GitLab

When a model is partially offloaded to system RAM, we can either
do the calculations on the CPU or we can temporarily transfer the
data to the GPU to do the calculations there. Small batches tend
to be better on the CPU, large batches on the GPU.

The llamarunner used the GPU in most cases and the ollamarunner
used the CPU. Although the ollamarunner saw an improvement in
token generation performance, there was a large performance hit
in prompt processing (3-10x).

There is an existing heuristic to dynamically switch between these
two modes but in practice it doesn't have enough information to
accurately make that decision. This adds authoritative data to make
the check work to get the best of both worlds.

Fixes #12037

This revamps how we discover GPUs in the system by leveraging the Ollama
runner.  This should eliminate inconsistency between our GPU discovery and the
runners capabilities at runtime, particularly for cases where we try to filter
out unsupported GPUs.  Now the runner does that implicitly based on the actual
device list.  In some cases free VRAM reporting can be unreliable which can
leaad to scheduling mistakes, so this also includes a patch to leverage more
reliable VRAM reporting libraries if available.

Automatic workarounds have been removed as only one GPU leveraged this, which
is now documented. This GPU will soon fall off the support matrix with the next
ROCm bump.

Additional cleanup of the scheduler and discovery packages can be done in the
future once we have switched on the new memory management code, and removed
support for the llama runner.

For each memory allocation we report the size of the (attempted)
allocation and whether it succeeded or failed. The latter status
reporting proved to be not that useful in practice as systems
such as Windows can automatically overflow from VRAM into RAM,
resultings in successful allocations even when there isn't
enough memory where we wanted.

As a result, this information is only used for debug logging,
which isn't worthwhile enough for the amount of code. It
also isn't fully accurate, as multiple allocations may result
in partial failures.

* use ggml_*_split activations when possible

* forward qkv

* fix truncate

* s/SentencePieceModel/SentencePiece/

* bert

* wordpiece

* refactor pooling

* more tokenizers

* normalize embeddings

* perf: build graph for next batch in parallel to keep GPU busy

This refactors the main run loop of the ollama runner to perform the main GPU
intensive tasks (Compute+Floats) in a go routine so we can prepare the next
batch in parallel to reduce the amount of time the GPU stalls waiting for the
next batch of work.

* tests: tune integration tests for ollama engine

This tunes the integration tests to focus more on models supported
by the new engine.

Flash attention kernels require the mask of the KV cache be a F16
rather than an F32. We can use the GGML operation ggml_cast to do
this rather than doing it ourselves, which allows reuse of a
preallocated buffer in the graph rather than allocating a new one
for each batch. This improves token generation performance with
flash attention by 10-30% (with gpt-oss). This also makes performance
with flash attention better than without it, as expected.

This changes the memory allocation strategy from upfront estimation to
tracking actual allocations done by the engine and reacting to that. The
goal is avoid issues caused by both under-estimation (crashing) and
over-estimation (low performance due to under-utilized GPUs).

It is currently opt-in and can be enabled for models running on the
Ollama engine by setting OLLAMA_NEW_ESTIMATES=1. Behavior in other
cases is unchanged and will continue to use the existing estimates.

* TEMPORARY: Update the llama.cpp upstream to my fork's Granite Four branch

This will be redone once my branch is merged upstream in llama.cpp

* feat: Update all patches

There are a number that are no longer needed at all:

- 0003-embeddings: Embeddings entirely overhauled on master
- 0008-ensure-KV-cache-is-fully-defragmented: KV caching entirely
    overhauled on master
- 0019-metal-add-mean-kernel-14267: Merged upstream
- 0020-CUDA-add-mean-operation-14313: Merged upstream

* feat: Sync llama.cpp and ggml

* fix: Update rsync-filter for all moved/new/removed files

* fix: Add files missing from sync

* fix: Update ggml rsync-filter for new ggml-cpu/arch subdirs

* fix: Add ggml files missing from sync

* fix: Narrow llama.cpp rsync-filter to not include mtmd main tool cpp files

* fix: Remove mtmd main cpp files

* fix: Add missing include in sampling_ext.cpp

* fix: Update llama.go to use mtmd instead of clip/llava

* fix: Add patch for mtmd_input_text

* chore: Ignore *.patched in the patch directory

* fix: Fix support for arch-specific ggml-cpu source files with new arrangement

In https://github.com/ggml-org/llama.cpp/pull/13892, all arch-specific
implementations were split out into a nested tree structure under
ggml-cpu/arch. This conflicts with standard CGO layout where all
arch-specific source files are expected to live in the same directory as
the parent go module and use suffixes based on GOOS and GOARCH. As such,
there were really two options for getting this to work:

