Upload flash-attn3/README.md with huggingface_hub

flash-attn3/README.md

@@ -0,0 +1,35 @@
+# flash-attn3
+
+Flash Attention 3 wheel compiled from [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention) (`hopper/` directory).
+
+## Build environment
+
+| Component | Version |
+|-----------|---------|
+| Base image | NGC 25.11 |
+| CUDA | 13.0 |
+| PyTorch | 2.10 |
+| Python | 3.12 |
+| GPU target | NVIDIA H200 (SM90) |
+
+## Wheel naming
+
+`flash_attn_3-3.0.0-cp39-abi3-linux_x86_64.whl`
+
+- `cp39-abi3` = Python **Stable ABI**, compatible with Python >= 3.9 (including 3.12).
+- CUDA and PyTorch versions are linked at compile time and **not** encoded in the filename.
+
+## Usage
+
+```python
+from flash_attn_interface import flash_attn_func
+```
+
+In HuggingFace Transformers:
+
+```python
+model = AutoModel.from_pretrained(
+    model_id,
+    attn_implementation="flash_attention_3",
+)
+```