Repository navigation
refactor(tests): reuse compilation caches and defer torch imports - #246
Merged
Merged
Conversation
Local compilation previously found /usr/include/cuda/std from the system CUDA 11.5 apt package: libcu++ 1.6.0 with _LIBCUDACXX_CUDA_API_VERSION=001006000. Those headers did not match the CUTLASS and CUDA 13.3 CCCL versions, while CI had no system fallback and failed to find cuda/std/utility. Add the selected toolkit's include/cccl directory ahead of the system include search so local and CI compilation use CUDA 13.3's shipped CCCL.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why
What
Reuse ccache and torch extension caches locally and in CI, with pull request cache isolation and per-job cache statistics.
Compare MMA layouts against host-compiled CuTe references in two tests, replacing handwritten references and thousands of parametrized cases. Use the selected CUDA toolkit's CCCL headers consistently on local and CI.
Import torch through lazy proxies and defer dtype resolution until use.
Configure local artifact storage and a PM2 cleanup task outside the repository.
Local full verification at
1da1c0c: source 2141 passed / 12 skipped (266s cold, 188s warm versus 335s baseline); installed 161 passed (315s versus 410s). Warm ccache: 118 direct hits, zero misses.Header-path follow-up:
tests/ir/cudapassed (2 tests); header tracing and a compile/link/run probe passed in the CI image.PR CI passed: source 2141 passed / 12 skipped (533s step), installed 161 passed (330s step), and published-checkpoint decoding passed. Read-only cache: 120 misses, no shared cache result or torch extension file changes.
Contract
Risk
tilefoundry-ci-runner:pre-ccache.