mirror of
https://github.com/mudler/LocalAI.git
synced 2026-07-31 02:18:50 -04:00
35922ac578
Plain 'flash-attn' in requirements-cublas13-after.txt cannot install: uv resolves the PyPI sdist and falls back to a source build, but the backend image builds FROM ubuntu:24.04 with no CUDA toolkit (nvcc), and Dao-AILab ships no prebuilt wheel for this venv (cu13 wheels exist only for cp312/torch2.9; the venv is Python 3.10 + torch 2.13 + cu130). The cublas12 variant only works because flash-attn's setup.py downloads a matching official cu12/cp310 wheel at build time. Pin the community prebuilt wheel (mjun0812/flash-attention-prebuild-wheels v0.9.47, flash_attn-2.8.3+cu130torch2.13-cp310-cp310-linux_x86_64) - the same pinned-URL pattern vllm-omni already uses for its flash-attn wheel - and pin torch to 2.13.* in requirements-cublas13.txt since flash-attn wheels are ABI-tied to the torch minor version. Fixes the cublas13 backend-build job failure; the cublas12/hipblas/intel failures on the previous run were already fixed on master (all qwen-tts variants built green on master on 2026-07-14). Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Assisted-by: Claude:opus-4.8 [Claude Code]