mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-25 20:12:07 +00:00

Large Language Model Text Generation Inference

bloom deep-learning falcon gpt inference nlp pytorch starcoder transformer

Go to file

Nicolas Patry 120d5773e8 Rebase TRT-llm (#2331 ) * wip wip refacto refacto Initial setup for CXX binding to TRTLLM Working FFI call for TGI and TRTLLM backend Remove unused parameters annd force tokenizer name to be set Overall build TRTLLM and deps through CMake build system Enable end to end CMake build First version loading engines and making it ready for inference Remembering to check how we can detect support for chunked context Move to latest TensorRT-LLM version Specify which default log level to use depending on CMake build type make leader executor mode working unconditionally call InitializeBackend on the FFI layer bind to CUDA::nvml to retrieve compute capabilities at runtime updated logic and comment to detect cuda compute capabilities implement the Stream method to send new tokens through a callback use spdlog release 1.14.1 moving forward update trtllm to latest version a96cccafcf6365c128f004f779160951f8c0801c correctly tell cmake to build dependent tensorrt-llm required libraries create cmake install target to put everything relevant in installation folder add auth_token CLI argument to provide hf hub authentification token allow converting huggingface::tokenizers error to TensorRtLlmBackendError use correct include for spdlog include guard to build example in cmakelists working setup of the ffi layer remove fmt import use external fmt lib end to end ffi flow working make sure to track include/ffi.h to trigger rebuild from cargo impl the rust backend which currently cannot move the actual computation in background thread expose shutdown function at ffi layer impl RwLock scenario for TensorRtLllmBackend oops missing c++ backend definitions compute the number of maximum new tokens for each request independently make sure the context is not dropped in the middle of the async decoding. remove unnecessary log add all the necessary plumbery to return the generated content update invalid doc in cpp file correctly forward back the log probabilities remove unneeded scope variable for now refactor Stream impl for Generation to factorise code expose the internal missing start/queue timestamp forward tgi parameters rep/freq penalty add some more validation about grammar not supported define a shared struct to hold the result of a decoding step expose information about potential error happening while decoding remove logging add logging in case of decoding error make sure executor_worker is provided add initial Dockerfile for TRTLLM backend add some more information in CMakeLists.txt to correctly install executorWorker add some more information in CMakeLists.txt to correctly find and install nvrtc wrapper simplify prebuilt trtllm libraries name definition do the same name definition stuff for tensorrt_llm_executor_static leverage pkg-config to probe libraries paths and reuse new install structure from cmake fix bad copy/past missing nvinfer linkage direction align all the linker search dependency add missing pkgconfig folder for MPI in Dockerfile correctly setup linking search path for runtime layer fix missing / before tgi lib path adding missing ld_library_path for cuda stubs in Dockerfile update tgi entrypoint commenting out Python part for TensorRT installation refactored docker image move to TensorRT-LLM v0.11.0 make docker linter happy with same capitalization rule fix typo refactor the compute capabilities detection along with num gpus update TensorRT-LLM to latest version update TensorRT install script to latest update build.rs to link to cuda 12.5 add missing dependant libraries for linking clean up a bit install to decoder_attention target add some custom stuff for nccl linkage fix envvar CARGO_CFG_TARGET_ARCH set at runtime vs compile time use std::env::const::ARCH make sure variable live long enough... look for cuda 12.5 add some more basic info in README.md * Rebase. * Fix autodocs. * Let's try to enable trtllm backend. * Ignore backends/v3 by default. * Fixing client. * Fix makefile + autodocs. * Updating the schema thing + redocly. * Fix trtllm lint. * Adding pb files ? * Remove cargo fmt temporarily. * ? * Tmp. * Remove both check + clippy ? * Backporting telemetry. * Backporting `457fb0a1` * Remove PB from git. * Fixing PB with default member backends/client * update TensorRT-LLM to latest version * provided None for api_key * link against libtensorrt_llm and not libtensorrt-llm --------- Co-authored-by: OlivierDehaene <23298448+OlivierDehaene@users.noreply.github.com> Co-authored-by: Morgan Funtowicz <morgan@huggingface.co>		2024-09-25 05:55:39 +00:00
.github	GPTQ CI improvements (#2151 )	2024-09-25 05:21:03 +00:00
assets	Update grafana template (#1918 )	2024-07-17 05:36:58 +00:00
backends	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
benchmark	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
clients/python	feat: add ruff and resolve issue (#2262 )	2024-09-25 05:46:24 +00:00
docs	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
examples	Code expects newer huggingface_hub versions, tested and this resolves issues with streaming response format (#190 )	2024-08-08 13:07:27 +02:00
integration-tests	fix: adjust test snapshots and small refactors (#2323 )	2024-09-25 05:50:17 +00:00
launcher	Run ci api key (#2315 )	2024-09-25 05:46:41 +00:00
load_tests	feat: add ruff and resolve issue (#2262 )	2024-09-25 05:46:24 +00:00
proto	Enable multiple LoRa adapters (#2010 )	2024-09-24 03:55:04 +00:00
router	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
server	server quantize: store quantizer config in standard format (#2299 )	2024-09-25 05:50:17 +00:00
.dockerignore	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
.gitignore	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
.pre-commit-config.yaml	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
Cargo.lock	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
Cargo.toml	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
CODE_OF_CONDUCT.md	Set maximum grpc message receive size to 2GiB (#2075 )	2024-09-24 03:44:36 +00:00
CONTRIBUTING.md	Set maximum grpc message receive size to 2GiB (#2075 )	2024-09-24 03:44:36 +00:00
Dockerfile	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
Dockerfile_amd	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
Dockerfile_intel	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
Dockerfile.trtllm	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
LICENSE	Revert license to Apache 2.0 (#1714 )	2024-04-25 15:11:10 +03:00
Makefile	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00
README.md	Update README with Docker image v2.0.5	2024-09-07 17:56:52 +00:00
rust-toolchain.toml	Set maximum grpc message receive size to 2GiB (#2075 )	2024-09-24 03:44:36 +00:00
sagemaker-entrypoint.sh	feat(sagemaker): add trust remote code to entrypoint (#394 )	2023-06-02 09:51:06 +02:00
update_doc.py	Rebase TRT-llm (#2331 )	2024-09-25 05:55:39 +00:00

