GPTQ / Quantization support?

[nikshepsvn]

Will vLLM support 4-bit GPTQ models?

[WoosukKwon]

Thanks for the feature request! Quantization is not currently supported, but it's definitely on our roadmap. Please stay tuned.

[nikshepsvn]

How do I best go about tracking this? Is there a discord or public roadmap somewhere I can look at?

[Symbolk]

How do I best go about tracking this? Is there a discord or public roadmap somewhere I can look at?

See Roadmap here: #244

[chu-tianxiang]

I looked into this a bit today and it seems straight forward to integrate AutoGPTQ into vllm, so I implemented a preliminary version for LLaMA (see this commit) and did a few benchmarks on single A100-80G. I don't know why but It's slower than expected.

python benchmark_throughput.py --model TheBloke/Llama-2-13B-chat-GPTQ --dataset ShareGPT_V3_unfiltered_cleaned_split.json

Model	Throughout (requests/s)	Throughout (tokens/s)
meta-llama/Llama-2-13b-chat-hf	4.00	1915
TheBloke/Llama-2-13B-chat-GPTQ	3.32	1587
TheBloke/Llama-2-70B-chat-GPTQ	1.09	519

[osilverstein]

I looked into this a bit today and it seems straight forward to integrate AutoGPTQ into vllm, so I implemented a preliminary version for LLaMA (see this commit) and did a few benchmarks on single A100-80G. I don't know why but It's slower than expected.

python benchmark_throughput.py --model TheBloke/Llama-2-13B-chat-GPTQ --dataset ShareGPT_V3_unfiltered_cleaned_split.json

Model Throughout (requests/s) Throughout (tokens/s)
meta-llama/Llama-2-13b-chat-hf 4.00 1915
TheBloke/Llama-2-13B-chat-GPTQ 3.32 1587
TheBloke/Llama-2-70B-chat-GPTQ 1.09 519

what's the baseline with normal version?

[chu-tianxiang]

I looked into this a bit today and it seems straight forward to integrate AutoGPTQ into vllm, so I implemented a preliminary version for LLaMA (see this commit) and did a few benchmarks on single A100-80G. I don't know why but It's slower than expected.

python benchmark_throughput.py --model TheBloke/Llama-2-13B-chat-GPTQ --dataset ShareGPT_V3_unfiltered_cleaned_split.json

Model Throughout (requests/s) Throughout (tokens/s)
meta-llama/Llama-2-13b-chat-hf 4.00 1915
TheBloke/Llama-2-13B-chat-GPTQ 3.32 1587
TheBloke/Llama-2-70B-chat-GPTQ 1.09 519

what's the baseline with normal version?

If you mean the throughput, in the above table TheBloke/Llama-2-13B-chat-GPTQ is quantized from meta-llama/Llama-2-13b-chat-hf and the throughput is about 17% less.

I dug into the kernel code of quant linear layer and found that it falls back to dequantization followed by fp16 matrix multiplication when the batch size is bigger than 8, so the performance degradation is understandable.

[chu-tianxiang]

As an update, I added tensor parallel QuantLinear layer and supported most AutoGPT compatible models in this branch. The code has not been thoroughly tested yet because the combinations of model architectures and GPTQ settings are way too many.

[singularity-sg]

@chu-tianxiang I tried forking your vllm-gptq branch and was successful deploying the TheBloke/Llama-2-13b-Chat-GPTQ model. However, when I tried the TheBloke/Llama-2-7b-Chat-GPTQ model, it threw the following exception whenever I made a query to the model. I wonder if the issue is with the model itself or something else. I'll dig further into this when I have the chance but it's likely the Sampler was generating the probability tensor with invalid values

INFO:     13.229.18.8:52663 - "POST /v1/completions HTTP/1.1" 500 Internal Server Error
ERROR:    Exception in ASGI application
Traceback (most recent call last):
  File "/home/ubuntu/vllm-gptq/vllm/engine/async_llm_engine.py", line 28, in _raise_exception_on_finish
    task.result()
  File "/home/ubuntu/vllm-gptq/vllm/engine/async_llm_engine.py", line 351, in run_engine_loop
    has_requests_in_progress = await self.engine_step()
                               ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/engine/async_llm_engine.py", line 330, in engine_step
    request_outputs = await self.engine.step_async()
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/engine/async_llm_engine.py", line 191, in step_async
    output = await self._run_workers_async(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/engine/async_llm_engine.py", line 220, in _run_workers_async
    all_outputs = await asyncio.gather(*all_outputs)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/.pyenv/versions/3.11.3/lib/python3.11/asyncio/tasks.py", line 684, in _wrap_awaitable
    return (yield from awaitable.__await__())
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ray.exceptions.RayTaskError(RuntimeError): ray::RayWorker.execute_method() (pid=68296, ip=172.31.40.30, actor_id=3b90ca9f90ebf20a67ae6c2c01000000, repr=<vllm.engine.ray_utils.RayWorker object at 0x7ef09122dad0>)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/engine/ray_utils.py", line 32, in execute_method
    return executor(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/.pyenv/versions/vllm-gptq/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/worker/worker.py", line 305, in execute_model
    output = self.model(
             ^^^^^^^^^^^
  File "/home/ubuntu/.pyenv/versions/vllm-gptq/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/model_executor/models/llama.py", line 296, in forward
    next_tokens = self.sampler(self.lm_head.weight, hidden_states,
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/.pyenv/versions/vllm-gptq/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/model_executor/layers/sampler.py", line 85, in forward
    return _sample(probs, logprobs, input_metadata)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/model_executor/layers/sampler.py", line 451, in _sample
    sample_results = _random_sample(seq_groups, is_prompts,
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/vllm-gptq/vllm/model_executor/layers/sampler.py", line 342, in _random_sample
    random_samples = torch.multinomial(probs,
                     ^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: probability tensor contains either `inf`, `nan` or element < 0

[wejoncy]

Hi,
If anyone wants try GPTQ quantizationo in vLLM.
Please use this repo QLLM to quantize model(LLama) and it would compatiable AWQ in vLLM.
And Of courcr you can select AWQ to quantize it as well.

[David-Lee-1990]

is baichuan-gptq supported?

[sssuperrrr]

支持Qwen-72B-Chat-Int4加速吗？

[uncensorie]

Hi, If anyone wants try GPTQ quantizationo in vLLM. Please use this repo QLLM to quantize model(LLama) and it would compatiable AWQ in vLLM. And Of courcr you can select AWQ to quantize it as well.

Something is off with this QLLM gptq quantization @wejoncy ... all the dependencies aren't specified in requirements file. Also, 3 times tried quantizing and every time it breaks when it tries to save the file or quant is done. Tried Llama2-70b and mistral 7b