How to Troubleshoot Mixed NVIDIA Package Versions in uv pip list Output

Clear the uv cache, delete uv.lock and .venv, then run uv sync --extra cu124 to resolve mixed NVIDIA package versions caused by stale wheel metadata when switching between CPU and CUDA PyTorch extras.

When managing PyTorch installations with Astral's uv package manager, switching between CPU-only and CUDA-enabled extras can leave your environment in an inconsistent state. The baonguyen6742/uv-install-torch repository demonstrates how optional dependencies and index-specific sources configured in pyproject.toml lead to mixed NVIDIA package versions appearing in uv pip list. This guide extracts the exact remediation steps from the repository's source code to restore a clean, CUDA-compatible environment.

Understanding the Root Cause

Mixed NVIDIA package versions typically occur when uv resolves dependencies using cached metadata from a previous installation with a different extra (e.g., cpu instead of cu124).

Extra-Aware Index Selection

In the repository's pyproject.toml, torch-related packages are defined as optional dependencies tied to specific PyPI indexes:

[project.optional-dependencies]
cpu = ["torch==2.4.1", "torchvision==0.19.1", "torchaudio==2.4.1"]
cu124 = ["torch==2.4.1", "torchvision==0.19.1", "torchaudio==2.4.1"]

[tool.uv.sources]
torch = [
  { index = "pytorch-cpu", extra = "cpu" },
  { index = "pytorch-cu124", extra = "cu124" },
]

When you run uv sync --extra cu124, uv pulls CUDA-enabled wheels from the pytorch-cu124 index. However, transitive dependencies like nvidia-cublas-cu12 are not part of the extra definition and are pulled from the default PyPI source. If a previous cpu extra installation left cached wheels or lock file references, uv may reuse CPU-only torch wheels while still installing CUDA-specific NVIDIA packages, creating a hybrid environment.

Cache Reuse and Lock File Stalemates

uv caches wheels globally to speed up repeated installations. If you previously synced with --extra cpu, the CPU-only torch wheel remains in the cache. Subsequent attempts to sync with --extra cu124 may reference the stale lock file or cache entries, causing uv to install torch compiled for CPU alongside nvidia-cublas-cu12 and nvidia-cudnn-cu12 wheels intended for CUDA 12.4.

Identifying Mixed Version Symptoms

The conflict manifests in uv pip list output showing NVIDIA packages (e.g., nvidia-cublas-cu12, nvidia-cudnn-cu12) alongside a torch package lacking CUDA suffixes. Runtime errors include:

• torch.cuda.is_available() returning False despite CUDA drivers being installed
• Import failures or undefined symbol errors when importing torch
• Version mismatches where the torch wheel reports 2.4.1+cpu while NVIDIA packages report cu12 suffixes

Step-by-Step Resolution

Follow the remediation protocol documented in the repository's README.md (lines 101-106) to purge stale state and force a clean resolution.

1. Clear the Package Cache

Remove stale wheel files that could be reused mistakenly:

# Clear all unused cached wheels

uv cache prune

# Or target only torch-related artifacts

uv cache clean torch

2. Delete Lock File and Virtual Environment

Force uv sync to resolve the dependency graph from scratch, guaranteeing that only the selected extra is installed:

rm -f uv.lock
rm -rf .venv

3. Re-install the Torch Stack

Invoke uv sync with the correct CUDA extra to pull wheels exclusively from the pytorch-cu124 index:

uv sync --extra cu124

If leftover packages persist, force an explicit uninstall before syncing:

uv pip uninstall -y torch torchvision torchaudio
uv sync --extra cu124

4. Verify CUDA Compatibility

Confirm the environment contains only CUDA-compatible packages:

uv pip list | grep -E 'torch|cuda|nvidia'

Run a Python sanity check to validate CUDA availability:

uv run python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

Expected output: 2.4.1+cu124 True.

Distinguishing Normal from Problematic NVIDIA Packages

After a clean reinstall, seeing nvidia-cublas-cu12, nvidia-cudnn-cu12, and similar packages is normal. These are legitimate transitive dependencies of the CUDA-enabled torch wheel and should all share the same CUDA version suffix (e.g., cu12).

Mismatched versions—such as nvidia-cublas-cu11 alongside torch==2.4.1+cu124—indicate that the wrong index was consulted or remnants of a previous installation remain. Similarly, if torch reports a +cpu build tag while NVIDIA packages are present, the mixed state persists and requires repeating the cache clearing steps.

Summary

• Clear the uv cache using uv cache prune to remove stale CPU-only wheels that contaminate CUDA installations.
• Regenerate lock files by deleting uv.lock and .venv to force a fresh dependency resolution for the target extra.
• Use explicit extras (--extra cu124 or --extra cpu) to ensure uv selects wheels from the correct index defined in pyproject.toml.
• Verify with torch.cuda.is_available() to confirm the runtime environment matches the expected CUDA version.

Frequently Asked Questions

Why do I see NVIDIA packages after installing the CPU extra?

NVIDIA packages should not appear when using --extra cpu. If they do, a previous CUDA installation left artifacts in the cache or lock file. Run uv cache clean and delete uv.lock to remove these remnants, then resync with --extra cpu.

Is it safe to delete the uv.lock file and .venv directory?

Yes. uv.lock is a generated file that locks resolved dependency versions, and .venv contains the installed packages. Deleting both forces uv to re-resolve and re-download dependencies according to your current pyproject.toml configuration, which is the recommended fix for mixed-version scenarios.

Why does uv reuse cached wheels from different extras?

uv caches wheels globally based on package name and version, not by extra. A torch==2.4.1 wheel cached from the CPU index appears identical to the resolver as one from the CUDA index. Without clearing the cache or lock file, uv may install the cached CPU wheel when you request the CUDA extra, leading to the mixed NVIDIA package state.

How do I know if my NVIDIA packages are the correct version?

Run uv pip list and check that all nvidia-* packages share the same CUDA version suffix as your torch extra (e.g., cu12 for cu124). Then execute python -c "import torch; print(torch.version.cuda)". The output should match the CUDA version (e.g., 12.4) associated with your installed NVIDIA packages.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}