[Triton/Gluon] [Config] triton: add Qwen3.8-27B MXFP4 GDN in_proj_ba AFP4WFP4 tiles (gfx950)

[abrahamzewoudie]

Motivation

Follow-up to #5350 (A4W4 CSV for gemm_a4w4). Qwen3.8-27B MXFP4 GDN in_proj_ba is dispatched through AITER Triton gemm_afp4wfp4. Those kernels load GEMM-AFP4WFP4-N=<N>-K=<K>.json by shape. gfx950 currently has no tiles for this model's BA widths.

On TP2, in_proj_ba scale rows are not divisible by 32, so vLLM always falls back to Triton for this layer. A TP2 A/B on MI355X (ISL 8192 / OSL 1024 / c=4) measured +1.7% output tok/s with the N=48 JSON vs DEFAULT.json (434.91 vs 427.68 tok/s). TP1 N=96 is scale-swizzle legal and stays on ASM unless a per-layer Triton force is applied, so that JSON is for the fallback / correctness path, not the default ASM path.

Added tuning data

TP	(N,K)	Source
1	(96,5120)	Arete vllm-tp1-mxfp4-throughput/.../triton-ba/
2	(48,5120)	Arete vllm-tp2-mxfp4-throughput/.../triton-ba/
4	(24,5120)	Arete vllm-tp4-mxfp4-throughput/.../triton-ba/

Files:

• aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/GEMM-AFP4WFP4-N=96-K=5120.json
• aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/GEMM-AFP4WFP4-N=48-K=5120.json
• aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/GEMM-AFP4WFP4-N=24-K=5120.json

Dense hybrid GDN (Qwen3.5-style), not MoE. No CSV and no fused-MoE rows in this PR.

Data validation

• Tuned on MI355X (gfx950, 256 CUs).
• Triton BA configs passed reference comparisons at representative M=16, 128, 1024.
• Lookup is by filename N/K only. TP4 must be N=24, not a copy of the TP2 N=48 file.

This PR was developed with AI assistance. The human author reviewed the tuning provenance, the TP2 A/B, and validation, and owns the submitted data.

Test plan

• Confirm the three files land under aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/
• TP2 serve with the N=48 overlay: log shows Triton fallback for in_proj_ba (weight_scale rows not divisible by 32)
• TP4 uses N=24 JSON, not N=48

Made with Cursor

[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
multigpu	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: 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 5669 --add-label <label>

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 opt this PR out.

[Copilot]

🔵 Needs a closer look

The changes require final human review because they are too complex or risky for automated approval.

Pull request overview

Adds gfx950 Triton AFP4WFP4 tuning configs for Qwen3.8-27B GDN in_proj_ba across TP1/TP2/TP4 shapes.

Changes:

• Adds specialized N/K configs for (96, 5120), (48, 5120), and (24, 5120).
• Provides complete M-bucket, tile, and split-K parameters.
• Uses the existing logical-K filename lookup convention.

File summaries

File	Description
aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/GEMM-AFP4WFP4-N=96-K=5120.json	Updated as part of this pull request.
aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/GEMM-AFP4WFP4-N=48-K=5120.json	Updated as part of this pull request.
aiter/ops/triton/configs/gfx950/triton/gemm/gemm_afp4wfp4/GEMM-AFP4WFP4-N=24-K=5120.json	Updated as part of this pull request.

Review details

• Files reviewed: 3/3 changed files
• Comments generated: 0
• Review effort level: Lite

---

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

[abrahamzewoudie]

@Boss2002n Can you please have a look at this PR? Thanks

[abrahamzewoudie]

@Boss2002n Can this PR be merged since it is approved?