How to Force a Clean PyTorch Reinstall by Deleting `.venv` and `uv.lock`

Deleting both .venv and uv.lock forces uv to recreate the virtual environment and re-resolve dependencies from scratch, ensuring a completely clean PyTorch installation without stale CUDA binaries or corrupted wheels.

The baonguyen6742/uv-install-torch repository demonstrates the canonical workflow for managing PyTorch with uv. When CUDA wheels fail to load or dependency resolution drifts, wiping these two artifacts is the fastest path to a pristine environment.

Why Delete .venv and uv.lock?

The uv package manager relies on two persistent artifacts that can harbor stale state:

• .venv – The virtual environment directory containing compiled wheels, including the heavy CUDA binaries for torch, torchvision, and torchaudio. If a previous install pulled the wrong platform wheel or a corrupted file, uv will continue re-using those bad artifacts until the directory is purged.
• uv.lock – The lockfile that pins exact versions and platform-specific resolutions derived during the first sync. Deleting it forces uv sync to recompute the dependency graph from pyproject.toml from scratch, picking up updated indexes or changed extras.

According to the source at README.md lines 73‑74, this cleanup is explicitly recommended when installation problems persist or when switching CUDA versions.

Step-by-Step Clean Reinstall Procedure

1. Remove the Virtual Environment and Lock File

Execute a recursive delete to wipe the existing environment and resolution state:

rm -rf .venv uv.lock

This command eliminates the compiled Python environment and the frozen dependency graph, guaranteeing that the next sync starts fresh.

2. (Optional) Prune the uv Global Cache

If a corrupted wheel persists in uv’s global cache, clear it before reinstalling:

# Remove all unused caches

uv cache prune

# Or target only PyTorch-related artifacts

uv cache clean torch

Pruning prevents the resolver from re-injecting a bad wheel into the new .venv.

3. Reinstall Dependencies with CUDA Support

Run uv sync with the cu124 extra defined in pyproject.toml. This extra maps to the CUDA 12.4 builds of PyTorch and its companion libraries:

uv sync --extra cu124

The resolver reads the [project.optional-dependencies] table in pyproject.toml to pull the correct platform-specific wheels from the index.

4. Verify the Installation with main.py

Validate the stack by running the repository’s sanity-check script:

uv run python main.py

The script imports the scientific-Python stack, prints version metadata, and executes the basic_calculation function to confirm that CUDA tensors allocate successfully on the available device.

Code Examples

The main.py file contains a minimal validation routine that exercises PyTorch’s device placement:

# Excerpt from main.py

import torch

def basic_calculation(a: int, b: int) -> torch.Tensor:
    device = "cuda" if torch.cuda.is_available() else "cpu"
    tensor_a = torch.arange(start=1, end=a + 1).unsqueeze(0)
    tensor_b = torch.arange(start=1, end=b + 1).unsqueeze(0)
    return torch.matmul(tensor_a.T, tensor_b).to(device)

print("Test calculation")
print(basic_calculation(2, 3))

A successful reinstall produces output indicating the tensor resides on cuda:0 (or falls back to cpu on machines without a GPU):

Test calculation
tensor([[1, 2, 3],
        [2, 4, 6]], device='cuda:0')

Automate the entire clean-reinstall workflow in a single shell block:

# Clean previous state

rm -rf .venv uv.lock

# Re-resolve dependencies (CUDA 12.4)

uv sync --extra cu124

# Verify installation

uv run python main.py

Summary

• .venv holds compiled wheels; deleting it removes corrupted or mismatched CUDA binaries.
• uv.lock pins resolved versions; removing it forces a fresh dependency graph calculation.
• The baonguyen6742/uv-install-torch project uses the cu124 extra in pyproject.toml to select CUDA-enabled PyTorch builds.
• The basic_calculation function in main.py provides a quick runtime verification that the new environment is functional.
• Optional cache cleaning via uv cache clean or uv cache prune eliminates bad wheel artifacts from previous attempts.

Frequently Asked Questions

What does uv.lock do in a uv project?

uv.lock is a lockfile that records the exact versions and platform-specific hashes of every dependency resolved during the last uv sync operation. It ensures reproducible builds across machines and time, but it can become stale if pyproject.toml changes or if the initial resolution selected incompatible wheels. Deleting it triggers a full re-resolution from the current state of the package index.

Is it safe to delete the .venv directory?

Yes. The .venv directory is a disposable virtual environment created by uv. It contains only installed packages and symlinks; no source code or irreplaceable data. You can safely delete it at any time and recreate it with uv sync, which will re-install all dependencies declared in pyproject.toml.

How do I know if PyTorch is using CUDA after reinstalling?

Run the validation script (uv run python main.py) from the repository. The basic_calculation function checks torch.cuda.is_available() and moves the resulting tensor to the "cuda" device if possible. If the printed tensor shows device='cuda:0', PyTorch is utilizing your NVIDIA GPU; if it shows device='cpu', it is running on the CPU.

When should I use uv cache clean versus uv cache prune?

Use uv cache prune to remove unused or orphaned cache entries globally without affecting packages that are currently installed in any project. Use uv cache clean <package> (e.g., uv cache clean torch) when you suspect a specific package’s wheel is corrupted and you want to force a fresh download of only that artifact during the next sync.

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