-
Notifications
You must be signed in to change notification settings - Fork 1.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add AWQ quantization inference support (#1019) #1054
Conversation
# Add AWQ quantization inference support Fixes #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](https://github.com/mit-han-lab/llm-awq/tree/f084f40bd996f3cf3a0633c1ad7d9d476c318aaa). * 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 <[email protected]>
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
@abhinavkulkarni For visibility. |
They have released one new GEMV kernel and one new GEMM kernel. On their main branch, they use GEMV for token generation and GEMM for context processing. The GEMV kernel is 20% faster than the old GEMM kernel, but most importantly, the new GEMM kernel is 5-6 slower than the old GEMM kernel. My conclusion is that the current GEMM kernel already used in this PR is the optimal one for now. |
I didn't test the GEMV, but the GEMM seems to have similar speeds for me (A10G) which card did you test on ? |
There are multiple GEMM kernels. The new one is slower because it implements a different packed format for quantized models. I tested on RTX 3090, 4090, and A100. On RTX 3090, speed of context processing on LLaMa 7B: GEMM (original): 2400 tokens/s EDIT: What I mean by "new" GEMM kernel is this pull request that is about to be merged into the original llm-awq: mit-han-lab/llm-awq#90 |
Those seem like throughput values, which are relatively not important in general. We really care about latency much more. Ideally the whole curves gives a better story since we do want throughput, but at fixed latency. (2x throughput for 2x latency, is usually unacceptable in our deployments for instance). |
Add AWQ quantization inference support
Fixes
#781
This PR (partially) adds support for AWQ quantization for inference.
More information on AWQ here. 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:
Please note:
That needs to be done outside of TGI, instructions
here.
FlashLlama
models for now.maintainers are interested in this change.
here.
Please refer to the linked issue for benchmarks for
abhinavkulkarni/meta-llama-Llama-2-7b-chat-hf-w4-g128-awq
vs
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 athttps://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
What does this PR do?
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.