feat: Build pre-processor from GGUF (#344)

This lets us do: ``` dynamo-run out=llamacpp <gguf_file> ``` Previously a `--model-config <hf-repo>` was also required, to configure our tokenizer.

feat: Build pre-processor from GGUF (#344)
This lets us do: ``` dynamo-run out=llamacpp <gguf_file> ``` Previously a `--model-config <hf-repo>` was also required, to configure our tokenizer.
c7067fc2 · Graham King · GitHub · d29f7fcc · c7067fc2 · c7067fc2
Commit c7067fc2 authored Mar 24, 2025 by Graham King Committed by GitHub Mar 24, 2025
20 changed files
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -695,11 +695,11 @@ version = "0.8.0"
 source = "git+https://github.com/EricLBuehler/candle.git?rev=496a8d2b#496a8d2bf8f88e3be4ea27332a209d66e8b404f4"
 dependencies = [
 "byteorder",
- "candle-kernels",
+ "candle-kernels 0.8.0",
 "candle-metal-kernels",
 "cudarc 0.13.9 (registry+https://github.com/rust-lang/crates.io-index)",
 "float8",
- "gemm",
+ "gemm 0.17.1",
 "half",
 "memmap2",
 "metal",
@@ -714,6 +714,31 @@ dependencies = [
 "zip",
 ]
+[[package]]
+name = "candle-core"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "06ccf5ee3532e66868516d9b315f73aec9f34ea1a37ae98514534d458915dbf1"
+dependencies = [
+ "byteorder",
+ "candle-kernels 0.8.4",
+ "cudarc 0.13.9 (registry+https://github.com/rust-lang/crates.io-index)",
+ "gemm 0.17.1",
+ "half",
+ "memmap2",
+ "num-traits",
+ "num_cpus",
+ "rand 0.9.0",
+ "rand_distr",
+ "rayon",
+ "safetensors",
+ "thiserror 1.0.69",
+ "ug",
+ "ug-cuda",
+ "yoke",
+ "zip",
+]
 [[package]]
 name = "candle-kernels"
 version = "0.8.0"
@@ -722,6 +747,15 @@ dependencies = [
 "bindgen_cuda 0.1.5",
 ]
+[[package]]
+name = "candle-kernels"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a10885bd902fad1b8518ba2b22369aaed88a3d94e123533ad3ca73db33b1c8ca"
+dependencies = [
+ "bindgen_cuda 0.1.5",
+]
 [[package]]
 name = "candle-metal-kernels"
 version = "0.8.0"
@@ -738,7 +772,7 @@ name = "candle-nn"
 version = "0.8.0"
 source = "git+https://github.com/EricLBuehler/candle.git?rev=496a8d2b#496a8d2bf8f88e3be4ea27332a209d66e8b404f4"
 dependencies = [
- "candle-core",
+ "candle-core 0.8.0",
 "candle-metal-kernels",
 "half",
 "metal",
@@ -1383,6 +1417,15 @@ dependencies = [
 "crypto-common",
 ]
+[[package]]
+name = "digit-layout"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09157630eece4139f6cc5a457556d308c3465ecd5af492f0e5aadc043997e2ce"
+dependencies = [
+ "half",
+]
 [[package]]
 name = "dircpy"
 version = "0.3.19"
@@ -1448,10 +1491,20 @@ dependencies = [
 "reborrow",
 ]
+[[package]]
+name = "dyn-stack"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "490bd48eb68fffcfed519b4edbfd82c69cbe741d175b84f0e0cbe8c57cbe0bdd"
+dependencies = [
+ "bytemuck",
+]
 [[package]]
 name = "dynamo-llm"
 version = "0.1.0"
 dependencies = [
+ "akin",
 "anyhow",
 "async-openai",
 "async-stream",
@@ -1463,6 +1516,7 @@ dependencies = [
 "bs62",
 "bytemuck",
 "bytes",
+ "candle-core 0.8.4",
 "chrono",
 "cmake",
 "cudarc 0.13.9 (git+https://github.com/coreylowman/cudarc.git?rev=8c52e735b55bf8e979e1a16bd85e3dfe4f87c9fe)",
@@ -1473,12 +1527,14 @@ dependencies = [
 "erased-serde",
 "futures",
 "galil-seiferas",
+ "ggus",
 "hf-hub",
 "indexmap 2.8.0",
 "insta",
 "itertools 0.14.0",
 "libc",
 "llama-cpp-2",
+ "memmap2",
 "minijinja",
 "minijinja-contrib",
 "mistralrs",
@@ -1827,6 +1883,17 @@ dependencies = [
 "zune-inflate",
 ]
+[[package]]
+name = "fancy-regex"
+version = "0.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e24cb5a94bcae1e5408b0effca5cd7172ea3c5755049c5f3af4cd283a165298"
+dependencies = [
+ "bit-set",
+ "regex-automata 0.4.9",
+ "regex-syntax 0.8.5",
+]
 [[package]]
 name = "fastrand"
 version = "2.3.0"
@@ -2072,17 +2139,37 @@ version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6ab24cc62135b40090e31a76a9b2766a501979f3070fa27f689c27ec04377d32"
 dependencies = [
- "dyn-stack",
+ "dyn-stack 0.10.0",
- "gemm-c32",
+ "gemm-c32 0.17.1",
- "gemm-c64",
+ "gemm-c64 0.17.1",
- "gemm-common",
+ "gemm-common 0.17.1",
- "gemm-f16",
+ "gemm-f16 0.17.1",
- "gemm-f32",
+ "gemm-f32 0.17.1",
- "gemm-f64",
+ "gemm-f64 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab96b703d31950f1aeddded248bc95543c9efc7ac9c4a21fda8703a83ee35451"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-c32 0.18.2",
+ "gemm-c64 0.18.2",
+ "gemm-common 0.18.2",
+ "gemm-f16 0.18.2",
+ "gemm-f32 0.18.2",
+ "gemm-f64 0.18.2",
 "num-complex",
 "num-traits",
 "paste",
- "raw-cpuid",
+ "raw-cpuid 11.5.0",
 "seq-macro",
 ]
@@ -2092,12 +2179,27 @@ version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b9c030d0b983d1e34a546b86e08f600c11696fde16199f971cd46c12e67512c0"
 dependencies = [
- "dyn-stack",
+ "dyn-stack 0.10.0",
- "gemm-common",
+ "gemm-common 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-c32"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f6db9fd9f40421d00eea9dd0770045a5603b8d684654816637732463f4073847"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
 "num-complex",
 "num-traits",
 "paste",
- "raw-cpuid",
+ "raw-cpuid 11.5.0",
 "seq-macro",
 ]
@@ -2107,12 +2209,27 @@ version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fbb5f2e79fefb9693d18e1066a557b4546cd334b226beadc68b11a8f9431852a"
 dependencies = [
- "dyn-stack",
+ "dyn-stack 0.10.0",
- "gemm-common",
+ "gemm-common 0.17.1",
 "num-complex",
 "num-traits",
 "paste",
- "raw-cpuid",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-c64"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dfcad8a3d35a43758330b635d02edad980c1e143dc2f21e6fd25f9e4eada8edf"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
 "seq-macro",
 ]
@@ -2123,17 +2240,38 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a2e7ea062c987abcd8db95db917b4ffb4ecdfd0668471d8dc54734fdff2354e8"
 dependencies = [
 "bytemuck",
- "dyn-stack",
+ "dyn-stack 0.10.0",
+ "half",
+ "num-complex",
+ "num-traits",
+ "once_cell",
+ "paste",
+ "pulp 0.18.22",
+ "raw-cpuid 10.7.0",
+ "rayon",
+ "seq-macro",
+ "sysctl 0.5.5",
+]
+[[package]]
+name = "gemm-common"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a352d4a69cbe938b9e2a9cb7a3a63b7e72f9349174a2752a558a8a563510d0f3"
+dependencies = [
+ "bytemuck",
+ "dyn-stack 0.13.0",
 "half",
+ "libm",
 "num-complex",
 "num-traits",
 "once_cell",
 "paste",
- "pulp",
+ "pulp 0.21.4",
- "raw-cpuid",
+ "raw-cpuid 11.5.0",
 "rayon",
 "seq-macro",
- "sysctl",
+ "sysctl 0.6.0",
 ]
 [[package]]
@@ -2142,14 +2280,32 @@ version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7ca4c06b9b11952071d317604acb332e924e817bd891bec8dfb494168c7cedd4"
 dependencies = [
- "dyn-stack",
+ "dyn-stack 0.10.0",
- "gemm-common",
+ "gemm-common 0.17.1",
- "gemm-f32",
+ "gemm-f32 0.17.1",
+ "half",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "rayon",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f16"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cff95ae3259432f3c3410eaa919033cd03791d81cebd18018393dc147952e109"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "gemm-f32 0.18.2",
 "half",
 "num-complex",
 "num-traits",
 "paste",
- "raw-cpuid",
+ "raw-cpuid 11.5.0",
 "rayon",
 "seq-macro",
 ]
@@ -2160,12 +2316,27 @@ version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e9a69f51aaefbd9cf12d18faf273d3e982d9d711f60775645ed5c8047b4ae113"
 dependencies = [
- "dyn-stack",
+ "dyn-stack 0.10.0",
- "gemm-common",
+ "gemm-common 0.17.1",
 "num-complex",
 "num-traits",
 "paste",
- "raw-cpuid",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f32"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bc8d3d4385393304f407392f754cd2dc4b315d05063f62cf09f47b58de276864"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
 "seq-macro",
 ]
@@ -2175,12 +2346,27 @@ version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "aa397a48544fadf0b81ec8741e5c0fba0043008113f71f2034def1935645d2b0"
 dependencies = [
- "dyn-stack",
+ "dyn-stack 0.10.0",
- "gemm-common",
+ "gemm-common 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f64"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35b2a4f76ce4b8b16eadc11ccf2e083252d8237c1b589558a49b0183545015bd"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
 "num-complex",
 "num-traits",
 "paste",
- "raw-cpuid",
+ "raw-cpuid 11.5.0",
 "seq-macro",
 ]
@@ -2221,6 +2407,30 @@ dependencies = [
 "wasm-bindgen",
 ]
+[[package]]
+name = "ggml-quants"
+version = "0.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a27693512784e0786212eb0bef841779a6337d2d04520ed475b4d5a864f98366"
+dependencies = [
+ "digit-layout",
+ "half",
+ "rayon",
+]
+[[package]]
+name = "ggus"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3ac5654356c6f7f6116905aeaf92ab002c3d03414ada5dbe0bb2e32aa5fea173"
+dependencies = [
+ "fancy-regex",
+ "ggml-quants",
+ "indexmap 2.8.0",
+ "log",
+ "num_enum",
+]
 [[package]]
 name = "gif"
 version = "0.13.1"
@@ -3403,7 +3613,7 @@ version = "0.4.0"
 source = "git+https://github.com/EricLBuehler/mistral.rs.git?rev=aaafc2ef#aaafc2efc6305c1a79eee632b177d76586df1646"
 dependencies = [
 "anyhow",
- "candle-core",
+ "candle-core 0.8.0",
 "candle-nn",
 "clap",
 "either",
@@ -3431,7 +3641,7 @@ dependencies = [
 "bindgen_cuda 0.1.5",
 "bytemuck",
 "bytemuck_derive",
- "candle-core",
+ "candle-core 0.8.0",
 "candle-nn",
 "cfgrammar",
 "chrono",
@@ -3499,7 +3709,7 @@ source = "git+https://github.com/EricLBuehler/mistral.rs.git?rev=aaafc2ef#aaafc2
 dependencies = [
 "anyhow",
 "bindgen_cuda 0.1.6",
- "candle-core",
+ "candle-core 0.8.0",
 "float8",
 "half",
 "metal",
@@ -3514,7 +3724,7 @@ source = "git+https://github.com/EricLBuehler/mistral.rs.git?rev=aaafc2ef#aaafc2
 dependencies = [
 "bindgen_cuda 0.1.5",
 "byteorder",
- "candle-core",
+ "candle-core 0.8.0",
 "candle-nn",
 "float8",
 "half",
@@ -3539,7 +3749,7 @@ name = "mistralrs-vision"
 version = "0.4.0"
 source = "git+https://github.com/EricLBuehler/mistral.rs.git?rev=aaafc2ef#aaafc2efc6305c1a79eee632b177d76586df1646"
 dependencies = [
- "candle-core",
+ "candle-core 0.8.0",
 "image",
 ]
@@ -3781,6 +3991,20 @@ dependencies = [
 "rand 0.8.5",
 ]
+[[package]]
+name = "num"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35bd024e8b2ff75562e5f34e7f4905839deb4b22955ef5e73d2fea1b9813cb23"
+dependencies = [
+ "num-bigint",
+ "num-complex",
+ "num-integer",
+ "num-iter",
+ "num-rational",
+ "num-traits",
+]
 [[package]]
 name = "num-bigint"
 version = "0.4.6"
@@ -3827,6 +4051,28 @@ dependencies = [
 "num-traits",
 ]
+[[package]]
+name = "num-iter"
+version = "0.1.45"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1429034a0490724d0075ebb2bc9e875d6503c3cf69e235a8941aa757d83ef5bf"
+dependencies = [
+ "autocfg",
+ "num-integer",
+ "num-traits",
+]
+[[package]]
+name = "num-rational"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f83d14da390562dca69fc84082e73e548e1ad308d24accdedd2720017cb37824"
+dependencies = [
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
 [[package]]
 name = "num-traits"
 version = "0.2.19"
@@ -4456,6 +4702,20 @@ dependencies = [
 "reborrow",
 ]
+[[package]]
+name = "pulp"
+version = "0.21.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "95fb7a99b37aaef4c7dd2fd15a819eb8010bfc7a2c2155230d51f497316cad6d"
+dependencies = [
+ "bytemuck",
+ "cfg-if 1.0.0",
+ "libm",
+ "num-complex",
+ "reborrow",
+ "version_check",
+]
 [[package]]
 name = "pyo3"
 version = "0.23.5"
@@ -4754,6 +5014,15 @@ dependencies = [
 "bitflags 1.3.2",
 ]
+[[package]]
+name = "raw-cpuid"
+version = "11.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6df7ab838ed27997ba19a4664507e6f82b41fe6e20be42929332156e5e85146"
+dependencies = [
+ "bitflags 2.9.0",
+]
 [[package]]
 name = "rawpointer"
 version = "0.2.1"
@@ -5785,6 +6054,20 @@ dependencies = [
 "walkdir",
 ]
+[[package]]
+name = "sysctl"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "01198a2debb237c62b6826ec7081082d951f46dbb64b0e8c7649a452230d1dfc"
+dependencies = [
+ "bitflags 2.9.0",
+ "byteorder",
+ "enum-as-inner",
+ "libc",
+ "thiserror 1.0.69",
+ "walkdir",
+]
 [[package]]
 name = "sysinfo"
 version = "0.30.13"
@@ -6441,6 +6724,40 @@ version = "0.1.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2896d95c02a80c6d6a5d6e953d479f5ddf2dfdb6a244441010e373ac0fb88971"
+[[package]]
+name = "ug"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "03719c61a91b51541f076dfdba45caacf750b230cefaa4b32d6f5411c3f7f437"
+dependencies = [
+ "gemm 0.18.2",
+ "half",
+ "libloading",
+ "memmap2",
+ "num",
+ "num-traits",
+ "num_cpus",
+ "rayon",
+ "safetensors",
+ "serde",
+ "thiserror 1.0.69",
+ "tracing",
+ "yoke",
+]
+[[package]]
+name = "ug-cuda"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "50758486d7941f8b0a636ba7e29455c07071f41590beac1fd307ec893e8db69a"
+dependencies = [
+ "cudarc 0.13.9 (registry+https://github.com/rust-lang/crates.io-index)",
+ "half",
+ "serde",
+ "thiserror 1.0.69",
+ "ug",
+]
 [[package]]
 name = "unarray"
 version = "0.1.4"

