[XiaobingSuper]

Summary

• make unified_attention return deterministic zeros for all-decode rows whose KV length is zero
• neutralize split-KV partial output/max/expsum for the same padding rows
• add a gfx1250 Gluon regression test with lengths [2048, 0, 0, 0]

Root cause

CUDA-graph buckets can pad a decode batch with query rows whose seqused_k value is zero. The gfx1250 3D Gluon path returns before writing those rows, leaving NaN/uninitialized output. Later model layers consume the padding values and can produce repeated or non-terminating generations.

This was reproduced while bringing up MiniMax-M3 TP4 on MI455 for LLM-Booster #201 and the fallback in ATOM #1410. A standalone BF16 shuffled-cache check showed:

• before: finite_by_row=[True, False, False, False]
• after: finite_by_row=[True, True, True, True]
• active-row cosine remained 0.99999619

Validation

• test_unified_attention_zero_kv_decode_rows: passed
• existing BF16 shuffled 3D unified-attention case: passed
• MiniMax-M3 GSM8K debug, c4, first 8, max 2048: 0/8 before -> 3/8 after; the same result holds with the ATOM-side workaround removed
• ruff check: passed
• black --check: passed

Draft for correctness and performance review because the post-op mask adds a small decode kernel.

[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 (added automatically when gfx942 configs change); main branch always runs both MI35X and MI300X
ci:triton-355	Run the full Triton test suite on MI35X, not only the tests the change affects
ci:windows	Windows JIT op smoke test on the gfx1201 ROCm runner
ci:multi-gpu	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: DeepSeek-V4-Pro, MiniMax-M3-MXFP4
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 6566 --add-label <label>

One backend per PR:
A PR changes one kernel backend: [Triton/Gluon] (Triton and Gluon count as one), [HIP], [ASM], [CK], [OPUS] or [FlyDSL]. If the title ends up with two backend tags, split the PR -- as stacked pull requests when one part cannot merge without the other.

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 stop the title rewrites; labels stay in sync either way.

[XiaobingSuper]

MiniMax-M3 decode on gfx1250 now goes through FlyDSL (ROCm/ATOM#2710), so it no longer reaches unified attention with zero-length padding rows. Closing; happy to reopen if the unified-attention fallback needs this.