Improve DP attention (#4390)

Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: SangBin Cho <rkooo567@gmail.com>

python/sglang/srt/layers/dp_attention.py

 from __future__ import annotations
 
 import functools
+import logging
+from contextlib import contextmanager
 from typing import TYPE_CHECKING, Union
 
 import torch
@@ -14,6 +16,8 @@ from sglang.srt.distributed import (
     tensor_model_parallel_all_reduce,
 )
 
+logger = logging.getLogger(__name__)
+
 if TYPE_CHECKING:
     from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
@@ -86,6 +90,27 @@ def get_attention_dp_size():
     return _DP_SIZE
 
 
+@contextmanager
+def disable_dp_size():
+    """Patch the tp group temporarily until this function ends.
+
+    This method is for draft workers of speculative decoding to run draft model
+    with different tp degree from that of target model workers.
+
+    Args:
+        tp_group (GroupCoordinator): the tp group coordinator
+    """
+    global _DP_SIZE
+    assert _DP_SIZE is not None, "dp attention not initialized!"
+
+    old_dp_size = _DP_SIZE
+    _DP_SIZE = 1
+    try:
+        yield
+    finally:
+        _DP_SIZE = old_dp_size
+
+
 def get_dp_local_info(forward_batch: ForwardBatch):
     dp_rank = get_attention_dp_rank()
 
@@ -159,7 +184,8 @@ def dp_gather(
         layer_id != "embedding" or get_attention_tp_rank() == 0
     ):
         assert (
-            global_tokens.storage().data_ptr() != local_tokens.storage().data_ptr()
+            global_tokens.untyped_storage().data_ptr()
+            != local_tokens.untyped_storage().data_ptr()
         ), "aliasing between global_tokens and local_tokens not allowed"
         memcpy_triton(
             global_tokens, local_tokens, 0, local_start_pos, local_num_tokens, False
@@ -174,8 +200,9 @@ def dp_gather(
         torch.ops.sglang.inplace_all_reduce(
             global_tokens, group_name=get_tp_group().unique_name
         )
+
     else:
-        global_tokens = tensor_model_parallel_all_reduce(global_tokens)
+        global_tokens[:] = tensor_model_parallel_all_reduce(global_tokens)
 
 
 def dp_scatter(
@@ -186,6 +213,7 @@ def dp_scatter(
     # local_num_tokens is not necessarily the same as local_tokens.shape[0],
     # since local_tokens may be padded for cuda graph
     local_start_pos, local_num_tokens = get_dp_local_info(forward_batch)
+
     local_tokens.fill_(0)
     assert local_tokens.is_contiguous()
     assert global_tokens.is_contiguous()

python/sglang/srt/layers/logits_processor.py

@@ -23,6 +23,7 @@ import triton.language as tl
 from torch import nn
 
 from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_gather,
 )

python/sglang/srt/managers/data_parallel_controller.py

@@ -54,7 +54,7 @@ class LoadBalanceMethod(Enum):
 class DataParallelController:
     """A controller that dispatches requests to multiple data parallel workers."""
 
-    def __init__(self, server_args, port_args) -> None:
+    def __init__(self, server_args: ServerArgs, port_args: PortArgs) -> None:
         # Parse args
         self.max_total_num_tokens = None
         self.server_args = server_args

python/sglang/srt/managers/scheduler.py

@@ -997,7 +997,7 @@ class Scheduler(SchedulerOutputProcessorMixin):
 
         # Handle DP attention
         if self.server_args.enable_dp_attention:
-            ret = self.prepare_dp_attn_batch(ret)
+            ret, _ = self.prepare_dp_attn_batch(ret)
 
         return ret
 
@@ -1269,39 +1269,72 @@ class Scheduler(SchedulerOutputProcessorMixin):
         # Check if other DP workers have running batches
         if local_batch is None:
             num_tokens = 0
+            global_num_tokens_for_logprob = 0
         elif local_batch.forward_mode.is_decode():
             num_tokens = local_batch.batch_size()
+            if not self.spec_algorithm.is_none() and self.spec_algorithm.is_eagle():
+                num_tokens = num_tokens * self.server_args.speculative_num_draft_tokens
+            global_num_tokens_for_logprob = num_tokens
         else:
             num_tokens = local_batch.extend_num_tokens
+            global_num_tokens_for_logprob = sum(
+                [
+                    # We should have at least 1 token for sample in every case.
+                    max(extend_len - logprob_start_len, 1)
+                    for logprob_start_len, extend_len in zip(
+                        local_batch.extend_logprob_start_lens, local_batch.extend_lens
+                    )
+                ]
+            )
+
+        if local_batch is None or local_batch.forward_mode.is_decode_or_idle():
+            can_cuda_graph = 1
+        else:
+            can_cuda_graph = 0
+
+        if not self.spec_algorithm.is_none():
+            # TODO(sang): Support cuda graph when idle batch is there.
+            if local_batch is None or local_batch.forward_mode.is_idle():
+                can_cuda_graph = 0
 
