Support cuda graph for DP attention (#2061)

python/sglang/srt/managers/schedule_batch.py

@@ -455,6 +455,7 @@ class ScheduleBatch:
 
     # For DP attention
     global_num_tokens: Optional[List[int]] = None
+    can_run_dp_cuda_graph: bool = False
 
     # For processing logprobs
     return_logprob: bool = False
@@ -891,6 +892,13 @@ class ScheduleBatch:
         self.seq_lens = torch.empty(0, dtype=torch.int32).to(
             self.device, non_blocking=True
         )
+        self.out_cache_loc = torch.empty(0, dtype=torch.int32).to(
+            self.device, non_blocking=True
+        )
+        self.req_pool_indices = torch.empty(0, dtype=torch.int32).to(
+            self.device, non_blocking=True
+        )
+        self.seq_lens_sum = 0
         self.extend_num_tokens = 0
 
     def prepare_for_decode(self, enable_overlap: bool = False):
@@ -1032,6 +1040,7 @@ class ScheduleBatch:
             return_logprob=self.return_logprob,
             top_logprobs_nums=self.top_logprobs_nums,
             global_num_tokens=self.global_num_tokens,
+            can_run_dp_cuda_graph=self.can_run_dp_cuda_graph,
             extend_num_tokens=self.extend_num_tokens,
             extend_seq_lens=extend_seq_lens,
             extend_prefix_lens=extend_prefix_lens,
@@ -1093,6 +1102,7 @@ class ModelWorkerBatch:
 
     # For DP attention
     global_num_tokens: Optional[List[int]]
+    can_run_dp_cuda_graph: bool
 
     # For extend
     extend_num_tokens: Optional[int]

python/sglang/srt/managers/scheduler.py

@@ -337,7 +337,7 @@ class Scheduler:
 
         kill_parent_process()
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def event_loop_normal(self):
         """A normal blocking scheduler loop."""
         self.last_batch = None
@@ -375,7 +375,7 @@ class Scheduler:
 
             self.last_batch = batch
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def event_loop_overlap(self):
         """A scheduler loop that overlaps the CPU processing and GPU computation."""
         result_queue = deque()
@@ -411,16 +411,12 @@ class Scheduler:
         else:
             num_tokens = local_batch.extend_num_tokens
 
-        local_num_tokens = torch.tensor(
-            num_tokens, dtype=torch.int64, device=self.device
-        )
-        global_num_tokens = torch.empty(
-            self.tp_size, dtype=torch.int64, device=self.device
-        )
+        local_num_tokens = torch.tensor([num_tokens], dtype=torch.int64)
+        global_num_tokens = torch.empty(self.tp_size, dtype=torch.int64)
         torch.distributed.all_gather_into_tensor(
             global_num_tokens,
             local_num_tokens,
-            group=self.tp_worker.get_tp_device_group(),
+            group=self.tp_cpu_group,
         )
 
         if local_batch is None and global_num_tokens.max().item() > 0:
@@ -429,6 +425,24 @@ class Scheduler:
         if local_batch is not None:
             local_batch.global_num_tokens = global_num_tokens.tolist()
 
+            # 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 local_batch.forward_mode.is_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
+
         return local_batch
 
     def get_idle_batch(self):

python/sglang/srt/managers/tp_worker.py

@@ -128,9 +128,6 @@ class TpModelWorker:
     def get_tp_cpu_group(self):
         return self.model_runner.tp_group.cpu_group
 
-    def get_tp_device_group(self):
-        return self.model_runner.tp_group.device_group
-
     def get_memory_pool(self):
         return (
             self.model_runner.req_to_token_pool,

python/sglang/srt/managers/tp_worker_overlap_thread.py

@@ -83,9 +83,6 @@ class TpModelWorkerClient:
     def get_tp_cpu_group(self):
         return self.worker.get_tp_cpu_group()
 
