Text Generation Inference on Habana Gaudi

Running TGI on Gaudi

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

Pull the official Docker image with:

docker pull ghcr.io/huggingface/tgi-gaudi:2.0.1

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/Gaudi2 card

model=meta-llama/Llama-2-7b-hf
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 --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.1 --model-id $model --max-input-tokens 1024 --max-total-tokens 2048

For gated models such as LLama or 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/Gaudi2 card using pytorch eager mode with torch compile:

model=meta-llama/Llama-2-7b-hf
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 --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.1 --model-id $model --max-input-tokens 1024 --max-total-tokens 2048

iii. On 8 Gaudi/Gaudi2 cards:

model=meta-llama/Llama-2-70b-hf
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 --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.1 --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'

To run static benchmark test, 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.
To run continuous batching test, please refer to examples.

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-prefill-total-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 begginging 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 legnth 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.

Running TGI with FP8 precision

TGI supports FP8 precision runs within the limits provided by Habana Quantization Toolkit. Models with FP8 can be ran by properly setting QUANT_CONFIG environment variable. Detailed instruction on how to use that variable can be found in Optimum Habana FP8 guide. Summarising that instruction in TGI cases:

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

Note

Only models pointed in supported configurations are guaranteed to work with FP8

Additional hints to quantize model for TGI when using run_lm_eval.py:

use --limit_hpu_graphs flag to save memory
try to model your use case situation by adjusting --batch_size , --max_new_tokens 512 and --max_input_tokens 512; in case of memory issues, lower those values
use dataset/tasks suitable for your use case (see --help for defining tasks/datasets)

Currently supported configurations

Not all features of TGI are currently supported as this is still a work in progress. Currently supported and validated configurations (other configurations are not guaranted to work or ensure reasonable performance):

LLama 7b BF16 on 1 Gaudi2 card

model=meta-llama/Llama-2-7b-chat-hf
hf_token=YOUR_ACCESS_TOKEN   # Llama2 is a gated model and requires a special 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 $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e PREFILL_BATCH_BUCKET_SIZE=1 \
   -e BATCH_BUCKET_SIZE=16 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=128 \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.1 \
   --model-id $model \
   --max-input-tokens 1024 \
   --max-batch-prefill-tokens 4096 \
   --max-total-tokens 2048 \
   --max-batch-size 16

LLama 7b FP8 on 1 Gaudi2 card

model=meta-llama/Llama-2-7b-chat-hf
hf_token=YOUR_ACCESS_TOKEN   # Llama2 is a gated model and requires a special 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 $PWD/hqt_output:/usr/src/hqt_output \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e PREFILL_BATCH_BUCKET_SIZE=1 \
   -e BATCH_BUCKET_SIZE=64 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=128 \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.1 \
   --model-id $model \
   --max-input-tokens 1024 \
   --max-batch-prefill-tokens 4096 \
   --max-total-tokens 2048 \
   --max-batch-size 64

LLama 70b BF16 on 8 Gaudi2 card

model=meta-llama/Llama-2-70b-chat-hf
hf_token=YOUR_ACCESS_TOKEN   # Llama2 is a gated model and requires a special 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 $volume:/data \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e PREFILL_BATCH_BUCKET_SIZE=1 \
   -e BATCH_BUCKET_SIZE=256 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=128 \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.1 \
   --model-id $model \
   --max-input-tokens 1024 \
   --max-batch-prefill-tokens 16384 \
   --max-total-tokens 2048 \
   --max-batch-size 256 \
   --max-concurrent-requests 400 \
   --sharded true \
   --num-shard 8

LLama 70b FP8 on 8 Gaudi2 card

model=meta-llama/Llama-2-70b-chat-hf
hf_token=YOUR_ACCESS_TOKEN   # Llama2 is a gated model and requires a special 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 $PWD/hqt_output:/usr/src/hqt_output \
   -e HABANA_VISIBLE_DEVICES=all \
   -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
   -e HF_HUB_ENABLE_HF_TRANSFER=1 \
   -e HUGGING_FACE_HUB_TOKEN=$hf_token \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e PREFILL_BATCH_BUCKET_SIZE=1 \
   -e BATCH_BUCKET_SIZE=512 \
   -e PAD_SEQUENCE_TO_MULTIPLE_OF=128 \
   -e QUANT_CONFIG=./quantization_config/maxabs_quant.json \
   --cap-add=sys_nice \
   --ipc=host \
   ghcr.io/huggingface/tgi-gaudi:2.0.1 \
   --model-id $model \
   --max-input-tokens 1024 \
   --max-batch-prefill-tokens 16384 \
   --max-total-tokens 2048 \
   --max-batch-size 512 \
   --max-concurrent-requests 700 \
   --sharded true \
   --num-shard 8

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

Other sequence lengths can be used with proportionally decreased/increased batch size (the higher sequence length, the lower batch size). Support for other models from Optimum Habana will be added successively.

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

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 [email protected] if you have any question.

Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
assets		assets
benchmark		benchmark
clients/python		clients/python
docs		docs
examples		examples
integration-tests		integration-tests
launcher		launcher
load_tests		load_tests
proto		proto
router		router
server		server
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
Dockerfile		Dockerfile
Dockerfile_amd		Dockerfile_amd
Dockerfile_intel		Dockerfile_intel
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
rust-toolchain.toml		rust-toolchain.toml
sagemaker-entrypoint.sh		sagemaker-entrypoint.sh
update_doc.py		update_doc.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Text Generation Inference on Habana Gaudi

Table of contents

Running TGI on Gaudi

Adjusting TGI parameters

Running TGI with FP8 precision

Currently supported configurations

LLama 7b BF16 on 1 Gaudi2 card

LLama 7b FP8 on 1 Gaudi2 card

LLama 70b BF16 on 8 Gaudi2 card

LLama 70b FP8 on 8 Gaudi2 card

Environment variables

Profiler

About

Releases

Packages

Languages

License

bjornrun/tgi-gaudi-fixed-llama3.1

Folders and files

Latest commit

History

Repository files navigation

Text Generation Inference on Habana Gaudi

Table of contents

Running TGI on Gaudi

Adjusting TGI parameters

Running TGI with FP8 precision

Currently supported configurations

LLama 7b BF16 on 1 Gaudi2 card

LLama 7b FP8 on 1 Gaudi2 card

LLama 70b BF16 on 8 Gaudi2 card

LLama 70b FP8 on 8 Gaudi2 card

Environment variables

Profiler

About

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages