Repository navigation
fix(rccl): restore the gfx1250 LSA flat AddressFree skip - #14997
Merged
dlamd1dai merged 2 commits intoNov 6, 2026
Merged
Conversation
The NCCL 2.31.2 sync dropped the ncclDevrFinalize guard, so gfx1250 teardown can hang in cuMemAddressFree and SkipCuMemFreePolicy.IsolatedArchAndEnvBranches fails. Co-authored-by: Cursor <cursoragent@cursor.com>
1 of 4 tasks
✅ All Policy Checks Passed
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
![amd-argus[bot]](https://avatars.githubusercontent.com/in/4740020?s=60&v=4){kind=small}
The ROCM-30633 rationale already lives next to rcclSkipLsaFlatAddressFree. Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
pr-rocm-systems-14997-oci-rccl-pr-buildBuilds
FailuresNo failing tests in this run. 189 tests skipped. pr-rocm-systems-14997-oci-rccl-pr-smokeBuilds
Failures
15 tests skipped. Performance regressions
|
Contributor
pr-rocm-systems-14997-ruby64-rccl-pr-smokeBuilds
FailuresNo failing tests in this run. 15 tests skipped. Performance regressions
|
Contributor
pr-rocm-systems-14997-oci-rccl-pr-buildBuilds
FailuresNo failing tests in this run. 189 tests skipped. pr-rocm-systems-14997-oci-rccl-pr-smokeBuilds
Failures
15 tests skipped. Performance regressions
|
Contributor
pr-rocm-systems-14997-oci-rccl-pr-smokeBuilds
FailuresNo failing tests in this run. 15 tests skipped. Performance regressions
|
Contributor
pr-rocm-systems-14997-ruby64-rccl-pr-smokeBuilds
FailuresNo failing tests in this run. 15 tests skipped. Performance regressions
|
Kapil-Shyam-Pawar
approved these changes
Nov 5, 2026
abouteiller
approved these changes
Nov 5, 2026
abouteiller
left a comment
abouteiller
left a comment
Contributor
There was a problem hiding this comment.
Fixes a bug
Envar discussion does not prevent merging.
dlamd1dai
deleted the
users/dondai/aicomrccl-3009-lsa-flat-address-free
branch
November 6, 2026 07:18
Contributor
TheRock Submodule Bump ActivityNewest first
|
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
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.
Motivation
ncclDevrFinalizealways frees the LSA flat virtual-address reservation. On gfx1250 thatcuMemAddressFreecan hang or double-free while peer maps are still live. The NCCL 2.31.2 sync dropped the guard, soSkipCuMemFreePolicy.IsolatedArchAndEnvBranchesfails on the RCCL unit nightly.Technical Details
Restore the gfx1250-only skip in
ncclDevrFinalize. WhenrcclSkipLsaFlatAddressFree()is true (NCCL_CUMEM_SKIP_FREE=1, or gfx1250 with the variable unset), finalize clearslsaFlatBaseand leaves the reservation for process exit. gfx950 still frees it.NCCL_CUMEM_SKIP_FREE=0forces the free on every architecture.The host unit test and
rcclSkipLsaFlatAddressFree()were already ondevelop; only the call site was missing.Issue Tracking
JIRA ID: AICOMRCCL-3009
Test Plan
rccl-UnitTestsDevRuntimeand run--gtest_filter=SkipCuMemFreePolicy.IsolatedArchAndEnvBranches.ForceOn_SkipsLsaFlatandGfx1250_SkipsLsaFlatreportrcclTestHipMemAddressFreeCount == 0.ForceOff_FreesLsaFlat,Gfx950_PeerSkipKeepsLsaFlatFree, andGfx900_SkipFreeOffreport1.ncclCommInitAll, symmetric window register,ncclDevCommCreate,ncclDevCommDestroy,ncclCommDestroy) with the skip on and withNCCL_CUMEM_SKIP_FREE=0.Test Result
All runs below are at 618a5bf on a 4x MI450 (gfx1250) node, ROCm 10.1 (HIP 7.17), built with
GPU_TARGETS=gfx1250.Unit test.
SkipCuMemFreePolicy.IsolatedArchAndEnvBranchespasses, with all five isolated cases passing. The wholerccl-UnitTestsDevRuntimebinary passes (9 tests). The test fakes the architecture withRCCL_TEST_GCN_ARCH, so the result does not depend on the GPU it runs on.Device-API comm create/destroy loop, 2 GPUs, 10 iterations per run:
ncclDevrFinalizeskip logNCCL_CUMEM_SKIP_FREE=0(forced free)With the default setting, every finalize logs
skipping lsaFlatBase hipMemAddressFreeand teardown completes. The forced-free run did not hang or double-free in 10 cycles on this node, so the original ROCM-30633 failure did not reproduce here.The 4-GPU loop was stopped during its first init and has no result. The first init in each process takes about 490 s on this node because HIP translates the gfx1250 code object for the node's A0 silicon. For the same reason, the in-tree
DeviceApi.LsaRemoteReadtimes out at its 60 s limit duringncclCommInitAll, before any device-API code runs. That is a property of this node, not of the change.No local results yet on MI300 (gfx942) or MI350 (gfx950).
Submission Checklist
Made with Cursor