fix: correct MXFP4 MoE weight shuffle for Quark models (MiniMax-M2.1)

[thpereir]

Fix the Quark branch of Mxfp4MoEMethod.process_weights_after_loading() to use the correct pre-shuffle format expected by CK JIT MoE kernels (DeviceMoeGemmMXBPreShuffle).

Previously the Quark path called shuffle_weights() (a16w4 interleave format) and e8m0_shuffle() for scales. The CK JIT a4w4 kernels (ck_moe_stage1/ck_moe_stage2_fwd) expect shuffle_weight_a16w4(gate_up= False) format for both w1 and w2, and shuffle_scale_a16w4() for scales.

Changes:

• w13 (gate+up): shuffle_weight_a16w4(w13, 16, gate_up=False) + shuffle_scale_a16w4(scale, E, gate_up=False). The a4w4 kernel uses gate_up=False even for w1 which has 2*N rows (unlike the a16w4 path which uses gate_up=True for w1).
• w2 (down-proj): shuffle_weight_a16w4(w2, 16, gate_up=False) + shuffle_scale_a16w4(scale, E, gate_up=False).
• Set is_shuffled=True on both w13_weight and w2_weight so that ck_moe_stage2_fwd selects the preshuffle_on module variant.
• Also set is_shuffled=True on w2_weight in the existing gpt_oss/Swiglu branch (latent bug: weight was shuffled but attribute was missing).

Scope: Only the Quark branch is new code. Currently only MiniMax-M2.1- MXFP4 uses quant_method="quark" with fp4x2 MoE. The gpt_oss is_shuffled fix is a one-line correctness fix for DeepSeek-R1 FP4 and similar models.

Depends on: ROCm/aiter thpereir/cktile_a4w4 branch (CK JIT a4w4 dispatch)

Tested: MiniMax-M2.1-MXFP4 TP=8 server starts and produces correct output.

Motivation

Technical Details

Test Plan

Test Result

Serving with ATOM:

python -m atom.entrypoints.openai_server \
  --model amd/MiniMax-M2.1-MXFP4 \
  --trust-remote-code \
  -tp 8

lm_eval \
  --model local-completions \
  --model_args model=amd/MiniMax-M2.1-MXFP4,base_url=http://localhost:8000/v1/completions,num_concurrent=32,max_retries=3,tokenized_requests=False \
  --tasks gsm8k \
  --num_fewshot 5 \
  --batch_size 1

lm-eval before changes:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0	±	0
		strict-match	5	exact_match	↑	0	±	0

lm-eval after changes:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9371	±	0.0067
		strict-match	5	exact_match	↑	0.9348	±	0.0068

Submission Checklist

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