[voltjia]

Summary

• Update the submodules/InfiniOps gitlink to the merged vLLM-prefix-aligned NineToothed fused Q/K normalization revision.
• Keep this manifest PR to a single component revision change and introduce no component source copies or build logic.

Revision Update

Component	Old revision	New revision	Upstream change
submodules/InfiniOps	640475a658de1f26bbf7c018fd0e0f8d5303a74f	1a68b261d19f0ed6f76bbabeab7a25f479f05e07	InfiniOps #1023

The new revision adds FusedQkNorm, including separate Q/K weights, packed-QKV in-place semantics, and an untouched V section. Its argument order is the direct prefix of vLLM's fused_qk_norm_rope API; RoPE remains a separate operation in this performance path.

Merge Status

InfiniOps #1023 was merged as master commit 1a68b261d19f0ed6f76bbabeab7a25f479f05e07. This manifest now pins that final master revision rather than the PR branch head.

Validation

Manifest validation on InfiniCore head 03db63ae:

git diff --check                 passed
git ls-files --stage submodules  passed
git submodule sync --recursive   passed
git submodule status --recursive passed

The recursive status confirms:

• submodules/InfiniOps is a gitlink at 1a68b261d19f0ed6f76bbabeab7a25f479f05e07.
• submodules/InfiniOps/.ci is initialized at b45d360c.
• submodules/InfiniCCL and submodules/InfiniRT remain at their existing revisions.

Component validation for InfiniOps #1023:

• Full and explicit-op Release builds/installs with CPU, NVIDIA, Torch, and NineToothed enabled.
• A100-SXM4-80GB, CUDA 12.6.77, driver 580.105.08.
• pytest -q tests/test_fused_qk_norm.py --devices cuda: 2 passed.
• Ruff, clang-format, git diff --check, and all configured InfiniOps CI and CI v2 shadow matrix platforms passed on PR head 77cd5ea7.
• GitHub merged the reviewed PR as final master commit 1a68b261.

Downstream InfiniLM validation:

• Full InfiniLM Release build/install with C++11 ABI=1 passed on the same A100.
• One-layer BF16 Qwen3 prefill plus decode smoke passed.
• INFINI_OPS_TRACE_CALLS=1 recorded FusedQkNorm, NVIDIA, implementation 9 during Qwen3 forward.
• Long-input prefill comparison against the prior merged-QK stack showed a 13.2% median TTFT reduction at 1024, approximately parity at 4096 and 8192, and no regression outside run-to-run variance.
• The measurements used the reviewed issue/1022 - patch metax hpcc hrc include #1023 operator patch before its final squash merge. The operator patch was unchanged by rebases; the only later PR-head addition was the NVIDIA CI runtime PyPI mirror setting, which is preserved in the merge commit.

Checks Not Run

• InfiniCore has no native build or test suite, so no repository-local unit tests are applicable.
• No non-NVIDIA InfiniLM integration was run; the downstream fast path is NVIDIA-only and other configurations retain their existing path.