Commit 533254a
Migrate check_op_args.h validations to meta kernels and remove it
Summary:
Move operator argument validation from runtime C++ checks (check_op_args.h) to
AOT meta kernels (ops_registrations.py), enabling compile-time error detection
and reducing on-device binary size.
- Delete check_op_args.h (~3,244 lines) and check_macros.h (~175 lines)
- Add torch._check() validation to ~50 meta kernels in ops_registrations.py
covering dtype, dimension, shape, and scalar parameter constraints
- Inline the 4 data-dependent runtime checks (division-by-zero, attention_mask
bounds) directly into kernel implementations across generic, HiFi, and G3
backends
- Remove codegen wrapper generation from codegen.py (generate_wrapper_code,
insert_runtime_context, etc.) since checks no longer need wrapper functions
- Update ops.bzl to remove wrapper.cpp references and wrapper library creation
- Update targets.bzl to remove deleted header references
Reviewed By: abeakkas
Differential Revision: D1026241631 parent ddd8ac6 commit 533254a
3 files changed
Lines changed: 210 additions & 0 deletions
File tree
- backends/cadence
- aot
- fusion_g3/operators
- hifi/operators
0 commit comments