[ATOMesh][CI][benchmark] add GLM-5.2 CPP4+DCP4 ATOMesh coverage - #2134
Open
MengqingCao wants to merge 3 commits into
Open
[ATOMesh][CI][benchmark] add GLM-5.2 CPP4+DCP4 ATOMesh coverage#2134MengqingCao wants to merge 3 commits into
MengqingCao wants to merge 3 commits into
Conversation
Wire single-node PD CPP4/DCP4 jobs, isolate prefill/decode env so PP partitions do not leak across roles, and divide per-GPU throughput by total cluster GPUs. Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
🏷️ CI GuideRuns automatically on every eligible PR before approval:
Heavy model tests:
|
Give glm-52-mxfp4-1p1d-tp4-agentic-1m c1/c2/c4/c8 an MTP draft of 4 tokens at a forced 0.5825 acceptance rate, so the small-concurrency agentic points are measured on the same speculative path the cpp4-dcp4 cases already use. The DPA lmcache cases inherit this group's anchor without overriding `server`, so the flags live in a named copy of the shared common_args rather than in the anchor itself; the DPA group keeps running without speculation. Forced acceptance replaces every generated token's text with a placeholder, which leaves SWE-bench Lite unscoreable, so c1 drops its accuracy gate and reverts to the default runner time limit. Drop the tp4 c48 case.
The prefill role already offloads to CPU DRAM, but the L3 tier that PR #1851 wired up was never enabled in the matrix, so a working set larger than the 256GiB hot cache simply fell out of the cache. LMCACHE_LOCAL_DISK and LMCACHE_MAX_LOCAL_DISK_SIZE are required together, and the size is per rank because scale_cpu_size_for_pp splits only the CPU budget across PP stages: 4 x 500GiB on a pp4 prefill. The tier lives on a host bind mount, so it outlives `docker run --rm` and would otherwise serve one concurrency's KV to the next while holding its disk. The launcher now empties the directory before the first prefill worker starts and removes it on exit.
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
2 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.
Motivation
rely on #2121
Add ATOMesh CI coverage for GLM-5.2 MXFP4 on a single-node 1P1D CPP4+DCP4 topology (prefill PP4×TP1 + decode TP4+DCP4, with LMCache offload), and fix the supporting launch/result-processing bugs that blocked reliable runs and dashboard metrics.
Main Changes
workflow_dispatchand add extra single-node node orderings so scheduling can pick any eligible MI355X host.Submission Checklist