[Bugfix] Add int8 torch dtype for KVCache

[shen-shanshan]

Some attention backend requires int8 kvcache dtype (e.g., quantization). It is used in initialization of CacheConfig:

if cache_config.cache_dtype == "auto":
    self.dtype = model_config.dtype
else:
    self.dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]

But there are no int8 dtype in STR_DTYPE_TO_TORCH_DTYPE:

STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.half,
    "bfloat16": torch.bfloat16,
    "float": torch.float,
    "fp8": torch.uint8,
    "fp8_e4m3": torch.uint8,
    "fp8_e5m2": torch.uint8,
}

So, I think maybe it's better to add int8 into this STR_DTYPE_TO_TORCH_DTYPE.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[houseroad]

Have we ever tested any thing with int8 KV cache?

Is adding an item to a map enough? I am wondering how int8 KV works here.

[Isotr0py]

Have we ever tested any thing with int8 KV cache?
Is adding an item to a map enough? I am wondering how int8 KV works here.

I think no quantization in main repo support int8 kv cache currently, but some OOT hardware like vllm-ascend indeed can support int8 kv_cache: https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/quantization/quant_config.py#L60-L96

[shen-shanshan]

Have we ever tested any thing with int8 KV cache?
Is adding an item to a map enough? I am wondering how int8 KV works here.

I think no quantization in main repo support int8 kv cache currently, but some OOT hardware like vllm-ascend indeed can support int8 kv_cache: https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/quantization/quant_config.py#L60-L96

Yes, thanks for your explaination~