--- a/docs/guides/dynamo_run.md
+++ b/docs/guides/dynamo_run.md
@@ -177,12 +177,7 @@ dynamo-run in=none out=sglang --model-path ~/llm_models/DeepSeek-R1-Distill-Llam
 - `cargo build --features llamacpp,cuda`
- `dynamo-run out=llama_cpp --model-path ~/llm_models/Llama-3.2-3B-Instruct-Q6_K.gguf --model-config ~/llm_models/Llama-3.2-3B-Instruct/`
+- `dynamo-run out=llama_cp ~/llm_models/Llama-3.2-3B-Instruct-Q6_K.gguf`
-The extra `--model-config` flag is because:
- llama_cpp only runs GGUF
- We send it tokens, meaning we do the tokenization ourself, so we need a tokenizer
- We don't yet read it out of the GGUF (TODO), so we need an HF repo with `tokenizer.json` et al
 If the build step also builds llama_cpp libraries into the same folder as the binary ("libllama.so", "libggml.so", "libggml-base.so", "libggml-cpu.so", "libggml-cuda.so"), then `dynamo-run` will need to find those at runtime. Set `LD_LIBRARY_PATH`, and be sure to deploy them alongside the `dynamo-run` binary.
@@ -215,7 +210,7 @@ Run (still inside that virtualenv) - HF repo:
 Run (still inside that virtualenv) - GGUF:
 ```
-./dynamo-run in=http out=vllm --model-path ~/llm_models/Llama-3.2-3B-Instruct-Q6_K.gguf --model-config ~/llm_models/Llama-3.2-3B-Instruct/
+./dynamo-run in=http out=vllm ~/llm_models/Llama-3.2-3B-Instruct-Q6_K.gguf
 ```
 + Multi-node:

--- a/launch/dynamo-run/src/input/common.rs
+++ b/launch/dynamo-run/src/input/common.rs
@@ -74,7 +74,9 @@ pub async fn prepare_engine(
            let preprocessor = OpenAIPreprocessor::new(*card.clone())
                .await?
                .into_operator();
-            let backend = Backend::from_mdc(*card.clone()).await?.into_operator();
+            let backend = Backend::from_tokenizer(card.tokenizer_hf()?)
+                .await?
+                .into_operator();
            let engine = ServiceBackend::from_engine(inner_engine);
            let pipeline = frontend

--- a/launch/dynamo-run/src/lib.rs
+++ b/launch/dynamo-run/src/lib.rs
@@ -126,46 +126,49 @@ pub async fn run(
    // Load the model deployment card, if any
    // Only used by some engines, so without those feature flags it's unused.
    #[allow(unused_variables)]
-    let (maybe_card_path, maybe_card) = match (&model_path, &flags.model_config) {
+    let maybe_card = match (&model_path, &flags.model_config) {
        // --model-config takes precedence
        (_, Some(model_config)) => {
-            let card =
+            match ModelDeploymentCard::from_local_path(model_config, model_name.as_deref()).await {
-                match ModelDeploymentCard::from_local_path(model_config, model_name.as_deref())
-                    .await
-                {
                Ok(card) => Some(card),
                Err(e) => {
                    tracing::error!(
-                            "Failed to load model card from config path {}: {}",
+                        "Failed to load model card from --model-config path {}: {e}",
                        model_config.display(),
-                            e
                    );
                    None
                }
-                };
+            }
-            (Some(model_config.clone()), card)
        }
        // If --model-path is an HF repo use that
        (Some(model_path), _) if model_path.is_dir() => {
-            let card = match ModelDeploymentCard::from_local_path(model_path, model_name.as_deref())
+            match ModelDeploymentCard::from_local_path(model_path, model_name.as_deref()).await {
-                .await
-            {
                Ok(card) => Some(card),
                Err(e) => {
                    tracing::error!(
-                        "Failed to load model card from model path {}: {}",
+                        "Failed to load model card from --model-path {}: {e}",
                        model_path.display(),
-                        e
                    );
                    None
                }
-            };
+            }
-            (Some(model_path.clone()), card)
+        }
+        (Some(model_path), _) if model_path.is_file() => {
+            match ModelDeploymentCard::from_gguf(model_path, model_name.as_deref()).await {
+                Ok(card) => Some(card),
+                Err(e) => {
+                    tracing::error!(
+                        "Failed to load model card from GGUF {}: {e}",
+                        model_path.display(),
+                    );
+                    None
+                }
+            }
        }
        // Otherwise we don't have one, but we only need it if we're tokenizing
        _ => {
            tracing::debug!("No model card path provided (neither --model-config nor a directory in --model-path)");
-            (None, None)
+            None
        }
    };
@@ -276,17 +279,9 @@ pub async fn run(
                    "out=vllm requires flag --model-path=<full-path-to-hf-repo-or-model-gguf>"
                );
            };
-            let Some(card_path) = maybe_card_path else {
-                // If we have a gguf we also need a model card because we don't currently parse
-                // tokenizer et al out of gguf.
-                anyhow::bail!(
-                    "Running GGUF files also requires a `--model-config` for the tokenizer et al."
-                );
-            };
            let Some(card) = maybe_card.clone() else {
                anyhow::bail!(
-                    "Failed to load model card: either unsupported HuggingFace repo format \
+                    "Unable to build tokenizer. out=vllm requires --model-path to be an HF repo with fast tokenizer (tokenizer.json) or a GGUF file"
-                    or for GGUF files --model-config is missing."
                );
            };
            let Some(sock_prefix) = zmq_socket_prefix else {
@@ -311,7 +306,6 @@ pub async fn run(
                // vllm multi-node only the leader runs vllm
                let (engine, vllm_future) = vllm::make_leader_engine(
                    cancel_token.clone(),
-                    &card_path,
                    &model_path,
                    &sock_prefix,
                    node_conf,

--- a/launch/dynamo-run/src/main.rs
+++ b/launch/dynamo-run/src/main.rs
@@ -80,12 +80,6 @@ fn main() -> anyhow::Result<()> {
            let Some(model_path) = flags.model_path_flag else {
                anyhow::bail!("vllm subprocess requires --model-path flag");
            };
-            let Some(model_config) = flags.model_config else {
-                anyhow::bail!("vllm subprocess requires --model-config");
-            };
-            if !model_config.is_dir() {
-                anyhow::bail!("vllm subprocess requires model config path to be a directory containing tokenizer.json, config.json, etc");
-            }
            if cfg!(feature = "vllm") {
                #[cfg(feature = "vllm")]
                {
@@ -97,7 +91,6 @@ fn main() -> anyhow::Result<()> {
                    };
                    return vllm::run_subprocess(
                        ZMQ_SOCKET_PREFIX,
-                        &model_config,
                        &model_path,
                        node_config,
                        flags.tensor_parallel_size,