1. Add a patch on top of the GGML sync to rearrange the files to match the
GO layout convention
2. Use CGO directives to conditionally include the nested source files in
the compilation units

This commit does (2) in order to minimize the set of changes needed on top
of the upstream file layout. To get this to work, there are two key things
needed:

1. In cpu.go, #cgo directives are added to explicitly set __${GOARCH}__ in
the preprocessor directives
2. In arch-impls.c|cpp, use an #ifdef | #elif defined | #endif chain to
explicitly include the .c|.cpp files for the given architecture from the
nested directory

* fix: Use mtmd_helper to correctly load the bitmap for the image

* fix: Apply patch for mtmd_text_input

* fix: Add missing stb to llama.cpp rsync-filter

* fix: Add sync'ed stb vendored header

* fix: Use c++17 and include vendor for go wrapper modules

* fix: Update patch 0015 for upstream implementation of uuid

* feat: Bump to the latest tip of the branch

* fix: Update patches for bump

* feat: Bump back to the cenral repo and point at the latest master

This includes granite 4 and a number of other model architectures!

* fix: Revert changes to ggml export GPU UUID patch

* fix: Add patch for GGML_VERSION and GGML_COMMIT constants

* feat: Sync all patched code

* build: Include cmake/common.cmake in ggml sync

* build: Add top-level include for GNUINstallDirs in CMakeLists.txt

This is used to populate CMAKE_INSTALL_BINDIR

* fix: Add a patch to avoid power throttling API on non-msvc windows builds

* fix: Sync patch changes for ggml-cpu.c

* feat: Bump llama.cpp to 4a4f42

This picks up support for Kimi K2 and PLaMO-2

* feat: Sync llama.cpp

* fix: Handle multi-chunk image encodings from mtmd

* fix: Re-number patches after merge with `main`

* feat: Bump to 41e78c in the makefile

* fix: Fix Solar and argsort/copy patches after bump

* fix: Remove Gemma3n CUDA Graphs patch

It was implemented upstream:
https://github.com/ggml-org/llama.cpp/pull/14741

* feat: Sync llama.cpp / ggml after latest bump

* build: Remove unnecessary CFLAGS definitions in cpu.go

* fix: Remove unnecessary additions in the rsync-filter

* fix: Remove unused vendored code for chat template parsing

* Revert "fix: Remove Gemma3n CUDA Graphs patch"

This reverts commit d724caced3ce21f08924d4b7801f94ce6638f6ea.

* fix: Update 0020 CUDA Graphs for gemma3n to keep both llama.cpp and ollama fixes

https://github.com/ollama/ollama/pull/11195#issuecomment-3137312394



* fix: Sync ggml-cuda.cu after keeping both style cuda graph fixes for gemma3n

* unwind mxfp4 patch

Prepare to bump ggml with their impl for mxfp4

* bump

* fix windows build error

* Convert tensors at load time

Repack the mxfp4 tensors as ggmls kernels expect them to be.

* convert mlp bf16 to f32

* buffer the conversion better

* reshape earlier

* openai swiglu

* add ids

* split qkv, gate_up

* fix nested alt tags

* fast attention

* remove debug messages

* fix lint

* remove redundant test

* remap values only if source/target are different

* add back i32->i32 copy

* refactor cpu quants

* clean up vendor

* update patch instructions

* clean up patches

* remove webgpu

* update mem

* also handle gpt-oss

* revert convert changes

---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

In order to iteratively find the best memory allocation, we need to
be able to free backend memory so we can try again.

* bf16

* tests

* gpt-oss

* enable gptoss for engine

* rough estimate

* convert to mxfp4

* handle safetensors U8

* clamp glu/linear

* update tokenizer

* MXFP4 support

This implements the Open Compute Microscaling (MX) FP4 format
as a tensor type with backend implementations focusing
on mulmat and mulmatid on CPU, CUDA, and Metal.

