[TRITON][GLUON][GFX950] Add Unified Attention Gluon Kernel

[cagrikymk]

Motivation

This PR adds gfx950 gluon kernel for unified attention. It uses single kernel that supports both 2d and 3d grid.

Performance is tuned for head sizes [64, 128, 256], dtype fp8 x fp8 and bf16 x bf16.

Remaining Steps

• Test e2e performance, at least for GPT-OSS-120b
• Extend the gluon routing to make it more bullet-proof

Technical Details

Performance numbers are collected using Triton 3.7.1.

Geomean speedup by head_size against Triton

This is the geomean speedup of various run configs with varrying concurrency, kv/seq. length, number of KV/Q heads,

head_size	decode bf16	decode fp8	prefill bf16	prefill fp8	mixed bf16	mixed fp8
256	1.58x	1.02x	2.17x	2.22x	1.90x	1.81x
128	1.10x	1.08x	1.19x	1.38x	1.23x	1.41x
64	1.11x	1.24x	1.51x	1.55x	1.54x	1.62x

Detailed performance report:
gluon_gfx950_perf_report.md

Test Plan

Existing tests are routed to the gluon kernel if conditions are met

Test Result

Tests pass

[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 4614 --add-label <label>

[cagrikymk]

GPT-OSS 120b benchmark, bf16 kv cache:

ISL/OSL 1024/1024

C	gluon tok/s	triton tok/s	delta
8	1979	1958	+1.1%
16	3937	3921	+0.4%
32	7014	6941	+1.1%
64	12090	11960	+1.1%
128	19208	19003	+1.1%

ISL/OSL 8192/1024

C	gluon tok/s	triton tok/s	delta
8	1836	1816	+1.1%
16	3419	3384	+1.0%
32	5528	5434	+1.7%
64	8306	8144	+2.0%
128	11189	10945	+2.2%

[cagrikymk]

@vgokhale Due to triton version difference between CI and my local setup (both are 3.7 but at different commits), I was getting LLVM compilation error when using waves_per_eu=3 with sliding window in CI:

llvm/ADT/Sequence.h:275  llvm::iota_range<unsigned>::iota_range(T, T, bool):
  Assertion `Begin <= End && "Begin must be less or equal to End."' failed.
 (anonymous namespace)::SIInsertWaitcnts::run()
translateLLVMIRToASM   (python/src/llvm.cc:399)

I changed occupancy to 0 for that case, and it compiles to a kernel with occupancy 3 without any error in CI triton now.