[mattwittwer]

What does the PR do?

Serializes torch::inductor::AOTIModelPackageLoader construction in the pt2 path behind a process-wide mutex.

Checklist

• PR title reflects the change and is of format <commit_type>: <Title>
• Changes are described in the pull request.
• Related issues are referenced.
• Populated github labels field
• Added test plan and verified test passes.
• Verified that the PR passes existing CI.
• Verified copyright is correct on all changed files.
• Added succinct git squash message before merging ref.
• All template sections are filled out.
• Optional: Additional screenshots for behavior/output changes with before/after.

Commit Type:

Check the conventional commit type
box here and add the label to the github PR.

• build
• ci
• docs
• feat
• fix
• perf
• refactor
• revert
• style
• test

Related PRs:

Where should the reviewer start?

Test plan:

• CI Pipeline ID:

62435323

Caveats:

Background

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• closes GitHub issue: #xxx

[greptile-apps]

Greptile Summary

The PR serializes AOTI package-loader construction to prevent concurrent access to libtorch’s shared JSON parser state.

• Adds a process-wide mutex for package-loader construction.
• Limits the critical section to TorchModelLoader construction for both CPU and GPU models.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
src/pt2/model_state.cc	Adds a narrowly scoped constructor lock that addresses concurrent metadata parsing without serializing subsequent call-spec processing.

_{Reviews (2): Last reviewed commit: "narrow mutex scope" | Re-trigger Greptile}

[whoisj]

Can only a single model be loaded safely at a time?

[mattwittwer]

Can only a single model be loaded safely at a time?

@whoisj - Good clarification, model loading overall is still parallel. The mutex covers only the construction of torch::inductor::AOTIModelPackageLoader. Non-AOTI models are unaffected, and for AOTI models everything after construction (call-spec parsing, instance setup) runs outside the lock.

The issue is in libtorch here: torch/csrc/inductor/aoti_package/model_package_loader.cpp this parses each package's metadata into a function-local static nlohmann::json and returns a reference to it, so two loaders constructing concurrently corrupt each other's view.