        ".inst 0x4e8ea47d  // smmla v29.4s, v3.16b, v14.16b\n"
+        "uzp1 v19.2d, v8.2d, v27.2d\n"
+        "uzp2 v18.2d, v8.2d, v27.2d\n"
+        "scvtf v19.4s, v19.4s, #0x4\n"
+        "uzp1 v17.2d, v0.2d, v29.2d\n"
+        "uzp2 v16.2d, v0.2d, v29.2d\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "fmla v2.4s, v19.4s, v23.4s\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "scvtf v16.4s, v16.4s, #0x4\n"
+        "fmla v10.4s, v18.4s, v21.4s\n"
+        "fmla v12.4s, v17.4s, v1.4s\n"
+        "fmla v28.4s, v16.4s, v20.4s\n"
+        "bgt 7b\n"
+        "mov x20, %x[res_ptr]\n"
+        "cmp x10, #0x1\n"
+        "str q2, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x2\n"
+        "str q10, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x3\n"
+        "str q12, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "str q28, [x20, #0x0]\n"
+        "8:"  // Row tail: Accumulator store skip
+        "subs x23, x23, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "bne 6b\n"
+        "subs x10, x10, #0x4\n"
+        "add %x[a_ptr], %x[a_ptr], x9\n"
+        "mov %x[res_ptr], x22\n"
+        "bgt 5b\n"
+        "9:"  // Row tail: Row loop skip
+        : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+        : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+#endif
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    if (ggml_sve_cnt_b == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x20, #0x4\n"
+            "mov x13, %x[nr]\n"
+            "mov z28.s, #-0x4\n"
+            "mov x12, #0x88\n"
+            "ptrue p1.b\n"
+            "whilelt p0.s, XZR, x20\n"
+            "cmp x13, #0x10\n"
+            "mul x12, %x[nb], x12\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x11, %x[b_ptr], #0x10\n"
+            "mov x10, %x[nc]\n"
+            "add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "mov x27, %x[nb]\n"
+            "add x26, x28, x12\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "add x25, x26, x12\n"
+            "mov z13.b, #0x0\n"
+            "mov z1.b, #0x0\n"
+            "add x24, x25, x12\n"
+            "mov z20.b, #0x0\n"
+            "mov z25.b, #0x0\n"
+            "mov z11.b, #0x0\n"
+            "mov z16.b, #0x0\n"
+            "mov z19.b, #0x0\n"
+            "mov z26.b, #0x0\n"
+            "mov z8.b, #0x0\n"
+            "mov z29.b, #0x0\n"
+            "mov z27.b, #0x0\n"
+            "mov z10.b, #0x0\n"
+            "3:"  // Block loop
+            "ld1b { z30.b }, p1/Z, [x11]\n"
+            "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
+            "mov z18.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            "ld1rqb { z3.b }, p1/Z, [x28]\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
+            "mov z9.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
+            "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
+            "sub x20, x11, #0x10\n"
+            "sub x23, x28, #0x8\n"
+            "lsl z31.b, z30.b, #0x4\n"
+            "lsl z6.b, z21.b, #0x4\n"
+            "ld1h { z23.s }, p1/Z, [x20]\n"
+            "sub x22, x26, #0x8\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z21.b, z21.b, #0xf0\n"
+            "sub x21, x25, #0x8\n"
+            "sub x20, x24, #0x8\n"
+            "lsl z14.b, z4.b, #0x4\n"
+            "lsl z2.b, z17.b, #0x4\n"
+            "subs x27, x27, #0x1\n"
+            "add x11, x11, #0x90\n"
+            ".inst 0x451f9872  // smmla z18.s, z3.b, z31.b\n"
+            ".inst 0x45069867  // smmla z7.s, z3.b, z6.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
+            "and z4.b, z4.b, #0xf0\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
+            "and z17.b, z17.b, #0xf0\n"
+            "fcvt z23.s, p1/m, z23.h\n"
+            ".inst 0x450e9872  // smmla z18.s, z3.b, z14.b\n"
+            ".inst 0x45029867  // smmla z7.s, z3.b, z2.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
+            "fscale z23.s, p1/m, z23.s, z28.s\n"
+            ".inst 0x451e9872  // smmla z18.s, z3.b, z30.b\n"
+            ".inst 0x45159867  // smmla z7.s, z3.b, z21.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
+            "add x28, x28, #0x88\n"
+            ".inst 0x45049872  // smmla z18.s, z3.b, z4.b\n"
+            ".inst 0x45119867  // smmla z7.s, z3.b, z17.b\n"
+            "ld1h { z3.s }, p0/Z, [x23]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            "uzp1 z5.d, z18.d, z7.d\n"
+            "uzp2 z18.d, z18.d, z7.d\n"
+            "mov z3.q, z3.q[0]\n"
+            "uzp1 z7.d, z9.d, z22.d\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z3.s[0]\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z24.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z5.b }, p1/Z, [x26]\n"
+            "fmul z9.s, z23.s, z3.s[1]\n"
+            "fmla z15.s, p1/M, z18.s, z9.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
+            "fmul z9.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "fmla z12.s, p1/M, z7.s, z9.s\n"
+            "mov z9.s, #0x0\n"
+            "ld1h { z7.s }, p0/Z, [x22]\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            "fmla z0.s, p1/M, z22.s, z3.s\n"
+            "mov z22.s, #0x0\n"
+            "ld1h { z3.s }, p0/Z, [x21]\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
+            "fcvt z7.s, p1/m, z7.h\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
+            "mov z7.q, z7.q[0]\n"
