[yadaish]

Motivation

Add a CSV-selectable A-preshuffled activation layout to gfx1250 A4W4 grouped MoE and make the existing TP/EP test entry points validate every applicable tuned configuration. CSV sweeps now use each selected file for runtime configuration lookup, including when switching between model-specific files.

Technical Details

• Thread a_preshuffle from the matched tuning row through activation preparation and both grouped GEMMs. Add compact MXFP4 quantization followed by routed scatter into M16 payload tiles and M16K4 scales; prequantized inputs can enter scatter directly.
• Update TDM/LDS addressing and fused GEMM1 quantized output for the packed layout. Validate layout constraints and include the layout in kernel specialization/cache identities.
• Align AOT generation with runtime launch signatures, stage-specific tile settings, A-preshuffle variants, and compact/aligned scale strides. Keep EP's topk=-1 tuning sentinel out of compile-time route-count arithmetic.
• Add format-specific scheduling defaults for 256x256x256 prefill tiles, enable MoE expert scheduling by default, refresh main/Qwen tuning, and add Kimi gfx1250 tuning.
• Extend the existing CSV runners: TP defaults to main + Qwen, EP defaults to main + Kimi, and --csv-path selects one file. Clear configuration lookup caches between files, support SiLU/Swiglu/Situv2, fix shared-expert reference mapping, and aggregate failures across selected files. Default sweeps exclude untuned configurations; TP skips EP rows.

Test Plan

Validated on gfx1250 using the existing test entry points without --csv-path:

• TP: op_tests/flydsl_tests/test_flydsl_grouped_gemm.py --scenario csv --no-bias --iters 16 --no-check-aot-cache, with grouped GEMM enabled, CK disabled, and balanced routing. Run constant/constant and uniform/auto data/scale initialization with CUDA graphs and JIT allowed.
• EP: op_tests/test_moe_ep.py -t ep_csv --ep-mode real, with four simulated ranks, top-k 8, and balanced routing. Validate local outputs only; distributed EP scatter-combine is outside this validation.
• Ruff 0.16.0: ruff check .; whitespace: git diff --check.

Test Result

Mode	Tuned CSV	Correctness checks passed
TP	Main	190/190
TP	Qwen3.5-397B	72/72
EP	Main	42/42
EP	Kimi K3 gfx1250	27/27

All 200 applicable CSV rows passed, totaling 331 correctness checks. Both CSV processes exited successfully. The two ep_fused=1 rows were tested as local EP shapes; their fused scatter behavior was not validated. Ruff and whitespace checks passed.

Performance comparison remains pending: existing priority-script measurements had concurrent GPU activity or substantial timing variation and do not support a stable speedup/regression claim.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

[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 (added automatically when gfx942 configs change); main branch always runs both MI35X and MI300X
ci:triton-355	Run the full Triton test suite on MI35X, not only the tests the change affects
ci:windows	Windows JIT op smoke test on the gfx1201 ROCm runner
ci:multi-gpu	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: DeepSeek-V4-Pro, MiniMax-M3-MXFP4
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 6709 --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.

[coderfeli]

LGTM