Repository navigation
Conversation
1 task
3 of 5 tasks
This was referenced Oct 3, 2026
…ault DSpark draft TARGET_VERIFY bucket now holds DSV4RawDSparkDraftMetadata until init_forward_metadata_in_graph materializes it. The two tests that read core_attn_metadata straight after init_forward_metadata_out_graph failed with AttributeError on the mi35x small lane. Run the block-window replay check on both the eager and raw paths, and pin the length-fold test to the eager path, the only one whose bucket keeps a core object across steps.
michaelzhang-ai
force-pushed
the
fix/amd-dsv4-bcg-test-raw-dspark-metadata
branch
from
October 4, 2026 05:40
a4b37c6 to
faec358
Compare
This was referenced Oct 7, 2026
This was referenced Oct 17, 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
#42014 turned on
SGLANG_HIP_DSPARK_DRAFT_RAW_METADATAby default on HIP. Since then, a DSpark draftTARGET_VERIFYbucket holdsDSV4RawDSparkDraftMetadatauntilinit_forward_metadata_in_graphmaterializes it. Two tests intest/registered/amd/test_dsv4_hip_bcg_metadata.pystill readcore_attn_metadatastraight afterinit_forward_metadata_out_graph, sostage-b-test-1-gpu-small-amd-mi35xnow fails on every PR:Example: #42061's mi35x job, where this file was 2/54 and the lane stopped there. #42014's own mi35x lane was cancelled, so CI did not catch the failure before merge.
Modifications
Test-only,
test/registered/amd/test_dsv4_hip_bcg_metadata.py:TestDsparkDraftBlockWindowHip: the block-window replay check now runs with the flag both off and on. On the raw path it asserts that the bucket holdsDSV4RawDSparkDraftMetadata, callsinit_forward_metadata_in_graphthe way the graph would, and checks the same window rows, lengths and positions. This also adds coverage for the new default.TestAiterSparseLengthFoldPerStep: patched to the eager path. That is the only path whose bucket keeps one core object across steps, which is the cache-reuse bug this test guards. The raw path builds a fresh core inside the graph on every forward.No backend change.
Checklist
stage-b-test-1-gpu-small-amd-mi35xshard 0 failstest_dsv4_hip_bcg_metadata.pywith the same'DSV4RawDSparkDraftMetadata' object has no attribute 'core_attn_metadata'(job)stage-a-test-1-gpu-small-amdpasses 9/9 on this branch after merging main (job)stage-b-test-1-gpu-small-amd-mi35xpassestest_dsv4_hip_bcg_metadata.pyon this branch (job, shard 0, 24/27). A normal PR run stops shard 0 at its first failure,test_fused_kda_conv_recurrent_verify.py(file 5 of 27; this file is 18), so this is acontinue_on_errordispatch of the stage on the PR head. The three files that still fail there —test_quark_mxfp4.py(fixed by [AMD] Recalibrate test_quark_mxfp4 GSM8K thresholds to observed MI355X noise #41980),test_fused_kda_conv_recurrent_verify.pyandtest_compact_attention_hip.py— are exactly the ones that fail on main's shard 0, which failstest_dsv4_hip_bcg_metadata.pyon top of them.CI States
Latest PR Test (Base): ❌ Run #37557039147
Latest PR Test (Extra): ❌ Run #37557038987
Latest PR Test (AMD ROCm 10): ❌ Run #37557039112