-        local_num_tokens = torch.tensor([num_tokens], dtype=torch.int64)
-        global_num_tokens = torch.empty(self.tp_size, dtype=torch.int64)
+        is_extend_in_batch = (
+            local_batch.forward_mode.is_extend() if local_batch else False
+        )
+        local_info = torch.tensor(
+            [
+                num_tokens,
+                can_cuda_graph,
+                global_num_tokens_for_logprob,
+                is_extend_in_batch,
+            ],
+            dtype=torch.int64,
+        )
+        global_info = torch.empty(
+            (self.server_args.dp_size, self.attn_tp_size, 4),
+            dtype=torch.int64,
+        )
         torch.distributed.all_gather_into_tensor(
-            global_num_tokens,
-            local_num_tokens,
+            global_info.flatten(),
+            local_info,
             group=self.tp_cpu_group,
         )
+        global_num_tokens = global_info[:, 0, 0].tolist()
+        can_cuda_graph = min(global_info[:, 0, 1].tolist())
+        global_num_tokens_for_logprob = global_info[:, 0, 2].tolist()
+        is_extend_in_batch = global_info[:, 0, 3].tolist()
 
-        if local_batch is None and global_num_tokens.max().item() > 0:
+        if local_batch is None and max(global_num_tokens) > 0:
             local_batch = self.get_idle_batch()
 
         if local_batch is not None:
-            local_batch.global_num_tokens = global_num_tokens.tolist()
+            local_batch.global_num_tokens = global_num_tokens
+            local_batch.global_num_tokens_for_logprob = global_num_tokens_for_logprob
 
             # Check forward mode for cuda graph
             if not self.server_args.disable_cuda_graph:
-                forward_mode_state = torch.tensor(
-                    (1 if local_batch.forward_mode.is_decode_or_idle() else 0),
-                    dtype=torch.int32,
-                )
-                torch.distributed.all_reduce(
-                    forward_mode_state,
-                    op=torch.distributed.ReduceOp.MIN,
-                    group=self.tp_cpu_group,
-                )
-                local_batch.can_run_dp_cuda_graph = forward_mode_state.item() == 1
+                local_batch.can_run_dp_cuda_graph = can_cuda_graph
 
-        return local_batch
+        return local_batch, any(is_extend_in_batch)
 
     def get_idle_batch(self):
         idle_batch = ScheduleBatch.init_new(

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -33,7 +33,7 @@ from sglang.srt.model_executor.forward_batch_info import (
     ForwardBatch,
     ForwardMode,
 )
-from sglang.srt.utils import is_hip
+from sglang.srt.utils import get_available_gpu_memory, is_hip
 
 _is_hip = is_hip()
 
@@ -174,6 +174,7 @@ class CudaGraphRunner:
         self.disable_padding = model_runner.server_args.disable_cuda_graph_padding
         self.is_encoder_decoder = model_runner.model_config.is_encoder_decoder
         self.enable_dp_attention = model_runner.server_args.enable_dp_attention
+        self.speculative_algorithm = model_runner.server_args.speculative_algorithm
         self.tp_size = model_runner.server_args.tp_size
         self.dp_size = model_runner.server_args.dp_size
 
