Skip to content

Publish cuda-pathfinder to the torch CUDA arch indexes - #8515

Merged
atalman merged 1 commit into
pytorch:mainfrom
atalman:atalman/pathfinder-torch-targets
Aug 12, 2026
Merged

Publish cuda-pathfinder to the torch CUDA arch indexes#8515
atalman merged 1 commit into
pytorch:mainfrom
atalman:atalman/pathfinder-torch-targets

Conversation

@atalman

@atalman atalman commented Aug 12, 2026

Copy link
Copy Markdown
Contributor

https://download.pytorch.org/whl/test/cu134/cuda-pathfinder/ returns 403 while the cu134 index advertises it:

<a href="cuda-pathfinder/">cuda-pathfinder</a>   ->  403

Why it is needed

torch declares cuda-bindings in PYTORCH_EXTRA_INSTALL_REQUIREMENTS, and cuda-bindings in turn requires:

cuda-pathfinder>=1.4.2

Release smoke tests install from a single arch index with no PyPI fallback (Looking in indexes: https://download.pytorch.org/whl/test/cu134), so a transitive dependency of a declared dependency has to be resolvable in that same index.

Why it is missing

cuda-pathfinder was registered under the vllm project only:

"cuda-python": [{"project": "vllm"}],
"cuda-pathfinder": [{"project": "vllm"}],

so update_dependencies.py never writes whl/{test,nightly}/cu*/cuda-pathfinder/. Confirmed against a live run (job) — it uploaded the identical 18 packages to cu126, cu130, cu132 and cu134, with neither cuda-pathfinder nor cuda-python among them:

cu134   cuda-bindings cuda-toolkit nvidia-cublas nvidia-cuda-cccl nvidia-cuda-cupti
        nvidia-cuda-nvrtc nvidia-cuda-runtime nvidia-cudnn-cu13 nvidia-cufft ...

The older arches still serve cuda-pathfinder, but that is residue from before it stopped being produced — nothing regenerates it today, which is why cu134, the newest arch, is the one that broke.

The change

Give cuda-pathfinder the same torch targets as cuda-bindings (cu126, cu128, cu129, cu130, cu132, cu134), keeping the existing vllm entry. It exists to satisfy cuda-bindings, so it should follow it exactly.

cuda-bindings      before=[cu126, cu128, cu129, cu130, cu132, cu134, vllm]
                   after =[cu126, cu128, cu129, cu130, cu132, cu134, vllm]
cuda-pathfinder    before=[vllm]
                   after =[cu126, cu128, cu129, cu130, cu132, cu134, vllm]   <-- changed
cuda-python        before=[vllm]
                   after =[vllm]

cuda-python is deliberately left alone. torch never depends on it — it is vllm-only — so its absence from the torch arch indexes is correct, and its 403 on cu134 is harmless for torch smoke tests. Worth noting the cu134 listing still advertises it, so that dangling link remains; it just is not on the path to anything torch installs.

torch declares cuda-bindings, and cuda-bindings requires
cuda-pathfinder>=1.4.2. Release smoke tests install from a single arch index
with no PyPI fallback, so that transitive dependency has to be resolvable
there too.

cuda-pathfinder was registered under the vllm project only, so
update_dependencies.py never wrote whl/{test,nightly}/cu*/cuda-pathfinder/.
whl/test/cu134 advertises it in the PEP 503 listing while the target 403s. The
older arches still serve it, but that is residue from before it stopped being
produced -- nothing regenerates it today, so cu134 is simply the first arch
created after that.

Give it the same torch targets as cuda-bindings, since it exists to satisfy
cuda-bindings. cuda-python is left alone: torch never depends on it, so its
absence from the arch indexes is correct.
@vercel

vercel Bot commented Aug 12, 2026

Copy link
Copy Markdown

@atalman is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Aug 12, 2026
@atalman
atalman merged commit 80e5be8 into pytorch:main Aug 12, 2026
4 of 5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants