[shantipriya-amd]

Motivation

Repins the 3rdparty/composable_kernel submodule to a valid, fetchable upstream commit, fixing the CI submodule-sync breakage that led to #6142 reverting #5708.

Technical Details

#5708 originally pinned 3rdparty/composable_kernel to 1f66d862231703a0f354e5771e58573139bb6efb — a commit that existed only in a local CK checkout, never pushed to any branch on ROCm/composable_kernel. This made the pin unfetchable and broke CI's submodule sync, so #6142 reverted it back to the last-known-good pin, af9e1d1f1ae347c22feeb08fd2d42645075e0c5d.

The two commits that were stacked on the bad pin (Add MoeRelu2 activation functor for gate-only MoE flatmm kernel and Fix half_t intrinsics (neg/sin/exp/log) via float round-trip for gfx950) have since been properly submitted and merged upstream via ROCm/rocm-libraries#13071, landing on CK's develop branch at commit 96afae1d54fc79c26857de7e1683b0dde1ff3c6f.

This PR repins 3rdparty/composable_kernel from af9e1d1f1ae347c22feeb08fd2d42645075e0c5d to 96afae1d54fc79c26857de7e1683b0dde1ff3c6f.

Note on scope (updated after review feedback): this repin advances the submodule to the current tip of CK's develop branch, not a hand-picked commit isolating only the two fixes above. Since develop is a shared branch, the pin advance necessarily carries 132 other commits that landed upstream between the old pin (af9e1d1f) and now — build/CI changes, gfx1250 enablement, FMHA performance tuning, dependency bumps, docs, etc. — none of which touch MoE/relu2/half_t code. This is standard/expected behavior for a submodule bump against a tracked branch (there is no way to cherry-pick just two commits onto the old base and still call it "develop"). No functional changes to aiter's own source are introduced by this PR itself — only the submodule pointer.

Test Plan

• Confirmed the target commit (96afae1d54fc79c26857de7e1683b0dde1ff3c6f) is reachable via git fetch origin against ROCm/composable_kernel and resolves cleanly with git checkout.
• Confirmed git submodule status reports a clean, non-dirty pin after the update (no +/- prefix indicating divergence).
• Confirmed via git log --oneline af9e1d1f1ae347c22feeb08fd2d42645075e0c5d..96afae1d54fc79c26857de7e1683b0dde1ff3c6f that the 132 intervening commits are all unrelated upstream CK work (gfx1250 enablement, FMHA perf, build/CI, dependency bumps) with no overlap with MoE/relu2/half_t code paths.
• No source-level testing needed for aiter itself in this PR, since it only changes the submodule pointer; the two upstreamed fixes' own functional correctness (relu2 CK-Tile kernel, half_t intrinsics fix) was already validated on real gfx950 hardware as part of [HIP] [CK] Add relu2 (Squared ReLU) activation support to CK-Tile fused MoE GEMM #5708's test plan and will be re-exercised end-to-end when the relu2 Tier 2 PR is resubmitted on top of this repin.
• Regression coverage for the other 131 commits relies on this PR's own CI (aiter op tests on gfx942/gfx950) passing at the new pin, which is standard practice for submodule bumps of this kind rather than a commit-by-commit audit.

Test Result

git submodule status 3rdparty/composable_kernel after this change reports:

96afae1d54fc79c26857de7e1683b0dde1ff3c6f 3rdparty/composable_kernel (...)

with no dirty/divergence marker — the pin is clean and fetchable.

git log --oneline af9e1d1f1ae..96afae1d54f | wc -l → 132 commits, all confirmed unrelated to MoE/relu2/half_t.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

[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
multigpu	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: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
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 6178 --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]

Copilot review overview

🔵 Needs a closer look

The broad CK upgrade requires expanded documentation and regression validation beyond the two stated fixes.

Review effort: Balanced
Findings: None

What changed in this PR

Repins CK to a fetchable upstream commit, but also incorporates 132 CK commits rather than only the two described fixes.

Changes:

• Updates the CK submodule revision.
• Restores upstream Relu2 and half_t fixes.

File	Description
3rdparty/​composable_kernel	Advances the CK submodule pin.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[shantipriya-amd]

@yzhou103: Could one of you review this ?. It repins 3rdparty/composable_kernel from the last-known-good commit to 96afae1d54fc79c26857de7e1683b0dde1ff3c6f on CK's develop branch, now that the two orphaned commits (MoeRelu2 functor + gfx950 half_t intrinsics fix) have been properly upstreamed and merged via rocm-libraries#13071.

[shantipriya-amd]

@valarLip : Can you please review at your convenient time ?

[Copilot]

🔵 Needs a closer look

The pin incorporates 132 upstream commits, and the broad regression CI remains incomplete.

0 open findings

🧠 Review effort: Balanced