[Config] [Perf] Retune GLM5 MXFP4 FMoE for gfx950

[XiaobingSuper]

Summary

• retune all 64 GLM5 A4W4 FMoE shapes on gfx950 with the merged MXFP4 FlyDSL path
• update 21 rows that remain at least 2% faster under production Graph replay
• preserve the other 43 incumbent rows byte-for-byte

Tuning and selection

• 8x MI355X (gfx950, 256 CU/GPU), one shape per worker/GPU
• Mxfp4FlydslTuner: 5 warmups and 101 GPU-event iterations per coupled stage1/stage2 candidate
• production incumbent-vs-candidate replay with a 2% update threshold
• three balanced-routing HIP Graph rounds over all 64 shapes (300 MoE calls per graph)
• low-M random/hotspot replay; the route-regressing M=64, I=2048 candidate was rejected

Performance

• accepted-row geomean speedup: 1.077x
• projected geomean across all 64 shapes: 1.025x
• final all-shape Graph sanity replay: 1.016x
• the M=25 decode bucket (M=32, I=256/512) is unchanged

inter_dim	Updated rows	Updated-row geomean	All-16-shape projected geomean
256	6	1.060x	1.022x
512	4	1.054x	1.013x
1024	3	1.112x	1.020x
2048	8	1.088x	1.043x

Accepted rows

M	inter_dim	Incumbent us	Tuned us	Speedup
2	256	35.266	31.590	1.116x
256	256	138.201	132.211	1.045x
1024	256	181.332	173.039	1.048x
2048	256	277.561	263.216	1.054x
4096	256	374.033	361.133	1.036x
32768	256	2292.030	2157.160	1.063x
256	512	235.802	222.697	1.059x
512	512	249.477	243.600	1.024x
1024	512	297.111	282.400	1.052x
4096	512	580.359	536.329	1.082x
4	1024	88.590	74.830	1.184x
1024	1024	497.433	469.473	1.060x
32768	1024	5254.530	4791.180	1.097x
1	2048	47.167	45.819	1.029x
2	2048	96.403	76.191	1.265x
16	2048	476.406	447.295	1.065x
128	2048	829.480	804.122	1.032x
2048	2048	1188.490	1101.310	1.079x
4096	2048	1730.600	1646.990	1.051x
8192	2048	2760.690	2608.290	1.058x
32768	2048	9878.620	8630.950	1.145x

Validation

• all 64 tuner shapes completed and passed the configured cosine-error gate
• no NaN or runtime failure in final production Graph replay
• pytest -q op_tests/tuning_tests/test_csv_validation.py op_tests/tuning_tests/test_config_shape_collision.py: 37 passed, 45 subtests passed
• unique config keys and git diff --check

Only aiter/configs/model_configs/glm5_fp4_tuned_fmoe.csv is changed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
multigpu	Aiter multi-GPU tests on the 8-GPU runner
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 5902 --add-label <label>

One backend per PR:
A PR changes one kernel backend: [Triton/Gluon] (Triton and Gluon count as one), [HIP], [ASM], [CK], [OPUS] or [FlyDSL]. If the title ends up with two backend tags, split the PR -- as stacked pull requests when one part cannot merge without the other.

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to stop the title rewrites; labels stay in sync either way.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.