Agent Skills: TorchTalk Analyzer

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:

| Question Pattern | Tool | Example | |---|---|---| | "How does X work?" / "Trace X" | mcp__torchtalk__trace | trace("softmax", "full") | | "Find functions matching X" | mcp__torchtalk__search | search("conv", "CUDA") | | "Where are the CUDA kernels for X?" | mcp__torchtalk__cuda_kernels | cuda_kernels("softmax") | | "What does X call?" | mcp__torchtalk__calls | calls("at::native::add") | | "What calls X?" | mcp__torchtalk__called_by | called_by("at::native::add") | | "What breaks if I change X?" | mcp__torchtalk__impact | impact("at::native::add", 3) | | "How does nn.Linear work?" | mcp__torchtalk__trace_module | trace_module("Linear") | | "List all nn modules" | mcp__torchtalk__list_modules | list_modules("nn") | | "Find tests for X" | mcp__torchtalk__find_similar_tests | find_similar_tests("softmax") | | "What test utilities exist?" | mcp__torchtalk__list_test_utils | list_test_utils("all") | | "What tests are in file X?" | mcp__torchtalk__test_file_info | test_file_info("test_torch") |

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

| Tool | Parameters | Description | |---|---|---| | mcp__torchtalk__trace | function_name, focus? | Trace Python to C++ binding chain. Focus: "full", "yaml", "dispatch" | | mcp__torchtalk__search | query, backend?, limit? | Find bindings by name with optional backend filter | | mcp__torchtalk__cuda_kernels | function_name? | Find GPU kernel launches with file:line |

C++ Call Graph

| Tool | Parameters | Description | |---|---|---| | mcp__torchtalk__impact | function_name, depth? | Transitive callers + Python entry points (depth 1-5) | | mcp__torchtalk__calls | function_name | Functions this function invokes (outbound) | | mcp__torchtalk__called_by | function_name | Functions that invoke this (inbound) |

Python Modules

| Tool | Parameters | Description | |---|---|---| | mcp__torchtalk__trace_module | module_name | Trace torch.nn.Linear, torch.optim.Adam, etc. | | mcp__torchtalk__list_modules | category? | List modules: "nn" (default), "optim", "all", or search query |

Test Infrastructure

| Tool | Parameters | Description | |---|---|---| | mcp__torchtalk__find_similar_tests | query, limit? | Find tests for an operator or concept | | mcp__torchtalk__list_test_utils | category? | List test utilities: "all" (default), "fixtures", "assertions", "decorators" | | mcp__torchtalk__test_file_info | file_path | Details about a specific test file |

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