--- a/lib/bindings/python/Cargo.lock
+++ b/lib/bindings/python/Cargo.lock
@@ -26,6 +26,12 @@ dependencies = [
 "memchr",
 ]
+[[package]]
+name = "akin"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1763692fc1416554cf051efc56a3de5595eca47299d731cc5c2b583adf8b4d2f"
 [[package]]
 name = "android-tzdata"
 version = "0.1.1"
@@ -97,6 +103,15 @@ version = "1.0.97"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "dcfed56ad506cb2c684a14971b8861fdc3baaaae314b9e5f9bb532cbe3ba7a4f"
+[[package]]
+name = "arbitrary"
+version = "1.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dde20b3d026af13f561bdd0f15edf01fc734f0dafcedbaf42bba506a9517f223"
+dependencies = [
+ "derive_arbitrary",
+]
 [[package]]
 name = "arrayref"
 version = "0.3.9"
@@ -424,6 +439,21 @@ dependencies = [
 "syn 2.0.100",
 ]
+[[package]]
+name = "bit-set"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "08807e080ed7f9d5433fa9b275196cfc35414f66a0c79d864dc51a0d825231a3"
+dependencies = [
+ "bit-vec",
+]
+[[package]]
+name = "bit-vec"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e764a1d40d510daf35e07be9eb06e75770908c27d411ee6c92109c9840eaaf7"
 [[package]]
 name = "bitflags"
 version = "1.3.2"
@@ -480,6 +510,9 @@ name = "bytemuck"
 version = "1.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6b1fc10dbac614ebc03540c9dbd60e83887fda27794998c6528f1782047d540"
+dependencies = [
+ "bytemuck_derive",
+]
 [[package]]
 name = "bytemuck_derive"
@@ -507,6 +540,28 @@ dependencies = [
 "serde",
 ]
+[[package]]
+name = "candle-core"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "06ccf5ee3532e66868516d9b315f73aec9f34ea1a37ae98514534d458915dbf1"
+dependencies = [
+ "byteorder",
+ "gemm 0.17.1",
+ "half",
+ "memmap2",
+ "num-traits",
+ "num_cpus",
+ "rand 0.9.0",
+ "rand_distr",
+ "rayon",
+ "safetensors",
+ "thiserror 1.0.69",
+ "ug",
+ "yoke",
+ "zip",
+]
 [[package]]
 name = "cc"
 version = "1.2.16"
@@ -755,6 +810,12 @@ version = "0.8.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
+[[package]]
+name = "crunchy"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "43da5946c66ffcc7745f48db692ffbb10a83bfe0afd96235c5c2a4fb23994929"
 [[package]]
 name = "crypto-common"
 version = "0.1.6"
@@ -870,6 +931,17 @@ dependencies = [
 "syn 2.0.100",
 ]
+[[package]]
+name = "derive_arbitrary"
+version = "1.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "30542c1ad912e0e3d22a1935c290e12e8a29d704a420177a31faad4a601a0800"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
+]
 [[package]]
 name = "derive_builder"
 version = "0.20.2"
@@ -911,6 +983,15 @@ dependencies = [
 "crypto-common",
 ]
+[[package]]
+name = "digit-layout"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09157630eece4139f6cc5a457556d308c3465ecd5af492f0e5aadc043997e2ce"
+dependencies = [
+ "half",
+]
 [[package]]
 name = "dircpy"
 version = "0.3.19"
@@ -954,10 +1035,30 @@ dependencies = [
 "syn 2.0.100",
 ]
+[[package]]
+name = "dyn-stack"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "56e53799688f5632f364f8fb387488dd05db9fe45db7011be066fc20e7027f8b"
+dependencies = [
+ "bytemuck",
+ "reborrow",
+]
+[[package]]
+name = "dyn-stack"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "490bd48eb68fffcfed519b4edbfd82c69cbe741d175b84f0e0cbe8c57cbe0bdd"
+dependencies = [
+ "bytemuck",
+]
 [[package]]
 name = "dynamo-llm"
 version = "0.1.0"
 dependencies = [
+ "akin",
 "anyhow",
 "async-openai",
 "async-stream",
@@ -969,6 +1070,7 @@ dependencies = [
 "bs62",
 "bytemuck",
 "bytes",
+ "candle-core",
 "chrono",
 "cmake",
 "derive-getters",
@@ -978,9 +1080,11 @@ dependencies = [
 "erased-serde",
 "futures",
 "galil-seiferas",
+ "ggus",
 "indexmap 2.8.0",
 "itertools 0.14.0",
 "libc",
+ "memmap2",
 "minijinja",
 "minijinja-contrib",
 "prometheus",
@@ -1123,6 +1227,18 @@ version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "34aa73646ffb006b8f5147f3dc182bd4bcb190227ce861fc4a4844bf8e3cb2c0"
+[[package]]
+name = "enum-as-inner"
+version = "0.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a1e6a265c649f3f5979b601d26f1d05ada116434c87741c9493cb56218f76cbc"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
+]
 [[package]]
 name = "enum-ordinalize"
 version = "4.3.0"
@@ -1226,6 +1342,17 @@ dependencies = [
 "pin-project-lite",
 ]
+[[package]]
+name = "fancy-regex"
+version = "0.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e24cb5a94bcae1e5408b0effca5cd7172ea3c5755049c5f3af4cd283a165298"
+dependencies = [
+ "bit-set",
+ "regex-automata 0.4.9",
+ "regex-syntax 0.8.5",
+]
 [[package]]
 name = "fastrand"
 version = "2.3.0"
@@ -1407,6 +1534,243 @@ dependencies = [
 "unchecked-index",
 ]
+[[package]]
+name = "gemm"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ab24cc62135b40090e31a76a9b2766a501979f3070fa27f689c27ec04377d32"
+dependencies = [
+ "dyn-stack 0.10.0",
+ "gemm-c32 0.17.1",
+ "gemm-c64 0.17.1",
+ "gemm-common 0.17.1",
+ "gemm-f16 0.17.1",
+ "gemm-f32 0.17.1",
+ "gemm-f64 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab96b703d31950f1aeddded248bc95543c9efc7ac9c4a21fda8703a83ee35451"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-c32 0.18.2",
+ "gemm-c64 0.18.2",
+ "gemm-common 0.18.2",
+ "gemm-f16 0.18.2",
+ "gemm-f32 0.18.2",
+ "gemm-f64 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-c32"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9c030d0b983d1e34a546b86e08f600c11696fde16199f971cd46c12e67512c0"
+dependencies = [
+ "dyn-stack 0.10.0",
+ "gemm-common 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-c32"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f6db9fd9f40421d00eea9dd0770045a5603b8d684654816637732463f4073847"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-c64"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fbb5f2e79fefb9693d18e1066a557b4546cd334b226beadc68b11a8f9431852a"
+dependencies = [
+ "dyn-stack 0.10.0",
+ "gemm-common 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-c64"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dfcad8a3d35a43758330b635d02edad980c1e143dc2f21e6fd25f9e4eada8edf"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-common"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a2e7ea062c987abcd8db95db917b4ffb4ecdfd0668471d8dc54734fdff2354e8"
+dependencies = [
+ "bytemuck",
+ "dyn-stack 0.10.0",
+ "half",
+ "num-complex",
+ "num-traits",
+ "once_cell",
+ "paste",
+ "pulp 0.18.22",
+ "raw-cpuid 10.7.0",
+ "rayon",
+ "seq-macro",
+ "sysctl 0.5.5",
+]
+[[package]]
+name = "gemm-common"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a352d4a69cbe938b9e2a9cb7a3a63b7e72f9349174a2752a558a8a563510d0f3"
+dependencies = [
+ "bytemuck",
+ "dyn-stack 0.13.0",
+ "half",
+ "libm",
+ "num-complex",
+ "num-traits",
+ "once_cell",
+ "paste",
+ "pulp 0.21.4",
+ "raw-cpuid 11.5.0",
+ "rayon",
+ "seq-macro",
+ "sysctl 0.6.0",
+]
+[[package]]
+name = "gemm-f16"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7ca4c06b9b11952071d317604acb332e924e817bd891bec8dfb494168c7cedd4"
+dependencies = [
+ "dyn-stack 0.10.0",
+ "gemm-common 0.17.1",
+ "gemm-f32 0.17.1",
+ "half",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "rayon",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f16"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cff95ae3259432f3c3410eaa919033cd03791d81cebd18018393dc147952e109"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "gemm-f32 0.18.2",
+ "half",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
+ "rayon",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f32"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e9a69f51aaefbd9cf12d18faf273d3e982d9d711f60775645ed5c8047b4ae113"
+dependencies = [
+ "dyn-stack 0.10.0",
+ "gemm-common 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f32"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bc8d3d4385393304f407392f754cd2dc4b315d05063f62cf09f47b58de276864"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f64"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "aa397a48544fadf0b81ec8741e5c0fba0043008113f71f2034def1935645d2b0"
+dependencies = [
+ "dyn-stack 0.10.0",
+ "gemm-common 0.17.1",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 10.7.0",
+ "seq-macro",
+]
+[[package]]
+name = "gemm-f64"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35b2a4f76ce4b8b16eadc11ccf2e083252d8237c1b589558a49b0183545015bd"
+dependencies = [
+ "dyn-stack 0.13.0",
+ "gemm-common 0.18.2",
+ "num-complex",
+ "num-traits",
+ "paste",
+ "raw-cpuid 11.5.0",
+ "seq-macro",
+]
 [[package]]
 name = "generic-array"
 version = "0.14.7"
@@ -1444,6 +1808,30 @@ dependencies = [
 "wasm-bindgen",
 ]
+[[package]]
+name = "ggml-quants"
+version = "0.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a27693512784e0786212eb0bef841779a6337d2d04520ed475b4d5a864f98366"
+dependencies = [
+ "digit-layout",
+ "half",
+ "rayon",
+]
+[[package]]
+name = "ggus"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3ac5654356c6f7f6116905aeaf92ab002c3d03414ada5dbe0bb2e32aa5fea173"
+dependencies = [
+ "fancy-regex",
+ "ggml-quants",
+ "indexmap 2.8.0",
+ "log",
+ "num_enum",
+]
 [[package]]
 name = "gimli"
 version = "0.31.1"
@@ -1475,6 +1863,20 @@ dependencies = [
 "tracing",
 ]
+[[package]]
+name = "half"
+version = "2.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7db2ff139bba50379da6aa0766b52fdcb62cb5b263009b09ed58ba604e14bbd1"
+dependencies = [
+ "bytemuck",
+ "cfg-if 1.0.0",
+ "crunchy",
+ "num-traits",
+ "rand 0.9.0",
+ "rand_distr",
+]
 [[package]]
 name = "hashbrown"
 version = "0.12.3"
@@ -1493,6 +1895,12 @@ version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
+[[package]]
+name = "hermit-abi"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d231dfb89cfffdbc30e7fc41579ed6066ad03abda9e567ccafae602b97ec5024"
 [[package]]
 name = "hf-hub"
 version = "0.4.2"
@@ -1989,6 +2397,12 @@ dependencies = [
 "windows-targets 0.52.6",
 ]
+[[package]]
+name = "libm"
+version = "0.2.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8355be11b20d696c8f18f6cc018c4e372165b1fa8126cef092399c9951984ffa"
 [[package]]
 name = "libredox"
 version = "0.1.3"
@@ -2082,6 +2496,16 @@ version = "2.7.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "78ca9ab1a0babb1e7d5695e3530886289c18cf2f87ec19a575a0abdce112e3a3"
+[[package]]
+name = "memmap2"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fd3f7eed9d3848f8b98834af67102b720745c4ec028fcd0aa0239277e7de374f"
+dependencies = [
+ "libc",
+ "stable_deref_trait",
+]
 [[package]]
 name = "memo-map"
 version = "0.3.3"
@@ -2321,6 +2745,20 @@ dependencies = [
 "rand 0.8.5",
 ]
+[[package]]
+name = "num"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35bd024e8b2ff75562e5f34e7f4905839deb4b22955ef5e73d2fea1b9813cb23"
+dependencies = [
+ "num-bigint",
+ "num-complex",
+ "num-integer",
+ "num-iter",
+ "num-rational",
+ "num-traits",
+]
 [[package]]
 name = "num-bigint"
 version = "0.4.6"
@@ -2331,6 +2769,16 @@ dependencies = [
 "num-traits",
 ]
+[[package]]
+name = "num-complex"
+version = "0.4.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "73f88a1307638156682bada9d7604135552957b7818057dcef22705b4d509495"
+dependencies = [
+ "bytemuck",
+ "num-traits",
+]
 [[package]]
 name = "num-conv"
 version = "0.1.0"
@@ -2346,6 +2794,28 @@ dependencies = [
 "num-traits",
 ]
+[[package]]
+name = "num-iter"
+version = "0.1.45"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1429034a0490724d0075ebb2bc9e875d6503c3cf69e235a8941aa757d83ef5bf"
+dependencies = [
+ "autocfg",
+ "num-integer",
+ "num-traits",
+]
+[[package]]
+name = "num-rational"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f83d14da390562dca69fc84082e73e548e1ad308d24accdedd2720017cb37824"
+dependencies = [
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
 [[package]]
 name = "num-traits"
 version = "0.2.19"
@@ -2353,6 +2823,38 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841"
 dependencies = [
 "autocfg",
+ "libm",
+]
+[[package]]
+name = "num_cpus"
+version = "1.16.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4161fcb6d602d4d2081af7c3a45852d875a03dd337a6bfdd6e06407b61342a43"
+dependencies = [
+ "hermit-abi",
+ "libc",
+]
+[[package]]
+name = "num_enum"
+version = "0.7.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4e613fc340b2220f734a8595782c551f1250e969d87d3be1ae0579e8d4065179"
+dependencies = [
+ "num_enum_derive",
+]
+[[package]]
+name = "num_enum_derive"
+version = "0.7.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "af1844ef2428cc3e1cb900be36181049ef3d3193c63e43026cfe202983b27a56"
+dependencies = [
+ "proc-macro-crate",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
 ]
 [[package]]
@@ -2587,6 +3089,15 @@ dependencies = [
 "syn 2.0.100",
 ]
+[[package]]
+name = "proc-macro-crate"
+version = "3.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "edce586971a4dfaa28950c6f18ed55e0406c1ab88bbce2c6f6293a7aaba73d35"
+dependencies = [
+ "toml_edit",
+]
 [[package]]
 name = "proc-macro-error-attr2"
 version = "2.0.0"
@@ -2704,6 +3215,32 @@ version = "2.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "106dd99e98437432fed6519dedecfade6a06a73bb7b2a1e019fdd2bee5778d94"
+[[package]]
+name = "pulp"
+version = "0.18.22"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a0a01a0dc67cf4558d279f0c25b0962bd08fc6dec0137699eae304103e882fe6"
+dependencies = [
+ "bytemuck",
+ "libm",
+ "num-complex",
+ "reborrow",
+]
+[[package]]
+name = "pulp"
+version = "0.21.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "95fb7a99b37aaef4c7dd2fd15a819eb8010bfc7a2c2155230d51f497316cad6d"
+dependencies = [
+ "bytemuck",
+ "cfg-if 1.0.0",
+ "libm",
+ "num-complex",
+ "reborrow",
+ "version_check",
+]
 [[package]]
 name = "pyo3"
 version = "0.23.5"
@@ -2934,6 +3471,34 @@ dependencies = [
 "getrandom 0.3.2",
 ]
+[[package]]
+name = "rand_distr"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6a8615d50dcf34fa31f7ab52692afec947c4dd0ab803cc87cb3b0b4570ff7463"
+dependencies = [
+ "num-traits",
+ "rand 0.9.0",
+]
+[[package]]
+name = "raw-cpuid"
+version = "10.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6c297679cb867470fa8c9f67dbba74a78d78e3e98d7cf2b08d6d71540f797332"
+dependencies = [
+ "bitflags 1.3.2",
+]
+[[package]]
+name = "raw-cpuid"
+version = "11.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6df7ab838ed27997ba19a4664507e6f82b41fe6e20be42929332156e5e85146"
+dependencies = [
+ "bitflags 2.9.0",
+]
 [[package]]
 name = "rayon"
 version = "1.10.0"
@@ -2965,6 +3530,12 @@ dependencies = [
 "crossbeam-utils",
 ]
+[[package]]
+name = "reborrow"
+version = "0.5.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "03251193000f4bd3b042892be858ee50e8b3719f2b08e5833ac4353724632430"
 [[package]]
 name = "redox_syscall"
 version = "0.5.10"
@@ -3238,6 +3809,16 @@ version = "1.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "28d3b2b1366ec20994f1fd18c3c594f05c5dd4bc44d8bb0c1c632c8d6829481f"
+[[package]]
+name = "safetensors"
+version = "0.4.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44560c11236a6130a46ce36c836a62936dc81ebf8c36a37947423571be0e55b6"
+dependencies = [
+ "serde",
+ "serde_json",
+]
 [[package]]
 name = "same-file"
 version = "1.0.6"
@@ -3323,6 +3904,12 @@ dependencies = [
 "serde",
 ]
+[[package]]
+name = "seq-macro"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1bc711410fbe7399f390ca1c3b60ad0f53f80e95c5eb935e52268a0e2cd49acc"
 [[package]]
 name = "serde"
 version = "1.0.219"
@@ -3616,6 +4203,34 @@ dependencies = [
 "syn 2.0.100",
 ]
+[[package]]
+name = "sysctl"
+version = "0.5.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ec7dddc5f0fee506baf8b9fdb989e242f17e4b11c61dfbb0635b705217199eea"
+dependencies = [
+ "bitflags 2.9.0",
+ "byteorder",
+ "enum-as-inner",
+ "libc",
+ "thiserror 1.0.69",
+ "walkdir",
+]
+[[package]]
+name = "sysctl"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "01198a2debb237c62b6826ec7081082d951f46dbb64b0e8c7649a452230d1dfc"
+dependencies = [
+ "bitflags 2.9.0",
+ "byteorder",
+ "enum-as-inner",
+ "libc",
+ "thiserror 1.0.69",
+ "walkdir",
+]
 [[package]]
 name = "system-deps"
 version = "6.2.2"
@@ -4131,6 +4746,27 @@ version = "1.18.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1dccffe3ce07af9386bfd29e80c0ab1a8205a2fc34e4bcd40364df902cfa8f3f"
+[[package]]
+name = "ug"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "03719c61a91b51541f076dfdba45caacf750b230cefaa4b32d6f5411c3f7f437"
+dependencies = [
+ "gemm 0.18.2",
+ "half",
+ "libloading",
+ "memmap2",
+ "num",
+ "num-traits",
+ "num_cpus",
+ "rayon",
+ "safetensors",
+ "serde",
+ "thiserror 1.0.69",
+ "tracing",
+ "yoke",
+]
 [[package]]
 name = "uncased"
 version = "0.9.10"
@@ -4904,6 +5540,21 @@ dependencies = [
 "syn 2.0.100",
 ]
+[[package]]
+name = "zip"
+version = "1.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9cc23c04387f4da0374be4533ad1208cbb091d5c11d070dfef13676ad6497164"
+dependencies = [
+ "arbitrary",
+ "crc32fast",
+ "crossbeam-utils",
+ "displaydoc",
+ "indexmap 2.8.0",
+ "num_enum",
+ "thiserror 1.0.69",
+]
 [[package]]
 name = "zmq"
 version = "0.10.0"

