Add Kimi K2.5/K2.6 FP4 fused MoE tunings for TP2 (inter_dim=1024, 385 experts, top-9)

[xaguilar-amd]

Summary

Adds tuned fused-MoE kernel selections for the Kimi K2.5 / K2.6 FP4 TP2
shape group (cu_num=256, inter_dim=1024, expert=385, topk=9) to
aiter/configs/model_configs/kimik2_fp4_tuned_fmoe.csv, tuned on MI355X.

This shape group was previously missing from the config (the file only covered
the TP4 / TP8 shapes).

What changed

• Adds 32 new rows under (256, *, 7168, 1024, 385, 9):
  • 16 FlyDSL rows (one per token count: 1, 2, 4, 8, 16, 32, 64, 128, 256, 512,
    1024, 2048, 4096, 8192, 16384, 32768).
  • 16 matching flydsl_fallback rows (pure CK two-stage GEMMs) for runtime
    paths where FlyDSL is unavailable.

Motivation

Inference deployments of Kimi K2.5 / K2.6 MXFP4 on TP2 / 256 CU (MI355X)
currently fall through to untuned defaults for inter_dim=1024, since the
table only carries TP4 (inter_dim=512) and TP8 (inter_dim=256) entries
for the 385/9 expert geometry. Adding the TP2 rows lets the existing
fused-MoE dispatch pick the tuned FlyDSL kernels (or the CK fallback) for
this geometry instead.

Technical details

• Touches only aiter/configs/model_configs/kimik2_fp4_tuned_fmoe.csv.
• Strictly additive vs. current main: 1 file changed, 32 insertions(+), 0 deletions(-).
• No existing TP4 (inter_dim=512) or TP8 (inter_dim=256) row — neither
  the FlyDSL nor the flydsl_fallback variant — is modified, renamed, or
  reordered.

[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 3287 --add-label <label>