LyAhn 0e17c60bdc fix(server): guard flash-attn install behind OS platform check ...

The win_amd64 wheel URL was attempted on any OS with matching Python/
torch/CUDA tags. On Linux CUDA setups with VIBEPOD_ENABLE_FLASH_ATTN=1
this caused `uv pip install` to fail with an incompatible wheel; with
set -e the script then exited before launch instead of falling back to
SDPA.

- Add uname -s case statement inside the version-tag match: only set
  the wheel URL on MINGW*/CYGWIN*/MSYS* (Windows/Git Bash); all other
  platforms print a clear message and leave FLASH_ATTN_WHEEL_URL empty
- Move the install step into a separate `if [[ -n "$FLASH_ATTN_WHEEL_URL" ]]`
  block so non-Windows platforms skip it entirely
- Wrap `uv pip install` in an `if` so a wheel failure is non-fatal and
  falls through to SDPA regardless of set -e
- Update header comment to reflect cross-platform behaviour

2026-05-01 19:09:33 +01:00

..

.python-version

perf: improve streaming generation pipeline

2026-04-30 18:54:14 +01:00

download_model.py

style: apply ruff formatting and lint fixes to server

2026-05-01 19:06:13 +01:00

pyproject.toml

chore: add ruff linter/formatter for Python server

2026-05-01 19:05:39 +01:00

start.sh

fix(server): guard flash-attn install behind OS platform check

2026-05-01 19:09:33 +01:00

uv.lock

chore: add ruff linter/formatter for Python server

2026-05-01 19:05:39 +01:00

vibevoice_server.py

style: apply ruff formatting and lint fixes to server

2026-05-01 19:06:13 +01:00