torchtalk-analyzer

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

get_status()  # Check what's loaded and available

Tools

Tool	Use For
trace(name, focus?)	Trace any PyTorch op: Python → YAML → C++ → file:line
search(query, mode?, backend?)	mode="bindings": find dispatch registrations. mode="kernels": find CUDA kernel launches
graph(name, mode?, depth?, fuzzy_all_levels?, walk_python?, focus?)	mode="callers": inbound. mode="calls": outbound. mode="impact": transitive callers (depth/walk_python/focus apply to impact only)
modules(name, mode?, focus?)	mode="trace": class details (focus="full" adds bases/docstring). mode="list": browse by category ("nn", "optim", "all")
tests(query?, mode?, limit?, focus?)	mode="find": search tests (focus narrows to functions/classes/files). mode="utils": list utilities (query/focus ignored). mode="file_info": test file details
affected(funcs, depth?)	Map changed C++ functions (comma-separated) to impacted Python test files

Common Workflows

"How does torch.matmul work?"

trace("matmul")                    # Get binding chain
graph("matmul", mode="calls")     # See internal dependencies

"What breaks if I modify GEMM?"

graph("gemm", mode="impact", depth=4)

"Where is conv2d for CUDA?"

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

"How does nn.Linear work?"

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

"Find tests for softmax"

tests("softmax")

Requirements

• Always available: trace, search, modules, tests
• Requires PyTorch build: graph (needs compile_commands.json)

Run get_status() to check availability.