--- a/lib/llm/Cargo.toml
+++ b/lib/llm/Cargo.toml
@@ -36,7 +36,7 @@ sentencepiece = ["dep:sentencepiece"]
 trtllm = []
 vllm = []
-cuda = ["mistralrs/cuda", "llama-cpp-2/cuda"]
+cuda = ["mistralrs/cuda", "llama-cpp-2/cuda", "candle-core/cuda"]
 metal = ["mistralrs/metal", "llama-cpp-2/metal"]
 vulkan = ["llama-cpp-2/vulkan"]
@@ -68,9 +68,11 @@ uuid = { workspace = true }
 xxhash-rust = { workspace = true }
 strum = { workspace = true }
+akin = "0.4.0"
 async-openai = "0.27.2"
 blake3 = "1"
 bytemuck = "1.22"
+candle-core = { version = "0.8.0" }
 derive-getters = "0.5"
 regex = "1"
 rayon = "1"
@@ -78,7 +80,6 @@ rayon = "1"
 # kv_cuda
 cudarc = { git = "https://github.com/coreylowman/cudarc.git", rev = "8c52e735b55bf8e979e1a16bd85e3dfe4f87c9fe", features = ["cuda-12040"], optional = true }
 ndarray = { version = "0.16", optional = true }
-# candle-core = { version = "0.8.3", features = ["cuda"], optional = true }
 # half = "2.4.1"
 pyo3 = { version = "0.23.3", default-features = false, features = [
@@ -138,6 +139,9 @@ pyo3-async-runtimes = { version = "0.23.0", optional = true, default-features =
 ] }
 pythonize = { version = "0.23", optional = true }
+# GGUF
+ggus = "0.4.0"
+memmap2 = "0.9.5"
 [dev-dependencies]
 hf-hub = { workspace = true }

--- a/lib/llm/src/backend.rs
+++ b/lib/llm/src/backend.rs
@@ -52,9 +52,6 @@ use crate::protocols::{
 use crate::tokenizers::{DecodeStream, HuggingFaceTokenizer, Tokenizer};
 use tokenizers::Tokenizer as HfTokenizer;
-use toktrie::TokTrie;
-use toktrie_hf_tokenizers::ByteTokenizer;
 /// Represents the output stream from the execution engine
 pub type ExecutionOutputStream = Annotated<LLMEngineOutput>;
@@ -64,12 +61,8 @@ pub type ExecutionContext = ServerStreamingEngine<BackendInput, ExecutionOutputS
 /// Backend handles resource management and orchestrates LLM execution
 #[allow(dead_code)]
 pub struct Backend {
-    mdc: ModelDeploymentCard,
    pub tokenizer: Tokenizer,     // Handles token encoding/decoding
-    tok_trie: Arc<TokTrie>,          // Efficient token lookup structure
-    eos_token_ids: Vec<TokenIdType>, // End of sequence token IDs
    validate_engine_decode: bool, // Enable validation of engine decoding
-    mdcsum: String,                  // Model deployment checksum
 }
 /// Internal state for managing token decoding and stream processing
@@ -78,52 +71,40 @@ struct DecoderUnfoldState {
    stream: ManyOut<ExecutionOutputStream>,
    decoder: Decoder,
    validate_engine_decode: bool,
-    mdcsum: String,
 }
 impl Backend {
-    pub async fn from_mdc(mdc: ModelDeploymentCard) -> Result<Arc<Self>> {
+    pub async fn from_tokenizer(tokenizer: HfTokenizer) -> Result<Arc<Self>> {
-        let info = mdc.model_info.get_model_info().await?;
-        let tokenizer = match &mdc.tokenizer {
-            TokenizerKind::HfTokenizerJson(file) => {
-                HfTokenizer::from_file(file).map_err(Error::msg)?
-            }
-        };
-        let bt = ByteTokenizer::from_tokenizer(tokenizer.clone())?;
-        let toktrie = TokTrie::from(&bt.tokrx_info(), &bt.token_bytes());
-        let mdcsum = mdc.mdcsum();
        let tokenizer = HuggingFaceTokenizer::from_tokenizer(tokenizer);
        let tokenizer = Tokenizer::from(Arc::new(tokenizer));
        Ok(Arc::new(Self {
-            mdc,
            tokenizer,
-            tok_trie: Arc::new(toktrie),
-            eos_token_ids: info.eos_token_ids(),
            validate_engine_decode: false,
-            mdcsum,
        }))
    }
+    pub async fn from_mdc(mdc: ModelDeploymentCard) -> Result<Arc<Self>> {
+        let tokenizer = match &mdc.tokenizer {
+            TokenizerKind::HfTokenizerJson(file) => {
+                HfTokenizer::from_file(file).map_err(Error::msg)?
+            }
+            TokenizerKind::GGUF(t) => *t.clone(),
+        };
+        Self::from_tokenizer(tokenizer).await
+    }
    fn decoder(
        &self,
        stream: ManyOut<ExecutionOutputStream>,
        stop_conditions: StopConditions,
    ) -> DecoderUnfoldState {
-        let decoder = Decoder::new(
+        let decoder = Decoder::new(self.tokenizer.decode_stream(false), stop_conditions);
-            self.tokenizer.decode_stream(false),
-            stop_conditions,
-            self.mdcsum.clone(),
-        );
        DecoderUnfoldState {
            stream,
            decoder,
            validate_engine_decode: self.validate_engine_decode,
-            mdcsum: self.mdcsum.clone(),
        }
    }
 }
@@ -223,7 +204,7 @@ impl
        });
        // convert stream of processed Annotated<LLMEngineOutput> to Annotated<BackendOutput>
-        let mdcsum = self.mdcsum.clone();
+        //let mdcsum = self.mdcsum.clone();
        let stream = processed_stream.map(move |output| {
            output.map_data(|data| {
                Ok(BackendOutput {
@@ -233,7 +214,7 @@ impl
                    cum_log_probs: data.cum_log_probs,
                    log_probs: data.log_probs,
                    finish_reason: data.finish_reason,
-                    mdcsum: mdcsum.clone(),
+                    //mdcsum: mdcsum.clone(),
                })
            })
        });
@@ -275,9 +256,8 @@ pub struct Decoder {
    // the number of bytes currently jailed
    jailed_bytes: usize,
    // mdcsum
-    mdcsum: String,
+    //mdcsum: String,
 }
 #[allow(dead_code)]
@@ -331,7 +311,7 @@ impl Decoder {
    pub fn new(
        decode_stream: DecodeStream,
        stop_condition: StopConditions,
-        mdcsum: String,
+        //mdcsum: String,
    ) -> Self {
        let hidden_stop_ids: HashSet<TokenIdType> = stop_condition
            .stop_token_ids_hidden
@@ -364,7 +344,6 @@ impl Decoder {
            jail: String::new(),
            jail_max_bytes,
            jailed_bytes: 0,
-            mdcsum,
        }
    }

