GraphTok backend for NVIDIA Dynamo 1.3.0

graphtok-dynamo-1.3.0.patch adds a graphtok tokenizer backend to the Dynamo frontend. Prompt encoding runs on the GPU through libgtok.so (the C ABI of this repository); decoding stays with Hugging Face tokenizers.

It adds lib/tokenizers/src/graphtok.rs (backend, loaded with libloading), lib/tokenizers/tests/graphtok_dispatch.rs, wires the backend into lib/llm/src/model_card.rs, and adds graphtok to the frontend's --tokenizer-backend choices.

Build libgtok.so

From the root of this repository:

cmake --preset all && cmake --build build/all -j --target gtok
# -> build/all/libgtok.so

Apply and build Dynamo

git clone https://github.com/ai-dynamo/dynamo.git && cd dynamo
git checkout bae0051            # Dynamo 1.3.0
git apply /path/to/graphtok/integrations/dynamo/graphtok-dynamo-1.3.0.patch
cargo build --release           # or rebuild the Python bindings: cd lib/bindings/python && maturin develop --release
cargo test -p dynamo-tokenizers --test graphtok_dispatch   # needs a GPU and DYN_GRAPHTOK_LIB

Enable

export DYN_GRAPHTOK_LIB=/path/to/graphtok/build/all/libgtok.so
export DYN_TOKENIZER=graphtok          # or: python -m dynamo.frontend --tokenizer-backend graphtok
export DYN_GRAPHTOK_DISPATCH=1         # load-adaptive dispatcher (GPU batching + CPU route)
python -m dynamo.frontend ...

libgtok.so is located through DYN_GRAPHTOK_LIB; when unset, libgtok.so is resolved by the dynamic loader (LD_LIBRARY_PATH, rpath, system paths).

Other knobs:

On construction the backend encodes a probe containing chat-template special tokens with both GraphTok and Hugging Face and refuses to load if the ids differ (the frontend then falls back to Hugging Face).