Install a CUDA wheel that exists on a pre-12.4 driver - #650
Merged
Merged
Conversation
An RTX 4060 on Windows reported the NVIDIA build falling back to CPU with "No matching distribution found for torch==2.6.0+cu121" (#644). The tag comes from the driver and the torch version from a table, and nothing held the two together. cuda_tag sent a driver reporting CUDA 12.0-12.3 to cu121, torch_version_for_tag answers 2.6.0 for every non-cu128 tag, and the cu121 index stops at 2.5.1 -- for torch, torchaudio and torchvision alike, on every platform. So setup asked pip for a wheel that has never been published, the install failed, CPU torch was restored, and a working GPU sat unused. Every CUDA 12 driver now gets cu124. CUDA 12 is minor-version compatible, so a cu124 build runs on a 12.0 driver, and cu124 publishes the same 2.6.0 line the rest of the app is pinned to. cu121 is no longer reachable at all. cu118 follows it as a second candidate. wheel_candidates has returned a list since #502 precisely so a tag that fails falls through to the next, but every non-Blackwell branch returned one element, so the fallthrough had nowhere to go: cu124 failing meant CPU. cu118 runs on every 12.x driver and publishes the same torch line, which makes it a real second chance. The new test is the guard the original bug lacked: it walks every driver and compute-capability combination setup can see and asserts each tag offered appears in a table of tag/version pairs confirmed against download.pytorch.org. A tag whose index does not publish what setup installs now fails locally instead of on a user's machine. Re-running it with the cu121 branch restored reproduces #644 as a test failure. Verified: cargo fmt, clippy --tests -D warnings and test on Windows (99) and on the Linux target via WSL (109), plus pytest (1060 passing), ruff, bandit and the JS unit tests. Wheel availability checked live against download.pytorch.org/whl/<tag>/ for torch, torchaudio and torchvision, cp312, win_amd64 and linux_x86_64. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #644
Closes #649
An RTX 4060 on Windows reported the NVIDIA build falling back to CPU with
No matching distribution found for torch==2.6.0+cu121.What was wrong
The wheel tag comes from the driver and the torch version comes from a table, and nothing held the two together.
cuda_tagsent a driver reporting CUDA 12.0 to 12.3 tocu121,torch_version_for_taganswers 2.6.0 for every non-cu128 tag, and the cu121 index stops at 2.5.1. That is true for torch, torchaudio and torchvision alike, on every platform. So setup asked pip for a wheel that has never been published, the install failed, CPU torch was restored, and a working GPU went unused.What changed
Every CUDA 12 driver now gets
cu124. CUDA 12 is minor-version compatible, so a cu124 build runs on a 12.0 driver, and cu124 publishes the same 2.6.0 line the rest of the app is pinned to. cu121 is no longer reachable.cu118follows it as a second candidate, which is #649.wheel_candidateshas returned a list since #502 so that a tag which fails falls through to the next, but every non-Blackwell branch returned one element, so the fallthrough had nowhere to go and cu124 failing meant CPU. cu118 runs on every 12.x driver and publishes the same torch line, so the second attempt is now a real one.The guard that was missing
The new test walks every driver and compute capability combination setup can see, and asserts each tag it offers appears in a table of tag and version pairs confirmed against download.pytorch.org. A tag whose index does not publish what setup installs fails here instead of on someone's machine. Restoring the cu121 branch reproduces this bug as a test failure that names the tag.
Verified
cargo fmt,clippy --tests -D warningsandcargo teston Windows (99 passing) and on the Linux target via WSL (109 passing)uv.lockuntoucheddownload.pytorch.org/whl/<tag>/for torch, torchaudio and torchvision, cp312, win_amd64 and linux_x86_64