@@ -236,7 +237,7 @@ class CudaGraphRunner:
             if self.enable_dp_attention:
                 self.gathered_buffer = torch.zeros(
                     (
-                        self.max_bs * self.dp_size,
+                        self.max_bs * self.dp_size * self.num_tokens_per_bs,
                         self.model_runner.model_config.hidden_size,
                     ),
                     dtype=self.model_runner.dtype,
@@ -276,13 +277,12 @@ class CudaGraphRunner:
 
     def can_run(self, forward_batch: ForwardBatch):
         if self.enable_dp_attention:
-            min_num_tokens, max_num_tokens = min(
-                forward_batch.global_num_tokens_cpu
-            ), max(forward_batch.global_num_tokens_cpu)
+            total_global_tokens = sum(forward_batch.global_num_tokens_cpu)
+
             is_bs_supported = forward_batch.can_run_dp_cuda_graph and (
-                (min_num_tokens == max_num_tokens and max_num_tokens in self.graphs)
+                total_global_tokens in self.graphs
                 if self.disable_padding
-                else max_num_tokens <= self.max_bs
+                else total_global_tokens <= self.max_bs
             )
         else:
             is_bs_supported = (
@@ -304,6 +304,9 @@ class CudaGraphRunner:
     def capture(self):
         with graph_capture() as graph_capture_context:
             self.stream = graph_capture_context.stream
+            avail_mem = get_available_gpu_memory(
+                self.model_runner.device, self.model_runner.gpu_id, empty_cache=False
+            )
             # Reverse the order to enable better memory sharing across cuda graphs.
             capture_range = (
                 tqdm.tqdm(list(reversed(self.capture_bs)))
@@ -311,6 +314,16 @@ class CudaGraphRunner:
                 else reversed(self.capture_bs)
             )
             for bs in capture_range:
+                if get_tensor_model_parallel_rank() == 0:
+                    avail_mem = get_available_gpu_memory(
+                        self.model_runner.device,
+                        self.model_runner.gpu_id,
+                        empty_cache=False,
+                    )
+                    capture_range.set_description(
+                        f"Capturing batches ({avail_mem=:.2f} GB)"
+                    )
+
                 with patch_model(
                     self.model_runner.model,
                     bs in self.compile_bs,
@@ -345,8 +358,18 @@ class CudaGraphRunner:
         mrope_positions = self.mrope_positions[:, :bs]
 
         if self.enable_dp_attention:
-            global_num_tokens = [bs] * self.tp_size
-            gathered_buffer = self.gathered_buffer[: bs * self.tp_size]
+            self.global_num_tokens_gpu.copy_(
+                torch.tensor(
+                    [
+                        num_tokens // self.dp_size + (i < bs % self.dp_size)
+                        for i in range(self.dp_size)
+                    ],
+                    dtype=torch.int32,
+                    device=input_ids.device,
+                )
+            )
+            global_num_tokens = self.global_num_tokens_gpu
+            gathered_buffer = self.gathered_buffer[:num_tokens]
         else:
             global_num_tokens = None
             gathered_buffer = None
@@ -371,7 +394,7 @@ class CudaGraphRunner:
             encoder_lens=encoder_lens,
             return_logprob=False,
             positions=positions,
-            global_num_tokens_cpu=global_num_tokens,
+            global_num_tokens_gpu=global_num_tokens,
             gathered_buffer=gathered_buffer,
             mrope_positions=mrope_positions,
             spec_algorithm=self.model_runner.spec_algorithm,
@@ -392,6 +415,9 @@ class CudaGraphRunner:
 
         # Run and capture
         def run_once():
+            # Clean intermediate result cache for DP attention
+            forward_batch.dp_local_start_pos = forward_batch.dp_local_num_tokens = None
+
             logits_output = forward(input_ids, forward_batch.positions, forward_batch)
             return logits_output.next_token_logits, logits_output.hidden_states
 
@@ -426,7 +452,7 @@ class CudaGraphRunner:
             self.capture_hidden_mode = hidden_mode_from_spec_info
             self.capture()
 
-    def replay(self, forward_batch: ForwardBatch, skip_attn_backend_init: bool = False):
+    def replay_prepare(self, forward_batch: ForwardBatch):
         self.recapture_if_needed(forward_batch)
 
         raw_bs = forward_batch.batch_size
@@ -435,7 +461,7 @@ class CudaGraphRunner:
         # Pad
         if self.enable_dp_attention:
             index = bisect.bisect_left(
-                self.capture_bs, max(forward_batch.global_num_tokens_cpu)
+                self.capture_bs, sum(forward_batch.global_num_tokens_cpu)
             )
         else:
             index = bisect.bisect_left(self.capture_bs, raw_bs)
@@ -459,6 +485,8 @@ class CudaGraphRunner:
             self.encoder_lens[:raw_bs].copy_(forward_batch.encoder_lens)
         if forward_batch.mrope_positions is not None:
             self.mrope_positions[:, :raw_bs].copy_(forward_batch.mrope_positions)
+        if self.enable_dp_attention:
+            self.global_num_tokens_gpu.copy_(forward_batch.global_num_tokens_gpu)
 
         if hasattr(forward_batch.spec_info, "hidden_states"):
             self.hidden_states[:raw_num_token] = forward_batch.spec_info.hidden_states
@@ -475,14 +503,29 @@ class CudaGraphRunner:
             seq_lens_cpu=self.seq_lens_cpu,
         )
 
