Repository navigation
Per-thread grad mode and device-mode stack (#76, part 1) - #77
Merged
Merged
Conversation
…am's are (#76) Both were process-wide: torch.no_grad() in one thread stopped grad_fn recording in another, and torch.device('meta') on one thread moved another thread's factories to meta. Wrong with the GIL on, not only on 3.15t. tensor.rs GRAD_ENABLED is a thread_local Cell; the bootstrap's grad flag and _MODE_STACK sit on threading.local. tests/api/test_threadstate.py: 3 tests, green on upstream torch, red on the shim before, green after.
… out from under a live draw (#76) Streams were Box::leak'ed and reclaimed by Box::from_raw in stream_free, sound only while a kernel's guard and the generator's __del__ could not overlap under the GIL. stream() now returns a GenGuard that holds its own Arc clone; stream_free drops the table's reference. The new cargo test frees a locked stream and reuses its memory: SIGSEGV on the old code, green now.
Member
Author
|
Added: RNG streams are
|
Contributor
Gate on CI (issue #24)Full gate: 124 suites = 116 run on at least one CI runner + 8 local-only. The local gate (
Local-only: verified by no CI runner (8)
Suites that verified nothing, or failed, on a runner
Every suite, every runner
A SKIP is unverified on that runner, never passed. UNVERIFIED means the suite ran and every one of its tests skipped by name. |
…; test_threadstate sets TORCH_USE_RTLD_GLOBAL CAPTURING was cleared by whichever thread ended its capture first, so a thread still mid-capture stopped recording and its later ops vanished from its trace. It is now an AtomicUsize of open regions (one per thread at most, since capture_begin refuses nesting); the door still pays one relaxed load. test_threadstate.py never ran in the gate before this commit: it died at import torch for want of TORCH_USE_RTLD_GLOBAL, which every other suite that imports torch sets. Now 4 ok in the gate (2010 -> 2014).
Member
Author
|
Correction. My two earlier gate summaries on this PR were wrong. They said 'FAIL=2, both build.py self-tests'. In fact Fixed in the latest commit, which also adds the capture-flag fix (#76 item 3):
|
This was referenced Oct 5, 2026
Member
Author
|
CI evidence (run 37316717277, 992c69c): linux gate |
Member
Author
This was referenced Oct 16, 2026
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.
Refs #76 (the issue stays open: the RNG stream free race and the CAPTURING flag remain).
Defect fixed
thread_localintensor.rsand athreading.localin the bootstrap, as upstream'sGradModedoes.torch.device(...)) was process-wide. It is now a per-thread stack behind the same list-shaped API.Test added
tests/api/test_threadstate.py, 3 tests:Gate (worktree, local, 2026-10-05)
build.pyself-test suites that need the release machine's target CPythons under.caches/, which a worktree lacks. This branch does not touchbuild.py.