README.md

Text Generation Inference on Habana Gaudi

Text Generation Inference on Habana Gaudi

Running TGI on Gaudi

To use 🤗 text-generation-inference on Habana Gaudi/Gaudi2/Gaudi3, follow these steps:

Pull the official Docker image with:

docker pull ghcr.io/huggingface/tgi-gaudi:2.0.5

Note

Alternatively, you can build the Docker image using the Dockerfile located in this folder with:
docker build -t tgi_gaudi .

Launch a local server instance:

i. On 1 Gaudi card

model=meta-llama/Llama-2-7b-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$hf_token -e ENABLE_HPU_GRAPH=true -e LIMIT_HPU_GRAPH=true -e USE_FLASH_ATTENTION=true -e FLASH_ATTENTION_RECOMPUTE=true --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.5 --model-id $model --max-input-tokens 1024 --max-total-tokens 2048

For gated models such as StarCoder, you will have to pass -e HUGGING_FACE_HUB_TOKEN=<token> to the docker run command above with a valid Hugging Face Hub read token.

ii. On 1 Gaudi card using PyTorch eager mode with torch compile:

model=meta-llama/Llama-2-7b-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e PT_HPU_LAZY_MODE=0 -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$hf_token --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.5 --model-id $model --max-input-tokens 1024 --max-total-tokens 2048

iii. On 8 Gaudi cards:

model=meta-llama/Llama-2-70b-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 -v $volume:/data --runtime=habana -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$hf_token -e ENABLE_HPU_GRAPH=true -e LIMIT_HPU_GRAPH=true -e USE_FLASH_ATTENTION=true -e FLASH_ATTENTION_RECOMPUTE=true --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.5 --model-id $model --sharded true --num-shard 8 --max-input-tokens 1024 --max-total-tokens 2048

You can then send a simple request:

curl 127.0.0.1:8080/generate \
  -X POST \
  -d '{"inputs":"What is Deep Learning?","parameters":{"max_new_tokens":32}}' \
  -H 'Content-Type: application/json'

Please note that the model warmup can take several minutes, especially for FP8 inference. To minimize this time in consecutive runs, please refer to Disk Caching Eviction Policy.

TGI-Gaudi Benchmark

Static Batching Benchmark

To run static batching benchmark, please refer to TGI's benchmark tool.

To run it on the same machine, you can do the following:

docker exec -it <docker name> bash , pick the docker started from step 2 using docker ps
text-generation-benchmark -t <model-id> , pass the model-id from docker run command
after the completion of tests, hit ctrl+c to see the performance data summary.

