PyTorch adds DELETE_DEREF handler in Dynamo
PyTorch's Dynamo just got a critical bug fix for cell variable deletions in Python bytecode
PyTorch's Dynamo JIT compiler just received a critical bug fix that addresses how it handles Python's `del x` operation on cell and free variables. Previously, Dynamo lacked a handler for the DELETE_DEREF bytecode, which is emitted when deleting variables captured by nested functions or free variables.
The fix (PR #189216) implements this handler to match CPython's behavior, where DELETE_DEREF clears the cell's contents by setting it to NULL (`PyCell_SET(cell, NULL)`) while keeping the cell object itself alive. This matters because cell variables are commonly used in closures and nested functions, and improper handling could lead to memory leaks or incorrect behavior in optimized code.
The change was approved by Dynamo maintainer Ralf Timpe and merged into PyTorch's trunk. This fix is particularly important for developers working with Python decorators, nested functions, or any code that relies on cell variable cleanup.
- PyTorch's Dynamo compiler lacked DELETE_DEREF handling for cell/free variables in Python bytecode
- New PR #189216 implements CPython-compatible behavior by clearing cell contents while preserving cell objects
- Critical for proper cleanup in nested functions, closures, and decorator patterns
Why It Matters
Prevents memory leaks and undefined behavior in Python code optimized by PyTorch's Dynamo compiler