tune: add TP2 MXFP4 fused-MoE configs for Qwen3.5-397B

[yichiche]

Summary

• Add TP2 (inter_dim=512) MXFP4 fused-MoE tuned configs for Qwen3.5-397B on MI355X (gfx950).
• Covers both routing conventions — unfused expert=512, topk=10 and shared-expert-fused expert=513, topk=11 — for token buckets 1 … 32768.
• Pure data addition: existing TP4 (inter_dim=256) rows are untouched; only new rows are appended to qwen3_5_397b_fp4_{untuned,tuned}_fmoe.csv.

Details

The shipped qwen3_5_397b_fp4 fused-MoE configs only contained TP4 geometry (inter_dim=256). Serving Qwen3.5-397B at TP2 uses inter_dim=512, so per-shape lookups missed and fell back to untuned kernel selection.

This PR adds the TP2 shapes and their tuned kernels:

• Untuned (qwen3_5_397b_fp4_untuned_fmoe.csv): +32 rows — inter_dim=512 for both 512/10 and 513/11, tokens 1 → 32768. (The 16 existing TP4 rows are kept.)
• Tuned (qwen3_5_397b_fp4_tuned_fmoe.csv): +64 rows — the corresponding tuned best kernels, appended verbatim after the existing TP4 entries.

Quant/dtype is unchanged MXFP4: torch.float4_e2m1fn_x2 weights/activations, QuantType.per_1x32, ActivationType.Silu, model_dim=4096.

Generated with the in-tree tuner:

HIP_VISIBLE_DEVICES=0,1,2,3 python3 csrc/ck_gemm_moe_2stages_codegen/gemm_moe_tune.py \
  -i aiter/configs/model_configs/qwen3_5_397b_fp4_untuned_fmoe.csv \
  -o aiter/configs/model_configs/qwen3_5_397b_fp4_tuned_fmoe.csv  \
  --mp 4 --shape_grouped --sort True

All 32 shapes tuned successfully (0 failed shapes; tuner error tolerance errRatio=0.05).

Tuned MoE latency (TP2, unfused 512/10, best kernel per token)

Tokens	block_m	Winning family	Total (µs)
1	32	flydsl	25.0
2	32	flydsl	27.5
4	64	flydsl	37.2
8	32	flydsl	54.5
16	32	flydsl	91.5
32	32	flydsl	140.8
64	32	flydsl	204.1
128	32	flydsl	253.0
256	32	flydsl	278.2
512	32	flydsl	288.0
1024	32	flydsl	306.9
2048	64	flydsl	378.7
4096	64	flydsl	557.9
8192	64	flydsl	859.0
16384	64	flydsl	1507.1
32768	128	ck2stages	2770.5

The flydsl DSL kernels win across the small/medium token range; ck2stages takes over at the largest bucket (32768).

Accuracy

No accuracy impact — this PR only adds kernel-selection metadata. Each tuned kernel is validated by the tuner against the reference within errRatio=0.05; numerical behavior of the served model is unchanged.

Notes

• No code changes; CSV data only.
• Hardware: AMD Instinct MI355X (gfx950).

[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:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
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 3859 --add-label <label>

[lalala-sh]

LGTM