Import Verification for Torch, Torchvision, Torchaudio Simultaneously with uv: A Complete Guide

The baonguyen6742/uv-install-torch repository demonstrates a robust method to verify simultaneous imports of torch, torchvision, and torchaudio using uv's optional dependencies, custom indexes, and mutually exclusive extras.

Managing PyTorch's ecosystem—torch, torchvision, and torchaudio—often requires navigating complex CPU and CUDA wheel configurations. The uv-install-torch repository streamlines this process by leveraging uv's advanced dependency resolution to handle import verification for torch, torchvision, torchaudio simultaneously with uv while preventing conflicting installations.

Configuring Optional Dependencies for CPU and CUDA

The foundation of this setup resides in pyproject.toml at lines 24–28, where the three packages are declared as optional dependencies grouped into extras named cpu and cu124.

[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"]

• The cpu extra pins CPU-only wheels from the standard PyTorch index.
• The cu124 extra pins CUDA 12.4 wheels from the CUDA-specific index.
• Both extras use identical package versions (2.4.1, 0.19.1, 2.4.1), ensuring API compatibility across hardware targets.

Configuring Custom Package Indexes

To fetch the correct wheels, pyproject.toml defines custom indexes in the [tool.uv.sources] section (lines 37–50) and declares them in the [tool.uv] block (lines 53–62).

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

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"

[[tool.uv.index]]
name = "pytorch-cu124"
url = "https://download.pytorch.org/whl/cu124"

This configuration ensures that when you install .[cpu], uv pulls wheels exclusively from the pytorch-cpu index, while .[cu124] resolves against the pytorch-cu124 index.

Enforcing Mutually Exclusive Extras

To prevent ambiguous environments where both CPU and CUDA wheels might coexist, pyproject.toml at lines 30–33 declares the extras as mutually exclusive:

[tool.uv]
conflicts = [
  [{ extra = "cpu" }, { extra = "cu124" }],
]

Attempting to install both extras simultaneously—for example, uv pip install .[cpu,cu124]—triggers a resolution error. This safeguard ensures clean, deterministic environments for import verification.

Verifying Imports with main.py

The repository includes main.py to validate that all three packages import correctly and function as expected. The script imports torch, torchvision, and torchaudio, then prints version information, CUDA availability, and executes a tensor computation.

# main.py

import torch
import torchvision
import torchaudio

def main():
    print("torch version      :", torch.__version__)
    print("torchvision version:", torchvision.__version__)
    print("torchaudio version :", torchaudio.__version__)
    print("CUDA available?    :", torch.cuda.is_available())
    
    # Verify tensor operations work

    device = "cuda" if torch.cuda.is_available() else "cpu"
    x = torch.arange(1, 3).unsqueeze(0).to(device)
    y = torch.arange(1, 4).unsqueeze(0).to(device)
    result = torch.matmul(x.T, y)
    print("Test calculation:")
    print(result)

if __name__ == "__main__":
    main()

If any import fails, Python raises an ImportError immediately, providing instant feedback that the installation requires troubleshooting.

Installation and Verification Workflow

Execute the following commands to install the PyTorch stack and verify imports:

# Install CPU-only version

uv pip install .[cpu]

# OR install CUDA 12.4 version

uv pip install .[cu124]

# Run verification script using uv's managed environment

uv run python main.py

Successful execution produces output similar to:

torch version      : 2.4.1
torchvision version: 0.19.1
torchaudio version : 2.4.1
CUDA available?    : True
Test calculation:
tensor([[1, 2, 3],
        [2, 4, 6]])

Summary

• The baonguyen6742/uv-install-torch repository uses pyproject.toml optional dependencies to define cpu and cu124 extras at lines 24–28.
• Custom indexes in [tool.uv.sources] (lines 37–50) and [[tool.uv.index]] (lines 53–62) ensure uv fetches wheels from the correct PyTorch repository.
• Mutually exclusive extras declared at lines 30–33 prevent simultaneous installation of conflicting CPU and CUDA builds.
• The main.py script provides immediate import verification and functional testing through version printing and tensor operations.

Frequently Asked Questions

How do I verify that torch, torchvision, and torchaudio are correctly installed with uv?

Run the main.py script included in the repository using uv run python main.py. This script imports all three modules and prints their __version__ attributes along with torch.cuda.is_available(). If any package fails to import, Python raises an ImportError immediately.

Can I install both CPU and CUDA versions of PyTorch simultaneously using uv?

No. The pyproject.toml at lines 30–33 declares cpu and cu124 as mutually exclusive extras. Attempting to install both with uv pip install .[cpu,cu124] triggers a conflict error, ensuring you maintain a single, coherent environment.

What versions of PyTorch packages are pinned in the uv-install-torch repository?

According to lines 24–28 in pyproject.toml, the repository pins torch==2.4.1, torchvision==0.19.1, and torchaudio==2.4.1 for both CPU and CUDA 12.4 variants. These versions are synchronized to ensure API compatibility across hardware configurations.

How does uv know which PyTorch index to use for CPU vs CUDA wheels?

The [tool.uv.sources] section maps each package to specific indexes based on the active extra. When installing .[cpu], uv uses the pytorch-cpu index (defined at lines 53–62); when installing .[cu124], it uses the pytorch-cu124 index. This conditional routing ensures the correct wheel variants are resolved without manual index management.

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