+            "mov z3.q, z3.q[0]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "uzp1 z5.d, z9.d, z22.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z7.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z13.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x25]\n"
+            "fmul z5.s, z23.s, z7.s[1]\n"
+            "fmla z1.s, p1/M, z22.s, z5.s\n"
+            "mov z5.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            ".inst 0x451f9a45  // smmla z5.s, z18.b, z31.b\n"
+            ".inst 0x45069a56  // smmla z22.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
+            ".inst 0x450e9a45  // smmla z5.s, z18.b, z14.b\n"
+            ".inst 0x45029a56  // smmla z22.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
+            ".inst 0x451e9a45  // smmla z5.s, z18.b, z30.b\n"
+            ".inst 0x45159a56  // smmla z22.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
+            "add x26, x26, #0x88\n"
+            ".inst 0x45049a45  // smmla z5.s, z18.b, z4.b\n"
+            ".inst 0x45119a56  // smmla z22.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z5.d, z22.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z22.d, z5.d, z22.d\n"
+            "fmul z5.s, z23.s, z7.s[2]\n"
+            "fmul z7.s, z23.s, z7.s[3]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z20.s, p1/M, z18.s, z5.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
+            "ld1h { z5.s }, p0/Z, [x20]\n"
+            "fcvt z5.s, p1/m, z5.h\n"
+            "fmla z25.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9936  // smmla z22.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
+            "mov z5.q, z5.q[0]\n"
+            ".inst 0x450e9936  // smmla z22.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
+            ".inst 0x451e9936  // smmla z22.s, z9.b, z30.b\n"
+            ".inst 0x45159927  // smmla z7.s, z9.b, z21.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
+            ".inst 0x45049936  // smmla z22.s, z9.b, z4.b\n"
+            ".inst 0x45119927  // smmla z7.s, z9.b, z17.b\n"
+            "uzp1 z9.d, z22.d, z7.d\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "uzp2 z22.d, z22.d, z7.d\n"
+            "fmul z7.s, z23.s, z3.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z11.s, p1/M, z9.s, z7.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x24]\n"
+            "fmul z7.s, z23.s, z3.s[1]\n"
+            "fmla z16.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9a56  // smmla z22.s, z18.b, z31.b\n"
+            ".inst 0x45069a47  // smmla z7.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
+            ".inst 0x450e9a56  // smmla z22.s, z18.b, z14.b\n"
+            ".inst 0x45029a47  // smmla z7.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
+            ".inst 0x451e9a56  // smmla z22.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x45049a56  // smmla z22.s, z18.b, z4.b\n"
+            ".inst 0x45119a47  // smmla z7.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z22.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z7.d, z22.d, z7.d\n"
+            "fmul z22.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "fmla z19.s, p1/M, z18.s, z22.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
+            "fmul z22.s, z23.s, z5.s[0]\n"
+            "fmla z26.s, p1/M, z7.s, z3.s\n"
+            "mov z3.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9923  // smmla z3.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
+            ".inst 0x450e9923  // smmla z3.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "mov z9.s, #0x0\n"
+            ".inst 0x451f9a49  // smmla z9.s, z18.b, z31.b\n"
+            "mov z31.s, #0x0\n"
+            ".inst 0x45069a5f  // smmla z31.s, z18.b, z6.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
+            ".inst 0x450e98c9  // smmla z9.s, z6.b, z14.b\n"
+            "fmul z14.s, z23.s, z5.s[1]\n"
+            ".inst 0x450298df  // smmla z31.s, z6.b, z2.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
+            "fmul z2.s, z23.s, z5.s[2]\n"
+            "fmul z23.s, z23.s, z5.s[3]\n"
+            ".inst 0x451e9a43  // smmla z3.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
+            ".inst 0x451e98c9  // smmla z9.s, z6.b, z30.b\n"
+            ".inst 0x451598df  // smmla z31.s, z6.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
+            "add x24, x24, #0x88\n"
+            ".inst 0x450498a3  // smmla z3.s, z5.b, z4.b\n"
+            ".inst 0x451198a7  // smmla z7.s, z5.b, z17.b\n"
+            ".inst 0x45049a49  // smmla z9.s, z18.b, z4.b\n"
+            ".inst 0x45119a5f  // smmla z31.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z3.d, z7.d\n"
+            "uzp2 z5.d, z3.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp1 z6.d, z9.d, z31.d\n"
+            "uzp2 z9.d, z9.d, z31.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "fmla z8.s, p1/M, z18.s, z22.s\n"
+            "scvtf z6.s, p1/m, z6.s\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "fmla z29.s, p1/M, z5.s, z14.s\n"
+            "fmla z27.s, p1/M, z6.s, z2.s\n"
