[XiaobingSuper]

Problem

ATOM detects aiter's explicit FlyDSL PA-plan API (ROCm/aiter#5809) by requiring work_plan to be keyword-only with no default. ROCm/aiter#7056 gave work_plan a None default for its MXFP4 path (other paths still require a plan), so current aiter main is rejected as predating #5809. On gfx1250, MiniMax-M3 has no gluon fallback and fails model startup.

Change

• Probe presence only: plan_pa_decode, a keyword-only work_plan, and max_context_length. The default of work_plan is no longer checked: ATOM builds and passes the plan explicitly at every call site, so a default has no bearing on the call. This matches the other aiter capability probes (e.g. max_batch_size in the same file).
• An aiter before #5809 (positional work_plan=None, no max_context_length) is still rejected.
• The warning prints the observed pa_decode signature (or the import error) instead of always blaming an old aiter.

Validation

• tests/test_paged_attention_dispatch.py + tests/test_minimax_m3_sparse_pa_floor.py: 182 passed on gfx1250, against both an aiter with the required work_plan and one with work_plan=None (post-#7056)
• The probe test accepts required, None and other defaults, and rejects one fake per clause. Deleting any of the four clauses, or re-adding default is empty, default is None or default in (empty, None), fails it.
• With the post-#7056 aiter on gfx1250 and ATOM_PA_FLYDSL=1: on main the probe is False and run_pa_decode raises "no gluon fallback"; with this PR, one BF16 decode through FlyDSL matches a torch reference (max abs err 0.002 at qlen=1, 0.003 at qlen=2)
• aiter main (075695e8) pa_decode verified from source: keyword-only work_plan: PADecodePlan | None = None
• Black and Ruff pass

Tracked while validating MiniMax-M3 TP1/DP4/EP4 on MI455X.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 3052 --add-label <label>

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[valar-cop]

a few small things..

any startup log or accuracy from the MiniMax-M3 TP1/DP4/EP4 run on gfx1250?

ci green.

[yitingw1]

The fix is right and the breakage is real — I reproduced it against aiter origin/main (8e0000db3): the probe on main returns False for the installed signature, so FlyDSL is declined and gfx1250 has nothing to fall back to.

But I think the default clause should be deleted rather than widened, and the reason is in why aiter changed it.

aiter didn't make the plan optional — it made it optional for MXFP4. From origin/main:

# docstring
``work_plan`` is required and must be built with ``plan_pa_decode`` before calling decode.
``work_plan=None`` is accepted for MXFP4 only: one task per sequence

# body
if not isinstance(work_plan, PADecodePlan) and not (mxfp4 and work_plan is None):
    raise TypeError("work_plan must be a PADecodePlan")

So the two defaults map to post-#5809-pre-MXFP4 and post-#5809-post-MXFP4. Neither maps to "predates #5809" — an aiter that old has no work_plan and no max_context_length, and is already rejected by the other clauses. Running six signature shapes through the clause as it was, as this PR makes it, and with it removed:

signature	main	this PR	removed
required (#5809 shape)	True	True	True
optional_none (aiter main)	False	True	True
sentinel (some third default)	False	False	True
positional / missing max_context_length / missing work_plan	False	False	False

The only row where this PR and removal differ is sentinel — i.e. all the clause still does is reject a default we haven't seen before, which is exactly the failure being fixed here. If aiter ever writes work_plan=_UNSET, this breaks again the same way, and again as a gfx1250 startup failure. ATOM passes a plan at every call site, so whether aiter supplies a default has no bearing on whether the call works.

base_attention.py:385, 150 lines down and probing the same aiter module, already does it the other way: "max_batch_size" not in inspect.signature(planner).parameters. So do every other aiter capability probe in the tree (gemma_norm in layernorm.py:764, scale_row_major in mhc_pre_delayed.py:55, max_split_per_batch in aiter_mla.py:128, num_valid_rows in topk_compat.py:21) — all presence-only. Line 239 is the only one that reads a default, and the only one that broke.

