Add full 16-tier Qwen3-VL FP4 MoE tuned config for gfx950

[johannes-graner]

Motivation

The Qwen3-VL W4A4 FP4 MoE path on gfx950 had no tuned config for its dominant shapes, so
fused_moe fell back to a slower base stage2 kernel across most token counts. This ships a
complete, peer-matching tuned config so the FP4 MoE path dispatches its best-known kernel at
every padded token tier.

Technical Details

• aiter/configs/model_configs/qwen3_vl_fp4_tuned_fmoe.csv: expanded to the full 16-tier padded
  power-of-two grid (tokens 1..32768), matching the row density/completeness of peer fmoe
  configs. Every row is fully measured (us1/us2/us/tflops/bw populated).
• aiter/configs/model_configs/qwen3_vl_fp4_untuned_fmoe.csv: expanded to the same 16 tiers
  (superset of the tuned shapes).
• Kernel selection: small tiers (1–1024) use the isolated-tuner winners; large tiers
  (2048–32768) select the existing
  flydsl_moe2_afp4_wfp4_bf16_t64x128x256_atomic_persist_async_w4_cumul3 stage2 kernel, whose
  large-tier selection was validated against the real fused end-to-end fused_moe path (not
  just isolated per-kernel timing).
• Config-data only — no kernel or dispatch code changes.

Test Plan

• Fused end-to-end correctness: aiter.fused_moe(..., quant_type=QuantType.per_1x32) on the
  Qwen3-VL FP4 shapes, gated on cosine similarity against the torch reference.
• Collision guard: python3 -m unittest op_tests.tuning_tests.test_config_shape_collision.
• Fused-path performance validation (--run_config) across the large-tier grid (2048–32768)
  against the naive isolated-tuner pick.

Test Result

Correctness passes at every tuned tier (cosine-gated). Collision guard passes. Large tiers
match or beat the naive isolated-tuner pick when validated end-to-end, up to −22% at the
16384-token tier; dominant-shape stage2 improves ~19.8% over the untuned fallback.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.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; 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 4759 --add-label <label>

[Copilot]

Pull request overview

Adds complete per-model FP4 MoE (fused_moe) config coverage for Qwen3-VL on gfx950 by extending both tuned and untuned CSV shape grids to the full 16-tier padded power-of-two token set, enabling consistent best-kernel dispatch across token counts.

Changes:

• Added a full 16-tier (1..32768) untuned shape list for Qwen3-VL FP4 MoE.
• Added a full 16-tier tuned table with measured stage1/stage2 kernel selections and perf metrics for each tier.
• Standardized large-tier selections onto the validated stage2 kernel for 2048–32768 tokens.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
aiter/configs/model_configs/qwen3_vl_fp4_untuned_fmoe.csv	Adds the 16-tier padded token grid of Qwen3-VL FP4 MoE shapes for tuning/lookup coverage.
aiter/configs/model_configs/qwen3_vl_fp4_tuned_fmoe.csv	Adds measured tuned rows per tier so fused_moe can dispatch the intended best-performing kernels across token sizes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.