Can vllm become faster?

[godsakurapeng]

I find an artice Accelerating Generative AI with PyTorch II: GPT, Fast
The optimization used in this article is as shown below

I simply tried gpt-fast, the improvement is huge

codellama-python-7b, 2xA10(24G)

infer	speed(token/s)
vllm fp16	45.2
gpt-fast fp16	66.5
gpt-fast int8	105.1
gpt-fast int4	145.9

ps: results of int4 is terrible

I'm curious, can these optimizations be used on vllm?
I can see some discussion about these optimizations, but it doesn't look like they will be possible in the short term (because of some problems about vllm?)

torch.compile

+34% higher throughput?
Compiled model with torch.compile, unfortunately without performance improvements

quantization

Add GPTQ support (I tried a version before but it didn't work well

Speculative Decoding

Speculative Decoding

vllm is a great project!! I really hope to see these optimizations in vllm. I also want to know the difficulties that still exist :)

[yuxx0218]

hi, I want to run THUDM/chatglm-6b-int4 by vllm, but raising cuda oom error. Based on the log, it requires at least 10G gpu memory. Actually when I use huggingface transformers directly run this model, it only use 5G memory. Do you know why?

[godsakurapeng]

hi, I want to run THUDM/chatglm-6b-int4 by vllm, but raising cuda oom error. Based on the log, it requires at least 10G gpu memory. Actually when I use huggingface transformers directly run this model, it only use 5G memory. Do you know why?嗨，我想通过 vllm 运行 THUDM/chatglm-6b-int4，但引发 cuda oom 错误。根据日志，它至少需要 10G GPU 内存。其实我用 huggingface transformers 直接运行这个模型的时候，它只用了5G内存。你知道为什么吗？

#2176 maybe you can look at this issue

[yuxx0218]

hi, I want to run THUDM/chatglm-6b-int4 by vllm, but raising cuda oom error. Based on the log, it requires at least 10G gpu memory. Actually when I use huggingface transformers directly run this model, it only use 5G memory. Do you know why?嗨，我想通过 vllm 运行 THUDM/chatglm-6b-int4，但引发 cuda oom 错误。根据日志，它至少需要 10G GPU 内存。其实我用 huggingface transformers 直接运行这个模型的时候，它只用了5G内存。你知道为什么吗？

#2176 maybe you can look at this issue

THX, but it not works. I open a new issue, could you please have a look?
#2338

[hmellor]

Closing this as all the mentioned optimisations are either in progress or merged.