Continuous Batching Benchmark

To run continuous batching benchmark, please refer to README in examples folder.

Tested Models and Configurations

The following table contains models and configurations we have validated on Gaudi2.

Model	BF16	FP8	Single Card	Multi-Cards
Llama2-7B	✔	✔	✔	✔
Llama2-70B	✔	✔		✔
Llama3-8B	✔	✔	✔	✔
Llama3-70B	✔	✔		✔
Llama3.1-8B	✔	✔	✔	✔
Llama3.1-70B	✔	✔		✔
CodeLlama-13B	✔	✔	✔
Mixtral-8x7B	✔	✔	✔	✔
Mistral-7B	✔	✔	✔	✔
Llava-v1.6-Mistral-7B	✔	✔	✔	✔

Running TGI with BF16 Precision

The following are command examples for TGI models inference with BF16 precision.

Llama2-7B on 1 Card

model=meta-llama/Llama-2-7b-chat-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e PREFILL_BATCH_BUCKET_SIZE=2 \
   -e BATCH_BUCKET_SIZE=32 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=256 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 2048 --max-batch-total-tokens 65536 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 64

Llama2-70B on 8 cards

model=meta-llama/Llama-2-70b-chat-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e BATCH_BUCKET_SIZE=256 \
   -e PREFILL_BATCH_BUCKET_SIZE=4 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=64 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --sharded true --num-shard 8 \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 4096 --max-batch-total-tokens 524288 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 512

Llama3.1-8B on 1 card

model=meta-llama/Meta-Llama-3.1-8B-Instruct
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e PREFILL_BATCH_BUCKET_SIZE=2 \
   -e BATCH_BUCKET_SIZE=32 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=256 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 2048 --max-batch-total-tokens 65536 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 64

Llama3.1-70B 8 cards

model=meta-llama/Meta-Llama-3.1-70B-Instruct
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e BATCH_BUCKET_SIZE=256 \
   -e PREFILL_BATCH_BUCKET_SIZE=4 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=64 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --sharded true --num-shard 8 \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 4096 --max-batch-total-tokens 524288 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 512

Llava-v1.6-Mistral-7B on 1 card

In Llava-v1.6-Mistral-7B, an image usually accounts for 2000 input tokens. For example, an image of size 512x512 is represented by 2800 tokens. Thus, max-input-tokens must be larger than the number of tokens associated with the image. Otherwise the image may be truncated. We set BASE_IMAGE_TOKENS=2048 as the default image token value. This is the minimum value of max-input-tokens. You can override the environment variable BASE_IMAGE_TOKENS to change this value. The warmup will generate graphs with input length from BASE_IMAGE_TOKENS to max-input-tokens. For Llava-v1.6-Mistral-7B, the value of max-batch-prefill-tokens is 16384, which is calcualted as follows: prefill_batch_size = max-batch-prefill-tokens / max-input-tokens.

model=llava-hf/llava-v1.6-mistral-7b-hf
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
    -e PREFILL_BATCH_BUCKET_SIZE=1 \
    -e BATCH_BUCKET_SIZE=1 \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --max-input-tokens 4096 --max-batch-prefill-tokens 16384 \
   --max-total-tokens 8192 --max-batch-total-tokens 32768

Send the simple request.

curl -N 127.0.0.1:8080/generate_stream \
    -X POST \
    -d '{"inputs":"![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png)What is this a picture of?\n\n","parameters":{"max_new_tokens":16, "seed": 42}}' \
    -H 'Content-Type: application/json'

Running TGI with FP8 Precision

TGI-Gaudi supports FP8 precision inference with INC (Intel Neural Compressor) and HQT (Habana Quantization Toolkit). FP8 inference can be run by setting QUANT_CONFIG environment variable in the docker command. From TGI-Gaudi 2.0.4 release, INC is used by default for quantization. HQT will be removed in future releases. To use HQT, disable INC by setting -e USE_INC=0 in docker command.

To run FP8 Inference:

Measure statistics by using Optimum Habana measurement script
Run the model in TGI with QUANT_CONFIG setting - e.g. -e QUANT_CONFIG=./quantization_config/maxabs_quant.json.

The following are the commmand examples for FP8 inference based on the assumption that measurement is done in the first step above.

