viable/strict/1775219460: Add reuse_hash_fn escape hatch for nested_compile_region (#178239)
New parameter bypasses automatic fingerprinting, enabling O(1) cache lookup for subgraph reuse in Dynamo.
The PyTorch team has merged a significant pull request (#178239) that introduces a new `reuse_hash_fn` parameter to the `nested_compile_region` function. This technical enhancement provides developers with an "escape hatch" that bypasses PyTorch's automatic fingerprint and guard machinery for subgraph reuse. When this parameter is provided, Dynamo traces the hash function with guard installation suppressed (via `skip_guard_install`) to generate a constant integer key, enabling O(1) cache lookups instead of potentially expensive fingerprint calculations.
This change addresses performance bottlenecks in nested compilation scenarios where automatic fingerprinting could become computationally expensive. The implementation includes clear error handling—graph breaks within the hash function now raise explicit errors to help developers debug their implementations. Approved by core maintainers including aorenste and zou3519, this enhancement gives advanced users more fine-grained control over PyTorch's compilation caching behavior while maintaining backward compatibility for existing code.
- Adds `reuse_hash_fn` parameter to bypass automatic fingerprint/guard machinery for subgraph reuse
- Enables O(1) cache lookup via constant integer key from traced hash function
- Provides clear error messages for graph breaks in hash function implementations
Why It Matters
Gives PyTorch developers more control over compilation performance, potentially speeding up nested model compilation scenarios.