[Config] configs: GLM-5.3 fused shared MoE rows for gfx950

[Raiden-Makoto]

Why

GLM-5.3 shared-expert fusion changes the routed fused-MoE dispatch from E=288, topk=8 to E=289, topk=9. The ninth expert is always selected with weight 1.0, so one AITER launch performs the routed and shared gate/up GEMM, SiLU(gate) * up, down GEMM, and routed-weight accumulation.

The two TP-local signatures are:

• TP4: model/intermediate (4096,512), E=289, topk=9
• TP8: model/intermediate (4096,256), E=289, topk=9

The existing model-config set does not cover either signature. Without exact rows, both use fallback kernels.

What this adds

Sixteen gfx950/cu256 rows are added in a dedicated GLM-5.3 fused-shared table:

• TP4: M={4,8,1024,2048,4096,8192,16384}
• TP8: M={8,32,128,256,512,2048,4096,8192,16384}

The matching untuned table records the complete 15-shape power-of-two ladder for both TP4 and TP8 (30 source rows).

Both signatures were tuned with the standard fused-MoE tuner over the compatible one-stage ASM block-FP8 family. Only rows improving AITER's production operator by at least 3% were retained. The remaining buckets are intentionally absent and keep fallback dispatch.

Validation

• All 16 retained rows have err1=0.0%, run_1stage=1, kernelName2 empty, and us2=0.
• Production-operator execution passed for every retained row: 7 TP4 and 9 TP8, all status OK.
• TP4 retained rows improve production-operator latency by 10.98-48.12%.
• TP8 retained rows improve production-operator latency by 4.61-45.11%.
• Direct fused-versus-separate tensor comparison passed at TP4 and TP8 from M=1 through M=16384: all outputs finite, minimum cosine 0.999822, maximum mean absolute error 7.45e-6.
• test_csv_validation.py and test_config_shape_collision.py: 32 passed, 37 subtests passed.
• Every new dispatch key occurs exactly once across the complete current-main fused-MoE config set.
• No open AITER PR contains any new dispatch key.

Hardware: MI355X (gfx950, 256 CUs).
Image: rocm/sgl-dev:v0.5.20-rocm724-mi35x-20260921.
Tuning branch base: upstream ROCm/aiter main at commit 7d2f6a51a6364e0aad4d2e81dda5cc28c82bfdd9.

Submission Checklist

• Looked over the ROCm contributing guidelines.
• Targeting the repository default branch (main).
• Included successful config-validation and production-operator results.
• Commit includes the required DCO sign-off.
• CI green for the updated head.

[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 5735 --add-label <label>

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 opt this PR out.

[yifehuan]

LGTM