+        # Store fields
+        self.raw_bs = raw_bs
+        self.raw_num_token = raw_num_token
+        self.bs = bs
+
+    def replay(self, forward_batch: ForwardBatch, skip_attn_backend_init: bool = False):
+        if not skip_attn_backend_init:
+            self.replay_prepare(forward_batch)
+        else:
+            # In speculative decoding, these two fields are still needed.
+            self.input_ids[: self.raw_num_token].copy_(forward_batch.input_ids)
+            self.positions[: self.raw_num_token].copy_(forward_batch.positions)
+
         # Replay
-        self.graphs[bs].replay()
-        next_token_logits, hidden_states = self.output_buffers[bs]
+        self.graphs[self.bs].replay()
+        next_token_logits, hidden_states = self.output_buffers[self.bs]
 
         logits_output = LogitsProcessorOutput(
-            next_token_logits=next_token_logits[:raw_num_token],
+            next_token_logits=next_token_logits[: self.raw_num_token],
             hidden_states=(
-                hidden_states[:raw_num_token] if hidden_states is not None else None
+                hidden_states[: self.raw_num_token]
+                if hidden_states is not None
+                else None
             ),
         )
         return logits_output

python/sglang/srt/model_executor/forward_batch_info.py

@@ -38,7 +38,7 @@ import triton
 import triton.language as tl
 
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
-from sglang.srt.utils import get_compiler_backend, next_power_of_2
+from sglang.srt.utils import get_compiler_backend
 
 if TYPE_CHECKING:
     from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
@@ -263,15 +263,24 @@ class ForwardBatch:
             extend_input_logprob_token_ids_gpu=extend_input_logprob_token_ids_gpu,
         )
 
+        # For DP attention
         if batch.global_num_tokens is not None:
             ret.global_num_tokens_cpu = batch.global_num_tokens
-            max_len = max(ret.global_num_tokens_cpu)
+            ret.global_num_tokens_gpu = torch.tensor(
+                batch.global_num_tokens, dtype=torch.int64
+            ).to(device, non_blocking=True)
+
+            ret.global_num_tokens_for_logprob_cpu = batch.global_num_tokens_for_logprob
+            ret.global_num_tokens_for_logprob_gpu = torch.tensor(
+                batch.global_num_tokens_for_logprob, dtype=torch.int64
+            ).to(device, non_blocking=True)
+
+            sum_len = sum(batch.global_num_tokens)
             ret.gathered_buffer = torch.zeros(
-                (max_len * model_runner.tp_size, model_runner.model_config.hidden_size),
+                (sum_len, model_runner.model_config.hidden_size),
                 dtype=model_runner.dtype,
                 device=device,
             )
-
         if ret.forward_mode.is_idle():
             ret.positions = torch.empty((0,), device=device)
             return ret

python/sglang/srt/server_args.py

@@ -262,14 +262,14 @@ class ServerArgs:
 
         # Data parallelism attention
         if self.enable_dp_attention:
-            self.dp_size = self.tp_size
-            assert self.tp_size % self.dp_size == 0
-            self.chunked_prefill_size = self.chunked_prefill_size // 2
             self.schedule_conservativeness = self.schedule_conservativeness * 0.3
+            assert (
+                self.dp_size > 1
+            ), "Please set a dp-size > 1. You can use 1 < dp-size <= tp-size "
+            assert self.tp_size % self.dp_size == 0
+            self.chunked_prefill_size = self.chunked_prefill_size // self.dp_size
             logger.warning(
                 f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
-                f"The schedule conservativeness is adjusted to {self.schedule_conservativeness}. "
-                "Data parallel size is adjusted to be the same as tensor parallel size. "
             )
 
         # Speculative Decoding

test/srt/test_dp_attention.py

@@ -25,6 +25,8 @@ class TestDPAttention(unittest.TestCase):
                 "--tp",
                 "2",
                 "--enable-dp-attention",
+                "--dp",
+                "2",
             ],
         )