TorchTalk Analyzer

This skill enables cross-language analysis of PyTorch internals by leveraging the TorchTalk MCP server. It traces binding chains from Python through C++ to CUDA, analyzes dispatch mechanisms, maps call graphs, and locates test infrastructure.

When to Use

• When asked how a PyTorch operator works internally (e.g., "How does torch.add dispatch to CUDA?")
• When investigating where a function is implemented across CPU/CUDA/MPS backends
• When assessing what would break if a C++ function is modified (impact analysis)
• When tracing how torch.nn modules connect to native ATen operators
• When finding existing tests for a PyTorch operator or function
• When exploring PyTorch's dispatch architecture or autograd integration

Prerequisites

• TorchTalk MCP server must be running and registered with Claude Code
• PyTorch source code must be available locally
• Run /torchtalk:setup if TorchTalk is not yet installed

Verify availability:

mcp__torchtalk__get_status

If the status tool returns data, all tools below are ready.

Instructions

Step 1 - Verify MCP Server

Before using any tools, confirm the TorchTalk server is running:

mcp__torchtalk__get_status

Check that:

• Bindings are loaded (should show thousands of bindings)
• Native functions are parsed
• C++ call graph status is "Ready" (required for impact/calls/called_by)
• Python modules are loaded (required for trace_module/list_modules)

If the server is not available, direct the user to run /torchtalk:setup.

Step 2 - Identify the Analysis Type

Match the user's question to the appropriate tool:

Step 3 - Execute and Synthesize

For simple lookups, a single tool call suffices. For deeper questions, combine multiple tools:

"How does torch.softmax work end-to-end?"

1. mcp__torchtalk__trace("softmax", "full") - Get the binding chain
2. mcp__torchtalk__cuda_kernels("softmax") - Find GPU kernels
3. mcp__torchtalk__calls("at::native::softmax") - See internal dependencies

"What breaks if I modify at::native::add?"

1. mcp__torchtalk__impact("at::native::add", 3) - Transitive callers
2. mcp__torchtalk__find_similar_tests("add") - Affected tests

"How does nn.Linear connect to native code?"

1. mcp__torchtalk__trace_module("Linear") - Module definition
2. mcp__torchtalk__trace("linear", "full") - Native operator chain

Step 4 - Present Results

Format results with:

• Clear layer separation (Python -> YAML -> C++ -> CUDA)
• file:line references for every implementation location
• Architectural context explaining why the dispatch is structured this way
• Suggestions for further exploration if the user wants to go deeper

MCP Tools Reference

ATen Operators

C++ Call Graph

Python Modules

Test Infrastructure

Error Handling

• MCP server not running: Direct user to /torchtalk:setup
• "C++ call graph not available": PyTorch needs to be built once (python setup.py develop) to generate compile_commands.json
• "Function not found": Use mcp__torchtalk__search with partial names, or check spelling
• "Test infrastructure not loaded": Verify PyTorch source path is correct with torchtalk status