[coderfeli]

Summary

Remove redundant scalar and vector reconstruction across FlyDSL attention, MoE, GEMM, routing, communication, and convolution kernels. Use typed arithmetic, pointers, atomics, layouts, and quantization helpers while preserving kernel schedules, memory ordering, and numerical behavior.

• Remove same-type scalar casts, redundant wrappers around typed loads, duplicate element coercions, and singleton vectors used only for broadcasting. Use fx.full for vector fills. The repository-wide wrapper pass removes 719 scalar constructors, 60 vector constructors, and 30 from_elements calls, and shortens 145 fill calls.
• Keep conversions required for signedness, width, fragment shape, and raw IR boundaries. Preserve the BF16 packing form that keeps gfx950 code generation unchanged, target-specific assembly, invariant LLVM loads, and required alignment/scope/order on atomics.
• Merge main 5e3456494b877386d58c9ec6ca5e6a7fb37d40c7, including the updated persistent MLA and MegaMoE TP paths. Resolve the MLA conflict while retaining the new pipeline. The PR contains 115 source files and a net reduction of 612 lines. Validation scripts and generated artifacts are outside the diff.

Current merge validation

Against main 5e3456494b, using pinned FlyDSL 0.4.0:

• 745/745 gfx1250 matrix cases have byte-identical emitted GPU code objects: 273 configured GEMM jobs, 356 configured grouped-MoE jobs with their auxiliary kernels, 20 batched-GEMM CSV rows through the runtime wrapper, and 96 page-size-1 FP8 MLA variants. MLA covers supported head counts/query tile sizes, causal and context-parallel flags, and final-LSE modes.
• 54 additional unchanged-main comparisons match, covering quantization, FMHA, compression, routing, TDM helpers, page-size-64 MLA, gather, and combine sync/reduce.
• 24 batched A8W4 cluster cases match main with its missing multicast-mask helper supplied by the vendored helper. Twelve nonconfigured shapes fail on both sides: four negative tensor-wait counts at tile M=32/K=256, and eight LDS alignment errors at tile M=32 or 64/K=768. These are pre-existing failures and are not counted as successful comparisons.
• Across the 823 passing pairs, all 2929 actual code objects per side are byte-identical. Two standalone diagnostic .s dumps differ; their actual code objects and disassembly match. Repeated builds of unchanged main reproduce the dump-only variation because FlyDSL recompiles a clone to produce that diagnostic.
• Black passes for all 1,504 Python files. Ruff, Python syntax checks, and git diff --check pass for the 115 changed files. This gfx1250 validation is compile-only.

Compilation used main 1b1bc88f20; the only subsequent main change is a Gluon utility update. The FlyDSL runtime and AOT source trees are identical in 5e3456494b.

Earlier validation

Using the pinned FlyDSL 0.4.0, with final comparisons against main 9e2884545:

• 51/51 paired compilation cases produce byte-identical final ISA (54 generated kernel artifacts). Coverage includes mixed MoE stage 1/2, split-K and persistent paths, MXFP4 epilogues, preshuffle/HGEMM, Mega-MoE, routing and FP4/FP8 scatter, all-reduce, FMHA, MLA/MQA, GDR, compression, layouts, quantization, and conv3d reuse/padding. gfx942 a16w4 and gfx1250 TDM/FMHA/compression/routing paths are included.
• 175 repository pytest cases pass on gfx950: 111 PA decode, 21 FMHA, 14 Minimax index-score, 26 gather, and 3 shared-expert cases. These cover scale modes, paging and variable lengths, output layouts, graph replays, shared-weight K offsets, and clamped SiLU.
• 21 additional GPU cases pass: 11 compressor cases (BF16/FP8/FP4, decode/MTP, K-split and HCA), 3 QK norm/rope/quant cases, 3 Mega-MHC output modes with graph replay, and 4 convolution reference checks covering 2D/3D LDS reuse and reflect/replicate padding.
• All 65,536 BF16 bit patterns produce 851,968 integer quantization outputs identical to main, including special values. All 65,280 finite FP4 nibbles also match an independent nearest-even reference.
• HGEMM output checks pass with dirty output buffers and changed inputs across repeated launches and graph replays. Both BMM scale-block variants pass the exact all-ones check.
• A final repository scan found no further removable wrappers under the checked type/shape rules. Black, Ruff, Python compilation, and git diff --check pass for all 114 changed files.

Performance

On an otherwise idle gfx950 GPU, run main→PR and PR→main after warm-up. The table reports medians of repeated HIP-event measurements around graph replays, with compilation outside timing. The 19 cases range from −0.52% to +0.41%; no performance regression was observed.

The timing baseline is eb56b2d45, whose measured kernel sources are unchanged in main 9e2884545. Convolution is checked separately against the current main with identical ISA and GPU reference results.

Case	Main (µs)	PR (µs)	Change
BF16 GEMM, split-K 8	5.021	5.019	-0.03%
BF16 GEMM, split-K 8 + bias	4.955	4.943	-0.26%
BF16 GEMM, split-K 8, FP32 output	5.097	5.104	+0.15%
FP16 GEMM, split-K 8	4.927	4.926	-0.02%
BF16 GEMM, split-K 1	16.655	16.656	+0.00%
Gather GEMM, 512 tokens, K/V 128/128	7.459	7.445	-0.19%
Gather GEMM, 2048 tokens, K/V 192/256	14.214	14.185	-0.20%
MXFP8 BMM, scale block 32	2.449	2.437	-0.52%
MXFP8 BMM, scale block 128	2.466	2.455	-0.46%
Mixed FP8 × FP4 MoE stage 1	4.990	5.000	+0.20%
Mixed FP8 × FP4 MoE stage 2	2.999	2.996	-0.08%
MXFP4 GEMM stage 2	2.428	2.428	+0.02%
FP8 FMHA, GQA, split-K 1	33.563	33.577	+0.04%
FP8 FMHA, GQA, split-K 4	30.630	30.677	+0.16%
BF16 PA decode, MTP4	9.915	9.887	-0.29%
FP8 PA decode, per-token scale	7.554	7.585	+0.41%
CSA compressor, BF16	3.749	3.744	-0.14%
CSA compressor, FP4	3.151	3.147	-0.11%
HCA compressor, BF16	110.468	110.517	+0.04%

HGEMM: M/N/K = 32/384/7168. BMM: batch/M/N/K = 4/64/256/512. Mixed MoE: 32 tokens, 2 experts, top-k 2, model/intermediate dimensions 256; MXFP4 stage 2 uses 16 tokens. FMHA: batch 1, sequence 2048, 8 query heads/2 KV heads, dimensions 128. PA decode: context lengths 257/513, 2 partitions, head dimension 128. Compressors: batch 8 with MTP3, D=512 for BF16 and D=128 for FP4.

Runtime measurements cover representative shapes on gfx950. gfx942 and gfx1250 validation is compile-only; no runtime performance claim is made for those targets.

[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: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 6380 --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.

[yadaish]

LGTM