--- a/lib/llm/src/engines/vllm.rs
+++ b/lib/llm/src/engines/vllm.rs
@@ -38,8 +38,6 @@ mod worker;
 pub async fn make_leader_engine(
    cancel_token: CancellationToken,
-    // Where to find the tokenzier, and config.json
-    card_path: &Path,
    // Full path to the model, either a GGUF file or an HF repo dir
    model_path: &Path,
    // Unique string to name zmq sockets
@@ -63,7 +61,6 @@ pub async fn make_leader_engine(
    let mut engine = VllmEngine::new(
        cancel_token,
        sock_code,
-        card_path,
        model_path,
        node_conf,
        tensor_parallel_size,

--- a/lib/llm/src/engines/vllm/engine.rs
+++ b/lib/llm/src/engines/vllm/engine.rs
@@ -35,7 +35,6 @@ impl VllmEngine {
    pub async fn new(
        cancel_token: CancellationToken,
        sock_code: &str,
-        card_path: &Path,
        model_path: &Path,
        node_conf: MultiNodeConfig,
        tensor_parallel_size: u32,
@@ -43,7 +42,6 @@ impl VllmEngine {
        let w = worker::start(
            cancel_token.clone(),
            sock_code,
-            card_path,
            model_path,
            node_conf,
            tensor_parallel_size,

--- a/lib/llm/src/engines/vllm/subprocess.rs
+++ b/lib/llm/src/engines/vllm/subprocess.rs
@@ -30,9 +30,8 @@ from vllm.usage.usage_lib import UsageContext
 engine_args = AsyncEngineArgs(
    model=f"{model_path}",
    served_model_name=None,
-    tokenizer=f"{tokenizer_path}",
    task='generate',
-    tokenizer_mode='auto',
+    skip_tokenizer_init=True,
    seed=0,
    max_model_len=8192,
    max_seq_len_to_capture=8192,
@@ -43,7 +42,16 @@ engine_args = AsyncEngineArgs(
 ipc_path = f"ipc:///tmp/{socket_id}";
 engine_alive = multiprocessing.Value('b', True, lock=False)
+# 0.7.3
 run_mp_engine(engine_args, UsageContext.OPENAI_API_SERVER, ipc_path, engine_alive)
+# 0.8.1
+# TODO: In 0.8+ first argument is VllmConfig, not AsyncEngineArgs
+# disable_log_stats = False
+# disable_log_requests = True
+# run_mp_engine(engine_args, UsageContext.OPENAI_API_SERVER, ipc_path, disable_log_stats, disable_log_requests, engine_alive)
 "#;
 /// Start the Python vllm engine that listens on zmq socket
@@ -51,7 +59,6 @@ run_mp_engine(engine_args, UsageContext.OPENAI_API_SERVER, ipc_path, engine_aliv
 /// This does not return until vllm exits.
 pub fn run_subprocess(
    socket_id: &str,
-    model_card_path: &Path,
    model_path: &Path,
    node_config: MultiNodeConfig,
    tp_size: u32,
@@ -60,12 +67,10 @@ pub fn run_subprocess(
    if let Ok(venv) = env::var("VIRTUAL_ENV") {
        let _ = Python::with_gil(|py| crate::engines::fix_venv(venv, py));
    }
-    let card = model_card_path.display().to_string();
    let model_path_str = model_path.display().to_string();
    Python::with_gil(|py| {
        let locals = [
            ("socket_id", socket_id),
-            ("tokenizer_path", card.as_str()),
            ("model_path", model_path_str.as_str()),
            ("tp_size_str", &tp_size.to_string()),
            ("nnodes_str", &node_config.num_nodes.to_string()),

--- a/lib/llm/src/engines/vllm/worker.rs
+++ b/lib/llm/src/engines/vllm/worker.rs
@@ -160,7 +160,6 @@ struct Logprob {
 pub async fn start(
    cancel_token: CancellationToken,
    sock_code: &str,
-    card_path: &Path,
    model_path: &Path,
    _node_conf: MultiNodeConfig,
    tensor_parallel_size: u32,
@@ -180,14 +179,7 @@ pub async fn start(
        metrics,
    } = zmq_sockets(sock_code)?;
-    let vllm_process = start_vllm(
+    let vllm_process = start_vllm(model_path, &py_imports, data, tensor_parallel_size).await?;
-        card_path,
-        model_path,
-        &py_imports,
-        data,
-        tensor_parallel_size,
-    )
-    .await?;
    let vllm_join_handle = watch_vllm(cancel_token.clone(), vllm_process);
    tokio::spawn(heartbeat_loop(cancel_token.clone(), heartbeat));
@@ -308,7 +300,6 @@ fn zmq_sockets(sock_code: &str) -> anyhow::Result<Sockets> {
 /// Start the vllm python sub-process and wait for it to start
 async fn start_vllm(
-    card_path: &Path,
    model_path: &Path,
    python_imports: &Imports,
    mut data_socket: async_zmq::Dealer<IntoIter<Vec<u8>>, Vec<u8>>,
@@ -316,7 +307,6 @@ async fn start_vllm(
 ) -> anyhow::Result<tokio::process::Child> {
    let vllm_args = [
        "--internal-vllm-process",
-        &format!("--model-config={}", card_path.display()),
        &format!("--model-path={}", model_path.display()),
        &format!("--tensor-parallel-size={tensor_parallel_size}"),
    ];

