add pipeline phases in train_pipeline.types (#3443)

Summary:
Pull Request resolved: #3443

# context
* add pipeline phases in types for better describing a pipeline phase
* this can help parameterizing training pipeline options

Reviewed By: spmex

Differential Revision: D77062545

fbshipit-source-id: 92ad98235b59544d27a0bb277ee305888258377a

Author: TroyGarden

torchrec/distributed/train_pipeline/types.py

@@ -96,3 +96,72 @@ class PipelineState(Enum):
     IDLE = 0
     CALL_FWD = 1
     CALL_BWD = 2
+
+    def __str__(self) -> str:
+        return self.name
+
+
+@unique
+class PipelinePhase(Enum):
+    """
+    Pipeline phase for the train pipeline
+
+    please:
+        1. order the phases in the order of execution of base pipeline.
+        2. add notes to explain the phases if needed.
+
+    """
+
+    def __str__(self) -> str:
+        return self.value
+
+    def __eq__(self, obj: "PipelinePhase") -> bool:
+        return self.value == obj.value
+
+    # placeholder for empty
+    NULL = "null"
+
+    # usually the data is first available on CPU when loading from dataloader
+    # need to move/copy the input batch to device if using GPU training
+    COPY_BATCH_TO_DEVICE = "copy_batch_to_device"
+
+    # input post processing is needed for sparse data dist pipeline, where the sparse features
+    # are traced (built) from the ModelInput via fx tracing
+    INPUT_POST_PROC = "input_post_proc"
+
+    # the sparse features (AKA, KJTs) are in a jagged format so the data size are unknown to
+    # other ranks. so a comms is needed to exchange the data size info, i.e., the splits
+    INPUT_SPLITS_DIST = "input_splits_dist"
+
+    # once a rank knows the data size from other ranks (via splits dist), it can initialize
+    # a all-to-all comms to exchange the actual data of the sparse features
+    # NOTE: the splits have to be available on the host side
+    INPUT_DATA_DIST = "input_data_dist"
+
+    # embedding lookup is done in FBGEMM.TBE on each rank
+    EMBEDDING_LOOKUP = "embedding_lookup"
+
+    # the embedding lookup results (i.e., the embeddings) are needed in each rank, it's often done
+    # with the output dist with an all_to_all comms
+    EMBEDDING_OUTPUT_DIST = "embedding_output_dist"
+
+    # A typical DLRM model arch contains sparse arch and dense arch, here we treat the model excluding
+    # "sparse modules" as dense part. It actually also includes the dense-sharded embedding tables.
+    DENSE_FORWARD = "dense_forward"
+
+    # model's backward usually uses torch.autograd, the embedding modules' backward is handled by TBE
+    DENSE_BACKWARD = "dense_backward"
+
+    # on each rank, after dense arch's backward, the gradients are available for the embedding tables
+    # a backward of "embedding output dist" is needed to gather the embedding gradients from all ranks
+    # to the rank where the embedding table is hosted.
+    EMBEDDING_GRAD_DIST = "embedding_grad_dist"
+
+    # TBE backward usually update the embedding table weights inplace
+    EMBEDDING_BACKWARD = "embedding_backward"
+
+    # we decouple the embedding update from backward just in case the change is not coupled
+    EMBEDDING_UPDATE = "embedding_update"
+
+    # the optimizer step usually only includes the dense module weights
+    DENSE_OPTIMIZER_STEP = "dense_optimizer_step"