[voltjia]

Summary

• Update third_party/InfiniCore from 2b4f2f20584a69b074f62a455e3de7368e3446c2 to a6ba44fa568244c5dcc0f12b799f609483c06205.
• The new InfiniCore revision pins InfiniOps f937f49c181196c191993c27c47f3c30b5cb1739, which contains the NineToothed NVIDIA FusedAddRmsNorm provider.

Motivation

This is the InfiniLM dependency-stack step required to expose the faster residual-norm provider to the opt-in InfiniOps auto-tuning mechanism initialized by #605.

Closes N/A - dependency-stack change; no separate issue exists.

Type of Change

• feat — new feature / new model
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change

Test Results of Involved Models on Supported Platforms (Please attach screenshots)

Platform	Result
NVIDIA A100 / CUDA 12.6	Dependency build and runtime validation passed

Validation evidence:

InfiniOps f937f49c:
- CUDA smoke wheel build with WITH_NINETOOTHED=ON: passed
- Full CPU/NVIDIA/Torch/Flash-Inference/W4-Marlin install build: passed
- Focused FusedAddRmsNorm suite: 180 passed, 52 skipped
- InfiniOps NVIDIA CI: passed

InfiniLM:
- _infinilm build against the dependent full InfiniOps prefix: passed
- Native nvidia:0 runtime creation: passed
- Operator tuning initialization smoke from PR #605: passed
- Host-provided tuning cache regression from PR #605: pre-fix runtime failed as expected; fixed runtime passed

The InfiniOps provider benchmark motivating this dependency is documented in #605.

Benchmark / Performance Impact

N/A - this PR changes only the dependency revision and introduces no InfiniLM code. The referenced A100 (8192, 1024) BF16 provider benchmark measured 0.157911 ms for native index 0 and 0.093901 ms for NineToothed index 9.

Notes for Reviewers

This PR intentionally uses perf/initialize-ops-auto-tuning as its base and should be merged after InfiniLM #605. Its InfiniOps #1014 dependency is already merged. The default InfiniOps implementation selection remains unchanged; selection changes only when the existing opt-in tuning environment variables are enabled.

CI / ChatOps

Host behavior, Ruff, and format checks passed on rebased head 0295210. Native CI run 41965348600 was dispatched on this head and remains queued while the configured InfiniLM runner pool is offline to GitHub Actions. The dependent NVIDIA build and runtime validations above passed.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz.
• Each commit message follows Conventional Commits.
• Small PR is a single squashable commit.
• No stray merge commits from main.
• No fixup! / squash! / wip commits remain.
• N/A - Existing PR/branch/commit does not use the legacy issue format.

Scope and Design

• Changes are minimal and limited to the dependency manifest.
• No dead code, debug output, or TODOs.
• No unrelated formatting churn.
• N/A - No public InfiniLM API changes.

General Code Hygiene

• N/A - No source files were modified.
• N/A - No source-file newline changes were needed.
• N/A - No source whitespace or BOM changes were needed.
• N/A - No comments or identifiers were added.

C++ Specific

• N/A - No C++ files were changed.

Python Specific

• N/A - No Python files were changed.

Testing

• N/A - No requested platform was unavailable.
• N/A - Single request test: manifest-only dependency update; no model code path changed.
• N/A - Offline performance test: manifest-only dependency update; provider performance is documented by perf: initialize InfiniOps auto tuning #605.
• N/A - Sanity test: the dependent InfiniOps suite and InfiniLM native runtime smoke passed.
• N/A - Service test: manifest-only update with no service-layer changes.

Build, CI, and Tooling

• Dependency manifest and recursive submodule state were checked locally.
• The dependent project builds cleanly with the pinned component revision on NVIDIA.
• CI was triggered manually.

Documentation

• N/A - No developer workflow, user-facing behavior, or build flag changed beyond the dependency pin.
• N/A - No breaking change.

Security and Safety

• No secrets, tokens, internal URLs, customer data, or personal hardware identifiers were committed.
• No third-party code was copied into this repository.
• N/A - No executable InfiniLM code was changed.

[voltjia]

#606 was merged into its old stack base branch perf/initialize-ops-auto-tuning, so its dependency pin did not reach main.

Follow-up PR #609 re-targets the reviewed result at main:

• build: pin NineToothed fused add RMSNorm #609 base: main
• build: pin NineToothed fused add RMSNorm #609 head: perf/initialize-ops-auto-tuning@711054c
• Diff: only the intended third_party/InfiniCore gitlink change from build: pin NineToothed fused_add_rms_norm #606

#607 has also been rewritten as a clean single commit based on #609 head 711054c, and is mergeable again.