llgtrt (llguidance + TensorRT-LLM)

This project implements a REST HTTP server with OpenAI-compatible API, based on NVIDIA TensorRT-LLM and llguidance library for constrained output.

The server supports regular completions and chat endpoints with JSON schema enforcement ("Structured Output"), as well as full context-free grammars using the Guidance library.

This server is similar in spirit to the TensorRT-LLM OpenAI server example, but it is Python-free (implemented in Rust) and includes support for constrained output. Like the example above, it does not use the NVIDIA Triton Inference Server.

Structured Output

The sampling can be constrained by the Low-Level Guidance library, part of the Guidance project. While TensorRT computes logits (token probabilities) for the next token, the llguidance library computes a set of tokens allowed by the grammar (whether JSON schema, regular expression, or a full context-free grammar (CFG)) in the form of a bitmask. When both the logits and bitmask are ready, a custom CUDA kernel applies the mask to the logits, and the result is used for sampling inside of TensorRT-LLM.

There is no significant startup cost for all realistic sizes of grammars (no measurable impact on time to first token (TTFT)). The overhead on generation speed (median time between tokens (TBT)) is typically 1-3%. The mask computation takes on the order of 1 ms of single-core CPU time per token per sequence in the batch. Thus, with 16 cores and a TBT of around 10 ms, batch sizes of up to 160 are not CPU-bound. Typically, the unconstrained TBT is higher at such batch sizes, and more cores are available, so batch size is not a problem in production.

This approach differs from Outlines (which pre-computes masks, resulting in a startup cost and limits on schema complexity) and is more similar in spirit to llama.cpp grammars, though it is much faster due to the use of a custom lexer with derivative-based regexes, an Earley parser, and a highly optimized token prefix tree.

Requirements

You will need a Linux machine with an NVIDIA GPU and Docker set up to use the nvidia-docker runtime.

So far, we have only tested it on 4xA100 (and single A100).

Running

Overview of steps:

• Build or pull the llgtrt/llgtrt Docker container.
• Build a trtllm engine (likely using the container).
• Create configuration files.
• Use the container to run the engine.

Building or Pulling Docker Container

To use a pre-built container, run:

docker pull llgtrt/llgtrt

To build a container, use:

./docker/build.sh

The build script will initialize submodules if they are missing. It takes about 15 minutes on a GitHub runner and should typically be faster on a local machine.

Building the TensorRT-LLM Engine

This is based on the TensorRT-LLM Quick-start. Follow the steps here, and look into that guide if needed.

First, use the llgtrt/llgtrt container to run bash.

./docker/bash.sh --volume /path/to/hf-models:/models

The following steps are done inside the container:

# Convert the HF model to a checkpoint
python3 /opt/TensorRT-LLM-examples/llama/convert_checkpoint.py \
    --dtype bfloat16 \
    --model_dir /models/Meta-Llama-3.1-8B-Instruct \
    --output_dir /models/model-ckpt \
    --tp_size 1

# Then, run trtllm build
trtllm-build --checkpoint_dir /models/model-ckpt \
    --gemm_plugin bfloat16 \
    --output_dir /models/model-engine \
    --use_paged_context_fmha enable

# Clean up checkpoint (optional)
rm -rf /models/model-ckpt

# Finally, copy tokenizer.json and tokenizer_config.json
cp /models/Meta-Llama-3.1-8B-Instruct/tokenizer.json /models/model-engine
cp /models/Meta-Llama-3.1-8B-Instruct/tokenizer_config.json /models/model-engine

# Exit the container
exit

Make sure to modify the path to the input model (it needs to contain the HF Transformers config.json as well as the .safetensors files and tokenizer.json). If you're running on more than one GPU, modify the --tp_size argument.

Running the Engine

PORT=3001 ./docker/run.sh /path/to/hf-models/model-engine

The command will print out the actual docker run invocation on the first line if you want to invoke it directly later. PORT defaults to 3000.

Update Configuration (Optional)

The defaults should be mostly reasonable, but you can modify them. First, generate a template configuration file:

./docker/run.sh /path/to/hf-models/model-engine --print-config > llgtrt.json5

The file will contain commented-out defaults for all supported options (JSON5 is a superset of JSON, so you can use comments). Edit it and move it to the engine folder.

To modify the chat template, you can either use --print-complete-config, which will include the chat template from tokenizer_config.json, or preferably create a separate chat_template.j2 file in the engine folder:

./docker/run.sh /path/to/hf-models/model-engine --print-chat-template > chat_template.j2
mv chat_template.j2 /path/to/hf-models/model-engine

The paths to llgtrt.json5 and chat_template.j2 are controlled by command line arguments. See --help for more info:

./docker/run.sh /path/to/hf-models/model-engine --help

You can specify multiple JSON5 config files, and they will be merged in the order specified (with later ones overriding earlier ones). This way, you can separate configuration for the tokenizer, runtime, and guidance parser.

Running phi-3

The phi-3 tokenizer, while based on llama2 is slightly different. Drop the following llgtrt.json5 file in engine folder:

{
  "tokenizer": {
    "bos_token": null,
    "n_vocab_override": 32064
  }
}

Development

First, build the Docker container to be used in the dev container. If you have already followed the steps above, you can skip this. Otherwise, run ./docker/build.sh.

Next, in VSCode, reopen the folder in the container.

Credits

The basic structure of the server borrows inspiration from npuichigo/openai_trtllm, which has similar aims but uses NVIDIA Triton Server wrapping TensorRT-LLM.

TODO

• multi-LoRA?
• test phi-3.5
• multi-modal input
• when streaming, and stop is set, we need to buffer the output so as not to return the stop sequence itself
• logprobs (currently only work with TP>1; TRTLLM bug?)
• logprobs with argmax sampling and constraints
• expose the 'surprise' measure somehow