From b80aee1e46fea6bea620203d747d2435368eb6d0 Mon Sep 17 00:00:00 2001
From: Thales <>
Date: Wed, 24 Jun 2026 20:44:50 +0100
Subject: [PATCH] fix: remove orphaned nvidia-* CUDA packages from Linux bundle

Both Linux tarballs were 2.5 GB (over GitHub's 2 GiB asset limit) even
with CPU torch. Root cause: 'uv pip install <project>' pulls the default
Linux torch, which is the CUDA build, dragging in nvidia-* runtime
packages (cuDNN/cuBLAS/NCCL/...) and triton (~2.5 GB). The CPU torch swap
uses --force-reinstall --no-deps, so torch becomes CPU but those CUDA
packages stay installed and orphaned, bloating the tarball.

Uninstall the nvidia-* packages and triton after the swap. CPU torch does
not use them and the NVIDIA variant re-downloads CUDA at first run.
---
 scripts/linux/make-portable.sh | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/scripts/linux/make-portable.sh b/scripts/linux/make-portable.sh
index c0eec902..18412c82 100755
--- a/scripts/linux/make-portable.sh
+++ b/scripts/linux/make-portable.sh
@@ -114,6 +114,19 @@ echo "==> Baking CPU-only torch (NVIDIA variant downloads CUDA at first run)"
   --index-url https://download.pytorch.org/whl/cpu \
   --force-reinstall --no-deps
 
+# The project install above pulled the default Linux torch, which is the CUDA
+# build, dragging in nvidia-* CUDA runtime packages (cuDNN, cuBLAS, NCCL, ...) and
+# triton -- together ~2.5 GB. The CPU torch swap used --no-deps, so those packages
+# are now orphaned but still installed, bloating the tarball past GitHub's 2 GiB
+# asset limit. Remove them: CPU torch does not use them, and the NVIDIA variant
+# re-downloads CUDA at first run anyway.
+echo "==> Removing orphaned CUDA runtime packages"
+orphans=$("$BUNDLED_PYTHON" -m pip list --format=freeze 2>/dev/null \
+  | sed -n 's/^\(nvidia-[^=]*\)==.*/\1/p')
+orphans="$orphans triton"
+echo "    removing:$orphans"
+"$BUNDLED_PYTHON" -m pip uninstall -y $orphans 2>/dev/null || true
+
 echo "==> Verifying imports"
 "$BUNDLED_PYTHON" -c "import fastapi, uvicorn, yt_dlp, demucs, torch, torchaudio, librosa, pyloudnorm, soundfile; print('torch', torch.__version__, 'cuda', torch.version.cuda)"