mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-20 22:32:07 +00:00
c35f39cf83
# Add AWQ quantization inference support
Fixes
https://github.com/huggingface/text-generation-inference/issues/781
This PR (partially) adds support for AWQ quantization for inference.
More information on AWQ [here](https://arxiv.org/abs/2306.00978). In
general, AWQ is faster and more accurate than GPTQ, which is currently
supported by TGI.
This PR installs 4-bit GEMM custom CUDA kernels released by AWQ authors
(in `requirements.txt`, just one line change).
Quick way to test this PR would be bring up TGI as follows:
```
text-generation-server download-weights abhinavkulkarni/codellama-CodeLlama-7b-Python-hf-w4-g128-awq
text-generation-launcher \
--huggingface-hub-cache ~/.cache/huggingface/hub/ \
--model-id abhinavkulkarni/codellama-CodeLlama-7b-Python-hf-w4-g128-awq \
--trust-remote-code --port 8080 \
--max-input-length 2048 --max-total-tokens 4096 --max-batch-prefill-tokens 4096 \
--quantize awq
```
Please note:
* This PR was tested with FlashAttention v2 and vLLM.
* This PR adds support for AWQ inference, not quantizing the models.
That needs to be done outside of TGI, instructions
[here](f084f40bd9).
* This PR only adds support for `FlashLlama` models for now.
* Multi-GPU setup has not been tested.
* No integration tests have been added so far, will add later if
maintainers are interested in this change.
* This PR can be tested on any of the models released
[here](https://huggingface.co/abhinavkulkarni?sort_models=downloads#models).
Please refer to the linked issue for benchmarks for
[abhinavkulkarni/meta-llama-Llama-2-7b-chat-hf-w4-g128-awq](https://huggingface.co/abhinavkulkarni/meta-llama-Llama-2-7b-chat-hf-w4-g128-awq)
vs
[TheBloke/Llama-2-7b-Chat-GPTQ](https://huggingface.co/TheBloke/Llama-2-7b-Chat-GPTQ).
Please note, AWQ has released faster (and in case of Llama, fused)
kernels for 4-bit GEMM, currently at the top of the `main` branch at
https://github.com/mit-han-lab/llm-awq, but this PR uses an older commit
that has been tested to work. We can switch to latest commit later on.
## Who can review?
@OlivierDehaene OR @Narsil
---------
Co-authored-by: Abhinav Kulkarni <abhinav@concentric.ai>
23 lines
1.9 KiB
Markdown
23 lines
1.9 KiB
Markdown
# Preparing the Model
|
|
|
|
Text Generation Inference improves the model in several aspects.
|
|
|
|
## Quantization
|
|
|
|
TGI supports [bits-and-bytes](https://github.com/TimDettmers/bitsandbytes#bitsandbytes), [GPT-Q](https://arxiv.org/abs/2210.17323) and [AWQ](https://arxiv.org/abs/2306.00978) quantization. To speed up inference with quantization, simply set `quantize` flag to `bitsandbytes`, `gptq` or `awq` depending on the quantization technique you wish to use. When using GPT-Q quantization, you need to point to one of the models [here](https://huggingface.co/models?search=gptq) when using AWQ quantization, you need to point to one of the models [here](https://huggingface.co/models?search=awq). To get more information about quantization, please refer to (./conceptual/quantization.md)
|
|
|
|
|
|
## RoPE Scaling
|
|
|
|
RoPE scaling can be used to increase the sequence length of the model during the inference time without necessarily fine-tuning it. To enable RoPE scaling, simply pass `--rope-scaling`, `--max-input-length` and `--rope-factors` flags when running through CLI. `--rope-scaling` can take the values `linear` or `dynamic`. If your model is not fine-tuned to a longer sequence length, use `dynamic`. `--rope-factor` is the ratio between the intended max sequence length and the model's original max sequence length. Make sure to pass `--max-input-length` to provide maximum input length for extension.
|
|
|
|
<Tip>
|
|
|
|
We recommend using `dynamic` RoPE scaling.
|
|
|
|
</Tip>
|
|
|
|
## Safetensors
|
|
|
|
[Safetensors](https://github.com/huggingface/safetensors) is a fast and safe persistence format for deep learning models, and is required for tensor parallelism. TGI supports `safetensors` model loading under the hood. By default, given a repository with `safetensors` and `pytorch` weights, TGI will always load `safetensors`. If there's no `pytorch` weights, TGI will convert the weights to `safetensors` format.
|