Conversation
…sgl-project#30838) Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: BBuf <xiaoyu.zhang@radixark.ai> Co-authored-by: jessiewei7 <jessiewei747@gmail.com> Co-authored-by: root <root@GPUC5A6.maas>
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
…-project#31354) Co-authored-by: Elizaveta Martirosian <you@example.com> Co-authored-by: Elizaveta Martirosian <elizaveta.martirosian@gmail.com> Co-authored-by: ronnie_zheng <zl19940307@163.com>
Signed-off-by: Devashish Lal <devcode@fb.com> Co-authored-by: Devashish Lal <devcode@fb.com>
…gl-project#31619) Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… ahead (spec-v2 overlap) (sgl-project#31468) Co-authored-by: Hao Phan <htphan@nvidia.com>
…ect#25213) Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
…gl-project#24013) Co-authored-by: yhyang201 <yhyang201@gmail.com> Co-authored-by: Mick <mickjagger19@icloud.com>
sgl-project#29630) (sgl-project#31582) Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…cache mtp scenarios. (sgl-project#31659)
…epSeek-V3.2) (sgl-project#30514) Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
…gl-project#31346) Co-authored-by: mosya415 <263250241+mosya415@users.noreply.github.com>
github-actions
Bot
added
documentation
Owner
Author
|
Superseded by sgl-project#32374. This PR was opened against a stale fork base and includes unrelated upstream history, so please use the clean PR for review. |
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
20 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Scope\n\n- keep scheduler-visible sequence-length planning on host-reserved bounds while exact accepted lengths stay device-resident\n- make the confidence relay stream-safe and nonblocking\n- skip confidence D2D/D2H publication for plain verify-all when no policy or telemetry consumes it\n- remove the per-step pure-TP Gloo graph-tier broadcast; pure TP now uses a deterministic full-width tier while exact non-uniform verify lengths still broadcast on the GPU\n- preserve DP-attention CPU coordination as the conservative fallback\n\nGLM-5.2 runtime validation is stacked with sgl-project#31047; this branch does not duplicate that model-config change.\n\n## Evidence\n\nLatest-main focused ROCm tests:\n\n- 97 passed, 20 subtests passed\n- planner policy, evidence-mode, FutureMap host-sync, stream relay, TP/DP tier contracts\n\nExact GLM-5.2 MI350 TP4 verify-all runtime gate:\n\n- UltraChat natural ~2K ISL, seed 42, C16, 256 prompts x exact 512 OSL\n- 256/256 completed, no request errors\n- AR 49.90%, AL 4.493\n- target verify graph 2020/2020\n- folded accept 2020/2020\n- folded commit 2020/2020\n- one target forward 2020/2020\n\nMeasured relay-bypass signal on the exact validation stack:\n\n- C1 mean 212.07 output tok/s vs overlap-off 192.88: +9.95%\n- C16 891.04 output tok/s\n\n## Pending\n\nThe pure-TP dynamic/non-uniform runtime gate is still pending. GPU4-7 became occupied by an unrelated DSV4 validation container during the relaunch. Unit/static contracts are green; this PR stays draft until the exact dynamic gate passes.\n\nProfiler note: the earlier GPU-bubble percentage is intentionally excluded because ROCm Kineto did not materialize graph child kernels consistently across ranks/modes.
CI States
Latest PR Test (Base): ❌ Run #30134321551
Latest PR Test (Extra): ❌ Run #30134321420