metadata
license: bsd-3-clause
tags:
- wheels
- flash-attention
Prebuilt wheels for the OPD environment
Mirror of official upstream release artifacts, re-hosted so machines with slow
GitHub access can pull them through hf-mirror.com. Nothing is rebuilt or
modified — byte-identical to the GitHub release assets.
flash_attn-2.8.3+cu12torch2.9cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
- Upstream: Dao-AILab/flash-attention v2.8.3
- Size: 243 MiB
sha256: 4e2f9e39313266b1544b68138b15b91ee6221eccf14f7902b7c6620351340810- License: BSD-3-Clause (upstream)
Matches exactly one environment. All four coordinates have to line up:
| Coordinate | Value | Why it is fixed |
|---|---|---|
| torch | 2.9.0 | vllm==0.12.0 pins torch==2.9.0 (hard == in its metadata) |
| CUDA | cu12 | PyPI torch 2.9.0 is a cu128 build (nvidia-*-cu12==12.8.x) |
| cxx11abi | TRUE | the only variant published for torch2.9 |
| python | cp312 | the only variant published for torch2.9 — 3.10/3.11/3.13 do not exist |
| platform | linux_x86_64 | aarch64 exists upstream but is not mirrored here |
Requires compute capability >= 8.0 (Ampere or newer). FlashAttention-2 has no
kernels for sm70 (V100). Check with
nvidia-smi --query-gpu=compute_cap --format=csv before bothering.
Also note it is not needed for vLLM generation at all — verl's flash_attn
imports live under verl/models/*/megatron/*, which the generation path never
touches. It is only required once FSDP KL training starts.
Install
HF_ENDPOINT=https://hf-mirror.com hf download HzChen20/opd-wheels \
--repo-type dataset --include "*.whl" --local-dir ./wheels
pip install ./wheels/flash_attn-2.8.3+cu12torch2.9cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
python -c "import flash_attn; print(flash_attn.__version__)"
Verify the transfer first if the link is unreliable:
sha256sum -c <<< "4e2f9e39313266b1544b68138b15b91ee6221eccf14f7902b7c6620351340810 ./wheels/flash_attn-2.8.3+cu12torch2.9cxx11abiTRUE-cp312-cp312-linux_x86_64.whl"