Reduce computation and communication in DP attention (#4521)

python/sglang/srt/distributed/parallel_state.py

@@ -189,6 +189,9 @@ class GroupCoordinator:
     device_group: ProcessGroup  # group for device communication
     use_pynccl: bool  # a hint of whether to use PyNccl
     use_custom_allreduce: bool  # a hint of whether to use CustomAllreduce
+    use_message_queue_broadcaster: (
+        bool  # a hint of whether to use message queue broadcaster
+    )
     # communicators are only created for world size > 1
     pynccl_comm: Optional[Any]  # PyNccl communicator
     ca_comm: Optional[Any]  # Custom allreduce communicator
@@ -241,6 +244,7 @@ class GroupCoordinator:
         self.use_custom_allreduce = use_custom_allreduce
         self.use_hpu_communicator = use_hpu_communicator
         self.use_xpu_communicator = use_xpu_communicator
+        self.use_message_queue_broadcaster = use_message_queue_broadcaster
 
         # lazy import to avoid documentation build error
         from sglang.srt.distributed.device_communicators.custom_all_reduce import (
@@ -269,7 +273,7 @@ class GroupCoordinator:
             HpuCommunicator,
         )
 
-        self.hpu_communicator: Optional[HpuCommunicator]
+        self.hpu_communicator: Optional[HpuCommunicator] = None
         if use_hpu_communicator and self.world_size > 1:
             self.hpu_communicator = HpuCommunicator(group=self.device_group)
 
@@ -277,7 +281,7 @@ class GroupCoordinator:
             XpuCommunicator,
         )
 
-        self.xpu_communicator: Optional[XpuCommunicator]
+        self.xpu_communicator: Optional[XpuCommunicator] = None
         if use_xpu_communicator and self.world_size > 1:
             self.xpu_communicator = XpuCommunicator(group=self.device_group)
 

python/sglang/srt/layers/dp_attention.py

@@ -53,10 +53,8 @@ def initialize_dp_attention(
     )
 
     if enable_dp_attention:
-        local_rank = tp_rank % (tp_size // dp_size)
         _DP_SIZE = dp_size
     else:
-        local_rank = tp_rank
         _DP_SIZE = 1
 
     tp_group = get_tp_group()
@@ -65,7 +63,7 @@ def initialize_dp_attention(
             list(range(head, head + _ATTN_TP_SIZE))
             for head in range(0, tp_size, _ATTN_TP_SIZE)
         ],
-        local_rank,
+        tp_group.local_rank,
         torch.distributed.get_backend(tp_group.device_group),
         SYNC_TOKEN_IDS_ACROSS_TP,
         False,
@@ -180,20 +178,19 @@ def memcpy_triton(dst, src, dim, offset, sz, offset_src):
     memcpy_triton_kernel[grid](dst, src, offset, sz, offset_src, chunk_size, BLOCK_SIZE)
 
 
-def dp_gather(
+def _dp_gather(
     global_tokens: torch.Tensor,
     local_tokens: torch.Tensor,
     forward_batch: ForwardBatch,
-    layer_id: Union[str, int],
+    is_partial: bool,
 ):
     local_start_pos, local_num_tokens = get_dp_local_info(forward_batch)
 
     global_tokens.fill_(0)
     assert local_tokens.is_contiguous()
     assert global_tokens.is_contiguous()
-    if local_tokens.shape[0] > 0 and (
-        layer_id != "embedding" or get_attention_tp_rank() == 0
-    ):
+
+    if local_tokens.shape[0] > 0 and (is_partial or get_attention_tp_rank() == 0):
         assert (
             global_tokens.untyped_storage().data_ptr()
             != local_tokens.untyped_storage().data_ptr()
@@ -216,6 +213,22 @@ def dp_gather(
         global_tokens[:] = tensor_model_parallel_all_reduce(global_tokens)
 
 
+def dp_gather_partial(
+    global_tokens: torch.Tensor,
+    local_tokens: torch.Tensor,
+    forward_batch: ForwardBatch,
+):
+    _dp_gather(global_tokens, local_tokens, forward_batch, is_partial=True)
+
+
+def dp_gather_replicate(
+    global_tokens: torch.Tensor,
+    local_tokens: torch.Tensor,
+    forward_batch: ForwardBatch,
+):
+    _dp_gather(global_tokens, local_tokens, forward_batch, is_partial=False)
+
+
 def dp_scatter(
     local_tokens: torch.Tensor,  # output
     global_tokens: torch.Tensor,  # input
@@ -236,16 +249,3 @@ def dp_scatter(
         memcpy_triton(
             local_tokens, global_tokens, 0, local_start_pos, local_num_tokens, True
         )
-
-
-def get_do_logits_dp_scatter(forward_batch: ForwardBatch):
-    def do_logits_dp_scatter(logits: torch.Tensor):
-        local_logits = torch.empty(
-            (forward_batch.input_ids.shape[0], *logits.shape[1:]),
-            dtype=logits.dtype,
-            device=logits.device,
-        )
-        dp_scatter(local_logits, logits, forward_batch)
-        return local_logits
-
-    return do_logits_dp_scatter

