README.md

---
license: apache-2.0
---
  
# vllm-flash-attn3

This is an implementation of Flash Attention 3 CUDA kernels with support for attention sinks. The attention sinks implementation was contributed to Flash Attention by the [vLLM team](https://huggingface.co/vllm-project). The [transformers team](https://huggingface.co/transformers-community) packaged the implementation and pre-built it for use with the [kernels library](https://github.com/huggingface/kernels).

## How to Use

When loading your model with transformers, provide this repository id as the source of the attention implementation:

```diff
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "<your model id on the Hub>"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
+    # Flash Attention with Sinks
+    attn_implementation="kernels-community/vllm-flash-attn3”,
)
```

This will automatically resolve and download the appropriate code for your architecture. See more details in [this post](https://huggingface.co/blog/hello-hf-kernels).

## Credits

- [Tri Dao](https://huggingface.co/tridao) and team for Flash Attention and [Flash Attention 3](https://tridao.me/blog/2024/flash3/).
- The [vLLM team](https://huggingface.co/vllm-project) for their implementation and their contribution of attention sinks.
- The [transformers team](https://huggingface.co/transformers-community) for packaging, testing, building and making it available for use with the [kernels library](https://github.com/huggingface/kernels).