[RElbers]

Motivation

When setup.py precompiles FlyDSL kernels for a wheel (run_aot), it collects jobs from every kernel family. MoE and MegaMoE keep only the archs GPU_ARCHS names (#6362). The other families do not:

• GEMM filters by arch only in main(), the python -m aiter.aot.flydsl.gemm entry point. run_aot never goes through main(), so a gfx950 wheel also builds the gfx1250 kernels, and a gfx942 wheel builds 7885 kernels it can never run.
• chunk_gdn_h and fmha_fp8 do not filter at all, so a gfx950 wheel builds the gfx942-tuned chunk kernels and a gfx942 wheel builds the gfx950-only FP8 FMHA kernels.
• Conv filters inside parse_csv, but reads ARCH before GPU_ARCHS. Build images export ARCH as a machine type such as x86_64, so on those hosts the filter keeps nothing and the wheel ships no precompiled conv kernels.

An AOT miss is not fatal (the kernel is JIT-compiled at runtime), but the first two cases waste build time and wheel size, and the third turns every conv call into a runtime compile.

Technical details

• Every family run_aot pools now reads GPU_ARCHS through aot_target_archs() in common.py, the helper [OPUS] [FlyDSL] [JIT] [Feature] Build the MoE and MegaMoE AOT caches by replaying the runtime on CPU #6362 added for MoE, so all kinds agree on what a target list is. It does not read ARCH.
• GEMM, conv and chunk_gdn_h drop the jobs of unlisted archs in parse_csv. A job's arch is resolved the way its compile step resolves it: the CSV's gfx column with a cu_num fallback for GEMM and conv, cu_num for chunk_gdn_h, which compiles from it. chunk_gdn_h gets a job_arch() that both its filter and its compile step use.
• fmha_fp8 compiles for gfx950 only, so jobs_for_shape returns nothing unless GPU_ARCHS is unset or lists gfx950 (keeps_arch() in common.py).
• On the python -m entry points, collect_cli_jobs() / cli_arch_env() apply ARCH as GPU_ARCHS while collecting, so ARCH still takes precedence there, as on main, and parse_csv is the only filter. GEMM's second filter in main() is removed.

Behavior changes

FlyDSL jobs collected by run_aot from the in-tree tuned CSVs:

Environment	Build	GEMM	conv	chunk_gdn_h	fmha_fp8
GPU_ARCHS unset	main	7885	139	288	92
	this PR	7885	139	288	92
ARCH=x86_64	main	7885	0	288	92
	this PR	7885	139	288	92
GPU_ARCHS=gfx950	main	7885	139	288	92
	this PR	7151	139	0	92
GPU_ARCHS=gfx942	main	7885	0	288	92
	this PR	0	0	288	0
GPU_ARCHS=gfx1250	main	7885	0	288	92
	this PR	734	0	0	0

The 734 GEMM jobs a gfx950 build drops are the gfx1250 rows. Every chunk_gdn_h tuned row is gfx942. Unset and GPU_ARCHS=native keep every job, as before.

• GPU_ARCHS lists are ;-separated, as aot_target_archs() and the CK build read them. Conv also split on ,; it no longer does.
• On the python -m entry points, ARCH still overrides GPU_ARCHS, as on main.

Test plan

• New op_tests/test_flydsl_aot_arch_filter.py:
  • collect_cli_jobs with GPU_ARCHS=gfx942 ARCH=gfx950 collects for gfx950 and restores GPU_ARCHS afterwards (no FlyDSL needed).
  • GEMM and conv parse_csv on rows of two archs under unset and a single listed arch; conv also under ARCH=x86_64, which it now ignores (skips without FlyDSL).
  • Every kind run_aot pools collects nothing for an unlisted arch and something when GPU_ARCHS is unset, so a new kind cannot skip the filter.
  • 4 tests / 10 subtests pass.
• The job counts above come from _collect_aot_jobs_for, the collection step run_aot uses.
• ruff and black are clean on the touched files.

[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 6847 --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.

[Copilot]

🟢 Approval recommended

The scoped changes consistently filter build jobs while preserving CLI precedence and compile-target resolution, with no blocking issues identified.

No open issues 🎉

What changed in this PR

This PR makes FlyDSL GEMM and convolution AOT job collection respect wheel build targets in GPU_ARCHS, avoiding unrelated GEMM builds and convolution jobs being dropped because of a machine-type ARCH value.

Changes:

• Centralizes build-target parsing and applies filtering inside both CSV parsers.
• Preserves ARCH precedence for CLI entry points while restoring the environment after collection.
• Adds focused regression tests and documents the build-versus-CLI behavior.

File	Description
op_tests/​test_flydsl_aot_arch_filter.py	Tests target parsing, CLI overrides, and GEMM filtering.
aiter/​aot/​flydsl/​README.md	Documents architecture selection for builds and CLI usage.
aiter/​aot/​flydsl/​gemm.py	Moves architecture filtering into CSV job collection.
aiter/​aot/​flydsl/​conv.py	Uses shared filtering without build-time ARCH interference.
aiter/​aot/​flydsl/​common.py	Adds shared target parsing and CLI collection helpers.

_{Review effort: Balanced}

[RElbers]

Closing now that #6956 is merged