Skip to content

fix: Linux NVIDIA tarball exceeds GitHub's 2 GiB asset limit #222

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
thcp merged 1 commit into from
Jun 24, 2026
Merged

fix: Linux NVIDIA tarball exceeds GitHub's 2 GiB asset limit #222

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 10 additions & 7 deletions packaging/linux/README-LINUX.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,17 +38,20 @@ expects FFmpeg on your PATH. Install both with your package manager.
NVIDIA variant
--------------

The NVIDIA/CUDA build bundles a CUDA-enabled PyTorch. To use the GPU you also
need a working NVIDIA driver on the host such that `nvidia-smi` runs and reports
your GPU. The CUDA runtime itself is bundled — you do NOT need a separate CUDA
toolkit install, only the driver.
To use the GPU you need a working NVIDIA driver on the host such that
`nvidia-smi` runs and reports your GPU.

Check your driver:
nvidia-smi

If no usable GPU is detected, the NVIDIA build still runs but falls back to CPU.
If you do not have an NVIDIA GPU, use the CPU-only tarball instead — it is
smaller and avoids downloading the CUDA runtime.
On first launch, the NVIDIA build detects your GPU and downloads the matching
CUDA-enabled PyTorch (a few GB) into your data directory — so the first run
needs an internet connection and some disk space. You do NOT need a separate
CUDA toolkit install, only the driver.

If no usable GPU is detected, the NVIDIA build still runs and falls back to CPU.
If you do not have an NVIDIA GPU, use the CPU-only tarball instead — it skips
the CUDA download entirely.

Notes
-----
Expand Down
29 changes: 16 additions & 13 deletions scripts/linux/make-portable.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,19 +97,22 @@ if [[ -n "$PACKAGE_VERSION" ]]; then
fi
uv pip install --system --python "$BUNDLED_PYTHON" "$REPO_ROOT"

if [[ "$CPU_ONLY" == "1" ]]; then
# Linux PyPI torch wheels bundle CUDA. Force-reinstall the CPU-only variant
# afterwards: pip strips the local '+cpu' version when resolving, so the project
# install pulls the CUDA wheel even if a CPU wheel was requested. Mirrors the
# proven Windows swap (--force-reinstall --no-deps replaces just torch wheels).
echo "==> Forcing CPU-only torch"
"$BUNDLED_PYTHON" -m pip install \
"torch==${TORCH_VERSION}+cpu" "torchaudio==${TORCH_VERSION}+cpu" \
--index-url https://download.pytorch.org/whl/cpu \
--force-reinstall --no-deps
else
echo "==> Keeping CUDA torch (NVIDIA variant)"
fi
# Always bake the small CPU-only torch wheel — for BOTH variants. On Linux the
# default PyPI torch wheel bundles the full CUDA runtime (~2.5 GB), which makes
# the packaged tarball exceed GitHub's 2 GiB per-asset release limit. So we
# mirror what the Windows NVIDIA package actually does: ship CPU torch, and let
# the desktop shell download the matching CUDA wheel at first run on GPU
# machines (install_cuda_torch, gated cfg(not(macos)) so it covers Linux). The
# NVIDIA variant differs only by omitting the cpu-only marker below.
#
# pip strips the local '+cpu' version when resolving, so the project install
# pulls the CUDA wheel even if a CPU wheel was requested; --force-reinstall
# --no-deps replaces just the torch/torchaudio wheels (proven on Windows).
echo "==> Baking CPU-only torch (NVIDIA variant downloads CUDA at first run)"
"$BUNDLED_PYTHON" -m pip install \
"torch==${TORCH_VERSION}+cpu" "torchaudio==${TORCH_VERSION}+cpu" \
--index-url https://download.pytorch.org/whl/cpu \
--force-reinstall --no-deps

echo "==> Verifying imports"
"$BUNDLED_PYTHON" -c "import fastapi, uvicorn, yt_dlp, demucs, torch, torchaudio, librosa, pyloudnorm, soundfile; print('torch', torch.__version__, 'cuda', torch.version.cuda)"
Expand Down