python/sglang/srt/layers/logits_processor.py

@@ -28,7 +28,7 @@ from sglang.srt.distributed import (
     tensor_model_parallel_all_gather,
 )
 from sglang.srt.layers.dp_attention import (
-    dp_gather,
+    dp_gather_replicate,
     dp_scatter,
     get_attention_dp_rank,
     get_attention_dp_size,
@@ -428,7 +428,7 @@ class LogitsProcessor(nn.Module):
                 logits_metadata.gathered_buffer,
                 hidden_states.clone(),
             )
-            dp_gather(hidden_states, local_hidden_states, logits_metadata, "embedding")
+            dp_gather_replicate(hidden_states, local_hidden_states, logits_metadata)
 
         if hasattr(lm_head, "weight"):
             logits = torch.matmul(

python/sglang/srt/models/deepseek_v2.py

@@ -33,7 +33,7 @@ from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
     decode_attention_fwd_grouped_rope,
 )
 from sglang.srt.layers.dp_attention import (
-    dp_gather,
+    dp_gather_partial,
     dp_scatter,
     get_attention_dp_size,
     get_attention_tp_rank,
@@ -939,47 +939,58 @@ class DeepseekV2DecoderLayer(nn.Module):
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
     ) -> torch.Tensor:
-        if residual is None:
+        if hidden_states.shape[0] == 0:
             residual = hidden_states
-            hidden_states = self.input_layernorm(hidden_states)
         else:
-            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+            if residual is None:
+                residual = hidden_states
+                hidden_states = self.input_layernorm(hidden_states)
+            else:
+                hidden_states, residual = self.input_layernorm(hidden_states, residual)
 
-        # Scatter
-        if self.dp_size != 1:
-            # important: forward batch.gathered_buffer is used both after scatter and after gather.
-            # be careful about this!
-            hidden_states, global_hidden_states = (
-                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
-                hidden_states,
+            # Self Attention
+            hidden_states = self.self_attn(
+                positions=positions,
+                hidden_states=hidden_states,
+                forward_batch=forward_batch,
             )
-            dp_scatter(hidden_states, global_hidden_states, forward_batch)
-
-        # Self Attention
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-            forward_batch=forward_batch,
-        )
 
         # Gather
         if get_tensor_model_parallel_world_size() > 1:
             # all gather and all reduce
             if self.dp_size != 1:
+                if get_attention_tp_rank() == 0:
+                    hidden_states += residual
                 hidden_states, local_hidden_states = (
                     forward_batch.gathered_buffer,
                     hidden_states,
                 )
-                dp_gather(
-                    hidden_states, local_hidden_states, forward_batch, self.layer_id
-                )
+                dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
+                dp_scatter(residual, hidden_states, forward_batch)
+                hidden_states = self.post_attention_layernorm(hidden_states)
             else:
                 hidden_states = tensor_model_parallel_all_reduce(hidden_states)
-
-        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+                hidden_states, residual = self.post_attention_layernorm(
+                    hidden_states, residual
+                )
+        else:
+            hidden_states, residual = self.post_attention_layernorm(
+                hidden_states, residual
+            )
 
         # Fully Connected
         hidden_states = self.mlp(hidden_states)
+
+        # Scatter
+        if self.dp_size != 1:
+            # important: forward batch.gathered_buffer is used both after scatter and after gather.
+            # be careful about this!
+            hidden_states, global_hidden_states = (
+                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
+                hidden_states,
+            )
+            dp_scatter(hidden_states, global_hidden_states, forward_batch)
+
         return hidden_states, residual
 
 
@@ -1025,18 +1036,6 @@ class DeepseekV2Model(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
 
-        # Gather
-        if self.dp_size != 1:
-            input_ids, local_input_ids = (
-                torch.empty(
-                    (forward_batch.gathered_buffer.shape[0],),
-                    dtype=input_ids.dtype,
-                    device=input_ids.device,
-                ),
-                input_ids,
-            )
-            dp_gather(input_ids, local_input_ids, forward_batch, "embedding")
-
         if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
         else:
@@ -1087,15 +1086,6 @@ class DeepseekV2ForCausalLM(nn.Module):
 
         hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
 
-        if self.dp_size != 1:
-            # important: forward batch.gathered_buffer is used both after scatter and after gather.
-            # be careful about this!
-            hidden_states, global_hidden_states = (
-                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
-                hidden_states,
-            )
-            dp_scatter(hidden_states, global_hidden_states, forward_batch)
-
         return self.logits_processor(
             input_ids, hidden_states, self.lm_head, forward_batch
         )

test/srt/test_dp_attention.py

@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
 )
 
 
-class TestDPAttention(unittest.TestCase):
+class TestDPAttentionDP2TP2(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
@@ -59,7 +59,3 @@ class TestDPAttention(unittest.TestCase):
         metrics = run_eval(args)
         print(f"{metrics=}")
         self.assertGreater(metrics["score"], 0.8)
-
-
-if __name__ == "__main__":
-    unittest.main()