GLM-5.3-Flash cookbook: restore --disable-shared-experts-fusion to every cell - #36535
Closed
JustinTong0323 wants to merge 1 commit into
Closed
JustinTong0323 wants to merge 1 commit into
JustinTong0323 wants to merge 1 commit into
Conversation
…ery cell Without it (shared-experts fusion on) the model degenerates on the final weights: empty content, verbatim loops, never stops — isolated on 4x GB300 with the fusion flag as the only difference. Every measured serve behind the published numbers ran with the flag; the minimal-flags trim dropped it by mistake.
JustinTong0323
requested review from
sogalin,
wisclmy0611 and
zijiexia
as code owners
August 26, 2026 16:25
Collaborator
Author
|
Folded into #36519 per maintainer preference. |
github-actions
Bot
added
the
documentation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
The published cells are missing
--disable-shared-experts-fusion, and with shared-experts fusion ON the model degenerates on the final GLM-5.3-Flash weights: empty content, verbatim repetition loops, and no stop (runaway to the token cap). The flag was present in every measured serve behind the published numbers but was dropped by the minimal-flags trim.Isolation on 4x GB300 with the final weights (c5b82b63e37b): identical serves differing only in this flag — fusion ON degenerates (empty content + loops + finish_reason=length), fusion OFF answers correctly and stops. The failure is silent: no crash, no traceback.
This PR also disables the HiCache options for the Low Latency strategy: HiCache + MTP speculative decoding crashes at startup in the current build (the DSA draft pool lacks
full_kv_pool— a code fix is being validated separately). High Throughput keeps the HiCache options.Modifications
--disable-shared-experts-fusionto all twelve cells.Tests
node docs/scripts/check_cookbook_configs.mjs— OK.CI States
Latest PR Test (Base): ⏳ Run #32989879891
Latest PR Test (Extra): ⏳ Run #32989877568
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.