Llama2-7B on 1 Card

model=meta-llama/Llama-2-7b-chat-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -v $PWD/quantization_config:/usr/src/quantization_config \
   -v $PWD/hqt_output:/usr/src/hqt_output \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e PREFILL_BATCH_BUCKET_SIZE=2 \
   -e BATCH_BUCKET_SIZE=32 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=256 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 2048 --max-batch-total-tokens 65536 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 64

Llama2-70B on 8 Cards

model=meta-llama/Llama-2-70b-chat-hf
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -v $PWD/quantization_config:/usr/src/quantization_config \
   -v $PWD/hqt_output:/usr/src/hqt_output \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e BATCH_BUCKET_SIZE=256 \
   -e PREFILL_BATCH_BUCKET_SIZE=4 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=64 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --sharded true --num-shard 8 \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 4096 --max-batch-total-tokens 524288 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 512

Llama3.1-8B on 1 Card

model=meta-llama/Meta-Llama-3.1-8B-Instruct
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -v $PWD/quantization_config:/usr/src/quantization_config \
   -v $PWD/hqt_output:/usr/src/hqt_output \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e PREFILL_BATCH_BUCKET_SIZE=2 \
   -e BATCH_BUCKET_SIZE=32 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=256 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 2048 --max-batch-total-tokens 65536 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 64

Llama3.1-70B on 8 cards

model=meta-llama/Meta-Llama-3.1-70B-Instruct
hf_token=YOUR_ACCESS_TOKEN
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -v $PWD/quantization_config:/usr/src/quantization_config \
   -v $PWD/hqt_output:/usr/src/hqt_output \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   -e HABANA_VISIBLE_DEVICES=all \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e MAX_TOTAL_TOKENS=2048 \
   -e BATCH_BUCKET_SIZE=256 \
   -e PREFILL_BATCH_BUCKET_SIZE=4 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=64 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --sharded true --num-shard 8 \
   --max-input-length 1024 --max-total-tokens 2048 \
   --max-batch-prefill-tokens 4096 --max-batch-total-tokens 524288 \
   --max-waiting-tokens 7 --waiting-served-ratio 1.2 --max-concurrent-requests 512

Llava-v1.6-Mistral-7B on 1 Card

model=llava-hf/llava-v1.6-mistral-7b-hf
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -v $PWD/quantization_config:/usr/src/quantization_config \
   -v $PWD/hqt_output:/usr/src/hqt_output \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
    -e PREFILL_BATCH_BUCKET_SIZE=1 \
    -e BATCH_BUCKET_SIZE=1 \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --max-input-tokens 4096 --max-batch-prefill-tokens 16384 \
   --max-total-tokens 8192 --max-batch-total-tokens 32768

Llava-v1.6-Mistral-7B on 8 Cards

model=llava-hf/llava-v1.6-mistral-7b-hf
volume=$PWD/data   # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 \
   --runtime=habana \
   -v $volume:/data \
   -v $PWD/quantization_config:/usr/src/quantization_config \
   -v $PWD/hqt_output:/usr/src/hqt_output \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e TEXT_GENERATION_SERVER_IGNORE_EOS_TOKEN=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e ENABLE_HPU_GRAPH=true \
   -e LIMIT_HPU_GRAPH=true \
   -e USE_FLASH_ATTENTION=true \
   -e FLASH_ATTENTION_RECOMPUTE=true \
    -e PREFILL_BATCH_BUCKET_SIZE=1 \
    -e BATCH_BUCKET_SIZE=1 \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.5 \
   --model-id $model \
   --sharded true --num-shard 8 \
   --max-input-tokens 4096 --max-batch-prefill-tokens 16384 \
   --max-total-tokens 8192 --max-batch-total-tokens 32768

Adjusting TGI Parameters

Maximum sequence length is controlled by two arguments:

--max-input-tokens is the maximum possible input prompt length. Default value is 4095.
--max-total-tokens is the maximum possible total length of the sequence (input and output). Default value is 4096.

Maximum batch size is controlled by two arguments:

For prefill operation, please set --max-batch-prefill-tokens as bs * max-input-tokens, where bs is your expected maximum prefill batch size.
For decode operation, please set --max-batch-total-tokens as bs * max-total-tokens, where bs is your expected maximum decode batch size.
Please note that batch size will be always padded to the nearest multiplication of BATCH_BUCKET_SIZE and PREFILL_BATCH_BUCKET_SIZE.

