Add Sharding V1 broadcast and V2 allgather overlap optimize

llm/finetune_generation.py

@@ -51,6 +51,9 @@
 )
 from paddlenlp.utils.log import logger
 
+# Fine-tune Environment Variables to support sharding stage1 overlap optimization.
+os.environ["USE_CASUAL_MASK"] = "False"
+
 
 def add_start_docstrings(*docstr):
     def docstring_decorator(fn):

llm/run_pretrain.py

@@ -48,6 +48,9 @@
 from paddlenlp.utils.log import logger
 from paddlenlp.utils.tools import get_env_device
 
+# Pretaining Environment Variables to support sharding stage1 overlap optimization.
+os.environ["USE_CASUAL_MASK"] = "True"
+
 
 def add_start_docstrings(*docstr):
     def docstring_decorator(fn):

paddlenlp/trainer/trainer.py

@@ -1892,7 +1892,6 @@ def get_expected_keys(inputs, keys):
                     optimizer._set_broadcast_overlap(True, model)
 
                 self.optimizer = optimizer
-
         # pure tesnor parallel mode, no pipeline_parallel, no sharding.
         if (
             not in_pipeline_parallel_mode
@@ -1908,6 +1907,21 @@ def get_expected_keys(inputs, keys):
                 self.optimizer = mix_precision_utils.MixPrecisionOptimizer(self.optimizer)
             self.optimizer = fleet.distributed_optimizer(self.optimizer)
 
+        # stage1 has v1 and v2 version
+        if in_sharding_parallel_mode and ShardingOption.SHARD_OP in self.args.sharding:
+            if "split_param" in self.args.sharding_parallel_config:
+                if (
+                    hasattr(self.optimizer, "_set_all_gather_overlap_forward")
+                    and "enable_stage1_allgather_overlap" in self.args.sharding_parallel_config
+                ):
+                    self.optimizer._set_all_gather_overlap_forward(True, model)
+            else:
+                if (
+                    hasattr(self.optimizer, "_set_broadcast_overlap")
+                    and "enable_stage1_broadcast_overlap" in self.args.sharding_parallel_config
+                ):
+                    self.optimizer._set_broadcast_overlap(True, model)
+
         return model
 
     def _prepare_input(self, data: Union[paddle.Tensor, Any]) -> Union[paddle.Tensor, Any]:

paddlenlp/trainer/training_args.py

@@ -263,6 +263,8 @@ class TrainingArguments:
               enable_stage1_tensor_fusion, fuse small tensors into big tensor chunks to accelerate communications, may increase memory occupation
               enable_stage1_overlap, fuse small tensors into big tensor chunks to accelerate communications and do communication overlap with backward computation, may harm the backward speed
               enable_stage2_overlap, overlap stage2 NCCL communication with computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for broadcast overlap and no other sync could be called during the training for broadcast overlap.
+              enable_stage1_broadcast_overlap, overlap stage1 V1 broadcast with next step forward computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for broadcast overlap forward compute and no other sync could be called during the training for broadcast overlap.
+              enable_stage1_allgather_overlap, overlap stage1 V2 allgather with next step forward computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for allgather overlap forward compute and no other sync could be called during the training for allgather overlap.
               disable_stage1_reduce_avg, replace reduce_avg with original reduce_sum+scale in stage1, which can be used for accuracy verification.
         recompute (`bool`, *optional*, defaults to `False`):
             Recompute the forward pass to calculate gradients. Used for saving memory.
@@ -647,7 +649,9 @@ class TrainingArguments:
                 "enable_stage1_tensor_fusion, fuse small tensors into big tensor chunks to accelerate communications, may increase memory occupation\n"
                 "enable_stage1_overlap, fuse small tensors into big tensor chunks to accelerate communications and do communication overlap with backward computation, may harm the backward speed\n"
                 "disable_stage1_reduce_avg, replace reduce_avg with original reduce_sum+scale in stage1, which can be used for accuracy verification.\n"
-                "enable_stage2_overlap, overlap stage2 NCCL communication with computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for broadcast overlap and no other sync could be called during the training for broadcast overlap"
+                "enable_stage2_overlap, overlap stage2 NCCL communication with computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for broadcast overlap and no other sync could be called during the training for broadcast overlap\n"
+                "enable_stage1_broadcast_overlap, overlap stage1 V1 broadcast with next step forward computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for broadcast overlap forward compute and no other sync could be called during the training for broadcast overlap.\n"
+                "enable_stage1_allgather_overlap, overlap stage1 V2 allgather with next step forward computation. There are some constraints for the overlap, such as the logging_step should be bigger than 1 for allgather overlap forward compute and no other sync could be called during the training for allgather overlap."
             )
         },
     )
@@ -1186,10 +1190,12 @@ def is_segment_parallel_supported():
                                 "enable_stage2_overlap",
                                 "split_param",
                                 "disable_stage1_reduce_avg",
+                                "enable_stage1_broadcast_overlap",
+                                "enable_stage1_allgather_overlap",
                             ]:
                                 raise ValueError(
                                     f"Found unknown pipeline mode config {x}, "
-                                    f"accpet config is enable_stage1_tensor_fusion, enable_stage1_overlap, enable_stage2_overlap."
+                                    f"accpet config is enable_stage1_tensor_fusion, enable_stage1_overlap, enable_stage2_overlap, split_param, disable_stage1_reduce_avg, enable_stage1_broadcast_overlap, enable_stage1_allgather_overlap."
                                 )
                     if "disable_stage1_reduce_avg" in sharding_parallel_config:
                         assert self.sharding == [
@@ -1235,6 +1241,35 @@ def is_segment_parallel_supported():
                             "The logging_steps should be greater than 1 for stage2 overlap, "
                             f"but got logging_steps={self.logging_steps}."
                         )
+                    if "enable_stage1_broadcast_overlap" in sharding_parallel_config:
+                        assert (
+                            ShardingOption.SHARD_OP in self.sharding
+                        ), f"enable_stage1_broadcast_overlap expects sharding=stage1, but got {self.sharding}."
+
+                        assert (
+                            "split_param" not in sharding_parallel_config
+                        ), "split_param should not be set when enable_stage1_broadcast_overlap."
+                        use_casual_mask = os.getenv("USE_CASUAL_MASK", "False")
+                        assert use_casual_mask, "enable_stage1_broadcast_overlap requires USE_CASUAL_MASK=True."
+                        assert self.logging_steps > 1, (
+                            "The logging_steps should be greater than 1 for stage1_broadcast_overlap, "
+                            f"but got logging_steps={self.logging_steps}."
+                        )
+                    if "enable_stage1_allgather_overlap" in sharding_parallel_config:
+                        assert (
+                            ShardingOption.SHARD_OP in self.sharding
+                        ), f"enable_stage1_allgather_overlap expects sharding=stage1, but got {self.sharding}."
+
+                        assert (
+                            "split_param" in sharding_parallel_config
+                        ), "split_param should be set when enable_stage1_allgather_overlap."
+                        use_casual_mask = os.getenv("USE_CASUAL_MASK", "False")
+                        assert use_casual_mask, "enable_stage1_allgather_overlap requires USE_CASUAL_MASK=True."
+                        assert self.logging_steps > 1, (
+                            "The logging_steps should be greater than 1 for enable_stage1_allgather_overlap, "
+                            f"but got logging_steps={self.logging_steps}."
+                        )
+
                 fleet.init(is_collective=True, strategy=strategy)
                 logger.info(strategy)
 