--- a/lib/llm/src/gguf.rs
+++ b/lib/llm/src/gguf.rs
+// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+// Adapted from mistral.rs
+//
+// MIT License
+//
+// Copyright (c) 2025 Eric Buehler
+//
+// 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.
+mod content;
+mod gguf_metadata;
+mod gguf_tokenizer;
+use strum::EnumString;
+use anyhow::{Context, Result};
+pub(crate) use content::Content;
+pub(crate) use gguf_metadata::ContentConfig;
+pub(crate) use gguf_tokenizer::convert_gguf_to_hf_tokenizer;
+use std::str::FromStr;
+pub const GGUF_MULTI_FILE_DELIMITER: &str = " ";
+#[derive(Debug, EnumString, Clone, Copy, strum::Display)]
+#[strum(serialize_all = "lowercase")]
+pub enum GGUFArchitecture {
+    Llama,
+    Mpt,
+    Gptneox,
+    Gptj,
+    Gpt2,
+    Bloom,
+    Falcon,
+    Mamba,
+    Rwkv,
+    Phi2,
+    Phi3,
+    Starcoder2,
+    Qwen2,
+}
+// Wraps from_str() for some convenience:
+// - Case-insensitive variant matching (TODO: is this desirable?)
+// - Customized error until potential upstream support: https://github.com/Peternator7/strum/issues/332
+impl GGUFArchitecture {
+    pub fn from_value<T: AsRef<str> + std::fmt::Display>(value: T) -> Result<Self> {
+        Self::from_str(&value.as_ref().to_ascii_lowercase())
+            .with_context(|| format!("Unknown GGUF architecture `{value}`"))
+            .map_err(anyhow::Error::msg)
+    }
+}
--- a/lib/llm/src/gguf/content.rs
+++ b/lib/llm/src/gguf/content.rs
+// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+// Adapted from mistral.rs
+//
+// MIT License
+//
+// Copyright (c) 2025 Eric Buehler
+//
+// 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.
+use std::collections::HashMap;
+use anyhow::Context;
+use candle_core::{
+    quantized::gguf_file::{self, Value},
+    Result,
+};
+use tracing::info;
+use super::GGUFArchitecture;
+// Internal invariant: contents and readers must be paired.
+/// This abstracts the files for a GGUF model and enables multiple files to be used.
+pub struct Content {
+    _contents: Vec<gguf_file::Content>,
+    arch: GGUFArchitecture,
+    all_metadata: HashMap<String, Value>,
+}
+impl Content {
+    /// Create a `Content` from a set of file readers.
+    pub fn from_readers<R: std::io::Seek + std::io::Read>(readers: &mut [&mut R]) -> Result<Self> {
+        let mut contents = Vec::new();
+        let n_readers = readers.len();
+        for reader in readers.iter_mut() {
+            contents.push(gguf_file::Content::read(reader)?);
+        }
+        let n_splits = contents
+            .iter()
+            .filter_map(|ct| {
+                ct.metadata
+                    .get("split.count")
+                    .map(|val| val.to_u64().unwrap())
+            })
+            .fold(Vec::new(), |mut accum, x| {
+                if !accum.contains(&x) {
+                    accum.push(x);
+                }
+                accum
+            });
+        if n_splits.len() > 1 {
+            candle_core::bail!("GGUF files have differing `split.count` values: {n_splits:?}. Perhaps the GGUF files do not match?");
+        }
+        #[allow(clippy::cast_possible_truncation)]
+        if !n_splits.is_empty() && n_readers != n_splits[0] as usize {
+            candle_core::bail!(
+                "Number of GGUF files does not match the number of splits, expected {} files.",
+                n_splits[0]
+            );
+        } else if n_splits.len() == 1 {
+            info!("GGUF file has been split into {} shards", n_splits[0]);
+        }
+        let mut arch = None;
+        for ct in &contents {
+            if !ct.metadata.contains_key("general.architecture") {
+                continue;
+            }
+            arch = Some(
+                ct.metadata["general.architecture"]
+                    .to_string()
+                    .context("Model metadata should have declared an architecture")
+                    .and_then(GGUFArchitecture::from_value)
+                    .unwrap(),
+            );
+        }
+        let arch = arch.expect("GGUF files must specify `general.architecture`");
+        let mut all_metadata = HashMap::new();
+        for content in &contents {
+            all_metadata.extend(content.metadata.clone())
+        }
+        Ok(Self {
+            _contents: contents,
+            arch,
+            all_metadata,
+        })
+    }
+    pub fn arch(&self) -> GGUFArchitecture {
+        self.arch
+    }
+    /// Get all metadatas
+    pub fn get_metadata(&self) -> &HashMap<String, Value> {
+        &self.all_metadata
+    }
+}
--- a/lib/llm/src/gguf/gguf_metadata.rs
+++ b/lib/llm/src/gguf/gguf_metadata.rs
+// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+// Adapted from mistral.rs
+//
+// MIT License
+//
+// Copyright (c) 2025 Eric Buehler
+//
+// 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.
+use akin::akin;
+use anyhow::ensure;
+use anyhow::Result;
+use candle_core::quantized::gguf_file;
+use std::collections::HashMap;
+use tracing::warn;
+use crate::gguf::Content;
+#[allow(dead_code)]
+#[derive(Debug)]
+pub struct ContentConfig {
+    max_seq_len: usize,
+    hidden_size: usize,
+    num_attn_heads: usize,
+    num_kv_heads: usize,
+    num_layers: usize,
+    key_length: Option<usize>,
+    value_length: Option<usize>,
+}
+#[allow(clippy::cast_possible_truncation)]
+impl From<&Content> for ContentConfig {
+    fn from(value: &Content) -> Self {
+        let metadata = value.get_metadata();
+        let arch = metadata["general.architecture"].to_string().unwrap();
+        Self {
+            max_seq_len: metadata[&format!("{arch}.context_length")]
+                .to_u64()
+                .unwrap() as usize,
+            hidden_size: metadata[&format!("{arch}.embedding_length")]
+                .to_u64()
+                .unwrap() as usize,
+            num_attn_heads: metadata[&format!("{arch}.attention.head_count")]
+                .to_u64()
+                .unwrap() as usize,
+            num_kv_heads: metadata[&format!("{arch}.attention.head_count_kv")]
+                .to_u64()
+                .unwrap() as usize,
+            num_layers: metadata[&format!("{arch}.block_count")].to_u64().unwrap() as usize,
+            key_length: metadata
+                .get(&format!("{arch}.attention.key_length"))
+                .map(|x| x.to_u64().unwrap() as usize),
+            value_length: metadata
+                .get(&format!("{arch}.attention.value_length"))
+                .map(|x| x.to_u64().unwrap() as usize),
+        }
+    }
+}
+#[allow(dead_code)]
+impl ContentConfig {
+    pub fn max_seq_len(&self) -> usize {
+        self.max_seq_len
+    }
+    pub fn hidden_size(&self) -> usize {
+        self.hidden_size
+    }
+    pub fn num_attn_heads(&self) -> usize {
+        self.num_attn_heads
+    }
+    pub fn num_kv_heads(&self) -> usize {
+        self.num_kv_heads
+    }
+    pub fn num_layers(&self) -> usize {
+        self.num_layers
+    }
+    pub fn k_head_dim(&self) -> usize {
+        self.key_length
+            .unwrap_or(self.hidden_size / self.num_attn_heads)
+    }
+    pub fn v_head_dim(&self) -> usize {
+        self.value_length
+            .unwrap_or(self.hidden_size / self.num_attn_heads)
+    }
+}
+pub struct ContentMetadata<'a> {
+    pub path_prefix: &'a str,
+    pub metadata: &'a HashMap<String, gguf_file::Value>,
+}
+impl ContentMetadata<'_> {
+    // Retrieve a prop the struct needs by querying the metadata content:
+    pub fn get_value<T: TryFromValue>(&self, field_name: &str) -> Result<T, anyhow::Error> {
+        let prop_key = format!("{prefix}.{field_name}", prefix = self.path_prefix);
+        let value = self.metadata.get(&prop_key).cloned();
+        // Unwrap the inner value of the `Value` enum via trait method,
+        // otherwise format error with prop key as context:
+        value
+            .try_value_into()
+            .or_else(|e| anyhow::bail!("`{prop_key}` `{e}`"))
+    }
+    // Fail early - Catch all missing mandatory keys upfront:
+    pub fn has_required_keys(&self, fields: &[&str]) -> Result<()> {
+        let mut all_props_are_present = true;
+        for field_name in fields {
+            let prop_key = format!("{prefix}.{field_name}", prefix = self.path_prefix);
+            if !self.metadata.contains_key(&prop_key) {
+                all_props_are_present = false;
+                warn!("Expected GGUF metadata to have key: `{prop_key}`");
+            }
+        }
+        ensure!(all_props_are_present, "Tokenizer is missing required props");
+        Ok(())
+    }
+}
+// These traits below are a workaround for converting candles GGUF `Value` enum type wrapper.
+// A better upstream approach would instead be to provide serialize/deserialize support?
+pub trait TryFromValue {
+    fn try_from_value(value: gguf_file::Value) -> Result<Self, candle_core::Error>
+    where
+        Self: Sized;
+}
+// Value wrapped types, each has a different conversion method:
+// NOTE: Type conversion methods internally bail with "not a <into type> <input value>"
+// https://docs.rs/candle-core/latest/candle_core/quantized/gguf_file/enum.Value.html#variants
+akin! {
+    let &types = [String, bool, f32, f64, i8, i16, i32, i64, u8, u16, u32, u64];
+    let &to_type = [
+        value.to_string().cloned(),
+        value.to_bool(),
+        value.to_f32(),
+        value.to_f64(),
+        value.to_i8(),
+        value.to_i16(),
+        value.to_i32(),
+        value.to_i64(),
+        value.to_u8(),
+        value.to_u16(),
+        value.to_u32(),
+        value.to_u64(),
+    ];
+    impl TryFromValue for *types {
+        fn try_from_value(value: gguf_file::Value) -> Result<Self, candle_core::Error> {
+            *to_type.or_else(|_| candle_core::bail!("value is not a `*types`"))
+        }
+    }
+}
+// Vec<Value> to Vec<T> from above types:
+impl<T: TryFromValue> TryFromValue for Vec<T> {
+    fn try_from_value(value_vec: gguf_file::Value) -> Result<Self, candle_core::Error> {
+        value_vec
+            .to_vec()
+            .or_else(|_| candle_core::bail!("value is not a `Vec`"))?
+            .clone()
+            .into_iter()
+            .map(|item| T::try_from_value(item))
+            .collect()
+    }
+}
+pub trait TryValueInto<T>: Sized {
+    fn try_value_into(self) -> Result<T, candle_core::Error>;
+}
+impl<T: TryFromValue> TryValueInto<T> for gguf_file::Value {
+    fn try_value_into(self) -> Result<T, candle_core::Error> {
+        T::try_from_value(self)
+    }
+}
+impl<T: TryFromValue> TryValueInto<T> for Option<gguf_file::Value> {
+    fn try_value_into(self) -> Result<T, candle_core::Error> {
+        match self {
+            Some(value) => value.try_value_into(),
+            None => candle_core::bail!("Expected `Option<gguf_file::Value>` to contain a value"),
+        }
+    }
+}
--- a/lib/llm/src/gguf/gguf_tokenizer.rs
+++ b/lib/llm/src/gguf/gguf_tokenizer.rs
+// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+// Adapted from mistral.rs
+//
+// MIT License
+//
+// Copyright (c) 2025 Eric Buehler
+//
+// 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.
+// https://github.com/huggingface/transformers/blob/8685b3c5d2dd2550527773d2a02499495a759e31/src/transformers/convert_slow_tokenizer.py
+use std::collections::HashMap;
+use anyhow::Result;
+use itertools::Itertools;
+use tokenizers::{
+    decoders::{
+        self, byte_fallback::ByteFallback, byte_level::ByteLevel, fuse::Fuse, strip::Strip,
+    },
+    models::{bpe::BpeBuilder, unigram::Unigram},
+    normalizers::{self, Prepend, Replace},
+    pre_tokenizers,
+    processors::{
+        self,
+        template::{self, TemplateProcessing},
+    },
+    AddedToken, DecoderWrapper, ModelWrapper, NormalizerWrapper, Tokenizer,
+};
+use tracing::info;
+use crate::gguf::gguf_metadata::ContentMetadata;
+use super::Content;
+pub(crate) struct GgufTokenizerConversion {
+    pub tokenizer: Tokenizer,
+    // we most likely need those later
+    pub _bos: Option<String>,
+    pub _eos: Option<String>,
+    pub _unk: Option<String>,
+}
+struct PropsGGUF {
+    model: String,
+    tokens: Vec<String>,
+    added_tokens: Option<Vec<String>>,
+    scores: Option<Vec<f32>>,
+    merges: Option<Vec<String>>,
+    unk: Option<u32>,
+    eos: u32,
+    bos: u32,
+    add_bos_token: Option<bool>,
+}
+impl TryFrom<ContentMetadata<'_>> for PropsGGUF {
+    type Error = anyhow::Error;
+    fn try_from(c: ContentMetadata) -> Result<Self, Self::Error> {
+        let required = ["model", "tokens", "eos_token_id", "bos_token_id"];
+        c.has_required_keys(&required)?;
+        let props = Self {
+            model: c.get_value("model")?,
+            tokens: c.get_value("tokens")?,
+            added_tokens: c.get_value("added_tokens").ok(),
+            scores: c.get_value("scores").ok(),
+            merges: c.get_value("merges").ok(),
+            unk: c.get_value("unknown_token_id").ok(),
+            eos: c.get_value("eos_token_id")?,
+            bos: c.get_value("bos_token_id")?,
+            add_bos_token: c.get_value("add_bos_token").ok(),
+        };
+        Ok(props)
+    }
+}
+struct AddedTokensCollection {
+    bos: String,
+    eos: String,
+    unk: Option<String>,
+}
+pub fn convert_gguf_to_hf_tokenizer(content: &Content) -> Result<GgufTokenizerConversion> {
+    let metadata = ContentMetadata {
+        path_prefix: "tokenizer.ggml",
+        metadata: content.get_metadata(),
+    };
+    let props = PropsGGUF::try_from(metadata)?;
+    let (tokenizer, kind, special_tokens) = match props.model.as_str() {
+        "llama" | "replit" => unigram_tokenizer(&props)?,
+        "gpt2" => bpe_tokenizer(&props)?,
+        other => {
+            anyhow::bail!("Tokenizer model `{other}` not supported.");
+        }
+    };
+    info!(
+        "GGUF tokenizer model is `{model}`, kind: `{kind:?}`, num tokens: {}, num added tokens: {}, num merges: {}, num scores: {}",
+        tokenizer.get_vocab_size(true),
+        props.added_tokens.as_ref().map(|x| x.len()).unwrap_or(0),
+        props.merges.as_ref().map(|x| x.len()).unwrap_or(0),
+        props.scores.as_ref().map(|x| x.len()).unwrap_or(0),
+        model = props.model,
+    );
+    tracing::trace!("GGUF Tokenizer: {tokenizer:?}");
+    let AddedTokensCollection { bos, eos, unk } = special_tokens;
+    Ok(GgufTokenizerConversion {
+        tokenizer,
+        _bos: Some(bos),
+        _eos: Some(eos),
+        _unk: unk,
+    })
+}
+// TODO: Add support for additional tokenizer models: WordPiece, WordLevel
+// https://docs.rs/tokenizers/latest/tokenizers/models/enum.ModelWrapper.html
+#[derive(Debug)]
+enum TokenizerKind {
+    Unigram,
+    Bpe,
+}
+/// Add the special tokens and return their string representations
+fn add_special_tokens(
+    p: &PropsGGUF,
+    tokenizer: &mut Tokenizer,
+    bos: u32,
+    eos: u32,
+    unk: Option<u32>,
+) -> AddedTokensCollection {
+    // Add special tokens (bos, eos, unk):
+    let mut special_tokens: [Option<String>; 3] = Default::default();
+    // A little bit awkward here since eos/bos are assumed not options so we need to handle an Option
+    for (i, token_id) in [Some(bos), Some(eos), unk].into_iter().enumerate() {
+        if let Some(token_id) = token_id {
+            let token = p.tokens[token_id as usize].as_str();
+            special_tokens[i] = Some(token.to_string());
+            tokenizer.add_special_tokens(&[AddedToken::from(token.to_string(), true)]);
+        }
+    }
+    // Destructure array of options:
+    let [bos_str, eos_str, unk_str] = special_tokens;
+    // Would need to unwrap bos/eos here, or change the struct types
+    AddedTokensCollection {
+        bos: bos_str.unwrap(),
+        eos: eos_str.unwrap(),
+        unk: unk_str,
+    }
+}
+fn unigram_tokenizer(p: &PropsGGUF) -> Result<(Tokenizer, TokenizerKind, AddedTokensCollection)> {
+    let PropsGGUF { unk, eos, bos, .. } = *p;
+    // Unigram (SentencePiece) default UNK is 0
+    let unk = unk.unwrap_or(0);
+    // Create the Tokenizer model:
+    let model = {
+        let vocab: Vec<(String, f64)> = {
+            let Some(s) = p.scores.as_ref() else {
+                anyhow::bail!(
+                    "`llama` unigram tokenizer is missing required metadata `tokenizer.ggml.scores`"
+                );
+            };
+            let scores = s.iter().cloned().map(|f_32| f_32 as f64);
+            p.tokens.iter().cloned().zip(scores).collect()
+        };
+        Unigram::from(vocab, Some(unk as usize), true).map_err(anyhow::Error::msg)?
+    };
+    // Decoder + Normalizer config reference:
+    // https://github.com/EricLBuehler/mistral.rs/pull/389#discussion_r1630620763
+    let decoder = Decoder::Sequence(vec![
+        Decoder::Replace("▁", " "),
+        Decoder::ByteFallback,
+        Decoder::Fuse,
+        Decoder::Strip(' ', 1, 0),
+    ]);
+    let normalizer = Normalizer::Sequence(vec![
+        Normalizer::Prepend("▁"),
+        Normalizer::Replace(" ", "▁"),
+    ]);
+    let mut tokenizer: Tokenizer = TokenizerX::new(
+        ModelWrapper::Unigram(model),
+        Some(decoder),
+        Some(normalizer),
+    )?;
+    // Add special tokens (bos, eos, unk):
+    let special_tokens = add_special_tokens(p, &mut tokenizer, bos, eos, Some(unk));
+    Ok((tokenizer, TokenizerKind::Unigram, special_tokens))
+}
+fn bpe_tokenizer(p: &PropsGGUF) -> Result<(Tokenizer, TokenizerKind, AddedTokensCollection)> {
+    // BPE merges have each string item as a space-delimited pair:
+    // https://github.com/EricLBuehler/mistral.rs/pull/397#discussion_r1631988370
+    let merges = p
+        .merges
+        .as_ref()
+        .ok_or(anyhow::Error::msg("BPE tokenizer must include merges"))?
+        .iter()
+        .map(|merge| {
+            let split: (&str, &str) = merge
+                .splitn(2, ' ')
+                .collect_tuple()
+                .expect("Failed to convert split into 2-tuple");
+            (split.0.to_string(), split.1.to_string())
+        })
+        .collect::<Vec<_>>();
+    let mut vocab = HashMap::new();
+    for (i, token) in p.tokens.iter().enumerate() {
+        #[allow(clippy::cast_possible_truncation)]
+        vocab.insert(token.clone(), i as u32);
+    }
+    let PropsGGUF {
+        eos,
+        bos,
+        unk,
+        add_bos_token,
+        ..
+    } = *p;
+    let mut bpe = BpeBuilder::new().vocab_and_merges(vocab, merges);
+    if let Some(unk) = unk {
+        bpe = bpe.unk_token(p.tokens[unk as usize].to_string());
+    };
+    let bpe = bpe.build().map_err(anyhow::Error::msg)?;
+    let mut tokenizer = TokenizerX::new(
+        ModelWrapper::BPE(bpe),
+        Some(Decoder::ByteLevel(true, true, true)),
+        None,
+    )?;
+    tokenizer.with_pre_tokenizer(Some(pre_tokenizers::byte_level::ByteLevel::new(
+        false, true, true,
+    )));
+    if add_bos_token.is_some_and(|x| x) {
+        let mut special_toks = HashMap::new();
+        special_toks.insert(
+            p.tokens[bos as usize].clone(),
+            template::SpecialToken::new(
+                p.tokens[bos as usize].clone(),
+                vec![bos],
+                vec![p.tokens[bos as usize].clone()],
+            )
+            .unwrap(),
+        );
+        tokenizer.with_post_processor(Some(
+            TemplateProcessing::builder()
+                .try_single(format!("{}:0 $A:0", p.tokens[bos as usize]))
+                .unwrap()
+                .try_pair(format!("{}:0 $A:0 $B:1", p.tokens[bos as usize]))
+                .unwrap()
+                .special_tokens(special_toks)
+                .build()
+                .unwrap(),
+        ));
+    } else {
+        tokenizer.with_post_processor(Some(processors::byte_level::ByteLevel::new(
+            true, false, true,
+        )));
+    }
+    let special_tokens = add_special_tokens(p, &mut tokenizer, bos, eos, unk);
+    Ok((tokenizer, TokenizerKind::Bpe, special_tokens))
+}
+// This is a workaround to have a better builder API.
+// Upstream `TokenizerBuilder` is difficult to work with:
+// https://github.com/huggingface/tokenizers/issues/1549
+struct TokenizerX;
+impl TokenizerX {
+    #[allow(clippy::new_ret_no_self)]
+    fn new<'a>(
+        model: ModelWrapper,
+        decoder: Option<Decoder<'a>>,
+        normalizer: Option<Normalizer<'a>>,
+    ) -> Result<Tokenizer> {
+        let mut tokenizer = Tokenizer::new(model);
+        // Handle local enum to remote enum type:
+        if let Some(decoder) = decoder {
+            let d = DecoderWrapper::try_from(decoder)?;
+            tokenizer.with_decoder(Some(d));
+        }
+        if let Some(normalizer) = normalizer {
+            let n: NormalizerWrapper = NormalizerWrapper::try_from(normalizer)?;
+            tokenizer.with_normalizer(Some(n));
+        }
+        Ok(tokenizer)
+    }
+}
+// Convenient alternative to upstream:
+// https://docs.rs/tokenizers/latest/tokenizers/decoders/enum.DecoderWrapper.html
+enum Decoder<'a> {
+    ByteFallback,
+    Fuse,
+    Replace(&'a str, &'a str),
+    Strip(char, usize, usize),
+    Sequence(Vec<Self>),
+    ByteLevel(bool, bool, bool),
+}
+// Convert into upstream type wrapped enum variants:
+impl TryFrom<Decoder<'_>> for DecoderWrapper {
+    type Error = anyhow::Error;
+    fn try_from(variant: Decoder) -> Result<Self, Self::Error> {
+        let value: DecoderWrapper = match variant {
+            Decoder::ByteFallback => ByteFallback::default().into(),
+            Decoder::Fuse => Fuse::default().into(),
+            Decoder::Replace(pattern, content) => Replace::new(pattern, content)
+                .map_err(anyhow::Error::msg)?
+                .into(),
+            Decoder::Strip(content, start, stop) => Strip::new(content, start, stop).into(),
+            Decoder::Sequence(decoders) => {
+                let seq = decoders
+                    .into_iter()
+                    .map(DecoderWrapper::try_from)
+                    .collect::<Result<Vec<DecoderWrapper>>>()?;
+                decoders::sequence::Sequence::new(seq).into()
+            }
+            Decoder::ByteLevel(add_prefix_space, trim_offsets, use_regex) => {
+                ByteLevel::new(add_prefix_space, trim_offsets, use_regex).into()
+            }
+        };
+        Ok(value)
+    }
+}
+// Convenient alternative to upstream:
+// https://docs.rs/tokenizers/latest/tokenizers/normalizers/enum.NormalizerWrapper.html
+enum Normalizer<'a> {
+    Prepend(&'a str),
+    Replace(&'a str, &'a str),
+    Sequence(Vec<Self>),
+}
+impl TryFrom<Normalizer<'_>> for NormalizerWrapper {
+    type Error = anyhow::Error;
+    fn try_from(variant: Normalizer) -> Result<Self, Self::Error> {
+        let value: NormalizerWrapper = match variant {
+            Normalizer::Prepend(prepend) => Prepend::new(prepend.to_owned()).into(),
+            Normalizer::Replace(pattern, content) => Replace::new(pattern, content)
+                .map_err(anyhow::Error::msg)?
+                .into(),
+            Normalizer::Sequence(decoders) => {
+                let seq = decoders
+                    .into_iter()
+                    .map(NormalizerWrapper::try_from)
+                    .collect::<Result<Vec<NormalizerWrapper>>>()?;
+                normalizers::Sequence::new(seq).into()
+            }
+        };
+        Ok(value)
+    }
+}
+#[cfg(test)]
+mod tests {
+    use anyhow::Result;
+    use hf_hub::{api::sync::ApiBuilder, Repo, RepoType};
+    use tokenizers::Tokenizer;
+    #[allow(dead_code)]
+    #[derive(Debug)]
+    enum TokenizerType {
+        /// Mistral v0.1 tokenizer
+        Llama,
+        Replit,
+        Gpt2,
+        Rwkv,
+    }
+    fn get_gguf_tokenizer(tokenizer: TokenizerType) -> Result<Tokenizer> {
+        match tokenizer {
+            TokenizerType::Llama => {
+                let api = ApiBuilder::new().with_progress(true).build().unwrap();
+                let api = api.repo(Repo::with_revision(
+                    "EricB/mistralrs_tests".to_string(),
+                    RepoType::Model,
+                    "main".to_string(),
+                ));
+                let filename = api.get("llama_gguf_tokenizer.json").unwrap();
+                let tokenizer = Tokenizer::from_file(filename).expect("Valid tokenizer");
+                Ok(tokenizer)
+            }
+            TokenizerType::Gpt2 => {
+                let api = ApiBuilder::new().with_progress(true).build().unwrap();
+                let api = api.repo(Repo::with_revision(
+                    "EricB/mistralrs_tests".to_string(),
+                    RepoType::Model,
+                    "main".to_string(),
+                ));
+                let filename = api.get("gpt2_gguf_tokenizer.json").unwrap();
+                let tokenizer = Tokenizer::from_file(filename).expect("Valid tokenizer");
+                Ok(tokenizer)
+            }
+            other => anyhow::bail!("Cannot get testing HF tokenizer for type {other:?}"),
+        }
+    }
+    fn get_hf_tokenizer(tokenizer: TokenizerType) -> Result<Tokenizer> {
+        match tokenizer {
+            TokenizerType::Llama => {
+                let api = ApiBuilder::new().with_progress(true).build().unwrap();
+                let api = api.repo(Repo::with_revision(
+                    "EricB/mistralrs_tests".to_string(),
+                    RepoType::Model,
+                    "main".to_string(),
+                ));
+                let tokenizer_filename = api.get("tokenizer.json").unwrap();
+                Ok(Tokenizer::from_file(tokenizer_filename).unwrap())
+            }
+            TokenizerType::Gpt2 => {
+                let api = ApiBuilder::new().with_progress(true).build().unwrap();
+                let api = api.repo(Repo::with_revision(
+                    "EricB/mistralrs_tests".to_string(),
+                    RepoType::Model,
+                    "main".to_string(),
+                ));
+                let tokenizer_filename = api.get("tokenizer_gpt2.json").unwrap();
+                Ok(Tokenizer::from_file(tokenizer_filename).unwrap())
+            }
+            other => anyhow::bail!("Cannot get testing HF tokenizer for type {other:?}"),
+        }
+    }
+    // Content based upon https://github.com/ggerganov/llama.cpp/blob/master/tests/test-tokenizer-random.py#L99-L161
+    fn get_test_passage() -> String {
+        let passage = "Hello, world! \n🚀 (normal) 😶‍🌫️ (compound emoji, zwj sequence) ✅ (emoji as single token)\n你好世界！\nNǐ hǎo shìjiè!";
+        passage.to_owned()
+    }
+    // The provided passage should encode and decode back into the same passage string:
+    fn codec_roundtrip(
+        tokenizer: &Tokenizer,
+        passage: &str,
+        add_special_tokens: bool,
+    ) -> Result<String> {
+        let tokenized = tokenizer
+            .encode(passage, add_special_tokens)
+            .map_err(anyhow::Error::msg)?;
+        // NOTE: The special tokens bool param meaning differs between encode() / decode():
+        decode(tokenizer, tokenized.get_ids(), !add_special_tokens)
+    }
+    fn decode(
+        tokenizer: &Tokenizer,
+        token_ids: &[u32],
+        skip_special_tokens: bool,
+    ) -> Result<String> {
+        tokenizer
+            .decode(token_ids, skip_special_tokens)
+            .map_err(anyhow::Error::msg)
+    }
+    #[test]
+    fn test_encode_decode_llama() -> Result<()> {
+        use rand::seq::SliceRandom;
+        let passage = get_test_passage();
+        let hf_tokenizer = get_hf_tokenizer(TokenizerType::Llama)?;
+        let gguf_tokenizer = get_gguf_tokenizer(TokenizerType::Llama)?;
+        // Without adding special tokens
+        let hf_decoded = codec_roundtrip(&hf_tokenizer, passage.as_str(), false)?;
+        let gguf_decoded = codec_roundtrip(&gguf_tokenizer, passage.as_str(), false)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        assert_eq!(passage, gguf_decoded);
+        // With special tokens added
+        // SKIPPED:
+        // - Bugged the GGUF tokenizer does not prepend `<s> `
+        // - Due to HF tokenizer using BPE (tokenizer.json) while GGUF tokenizer uses Unigram (metadata)?
+        /*
+        let hf_decoded = codec_roundtrip(&hf_tokenizer, passage.as_str(), true)?;
+        let gguf_decoded = codec_roundtrip(&gguf_tokenizer, passage.as_str(), true)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        */
+        #[allow(clippy::cast_possible_truncation)]
+        let mut tokens = (0..hf_tokenizer.get_vocab_size(false) as u32).collect::<Vec<_>>();
+        tokens.shuffle(&mut rand::rng());
+        // Without skipping special tokens
+        let hf_decoded = decode(&hf_tokenizer, &tokens, false)?;
+        let gguf_decoded = decode(&gguf_tokenizer, &tokens, false)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        // With skipping special tokens
+        let hf_decoded = decode(&hf_tokenizer, &tokens, true)?;
+        let gguf_decoded = decode(&gguf_tokenizer, &tokens, true)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        Ok(())
+    }
+    #[test]
+    fn test_encode_decode_gpt2() -> Result<()> {
+        use rand::seq::SliceRandom;
+        let passage = get_test_passage();
+        let hf_tokenizer = get_hf_tokenizer(TokenizerType::Gpt2)?;
+        let gguf_tokenizer = get_gguf_tokenizer(TokenizerType::Gpt2)?;
+        // Without adding special tokens
+        let hf_decoded = codec_roundtrip(&hf_tokenizer, passage.as_str(), false)?;
+        let gguf_decoded = codec_roundtrip(&gguf_tokenizer, passage.as_str(), false)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        assert_eq!(passage, gguf_decoded);
+        // With special tokens added
+        // SKIPPED:
+        // - Bugged the GGUF tokenizer does not prepend `<s> `
+        // - Due to HF tokenizer using BPE (tokenizer.json) while GGUF tokenizer uses Unigram (metadata)?
+        /*
+        let hf_decoded = codec_roundtrip(&hf_tokenizer, passage.as_str(), true)?;
+        let gguf_decoded = codec_roundtrip(&gguf_tokenizer, passage.as_str(), true)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        */
+        #[allow(clippy::cast_possible_truncation)]
+        let mut tokens = (0..hf_tokenizer.get_vocab_size(false) as u32).collect::<Vec<_>>();
+        tokens.shuffle(&mut rand::rng());
+        // Without skipping special tokens
+        let hf_decoded = decode(&hf_tokenizer, &tokens, false)?;
+        let gguf_decoded = decode(&gguf_tokenizer, &tokens, false)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        // With skipping special tokens
+        let hf_decoded = decode(&hf_tokenizer, &tokens, true)?;
+        let gguf_decoded = decode(&gguf_tokenizer, &tokens, true)?;
+        assert_eq!(hf_decoded, gguf_decoded);
+        Ok(())
+    }
+}
--- a/lib/llm/src/lib.rs
+++ b/lib/llm/src/lib.rs
@@ -22,6 +22,7 @@ pub mod backend;
 pub mod common;
 pub mod disagg_router;
 pub mod engines;
+pub mod gguf;
 pub mod http;
 pub mod kv_router;
 pub mod model_card;