+            "fmla z10.s, p1/M, z9.s, z23.s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x10, x10, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z13.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z1.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z20.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z25.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z11.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z16.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z19.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z26.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z8.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z29.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z27.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z10.s }, p1, [x20]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x13, x13, #0x10\n"
+            "cmp x13, #0x10\n"
+            "mov %x[res_ptr], x9\n"
+            "madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x13, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x25, %x[b_ptr], #0x10\n"
+            "mov x24, %x[nc]\n"
+            "add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov x22, %x[nb]\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ld1b { z3.b }, p1/Z, [x25]\n"
+            "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
+            "mov z2.s, #0x0\n"
+            "mov z25.s, #0x0\n"
+            "ld1rqb { z26.b }, p1/Z, [x28]\n"
+            "ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
+            "mov z27.s, #0x0\n"
+            "mov z19.s, #0x0\n"
+            "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
+            "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
+            "sub x21, x25, #0x10\n"
+            "sub x20, x28, #0x8\n"
+            "lsl z20.b, z3.b, #0x4\n"
+            "lsl z4.b, z6.b, #0x4\n"
+            "ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
+            "ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
+            "and z3.b, z3.b, #0xf0\n"
+            "and z6.b, z6.b, #0xf0\n"
+            "ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
+            "ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
+            "lsl z8.b, z29.b, #0x4\n"
+            "lsl z14.b, z16.b, #0x4\n"
+            "ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
+            "ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
+            ".inst 0x45149b42  // smmla z2.s, z26.b, z20.b\n"
+            ".inst 0x45049b59  // smmla z25.s, z26.b, z4.b\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1h { z17.s }, p1/Z, [x21]\n"
+            ".inst 0x45149abb  // smmla z27.s, z21.b, z20.b\n"
+            ".inst 0x45049ab3  // smmla z19.s, z21.b, z4.b\n"
+            "and z16.b, z16.b, #0xf0\n"
+            "ld1h { z4.s }, p0/Z, [x20]\n"
+            "subs x22, x22, #0x1\n"
+            "add x28, x28, #0x88\n"
+            "fcvt z17.s, p1/m, z17.h\n"
+            "add x25, x25, #0x90\n"
+            ".inst 0x45089942  // smmla z2.s, z10.b, z8.b\n"
+            ".inst 0x450e9959  // smmla z25.s, z10.b, z14.b\n"
+            "fcvt z4.s, p1/m, z4.h\n"
+            ".inst 0x45089afb  // smmla z27.s, z23.b, z8.b\n"
+            ".inst 0x450e9af3  // smmla z19.s, z23.b, z14.b\n"
+            "fscale z17.s, p1/m, z17.s, z28.s\n"
+            "mov z4.q, z4.q[0]\n"
+            ".inst 0x45039962  // smmla z2.s, z11.b, z3.b\n"
+            ".inst 0x45069979  // smmla z25.s, z11.b, z6.b\n"
+            "fmul z23.s, z17.s, z4.s[0]\n"
+            "fmul z9.s, z17.s, z4.s[1]\n"
+            "fmul z21.s, z17.s, z4.s[2]\n"
+            "fmul z4.s, z17.s, z4.s[3]\n"
+            ".inst 0x450398fb  // smmla z27.s, z7.b, z3.b\n"
+            ".inst 0x450698f3  // smmla z19.s, z7.b, z6.b\n"
+            ".inst 0x451d9a42  // smmla z2.s, z18.b, z29.b\n"
+            ".inst 0x45109a59  // smmla z25.s, z18.b, z16.b\n"
+            ".inst 0x451d9bdb  // smmla z27.s, z30.b, z29.b\n"
+            ".inst 0x45109bd3  // smmla z19.s, z30.b, z16.b\n"
+            "uzp1 z31.d, z2.d, z25.d\n"
+            "uzp2 z13.d, z2.d, z25.d\n"
+            "scvtf z31.s, p1/m, z31.s\n"
+            "uzp1 z17.d, z27.d, z19.d\n"
+            "uzp2 z18.d, z27.d, z19.d\n"
+            "scvtf z13.s, p1/m, z13.s\n"
+            "fmla z24.s, p1/M, z31.s, z23.s\n"
+            "scvtf z17.s, p1/m, z17.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "fmla z15.s, p1/M, z13.s, z9.s\n"
+            "fmla z12.s, p1/M, z17.s, z21.s\n"
+            "fmla z0.s, p1/M, z18.s, z4.s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x13, #0x1\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x2\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x3\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x24, x24, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "bne 6b\n"
+            "subs x13, x13, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x12\n"
+            "mov %x[res_ptr], x23\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+    else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
+                    "performance");
+    }
+    else if (ggml_cpu_has_neon()) {
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
+                    "quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(ggml_cpu_has_sve() &&
+                "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+#endif
+}
--- a/llama/ggml-aarch64.h
+++ b/llama/ggml-aarch64.h
+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+// GEMV
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// GEMM
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#ifdef __cplusplus
+}
+#endif
+