ai-marketplace

Refactor a PyTorch test file to be device-agnostic. Walks through analyzing test classes, classifying them (GENERIC/DEVICE_GENERIC/DEVICE_SPECIFIC/MULTI_DEVICE_GENERIC/MULTI_DEVICE_SPECIFIC), splitting mixed classes, adding hw_classification attributes, converting hardcoded device references, and verifying the refactoring. Use when refactoring any test file in the PyTorch repo for the device-agnostic testing initiative.

Reference for PyTorch distributed training hang patterns, NCCL communicator architecture, flight recorder output format, debug logging levels, and timeout configuration. Covers barrier deadlocks, p2p hangs after communicator abort, init timeouts from store, collective mismatches, and DDP unused parameter hangs. Includes ARCHITECTURE.md with source-level ProcessGroupNCCL internals (communicator lifecycle, watchdog, work queue, store-based init). Use as knowledge base for diagnosing why distributed training freezes.

Test skill that runs a script from inside Claude. Use to verify that Claude can execute plugin-bundled scripts via ${CLAUDE_PLUGIN_ROOT}.

Expert guidance for vLLM's custom compiler - focusing on @support_torch_compile decorator, vllmBackend/Inductor Passes, PiecewiseBackend, and CudaGraphWrapper. Use when debugging torch.compile issues within vllm, implementing fusion passes, configuring graph splitting, investigating guard dropping, or working with CUDA graph capture in vLLM.

Create and manage Claude Code skills following Anthropic best practices. Use when creating new skills, modifying skill-rules.json, understanding trigger patterns, working with hooks, debugging skill activation, or implementing progressive disclosure. Covers skill structure, YAML frontmatter, trigger types (keywords, intent patterns, file paths, content patterns), enforcement levels (block, suggest, warn), hook mechanisms (UserPromptSubmit, PreToolUse), session tracking, and the 500-line rule.

Create new Claude Code plugins from scratch with proper directory structure (.claude-plugin/plugin.json, skills/, agents/), or add existing plugins to marketplace.json. Use when creating a plugin, setting up plugin metadata, configuring MCP servers, or adding plugins to a marketplace.

Create and update agent definition files following Anthropic best practices. Use when writing agent files, creating subagents, designing agent systems, or troubleshooting agent discovery and delegation issues.

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

Use PyTorch's compiler bisector to automatically find which backend/subsystem/operation causes compilation failures. Binary searches through backends (eager → aot_eager → inductor) and subsystems (passes, lowerings, etc.) to pinpoint exact failing operations. Outputs backend/subsystem/debug_info that routes you to the right stage-specific skill (compile-trace-dynamo for eager, compile-trace-aot for aot_*, compile-trace-inductor for inductor).

Reference documentation for torch.compile pipeline architecture, IR levels (Full ATen, Core ATen, Prims), stages, TORCH_LOGS flags, output files, and debugging tools. Use for understanding pipeline structure, operator IRs, and available debugging options.

Debug PyTorch AOT Autograd stage - functionalization, decompositions, IR transformations, joint forward+backward graph (when requires_grad=True), partitioning/recomputation, and post-grad passes. Use for tracing AOT stage and understanding decomposition application.

Debug PyTorch Dynamo stage - bytecode capture, FX graph construction, graph breaks, and pre-grad passes. Covers TORCH_LOGS for dynamo/graph_breaks/pre_grad_graphs, interpreting FX graph files, understanding graph break reasons, and pre-grad fusion patterns (Conv-BN, split-cat). Load after compile-bisect indicates backend='eager'.

Debug PyTorch Inductor compiler backend - IR lowering, scheduler/fusion, loopbody ops, and Triton/C++ codegen. Covers TORCH_LOGS for fusion/schedule/ir_post_fusion/output_code, config.trace.enabled for IR dumps, interpreting Inductor IR nodes (Pointwise/Reduction/etc), fusion decisions, kernel generation, and performance optimization. Load after compile-bisect indicates backend='inductor'.

Expert guidance for PyTorch functorch and AOT Autograd development. Covers torch/_functorch architecture, AOT Autograd pipeline (functionalization, IR transformations - Full ATen → Core ATen, joint graphs, partitioning, post-grad passes), vmap/batching, functional transforms (grad, vjp, jvp, jacrev, jacfwd), make_functional, functional_call, activation checkpointing, decompositions, and implementation patterns. Use for implementing AOT features, understanding functorch internals, Core ATen IR creation, and adding functionalization/partitioning support. For debugging, use compile-trace-aot.

Expert guidance for PyTorch Dynamo (torch.compile) development and debugging. Covers symbolic execution, VariableTracker system, pytree integration, guard failures, guard generation, FakeTensors, SymInts, symbolic shapes, dynamic shapes, C++ ATen ops, .size() vs .sym_size(), .numel() vs .sym_numel(), and bytecode tracing.

Expert guidance for PyTorch Inductor compiler backend development and optimization. Covers FX graph lowering, decomposition→lowering pipeline, kernel fusion, Triton codegen, TritonTemplate, TritonTemplateKernel, template system, Jinja2 templates, C++ codegen, scheduling, memory planning, select_algorithm, autotuning, IR nodes (Buffer, Pointwise, Reduction), and performance optimization. Use for understanding Inductor architecture, lowering vs decomposition order, and performance optimization.