Imported from TorchedHat/ai-marketplace (torch-compile/skills/pytorch-dynamo/SKILL.md). Install upstream with npx skills add TorchedHat/ai-marketplace --skill pytorch-dynamo. Copyright stays with the author.

PyTorch Dynamo Expert

Expert guidance for working with PyTorch's Dynamo compilation system (torch.compile).

Quick Start

Working with torch.compile? Start here:

• Debugging compilation errors → See debugging-guide.md
• Understanding architecture → See architecture.md
• Implementing features → See common-patterns.md
• Quick commands → See quick-reference.md

Working with pytree operations? See pytree-integration.md

Writing C++ ATen ops for dynamic shapes? See GUARD.md

What is Dynamo?

Dynamo is PyTorch's JIT compiler that intercepts Python bytecode execution to trace and optimize PyTorch models. It powers torch.compile().

Core idea: Intercept Python frame execution → Symbolically execute bytecode → Generate FX graph → Compile with backend

When to Use This Skill

Activate when:

• Implementing Dynamo features or fixes (VariableTracker, opcode handlers, etc.)
• Understanding Dynamo internals (symbolic execution, guards, bytecode)
• Developing new PyTorch compilation support
• Writing tests in test/dynamo/

For user-level debugging (graph breaks, TORCH_LOGS, FX graphs): Use compile-trace-dynamo skill instead. This skill is for Dynamo contributors implementing/fixing Dynamo code.

Implementation Workflow

When implementing a Dynamo feature or fixing an internal issue:

1. Understand the requirement - What behavior needs to change?
2. Locate the code - Use ARCHITECTURE.md to find relevant files
3. Write a test - Add failing test BEFORE implementing (test/dynamo/)
4. Implement - Follow patterns in COMMON-PATTERNS.md
5. Debug - Use DEBUGGING-GUIDE.md for implementation-level debugging
6. Verify - Run new test + related existing tests
7. Review - Check that implementation follows development principles

Implementation Quick Reference

Core Concepts (30-Second Version)

Frame Interception

Dynamo installs a custom frame evaluation hook to intercept Python execution.

File: torch/_dynamo/eval_frame.py

Symbolic Execution

Instead of executing Python bytecode normally, Dynamo executes it symbolically using InstructionTranslator.

File: torch/_dynamo/symbolic_convert.py

Variable Tracking

Python objects are represented as VariableTracker subclasses during symbolic execution.

Directory: torch/_dynamo/variables/

Guard System

Guards ensure compiled code only runs when assumptions hold (e.g., tensor shapes, types).

Files: guards.py, guard_failures.py

Graph Generation

Symbolic execution produces an FX GraphModule that backends can optimize.

File: output_graph.py

Common Tasks

Debug a Compilation Error

1. Create minimal repro
2. Enable logging: torch._logging.set_logs(dynamo=logging.INFO, bytecode=True)
3. Follow workflow in DEBUGGING-GUIDE.md

Add Support for a New Type

1. Create MyTypeVariable in torch/_dynamo/variables/
2. Update VariableBuilder to recognize it
3. See COMMON-PATTERNS.md

Fix a pytree Integration Bug

1. Understand fast-path optimization in PYTREE-INTEGRATION.md
2. Check call_tree_map_branch() implementation
3. Ensure both explicit and implicit registration checks

Key Files Quick Map

torch/_dynamo/
├── eval_frame.py          # Frame interception entry point
├── convert_frame.py       # Decides whether to trace
├── symbolic_convert.py    # InstructionTranslator (main bytecode interpreter)
├── output_graph.py        # Graph builder
├── guards.py              # Guard system
├── variables/
│   ├── base.py           # VariableTracker base class
│   ├── tensor.py         # TensorVariable
│   ├── lists.py          # List/Tuple/NamedTuple variables
│   ├── dicts.py          # Dict variables
│   ├── functions.py      # Function variables (includes tree_map fast-path)
│   └── user_defined.py   # UserDefinedObjectVariable (custom classes)
└── polyfills/
    └── pytree.py         # PyTree polyfills for tracing

Progressive Disclosure

• Getting started: This file
• Understanding internals: ARCHITECTURE.md
• Debugging issues: DEBUGGING-GUIDE.md
• Implementing features: COMMON-PATTERNS.md
• PyTree specifics: PYTREE-INTEGRATION.md
• Guard architecture: GUARD.md
• Quick lookups: QUICK-REFERENCE.md

Development Principles

1. Respect the architecture - Don't create parallel systems
2. Guard everything - Never bypass the guard system
3. Use existing abstractions - Don't invent new Dynamo APIs
4. Maintain consistency - VariableTrackers should behave uniformly
5. Place logic correctly - Opcode handlers belong in specific places

See ARCHITECTURE.md for details.

Getting Help

Compilation error? → DEBUGGING-GUIDE.md Guard failure? → DEBUGGING-GUIDE.md Adding feature? → COMMON-PATTERNS.md PyTree issue? → PYTREE-INTEGRATION.md C++ ATen ops & guards? → GUARD.md Need quick command? → QUICK-REFERENCE.md