[ZhiweiYan-96]

Problem

Since #6478 (88d7db6), ATOM's amd/MiniMax-M3-MXFP4 TP8 accuracy job generates garbage on every decode step and each GSM8K request runs to max_gen_toks=16384. The job then times out after ~35–54 min instead of passing in ~9 min. It has failed on every ATOM main run since aiter picked up 88d7db6: job 116798160693, 116822862840, 116835414847. The last pass (116770801850, aiter c118faf) scored GSM8K 0.9431. The other accuracy jobs in those runs still pass, apart from one unrelated 5-minute DeepSeek-R1-0528 MTP failure in one run.

Root cause

#6478 changed swiglu_mxfp4_bf16_cktile to require gate_up_interleave is False. Callers that never set w1.gate_up_interleave pass None; ATOM is one of them. With bf16/fp16 activations, these callers no longer take the CK-Tile split-k route with post_activation_layout="standard". They fall through to the "unshuffled weights" CK-Tile fallback (post_activation_layout="auto") instead. The failing CI server log shows SwiGLU MXFP4 with unshuffled weights not supported by CK2stages codegen; routing to CK-Tile (ROCM-25478); the passing log does not.

Only the TP8 shape reaches CK-Tile. A trace of get_2stage_cfgs inside the TP8 server shows the per-rank MoE call as model_dim 6144, inter_dim=512, intermediate_pad=0, E=129, top-5. Every decode step (M ≤ 256, bf16 activations) goes to CK-Tile. At TP4 the call has inter_dim=768, and FlyDSL handles every M, so a TP4 run on 88d7db6 stays healthy (GSM8K 64-sample 0.9375).

Route chosen for the TP8 decode calls, recorded by the same server-side trace:

aiter	route	post_activation_layout	bias
c118faf (before #6478)	CK-Tile split-k	standard	True/True
88d7db6	unshuffled-weights fallback → CK-Tile	auto (M ≤ 8) / unset (M ≥ 16)	False/False
88d7db6 + this PR	CK-Tile split-k	standard	False/False

Fix

Treat an untagged layout (None) with bf16/fp16 activations the way swiglu_mxfp4_bf16_cktile did before #6478. Tagged callers (True / False) take exactly the same route as on main.

Validation

All runs are on MI355X (gfx950).

End to end: ATOM main (ffd113b), MiniMax-M3-MXFP4, TP8, CI server flags and env. GSM8K 5-shot chat uses --limit 64 and max_gen_toks=4096. The direct probe sends 8 GSM8K questions at temperature=0.

aiter	GSM8K (64)	probe: requests ending on length	probe completion tokens
88d7db6	0.0000	8 / 8	4096 each (garbage text)
88d7db6 + this PR	0.9219	0 / 8	105–276, mean 195

Kernel: op_tests/test_moe_2stage.py::test_fmoe on the traced TP8 call shape. The shape is 6144 / 512 / pad 0 / E=129 / top-5, SwiGLU limit 7, no bias, preshuffled MXFP4 weights, AQDType=fp4x2. Every aiter build runs the same harness, the 88d7db6 copy. The table gives logits_diff = 1 − cosine similarity to the torch reference:

tokens	88d7db6	88d7db6 + this PR	c118faf
1	0.998	0.0810	0.0806
8	0.980	0.0751	0.0751
32	0.981	0.0722	0.0722
128	0.974	0.0737	0.0737

With this PR the error goes back to the c118faf baseline. The ~0.07 baseline is the pre-existing value #6478 already noted for this shape.

Tests. The new test_swiglu_mxfp4_untagged_layout_dispatch asserts the CK-Tile split-k route with post_activation_layout="standard" for the traced TP8 call. It passes with this PR and fails on main (got 'auto'). The existing test_swiglu_mxfp4_nonsplit_nobias_dispatch still passes. black and ruff (0.16.0) are clean.

[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 (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 6620 --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.