To ensure greatest performance results, at the beginning of each server run, warmup is performed. It's designed to cover major recompilations while using HPU Graphs. It creates queries with all possible input shapes, based on provided parameters (described in this section) and runs basic TGI operations on them (prefill, decode, concatenate).

Except those already mentioned, there are other parameters that need to be properly adjusted to improve performance or memory usage:

PAD_SEQUENCE_TO_MULTIPLE_OF determines sizes of input length buckets. Since warmup creates several graphs for each bucket, it's important to adjust that value proportionally to input sequence length. Otherwise, some out of memory issues can be observed.
ENABLE_HPU_GRAPH enables HPU graphs usage, which is crucial for performance results. Recommended value to keep is true .

For more information and documentation about Text Generation Inference, checkout the README of the original repo.

Environment Variables

Name	Value(s)	Default	Description	Usage
ENABLE_HPU_GRAPH	True/False	True	Enable hpu graph or not	add -e in docker run command
LIMIT_HPU_GRAPH	True/False	False	Skip HPU graph usage for prefill to save memory, set to `True` for large sequence/decoding lengths(e.g. 300/212)	add -e in docker run command
BATCH_BUCKET_SIZE	integer	8	Batch size for decode operation will be rounded to the nearest multiple of this number. This limits the number of cached graphs	add -e in docker run command
PREFILL_BATCH_BUCKET_SIZE	integer	4	Batch size for prefill operation will be rounded to the nearest multiple of this number. This limits the number of cached graphs	add -e in docker run command
PAD_SEQUENCE_TO_MULTIPLE_OF	integer	128	For prefill operation, sequences will be padded to a multiple of provided value.	add -e in docker run command
SKIP_TOKENIZER_IN_TGI	True/False	False	Skip tokenizer for input/output processing	add -e in docker run command
WARMUP_ENABLED	True/False	True	Enable warmup during server initialization to recompile all graphs. This can increase TGI setup time.	add -e in docker run command
QUEUE_THRESHOLD_MS	integer	120	Controls the threshold beyond which the request are considered overdue and handled with priority. Shorter requests are prioritized otherwise.	add -e in docker run command
USE_FLASH_ATTENTION	True/False	False	Whether to enable Habana Flash Attention, provided that the model supports it. Currently only llama and mistral supports this feature. Please refer to https://docs.habana.ai/en/latest/PyTorch/Model_Optimization_PyTorch/Optimization_in_PyTorch_Models.html?highlight=fusedsdpa#using-fused-scaled-dot-product-attention-fusedsdpa
FLASH_ATTENTION_RECOMPUTE	True/False	False	Whether to enable Habana Flash Attention in recompute mode on first token generation.

Profiler

To collect performance profiling, please set below environment variables:

Name	Value(s)	Default	Description	Usage
PROF_WAITSTEP	integer	0	Control profile wait steps	add -e in docker run command
PROF_WARMUPSTEP	integer	0	Control profile warmup steps	add -e in docker run command
PROF_STEP	integer	0	Enable/disable profile, control profile active steps	add -e in docker run command
PROF_PATH	string	/tmp/hpu_profile	Define profile folder	add -e in docker run command
PROF_RANKS	string	0	Comma-separated list of ranks to profile	add -e in docker run command
PROF_RECORD_SHAPES	True/False	False	Control record_shapes option in the profiler	add -e in docker run command

License

The license to use TGI on Habana Gaudi is the one of TGI: https://github.com/huggingface/text-generation-inference/blob/main/LICENSE

Please reach out to api-enterprise@huggingface.co if you have any question.

README.md

Text Generation Inference on Habana Gaudi

Table of contents

Running TGI on Gaudi

TGI-Gaudi Benchmark

Static Batching Benchmark

Continuous Batching Benchmark

Tested Models and Configurations

Running TGI with BF16 Precision

Llama2-7B on 1 Card

Llama2-70B on 8 cards

Llama3.1-8B on 1 card

Llama3.1-70B 8 cards

Llava-v1.6-Mistral-7B on 1 card

Running TGI with FP8 Precision

Llama2-7B on 1 Card

Llama2-70B on 8 Cards

Llama3.1-8B on 1 Card

Llama3.1-70B on 8 cards

Llava-v1.6-Mistral-7B on 1 Card

Llava-v1.6-Mistral-7B on 8 Cards

Adjusting TGI Parameters

Environment Variables

Profiler

License