--- a/lib/llm/src/model_card/create.rs
+++ b/lib/llm/src/model_card/create.rs
@@ -20,7 +20,7 @@ use anyhow::{Context, Result};
 use std::fs;
 use std::path::Path;
-use crate::model_card::model::{File, ModelInfoType, PromptFormatterArtifact, TokenizerKind};
+use crate::model_card::model::{ModelInfoType, PromptFormatterArtifact, TokenizerKind};
 impl ModelDeploymentCard {
    /// Creates a ModelDeploymentCard from a local directory path.
@@ -58,6 +58,33 @@ impl ModelDeploymentCard {
        Self::from_repo(&repo_id, model_name).await
    }
+    pub async fn from_gguf(gguf_file: &Path, model_name: Option<&str>) -> anyhow::Result<Self> {
+        let model_name = model_name.map(|s| s.to_string()).or_else(|| {
+            gguf_file
+                .iter()
+                .last()
+                .map(|n| n.to_string_lossy().to_string())
+        });
+        let Some(model_name) = model_name else {
+            // I think this would only happy on an empty path
+            anyhow::bail!(
+                "Could not extract model name from path '{}'",
+                gguf_file.display()
+            );
+        };
+        Ok(Self {
+            display_name: model_name.to_string(),
+            service_name: model_name.to_string(),
+            model_info: ModelInfoType::GGUF(gguf_file.to_path_buf()),
+            tokenizer: TokenizerKind::from_gguf(gguf_file)?,
+            prompt_formatter: Some(PromptFormatterArtifact::GGUF(gguf_file.to_path_buf())),
+            prompt_context: None, // TODO - auto-detect prompt context
+            revision: 0,
+            last_published: None,
+            requires_preprocessing: true,
+        })
+    }
    /// TODO: This will be implemented after nova-hub is integrated with the model-card
    /// TODO: Attempt to auto-detect model type and construct an MDC from a NGC repo
    pub async fn from_ngc_repo(_: &str) -> anyhow::Result<Self> {
@@ -127,7 +154,7 @@ impl TokenizerKind {
 }
 /// Checks if the provided path contains the expected file.
-async fn check_for_file(repo_id: &str, file: &str) -> anyhow::Result<File> {
+async fn check_for_file(repo_id: &str, file: &str) -> anyhow::Result<String> {
    let mut files = check_for_files(repo_id, vec![file.to_string()]).await?;
    let file = files
        .remove(file)
@@ -135,7 +162,7 @@ async fn check_for_file(repo_id: &str, file: &str) -> anyhow::Result<File> {
    Ok(file)
 }
-async fn check_for_files(repo_id: &str, files: Vec<String>) -> Result<HashMap<String, File>> {
+async fn check_for_files(repo_id: &str, files: Vec<String>) -> Result<HashMap<String, String>> {
    let dir_entries =
        fs::read_dir(repo_id).with_context(|| format!("Failed to read directory: {}", repo_id))?;
    let mut found_files = HashMap::new();

