[XinyuJiangCMU]

Summary

Skip empty inputs in the custom all-gather: return from dispatchAllGather before the buffer lookup, and skip the zero-byte hipMemcpyAsync in all_gather_unreg.

Motivation

With attention data parallelism (DP), a DP group that has no requests in a step still joins the collectives, with empty inputs. RL rollouts do this routinely: groups that finish early wait for the longest requests. When a DP group has more than one GPU (attention TP > 1), its all-gather then reaches AITER with an empty input, and dispatchAllGather launches its kernel with zero blocks. The next kernel that checks for errors fails with HIP runtime error: invalid configuration argument. should_custom_ag accepts size zero while the all-reduce and reduce-scatter checks reject it, so all-gather is the only path that gets here.

Changes

• dispatchAllGather returns before the buffer lookup when the input is empty, so no zero-block kernel is launched.
• all_gather_unreg skips the zero-byte hipMemcpyAsync, whose source pointer is null for an empty tensor.
• Non-empty inputs take the same path as before.

Testing

• Unit tests added or updated
• Performance benchmarks run
• Tested on MI250X
• Tested on MI300X

Tested on MI355X (gfx950); see "Validation and unverified scope" below.

Documentation

• Docstrings updated
• User guide updated
• Performance guide updated
• No documentation change is needed, with the reason stated below

This is an internal guard for empty inputs; no API changes.

Dependencies

• No new third-party dependencies added
• New dependencies are listed and justified below

Breaking Changes

None.

PR Self-Review

Reviewed revision

5a40117

Behavior and intent

An empty all-gather input now returns immediately instead of launching a kernel with zero blocks. Before, the next kernel that checked for errors failed with HIP runtime error: invalid configuration argument, which stopped RL rollouts with attention DP whenever a DP group was idle. Non-empty inputs are unchanged.

Validation and unverified scope

On MI355X with AITER e7d2453 plus this change:

• 72 two-rank API checks pass across FP32, FP16 and BF16, empty and non-empty inputs, registered and unregistered calls, first- and last-dimension gather, and graph replay on the copy-in (all_gather_unreg) path.
• A 4-GPU Qwen3.5 BF16 RL job with attention DP and MoRI that failed with this error completes two rollouts and four training steps.
• On main, the patched functions and the all-gather path they guard are identical to the tested build, and the patch applies to current main.

Unverified: graph replay on the registered (all_gather_reg) path, which was tested in eager mode only, and runs on MI250X and MI300X.

PR-specific question

The main boundary is that every rank has to skip together. A standard all-gather requires the same input size on every rank, so an empty collective is empty on all ranks of the group and they all return early; no rank waits at a barrier the others skipped. For non-empty inputs the only change is one extra size check.

Changed op tests

None.

Co-author-with: @JessicaJiang-123

[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.

[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 7302 --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.

[aiter-bot]

READY FOR HUMAN REVIEW

PR Self-Review

Revision 5a40117a3741 · mode shadow

Author self-review accepted. Readiness report:

Author answers

• PASS — current-SHA answers are present and structurally specific

Show the author's self-review answers

Behavior and intent

An empty all-gather input now returns immediately instead of launching a kernel with zero blocks. Before, the next kernel that checked for errors failed with `HIP runtime error: invalid configuration argument`, which stopped RL rollouts with attention DP whenever a DP group was idle. Non-empty inputs are unchanged.

Validation and unverified scope

On MI355X with AITER e7d2453f plus this change:

- 72 two-rank API checks pass across FP32, FP16 and BF16, empty and non-empty inputs, registered and unregistered calls, first- and last-dimension gather, and graph replay on the copy-in (`all_gather_unreg`) path.
- A 4-GPU Qwen3.5 BF16 RL job with attention DP and MoRI that failed with this error completes two rollouts and four training steps.
- On main, the patched functions and the all-gather path they guard are identical to the tested build, and the patch applies to current main.

Unverified: graph replay on the registered (`all_gather_reg`) path, which was tested in eager mode only, and runs on MI250X and MI300X.

PR-specific question

The main boundary is that every rank has to skip together. A standard all-gather requires the same input size on every rank, so an empty collective is empty on all ranks of the group and they all return early; no rank waits at a barrier the others skipped. For non-empty inputs the only change is one extra size check.

Changed op tests

None.

Co-author-with: @JessicaJiang-123

PR state

• PASS — pull request is marked ready for review

Copilot findings

• PASS — no active Copilot finding lacks a substantive author reply
• ADVISORY — Copilot could not review the current head (for example, quota or service failure)

Code and op-test hygiene

• PASS — no committed artifact or conflict marker detected

Required CI

• PASS — Black: passed
• PASS — Ruff: passed
• PASS — Repository dependency: passed
• PASS — Aiter Test Gate: passed

_{Shadow mode reports what would block but keeps the status successful. Set AITER_PR_SELF_REVIEW_MODE=enforce only after reviewing real false positives.}

[TennyWang1223]

LGTM