paddlenlp/transformers/llama/modeling.py

@@ -115,11 +115,15 @@ def _get_interleave_power_of_2(n):
         )
 
 
+def get_use_casual_mask():
+    """Get the value of the 'USE_CASUAL_MASK' environment variable."""
+    return os.getenv("USE_CASUAL_MASK", "False") == "True"
+
+
 def build_alibi_tensor(
     bool_attention_mask: Tensor, num_heads: int, dtype: paddle.dtype, tensor_parallel_degree=1
 ) -> Tensor:
-    attention_mask = bool_attention_mask.astype("float32")
-    batch_size, seq_length = attention_mask.shape[0], attention_mask.shape[-1]
+    batch_size, seq_length = bool_attention_mask.shape[0], bool_attention_mask.shape[-1]
     slopes = paddle.to_tensor(_get_interleave(num_heads), dtype="float32")
     alibi = slopes.unsqueeze(axis=[1, 2]) * paddle.arange(seq_length, dtype="float32").unsqueeze(axis=[0, 1]).expand(
         [num_heads, -1, -1]
@@ -307,7 +311,7 @@ def is_casual_mask(attention_mask):
 
 def _make_causal_mask(input_ids_shape, past_key_values_length):
     """
-    Make causal mask used for self-attention
+    Make casual mask used for self-attention
     """
     batch_size, target_length = input_ids_shape  # target_length: seq_len
 
@@ -1543,12 +1547,22 @@ def forward(
         if position_ids is None:
             position_ids = paddle.arange(seq_length, dtype="int64").expand((batch_size, seq_length))
 
-        attention_mask = self._prepare_decoder_attention_mask(
-            attention_mask, (batch_size, seq_length), cache_length, inputs_embeds.dtype
-        )  # [bs, 1, seq_len, seq_len]
+        use_casual_mask = get_use_casual_mask()
+
+        if use_casual_mask:
+            attention_mask = None
+        else:
+            attention_mask = self._prepare_decoder_attention_mask(
+                attention_mask, (batch_size, seq_length), cache_length, inputs_embeds.dtype
+            )  # [bs, 1, seq_len, seq_len]
+
         is_casual = False
+
         if self.config.use_flash_attention and get_env_device() != "gcu":
-            is_casual = is_casual_mask(attention_mask)
+            if use_casual_mask:
+                is_casual = True
+            else:
+                is_casual = is_casual_mask(attention_mask)
             if get_env_device() != "npu":
                 if is_casual and alibi is None:
                     attention_mask = None