uv sometimes can't handle PyTorch nightly's cuda dependencies correctly.

[YouJiacheng]

my pyproject.toml

[project]
name = "modded-nanogpt"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = ">=3.12"
dependencies = [
    "numpy>=2.1.3",
    "torch==2.6.0.dev20241222",
    "pytorch-triton>=3.2.0",
    "huggingface-hub>=0.26.2",
    "tqdm>=4.67.0",
]

[tool.uv]
environments = [
    "platform_system == 'Linux'",
]

[tool.uv.sources]
torch = [
    { index = "pytorch-nightly-cu126"},
]
pytorch-triton = [
    { index = "pytorch-nightly-cu126"},
]

[[tool.uv.index]]
name = "pytorch-nightly-cu126"
url = "https://download.pytorch.org/whl/nightly/cu126"
explicit = true

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

uv tree output after uv sync

Resolved 22 packages in 7ms
modded-nanogpt v0.1.0
├── huggingface-hub v0.27.0
│   ├── filelock v3.16.1
│   ├── fsspec v2024.10.0
│   ├── packaging v24.2
│   ├── pyyaml v6.0.2
│   ├── requests v2.32.3
│   │   ├── certifi v2024.12.14
│   │   ├── charset-normalizer v3.4.0
│   │   ├── idna v3.10
│   │   └── urllib3 v2.2.3
│   ├── tqdm v4.67.1
│   └── typing-extensions v4.12.2
├── numpy v2.2.0
├── pytorch-triton v3.2.0+git35c6c7c6
├── torch v2.6.0.dev20241222+cu126
│   ├── filelock v3.16.1
│   ├── fsspec v2024.10.0
│   ├── jinja2 v3.1.4
│   │   └── markupsafe v3.0.2
│   ├── networkx v3.4.2
│   ├── setuptools v75.6.0
│   ├── sympy v1.13.1
│   │   └── mpmath v1.3.0
│   └── typing-extensions v4.12.2
└── tqdm v4.67.1

In a new directory I run uv init uv venv then uv pip install torch --index-url https://download.pytorch.org/whl/nightly/cu126.
It can correctly install nvidia-*.

Resolved 24 packages in 2.83s
Installed 24 packages in 127ms
 + filelock==3.16.1
 + fsspec==2024.10.0
 + jinja2==3.1.4
 + markupsafe==2.1.5
 + mpmath==1.3.0
 + networkx==3.4.2
 + nvidia-cublas-cu12==12.6.4.1
 + nvidia-cuda-cupti-cu12==12.6.80
 + nvidia-cuda-nvrtc-cu12==12.6.77
 + nvidia-cuda-runtime-cu12==12.6.77
 + nvidia-cudnn-cu12==9.5.1.17
 + nvidia-cufft-cu12==11.3.0.4
 + nvidia-curand-cu12==10.3.7.77
 + nvidia-cusolver-cu12==11.7.1.2
 + nvidia-cusparse-cu12==12.5.4.2
 + nvidia-cusparselt-cu12==0.6.3
 + nvidia-nccl-cu12==2.21.5
 + nvidia-nvjitlink-cu12==12.6.85
 + nvidia-nvtx-cu12==12.6.77
 + pytorch-triton==3.2.0+git0d4682f0
 + setuptools==70.2.0
 + sympy==1.13.1
 + torch==2.6.0.dev20241222+cu126
 + typing-extensions==4.12.2

and uv pip tree gives

torch v2.6.0.dev20241222+cu126
├── filelock v3.16.1
├── fsspec v2024.10.0
├── jinja2 v3.1.4
│   └── markupsafe v2.1.5
├── networkx v3.4.2
├── nvidia-cublas-cu12 v12.6.4.1
├── nvidia-cuda-cupti-cu12 v12.6.80
├── nvidia-cuda-nvrtc-cu12 v12.6.77
├── nvidia-cuda-runtime-cu12 v12.6.77
├── nvidia-cudnn-cu12 v9.5.1.17
│   └── nvidia-cublas-cu12 v12.6.4.1
├── nvidia-cufft-cu12 v11.3.0.4
│   └── nvidia-nvjitlink-cu12 v12.6.85
├── nvidia-curand-cu12 v10.3.7.77
├── nvidia-cusolver-cu12 v11.7.1.2
│   ├── nvidia-cublas-cu12 v12.6.4.1
│   ├── nvidia-cusparse-cu12 v12.5.4.2
│   │   └── nvidia-nvjitlink-cu12 v12.6.85
│   └── nvidia-nvjitlink-cu12 v12.6.85
├── nvidia-cusparse-cu12 v12.5.4.2 (*)
├── nvidia-cusparselt-cu12 v0.6.3
├── nvidia-nccl-cu12 v2.21.5
├── nvidia-nvjitlink-cu12 v12.6.85
├── nvidia-nvtx-cu12 v12.6.77
├── pytorch-triton v3.2.0+git0d4682f0
├── setuptools v70.2.0
├── sympy v1.13.1
│   └── mpmath v1.3.0
└── typing-extensions v4.12.2
(*) Package tree already displayed

---

And torch==2.6.0.dev20241203+cu126 is okay:

Resolved 35 packages in 1.87s
Uninstalled 1 package in 139ms
Installed 14 packages in 132ms
 + nvidia-cublas-cu12==12.6.4.1
 + nvidia-cuda-cupti-cu12==12.6.80
 + nvidia-cuda-nvrtc-cu12==12.6.77
 + nvidia-cuda-runtime-cu12==12.6.77
 + nvidia-cudnn-cu12==9.5.1.17
 + nvidia-cufft-cu12==11.3.0.4
 + nvidia-curand-cu12==10.3.7.77
 + nvidia-cusolver-cu12==11.7.1.2
 + nvidia-cusparse-cu12==12.5.4.2
 + nvidia-cusparselt-cu12==0.6.3
 + nvidia-nccl-cu12==2.21.5
 + nvidia-nvjitlink-cu12==12.6.85
 + nvidia-nvtx-cu12==12.6.77
 - torch==2.6.0.dev20241222+cu126
 + torch==2.6.0.dev20241203+cu126

[YouJiacheng]

there are these requirements in METADATA