Conversation
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
This was referenced Jul 16, 2026
This was referenced Jul 16, 2026
tanth47
force-pushed
the
codex/glm52-dspark-rocm-dsa-stability-clean
branch
from
July 25, 2026 11:02
732323b to
5108141
Compare
tanth47
changed the title
tanth47
force-pushed
the
codex/glm52-dspark-rocm-dsa-stability-clean
branch
from
July 25, 2026 11:13
5108141 to
5229d7d
Compare
tanguofu
pushed a commit
to tanguofu/sglang
that referenced
this pull request
Jul 31, 2026
…on HIP PR sgl-project#31260 routes HIP target-verify through _get_topk_ragged, but the assert at line 1147 still required extend_seq_lens_cpu != None. target-verify batches don't populate extend_seq_lens_cpu, so the assert fired on every DSpark target-verify forward. Relax the assert to only require seq_lens_cpu (which target-verify does populate); the downstream code uses indexer_seq_lens_cpu from metadata, not extend_seq_lens_cpu. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stacked on #31047.
What this adds
DSparkDraftModelentry point.sgl-kernelwheel does not provide those ops.The DSA changes are needed because one request contributes multiple verify tokens. Its Q/K ranges must advance through the extended KV region rather than reuse prefix-only offsets.
Scope
This PR is the ROCm functionality/correctness slice. CUDA graph coverage, SPS/STS, overlap scheduling, and performance tuning stay out of scope.
MI350 validation
End-to-end TP4 validation head:
5108141f35229d7d0d:27 passed, plus22kernel-parity subtests.256/256requests, OSL 128 withignore_eos, AR59.62%, AL5.174.16/16requests attemperature=0.8,top_p=0.95, all returned 128 tokens.16/16requests, all returned 128 tokens.Target:
zai-org/GLM-5.2-FP8(ba978f7d...)Draft:
RedHatAI/GLM-5.2-speculator.dspark(a278cc09...)Image:
lmsysorg/sglang:v0.5.15.post1-rocm720-mi35xThe current checkpoint is distribution-sensitive, so this is not a workload-independent speedup claim. The goal here is to make the runtime contract correct and let future GLM speculators plug into the same path.
CI States
Latest PR Test (Base): ❌ Run #30155865578
Latest PR Test (Extra): ❌ Run #30155865593