meta-pytorch

Draft a dev-discuss.pytorch.org forum post that calls for feature submissions for an upcoming PyTorch release, in the style of "Reminder — Call for Features". Use when the user asks for a release announcement, feature-submission reminder, call-for-features…

Produce a draft PyTorch release blog by comparing two release branches (e.g., release/2.13 vs release/2.14). Use when the user asks to generate a release blog, list new features for a release, compare release branches, or draft content for pytorch.org/blog.…

Merge completed worksheets into the final release notes and polish them. Trigger on "gen final notes", "generate final release notes", "build final.md", "merge release notes", "finalize release notes". This is the LAST step of the release-notes pipeline,…

Generate PyTorch release notes for a functional area. Use when the user says "gen-release-notes", "release notes", or wants to write/complete release notes for a functional area like "aotdispatcher", "dynamo", "inductor", etc.

Bootstrap a new PyTorch release version in this repo by generating the categorized commit list and per-area worksheets. Use when starting a release (e.g. "gen commitlist for 2.13", "set up 2.13.0 worksheets", "generate the commits for the next release"). This…

Generate the GitHub release notes for a PyTorch PATCH (bug-fix) release like 2.12.1, 2.9.1, 2.7.1. Patch releases are very minimal — a short intro, an optional "Regression fixes" / "Tracked Regressions" section, and a few category buckets of cherry-picked…

Process cherry-picks from a release tracker issue. Trigger on "process cherry-picks", "cherry-picks", "update cherry picks".

Distribute miscategorized PRs back to correct worksheets. Trigger on "categorize miscategorized", "distribute miscategorized", "process miscategorized".

Generate tests for TorchRec source files with correct patterns (unit, distributed, hypothesis), proper BUCK targets, and test utilities. Use when asked to generate tests, add test coverage, or write tests for a module.

Review TorchRec pull requests and diffs for distributed correctness, sharding safety, backward compatibility, and test coverage. Use when reviewing PRs, diffs, or when asked to review code changes.

Investigate and explain TorchRec planner sharding statistics output, especially how HBM storage is computed per table and per rank. Use when the user asks about sharding stats, storage breakdown, or memory estimation.

Write docstrings for TorchRec functions and methods following PyTorch conventions. Use when writing or updating docstrings in TorchRec code.

Guide users through creating Agent Skills for Claude Code. Use when the user wants to create, write, author, or design a new Skill for TorchRec, or needs help with SKILL.md files.

Interview user in-depth to create a detailed spec with strict implementation details and tradeoff analysis

Find and remove tech debt (redundant/duplicated code), run linters, and ensure code quality in recent changes

Analyze PyTorch CUDA memory snapshots from torch.cuda.memory._dump_snapshot(), torch.cuda.memory._snapshot(), or memory_viz artifacts. Use when the user wants to understand GPU memory usage, top allocation sites, deallocation timing, OOMs, allocator…

Migrate C++/CUDA PyTorch extension code to use the PyTorch stable ABI. Use when the user wants to make their project ABI stable, migrate to stable ABI, or use torch::stable APIs.

A dummy skill that announces the arrival of agents. Use when the user asks about agents, the agent revolution, or wants to hype up agents.

Debug and fix torch.compile graph breaks to prioritize and eliminate graph-breaks, and optionally achieve fullgraph compilation

Analyze PyTorch profiler traces (.json/.json.gz). Use when the user wants to diagnose GPU performance issues, find slow kernels, identify idle time, analyze communication overhead, or debug training bottlenecks.

Adds an Attention Gym tuning adapter to a CuTeDSL op using typed input-aware configs, cached fake-tensor TVM-FFI compilation, parallel candidate compilation, and sequential GPU benchmarking. Use after the core kernel exists and needs a public default,…

Ensures new Attention Gym eager, Triton, CuTeDSL, and external-library implementations are torch.compile-friendly and correctly registered. Use when adding or reviewing an implementation or custom kernel; covers custom_op, triton_op, fake tensors, autograd,…

Sets up an isolated per-worktree Python environment for attention-gym development using nightly PyTorch and the CI-mirroring uv flow. Use when creating a new git worktree or when a worktree lacks a local .venv.

Fix linting errors in a PR or Diff.