Repository navigation
[Docs][AMD] Update GLM-5.2 MI355X MXFP4 MTP recipe to 20261008 daily - #43718
Closed
chunfangamd wants to merge 1 commit into
Closed
chunfangamd wants to merge 1 commit into
chunfangamd wants to merge 1 commit into
Conversation
Bump dockerImages["mi355x|mxfp4"] to lmsysorg/sglang-rocm:v0.5.21-rocm720-mi35x-20261008 and align the MI355X MXFP4 five-step MTP cells with the InferenceX AgentX recipe validated in SemiAnalysisAI/InferenceX#3851: 32K prefill chunks and a 0.85 static memory fraction on both cells, and a 180 GB/rank HiCache host pool (--hicache-size 180) instead of --hicache-ratio 1.0 on the TP4/EP4 cell.
chunfangamd
requested review from
JustinTong0323,
sogalin,
wisclmy0611 and
zijiexia
as code owners
October 11, 2026 17:57
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Keep the GLM-5.2 MI355X MXFP4 five-step MTP cookbook commands in sync with the InferenceX AgentX recipe and its latest validated SGLang ROCm daily image, SemiAnalysisAI/InferenceX#3851. The command builder still pinned
…-20260923and the page text…-20260930, and the two MTP cells did not carry the recipe's prefill chunk size, static memory fraction or HiCache host pool size.Modifications
dockerImages["mi355x|mxfp4"]:lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260923→lmsysorg/sglang-rocm:v0.5.21-rocm720-mi35x-20261008. As in [AMD] GLM-5.2 MI355X MXFP4: bump image to 20260923 daily #41109, this pin is shared by every MI355X MXFP4 cell.--chunked-prefill-size 131072→32768,--mem-fraction-static 0.80→0.85.--chunked-prefill-size 32768, and replace--hicache-ratio 1.0with--hicache-size 180(a 180 GB host KV pool per rank).write_throughstays the default write policy.GLM-5.2.mdx: update the AMD MTP note to the new image and these settings. Add a short source-reference comment in the config.mtp-314cells, other precisions and other hardware are otherwise unchanged.Accuracy Tests
Documentation only; no code changes. This image and flag set were validated by InferenceX run 38067578883: all nine points (TP8/EP1 conc 1/2/4 and TP4/EP4 HiCache conc 4–16) passed, and GSM8K at TP4/EP4 conc 16 scored 0.9689 strict / 0.9719 flexible.
Speed Tests and Profiling
No new measurements. The values come from the InferenceX recipe and the run above.
Validation
glm-5.2.jsxwith Node.js and compared it withupstream/main. Both versions have 52 cells. Only the two MI355X MXFP4 MTP cells change (the flags listed above), plusdockerImages["mi355x|mxfp4"]; every other field is identical.git -c core.whitespace=cr-at-eol diff --checkpasses.Checklist
AI model disclosure
Prepared with Claude Opus 5.5 via the Cursor agent at the author's request. No delegated models.
CI States
Latest PR Test (Base): ✅ Run #38161822686
Latest PR Test (Extra): ❌ Run #38161822244
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.