* Unit tests for MXFP4 support

This exercises various operations and shapes on both CPU and GPU (if detected
on the system)

* cuda graph

* unit test adjustments

* cuda: optimize memory access

Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4

* mac: fix crash on old macos versions

cblas_sgemm is only supported on v13.3 and up, however bf16 is
only supported on v14+ so we were falling back to ggml-blas and
crashing on bf16 tensors.  Checking for the function being null
seems to be the simplest way to condittionally avoid registering the
backend.

* server: Minimum context length for gptoss

This model requires a minimum context length of 8192 to function
effectively. Users can set higher values through all normal mechanisms
but lower values will be silently reset.

* ggml: Multiply by numParallel for gptoss sliding window

When computing the graph size estimate, the context size is already
multiplied by numParallel so estimates reflect that. However, since
sliding window models use a smaller, fixed context size, they need
to manually take numParallel into account.

* gpt-oss integration

includes harmony parser and thinking levels, etc.

* fix sync

* fix tests

* fix lint

---------
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Devon Rifkin <drifkin@drifkin.net>

We don't get valid UUIDs for AMD GPUs on Windows, so the best option
is to use the ordinal IDs. This brings us in line with what we currently
do on the Ollama server - the only exception is AMD GPUs on Linux, which
falls back to using ordinal IDs. The GGML implementation has no fallback
but it doesn't appear to occur for any of the GPUs that we support.

It's also possible that there are collisions between ordinal IDs for
different libraries - however the only places where we use them are
AMD on Windows and Metal on Mac, which can never occur on the same
system.

* update patches

* cherry pick metal mean kernel

* cherry pick cuda mean kernel

* gemma3n

Reverts PR #11115. The original change was mistakingly reverted instead of #10822

This reverts commit aaa78180.

This enables matching up devices and information reported by the backend
with system management libraries such as nvml to get accurate free
memory reporting.

FromFloatSlice and FromIntSlice return an error if the shape doesn't
match the passed data or if memory can't be allocated. Since these
are inputs, the memory being allocated is system memory rather than VRAM.

In many cases, the caller can't really handle the error and panics.

Empty and Zeros directly panic if they can't allocate memory.

This makes things consistent by panicing for the first two cases,
removing a fair amount of error handling code. This is also consistent
with how Go typically handles these situations.

This provides granular information about the backend memory allocations
required by the runner:
 - Per backend
 - Per layer
 - Weights, cache and graph
 - Allocation status

This can be used for debugging and validating memory estimates.

* feat: qwen3 dense
* feat: qwen3moe
* fix llama4 moe

Currently, when the backend is created, the tensors are loaded at the
same time, which is a slow operation. This separates them to be two
steps:
 - Create backend, including enumerating tensors and memory allocation
 - Loading tensor data

This allows more flexibility in managing model loading.

ml.Dump will preserve default values if not specified

Currently, the KV cache and graph are lazily allocated as needed.
The cache is fully allocated on first use of the corresponding
layer whereas the graph grows with the size of the context.

This can be an issue if another application allocates more VRAM
after we do our calculations - Ollama will crash in the middle of
inference. If we instead allocate the maximum needed memory at
startup of the runner, we will either succeed or fail at that point
rather than at some surprising time in the future.

Currently, this only generates a worst case batch for text, which
means that vision models may get a partial allocation and continue
to lazily allocate the rest.

Mistral is a popular research lab making open source models. This updates
the forward pass of llama architecture models to support both llama models
and mistral models by accounting for additional metadata present in mistral
models, and finding the correct dimensions for the output projection.

Model implementations should use Input for all of their tensors
supplied to the model. This includes tensors that relate to the
outputs, which is confusing since there is also an Output funciton.

Since Output is only used internally in GGML and not used by any
model implementations, we can remove it from the interface to
reduce confusion.

This enables the runner to report progress back to the Ollama server,
both for showing status to the user and also to prevent the server
from killing the runner if it thinks things have stalled.

Most of the infrastructure was already there, this extends it to
be available to the backends.