Add high-priority A2A stream and HybridEP preprocessing SMs

megatron/core/pipeline_parallel/combined_1f1b.py

@@ -52,7 +52,7 @@ def combined_1f1b_schedule_for_no_pipelining(
     Phases 4: 4th microbatch backward
     """
 
-    set_streams()
+    set_streams(high_priority=config.high_priority_a2a_comm_stream)
     # The forward step for the first microbatch is executed alone, no a2a overlapping
     output_tensor, num_tokens, _ = combined_forward_backward_step(
         forward_step_func,
@@ -178,7 +178,7 @@ def combined_1f1b_schedule_for_interleaved_pipelining():
                 # backward_step_helper_postprocess()
     """
 
-    set_streams()
+    set_streams(high_priority=config.high_priority_a2a_comm_stream)
     # forward prepare
     f_model_chunk_id = None
     f_microbatch_id = None

megatron/core/pipeline_parallel/utils.py

@@ -341,14 +341,18 @@ def run(
 _COMM_STREAM = None
 
 
-def set_streams(comm_stream=None):
+def set_streams(comm_stream=None, high_priority=False):
     """Set the stream for communication operations."""
     global _COMM_STREAM
 
     # Set communication stream
     if _COMM_STREAM is None:
         if comm_stream is None:
-            comm_stream = torch.cuda.Stream(device="cuda")
+            if high_priority:
+                _, high = torch.cuda.Stream.priority_range()
+                comm_stream = torch.cuda.Stream(device="cuda", priority=high)
+            else:
+                comm_stream = torch.cuda.Stream(device="cuda")
         _COMM_STREAM = comm_stream
 
 

megatron/core/transformer/moe/fused_a2a.py

@@ -345,6 +345,7 @@ def init_hybrid_ep_buffer(
     num_blocks_permute: Optional[int] = None,
     num_blocks_unpermute: Optional[int] = None,
     fp8_dispatch: bool = False,
+    num_sms_preprocessing_api: Optional[int] = None,
 ) -> None:
     '''
     Initialize the HybridEP buffer, including buffer allocation and metadata
@@ -373,6 +374,8 @@ def init_hybrid_ep_buffer(
             Number of blocks used by the unpermute part.
         fp8_dispatch (bool):
             Whether to use FP8 communication during the dispatch phase.
+        num_sms_preprocessing_api (Optional[int]):
+            Number of SMs used by the preprocessing (metadata scan) kernel.
     '''
     assert not fp8_dispatch, "HybridEP dispatcher does not support fp8 dispatch now"
     global _hybrid_ep_buffer
@@ -385,6 +388,8 @@ def init_hybrid_ep_buffer(
         kwargs['num_blocks_permute'] = num_blocks_permute
     if num_blocks_unpermute is not None:
         kwargs['num_blocks_unpermute'] = num_blocks_unpermute
+    if num_sms_preprocessing_api is not None:
+        kwargs['num_sms_preprocessing_api'] = num_sms_preprocessing_api
     _hybrid_ep_buffer = HybridEPBuffer(
         group=group,
         hidden_dim=hidden_dim,
@@ -423,6 +428,7 @@ def forward(
         fused=False,
         num_permuted_tokens=None,
         pad_multiple=None,
+        num_sms_preprocessing_api=108,
     ):
         '''
         Forward pass of fused dispatch of the HybridEP backend
@@ -458,6 +464,7 @@ def forward(
                 num_blocks_permute,
                 num_blocks_unpermute,
                 fp8_dispatch,
+                num_sms_preprocessing_api,
             )
         # If we provide the num_permuted_tokens, we do not need to use sync to
         # wait for the data in pinned memory ready
@@ -518,6 +525,7 @@ def backward(ctx, grad_x, grad_probs, grad_scaling_factor, grad_tokens_per_exper
             None,
             None,
             None,
+            None,
         )
 
 
@@ -577,6 +585,7 @@ def hybrid_ep_dispatch(
         fused=False,
         num_permuted_tokens=None,
         pad_multiple=None,
+        num_sms_preprocessing_api=108,
     ):
         '''
         Perform fused dispatch for "permute + dispatch a2a + permute" using the
@@ -608,6 +617,8 @@ def hybrid_ep_dispatch(
             pad_multiple (int):
                 Alignment multiple required for FP8 GEMM. If not provided, no padding
                 is performed.
+            num_sms_preprocessing_api (int):
+                Number of SMs used by the preprocessing (metadata scan) kernel.
         '''
         return HybridEPDispatch.apply(
             x,
@@ -622,6 +633,7 @@ def hybrid_ep_dispatch(
             fused,
             num_permuted_tokens,
             pad_multiple,
+            num_sms_preprocessing_api,
         )
 
     @internal_api

megatron/core/transformer/moe/token_dispatcher.py

@@ -1078,6 +1078,7 @@ def dispatch(
                 num_permuted_tokens=self.num_permuted_tokens,
                 pad_multiple=self.pad_multiple,
                 fused=self.config.moe_permute_fusion_into_hybridep,
+                num_sms_preprocessing_api=self.config.moe_hybridep_num_sms_preprocessing,
             )
         )
 

megatron/core/transformer/transformer_config.py

@@ -601,6 +601,10 @@ class TransformerConfig(ModelParallelConfig):
     """Python import path to a callable quantizer factory, e.g., package.module.quantizer_factory.
     Required when fp4_recipe is custom."""
 
+    high_priority_a2a_comm_stream: bool = False
+    """If True, the communication stream created by set_streams for combined 1f1b
+    a2a overlap is created with CUDA high priority."""
+
     ####################
     # MoE related
     ####################
@@ -821,6 +825,9 @@ class TransformerConfig(ModelParallelConfig):
     When permute_fusion_into_hybridep is True, this sets the number
     of SMs for the unpermute part (only 1 block per SM)."""
 
+    moe_hybridep_num_sms_preprocessing: int = 108
+    """Number of SMs to use for HybridEP preprocessing (metadata scan kernel)."""
+
     ##################
     # Context Parallel
     ##################

tests/unit_tests/models/test_hybrid_moe_model.py

@@ -127,6 +127,7 @@
     "hidden_dropout": 0.0,
     "hidden_size": 2688,
     "hierarchical_context_parallel_sizes": None,
+    "high_priority_a2a_comm_stream": False,
     "inference_fuse_tp_communication": False,
     "inference_rng_tracker": False,
     "inference_sampling_seed": 42,
@@ -163,6 +164,7 @@
     "moe_flex_dispatcher_backend": "deepep",
     "moe_grouped_gemm": True,
     "moe_hybridep_num_sms": None,
+    "moe_hybridep_num_sms_preprocessing": 108,
     "moe_hybridep_num_blocks_permute": None,
     "moe_hybridep_num_blocks_unpermute": None,
     "moe_input_jitter_eps": None,