--- a/lib/llm/src/model_card/model.rs
+++ b/lib/llm/src/model_card/model.rs
@@ -25,19 +25,21 @@
 //! - Prompt formatter settings (PromptFormatterArtifact)
 //! - Various metadata like revision, publish time, etc.
-use crate::protocols::TokenIdType;
-use anyhow::Result;
-use either::Either;
-use serde::{Deserialize, Serialize};
-use std::sync::Arc;
 use std::fmt;
-use std::path::Path;
+use std::fs::File;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
 use std::time::Duration;
+use anyhow::{Context, Result};
 use derive_builder::Builder;
 use dynamo_runtime::slug::Slug;
+use either::Either;
+use serde::{Deserialize, Serialize};
+use tokenizers::Tokenizer as HfTokenizer;
+use crate::gguf::{Content, ContentConfig};
+use crate::protocols::TokenIdType;
 pub const BUCKET_NAME: &str = "mdc";
@@ -48,18 +50,18 @@ pub const BUCKET_TTL: Duration = Duration::from_secs(5 * 60);
 /// If a model deployment card hasn't been refreshed in this much time the worker is likely gone
 const CARD_MAX_AGE: chrono::TimeDelta = chrono::TimeDelta::minutes(5);
-pub type File = String;
 #[derive(Serialize, Deserialize, Clone, Debug)]
 #[serde(rename_all = "snake_case")]
 pub enum ModelInfoType {
-    HfConfigJson(File),
+    HfConfigJson(String),
+    GGUF(PathBuf),
 }
 #[derive(Serialize, Deserialize, Clone, Debug)]
 #[serde(rename_all = "snake_case")]
 pub enum TokenizerKind {
-    HfTokenizerJson(File),
+    HfTokenizerJson(String),
+    GGUF(Box<HfTokenizer>),
 }
 /// Supported types of prompt formatters.
@@ -77,7 +79,8 @@ pub enum TokenizerKind {
 #[derive(Serialize, Deserialize, Clone, Debug)]
 #[serde(rename_all = "snake_case")]
 pub enum PromptFormatterArtifact {
-    HfTokenizerConfigJson(File),
+    HfTokenizerConfigJson(String),
+    GGUF(PathBuf),
 }
 #[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq, Hash)]
@@ -194,6 +197,15 @@ impl ModelDeploymentCard {
            false
        }
    }
+    pub fn tokenizer_hf(&self) -> anyhow::Result<HfTokenizer> {
+        match &self.tokenizer {
+            TokenizerKind::HfTokenizerJson(file) => {
+                HfTokenizer::from_file(file).map_err(anyhow::Error::msg)
+            }
+            TokenizerKind::GGUF(t) => Ok(*t.clone()),
+        }
+    }
 }
 impl fmt::Display for ModelDeploymentCard {
@@ -221,13 +233,14 @@ pub trait ModelInfo: Send + Sync {
 impl ModelInfoType {
    pub async fn get_model_info(&self) -> Result<Arc<dyn ModelInfo>> {
        match self {
-            Self::HfConfigJson(info) => HFConfigJsonFile::from_file(info).await,
+            Self::HfConfigJson(info) => HFConfig::from_json_file(info).await,
+            Self::GGUF(path) => HFConfig::from_gguf(path),
        }
    }
 }
 #[derive(Debug, Clone, Serialize, Deserialize)]
-struct HFConfigJsonFile {
+struct HFConfig {
    bos_token_id: TokenIdType,
    #[serde(with = "either::serde_untagged")]
@@ -253,15 +266,46 @@ struct HFConfigJsonFile {
    vocab_size: usize,
 }
-impl HFConfigJsonFile {
+impl HFConfig {
-    async fn from_file(file: &File) -> Result<Arc<dyn ModelInfo>> {
+    async fn from_json_file(file: &String) -> Result<Arc<dyn ModelInfo>> {
        let contents = std::fs::read_to_string(file)?;
        let config: Self = serde_json::from_str(&contents)?;
        Ok(Arc::new(config))
    }
+    fn from_gguf(gguf_file: &Path) -> Result<Arc<dyn ModelInfo>> {
+        let content = load_gguf(gguf_file)?;
+        let model_config_metadata: ContentConfig = (&content).into();
+        let num_hidden_layers =
+            content.get_metadata()[&format!("{}.block_count", content.arch())].to_u32()? as usize;
+        let bos_token_id = content.get_metadata()["tokenizer.ggml.bos_token_id"].to_u32()?;
+        let eos_token_id = content.get_metadata()["tokenizer.ggml.eos_token_id"].to_u32()?;
+        // to_vec returns a Vec that's already there, so it's cheap
+        let vocab_size = content.get_metadata()["tokenizer.ggml.tokens"]
+            .to_vec()?
+            .len();
+        let arch = content.arch().to_string();
+        Ok(Arc::new(HFConfig {
+            bos_token_id,
+            eos_token_id: Either::Left(eos_token_id),
+            architectures: vec![format!("{}ForCausalLM", capitalize(&arch))],
+            // "general.architecture"
+            model_type: arch,
+            // "llama.context_length"
+            max_position_embeddings: model_config_metadata.max_seq_len(),
+            // "llama.block_count"
+            num_hidden_layers,
+            // "llama.attention.head_count"
+            num_attention_heads: model_config_metadata.num_attn_heads(),
+            // "tokenizer.ggml.tokens".len()
+            vocab_size,
+        }))
+    }
 }
-impl ModelInfo for HFConfigJsonFile {
+impl ModelInfo for HFConfig {
    fn model_type(&self) -> String {
        self.model_type.clone()
    }
@@ -285,3 +329,33 @@ impl ModelInfo for HFConfigJsonFile {
        self.vocab_size
    }
 }
+impl TokenizerKind {
+    pub fn from_gguf(gguf_file: &Path) -> anyhow::Result<Self> {
+        let content = load_gguf(gguf_file)?;
+        let out = crate::gguf::convert_gguf_to_hf_tokenizer(&content)
+            .with_context(|| gguf_file.display().to_string())?;
+        Ok(TokenizerKind::GGUF(Box::new(out.tokenizer)))
+    }
+}
+fn load_gguf(gguf_file: &Path) -> anyhow::Result<Content> {
+    let filename = gguf_file.display().to_string();
+    let mut f = File::open(gguf_file).with_context(|| filename.clone())?;
+    // vec because GGUF can be split into multiple files (shards)
+    let mut readers = vec![&mut f];
+    crate::gguf::Content::from_readers(&mut readers).with_context(|| filename.clone())
+}
+fn capitalize(s: &str) -> String {
+    s.chars()
+        .enumerate()
+        .map(|(i, c)| {
+            if i == 0 {
+                c.to_uppercase().to_string()
+            } else {
+                c.to_lowercase().to_string()
+            }
+        })
+        .collect()
+}
--- a/lib/llm/src/preprocessor.rs
+++ b/lib/llm/src/preprocessor.rs
@@ -74,6 +74,9 @@ impl OpenAIPreprocessor {
        let tokenizer = match &mdc.tokenizer {
            TokenizerKind::HfTokenizerJson(file) => HuggingFaceTokenizer::from_file(file)?,
+            TokenizerKind::GGUF(tokenizer) => {
+                HuggingFaceTokenizer::from_tokenizer(*tokenizer.clone())
+            }
        };
        let tokenizer = Arc::new(tokenizer);