[autoparallel]integrate auto parallel feature with new tracer (#3408)

* [autoparallel] integrate new analyzer in module level

* unify the profiling method

* polish

* fix no codegen bug

* fix pass bug

* fix liveness test

* polish
This commit is contained in:
YuliangLiu0306
2023-04-04 17:40:45 +08:00
committed by GitHub
parent 573af84184
commit ffcdbf0f65
46 changed files with 396 additions and 470 deletions

View File

@@ -112,7 +112,7 @@ class MetaInfo:
# should keep the same whenever manipulated
# ============================= Invariant ==================================
to_recompute: Tuple[torch.Tensor] = () # (region_0, region_1, ...) support nested codegen
activation_checkpoint: Tuple[torch.Tensor] = () # (region_0, region_1, ...) support nested codegen
to_offload: Optional[bool] = False
sharding_spec: str = 'RR'