AgentSkillsCN

torchtalk-analyzer

深入分析 PyTorch 的内部机制，涵盖 Python、C++ 以及 CUDA 层面。适用于解答关于 PyTorch 算子内部运作原理、函数实现位置（CPU/CUDA 后端）、代码修改可能引发的故障（影响分析）、torch.nn 模块如何与原生代码对接，或为 PyTorch 算子寻找测试用例等问题时使用。内容覆盖 ATen 操作、nn.Module 类、分发机制以及测试基础设施。

SKILL.md
--- frontmatter
name: torchtalk-analyzer
description: Analyze PyTorch internals across Python, C++, and CUDA layers. Use when asked about how PyTorch operators work internally, where functions are implemented (CPU/CUDA backends), what would break if code is modified (impact analysis), how torch.nn modules connect to native code, or finding tests for PyTorch operators. Covers ATen ops, nn.Module classes, dispatch mechanisms, and test infrastructure.
allowed-tools: mcp__torchtalk__get_status, mcp__torchtalk__trace, mcp__torchtalk__search, mcp__torchtalk__impact, mcp__torchtalk__calls, mcp__torchtalk__called_by, mcp__torchtalk__cuda_kernels, mcp__torchtalk__trace_module, mcp__torchtalk__list_modules, mcp__torchtalk__find_similar_tests, mcp__torchtalk__list_test_utils, mcp__torchtalk__test_file_info, Read, Grep, Glob

TorchTalk PyTorch Analyzer

When to Use

  • •"How does torch.X work internally?"
  • •"Where is X implemented?" (Python/C++/CUDA)
  • •"What would break if I change X?"
  • •"How does nn.Linear connect to native code?"
  • •"Find tests for the softmax operator"

Quick Start

code
get_status()  # Check what's loaded and available

Tools by Category

ATen Operators (torch.add, torch.matmul, etc.)

ToolUse For
trace(name, focus?)Python → YAML → C++ → file:line
search(query, backend?)Find ops by name, filter by CPU/CUDA
cuda_kernels(name?)GPU kernel launches

Call Graph Analysis (requires PyTorch build)

ToolUse For
impact(name, depth?)What breaks if I change this? (transitive callers)
calls(name)What does this function call?
called_by(name)What calls this function?

Python Modules (torch.nn, torch.optim)

ToolUse For
trace_module(name)Trace nn.Linear, optim.Adam, etc.
list_modules(category)List available modules ("nn", "optim", "all")

Test Infrastructure

ToolUse For
find_similar_tests(query)Find tests for an operator/concept
list_test_utils(category)Available test utilities and patterns
test_file_info(path)Details about a specific test file

Common Workflows

"How does torch.matmul work?"

code
trace("matmul")           # Get binding chain
calls("matmul")           # See internal dependencies

"What breaks if I modify GEMM?"

code
impact("gemm", depth=4)   # Transitive callers + Python entry points

"Where is conv2d for CUDA?"

code
search("conv2d", backend="CUDA")
# or
trace("conv2d", focus="dispatch")

"How does nn.Linear work?"

code
trace_module("Linear")    # Python class details
trace("linear")           # Underlying ATen op

"Find tests for softmax"

code
find_similar_tests("softmax")

Requirements

  • •Always available: trace, search, cuda_kernels, trace_module, list_modules, test tools
  • •Requires PyTorch build: impact, calls, called_by (need compile_commands.json)

Run get_status() to check availability.