→ Change: drop the clause so the probe matches the idiom used in the rest of the file.

Neither accepted shape is pinned by a test. Two mutations, both green on test_paged_attention_dispatch.py + test_minimax_m3_sparse_pa_floor.py (182 passed, same as baseline):

• remove the clause entirely → green
• narrow it to plan.default is None, i.e. stop accepting the #5809 shape → green

no_bound is rejected for missing max_context_length, not for its default, so the required half of the new test's name has no fake behind it — it rests only on the trailing assertion against the installed aiter, which today is the optional shape.

→ Change: if the clause stays, add a def required(output, *a, work_plan, max_context_length=None) fake expecting True. If it goes, this need goes with it.

The warning misdiagnoses the failure it was just hit by. "this aiter predates the #5809 decode API" — here aiter was too new. On gfx1250 there is no fallback and startup fails, so that line is the only clue an operator gets.

→ Change: log which clause failed and the observed signature, instead of attributing every failure to an old aiter.

One more on the new test's docstring — "aiter may also allow callers to omit it" reads as a general loosening; per aiter's own docstring and body it holds for MXFP4 only, and omitting it elsewhere raises. Suggest something like "aiter gives it a default for its MXFP4 path; ATOM still always passes one."

Not verified: I have no gfx1250, so I did not reproduce the MiniMax-M3 startup failure or run TP1/DP4/EP4 — the data valar-cop is asking for. Everything above is at the signature level and does not exercise the kernel. I'm also not repeating valar-cop's two points; both still look open.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[yitingw1]

All four points are in f50114b4a, and the first one landed as removal rather than widening — thanks.

The new warning is a real improvement. It showed up in my own mutation output, which is a decent demonstration of it:

WARNING ... aiter lacks the #5809 explicit-plan decode API
(pa_decode signature: (output, query, work_plan=None)); using gluon.

Someone debugging a gfx1250 startup now sees the signature instead of being told their aiter is too old.

One thing: the test grew from 2 cases to 5, but it still pins only one of the four clauses. Deleting each clause in turn and running the file (baseline 169 passed):

clause	with it deleted
hasattr(mod, "plan_pa_decode")	169 passed
plan is not None	169 passed
plan.kind is inspect.Parameter.KEYWORD_ONLY	169 passed
"max_context_length" in sig.parameters	1 failed

The reason is in the two cases that expect False:

def pre_5809(output, query, work_plan=None):      # no max_context_length
def no_bound(output, *a, work_plan, sinks=None):  # no max_context_length

Both are rejected for the missing max_context_length, never reaching the clause they look like they're about — pre_5809 is written positionally as if to exercise KEYWORD_ONLY, but it fails before that. So the suite reads as "five signature shapes covered" while establishing only that a missing max_context_length is refused.

That matters here more than it usually would: this probe's whole history is breaking because it checked one thing too many. If anyone tightens or relaxes the remaining three, nothing goes red.

→ Change: three one-line fakes, each isolating one clause, all expecting False:

def positional_plan(output, work_plan=None, max_context_length=None):  # kind
def no_plan(output, *a, max_context_length=None):                      # plan is not None
# plus one case with monkeypatch.delattr(mod, "plan_pa_decode")

Minor: in the ImportError branch sig stays None, so the warning reads "aiter lacks the ... API (pa_decode signature: None)" — which looks like the function exists with an odd signature rather than the module being absent. Worth separating the two, since diagnosis is the whole point of that line.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[XiaobingSuper]

@yitingw1 Thanks, both addressed in 210abf0:

• Each rejected fake now misses exactly one piece (positional_plan, no_plan, no_bound, plus a case with plan_pa_decode deleted). Deleting any of the four clauses, or re-adding any of the three default checks, now fails the test.
• An import failure now logs import failed: <error> instead of pa_decode signature: None.