-    def get_tp_device_group(self):
-        return self.worker.get_tp_device_group()
-
     def get_memory_pool(self):
         return (
             self.worker.model_runner.req_to_token_pool,
@@ -96,7 +93,7 @@ class TpModelWorkerClient:
         with torch.cuda.stream(self.forward_stream):
             self.forward_thread_func_()
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def forward_thread_func_(self):
         while True:
             model_worker_batch, future_token_ids_ct = self.input_queue.get()

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -111,6 +111,8 @@ class CudaGraphRunner:
         self.use_torch_compile = model_runner.server_args.enable_torch_compile
         self.disable_padding = model_runner.server_args.disable_cuda_graph_padding
         self.is_encoder_decoder = self.model_runner.model_config.is_encoder_decoder
+        self.enable_dp_attention = self.model_runner.server_args.enable_dp_attention
+        self.tp_size = self.model_runner.tp_size
 
         # Batch sizes to capture
         if model_runner.server_args.disable_cuda_graph_padding:
@@ -165,6 +167,16 @@ class CudaGraphRunner:
             else:
                 self.encoder_lens = None
 
+            if self.enable_dp_attention:
+                self.global_num_tokens = [0] * self.tp_size
+                self.gathered_buffer = torch.zeros(
+                    (
+                        self.max_bs * self.tp_size,
+                        self.model_runner.model_config.hidden_size,
+                    ),
+                    dtype=self.model_runner.dtype,
+                )
+
         # Capture
         try:
             with self.model_capture_mode():
@@ -190,6 +202,16 @@ class CudaGraphRunner:
             self.model_runner.model.capture_mode = False
 
     def can_run(self, forward_batch: ForwardBatch):
+        if self.enable_dp_attention:
+            min_num_tokens, max_num_tokens = min(forward_batch.global_num_tokens), max(
+                forward_batch.global_num_tokens
+            )
+            is_bs_supported = forward_batch.can_run_dp_cuda_graph and (
+                (min_num_tokens == max_num_tokens and max_num_tokens in self.graphs)
+                if self.disable_padding
+                else max_num_tokens <= self.max_bs
+            )
+        else:
             is_bs_supported = (
                 forward_batch.batch_size in self.graphs
                 if self.disable_padding
@@ -239,6 +261,13 @@ class CudaGraphRunner:
         seq_lens_sum = seq_lens.sum().item()
         mrope_positions = self.mrope_positions[:, :bs]
 
+        if self.enable_dp_attention:
+            self.global_num_tokens[:] = [bs] * self.tp_size
+            gathered_buffer = self.gathered_buffer[: bs * self.tp_size]
+        else:
+            self.global_num_tokens = None
+            gathered_buffer = None
+
         # Attention backend
         self.model_runner.attn_backend.init_forward_metadata_capture_cuda_graph(
             bs,
@@ -265,6 +294,8 @@ class CudaGraphRunner:
                 top_logprobs_nums=[0] * bs,
                 positions=clamp_position(seq_lens),
                 mrope_positions=mrope_positions,
+                global_num_tokens=self.global_num_tokens,
+                gathered_buffer=gathered_buffer,
             )
             logits_output = forward(input_ids, forward_batch.positions, forward_batch)
             return logits_output.next_token_logits
@@ -295,6 +326,11 @@ class CudaGraphRunner:
         raw_bs = forward_batch.batch_size
 
         # Pad
+        if self.enable_dp_attention:
+            index = bisect.bisect_left(
+                self.capture_bs, max(forward_batch.global_num_tokens)
+            )
+        else:
             index = bisect.bisect_left(self.capture_bs, raw_bs)
         bs = self.capture_bs[index]
         if bs != raw_bs:
@@ -310,6 +346,8 @@ 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[:] = [bs] * self.tp_size
 
         # Attention backend
         self.model_runner.attn_backend.init_forward_metadata_replay_cuda_graph(

python/sglang/srt/model_executor/forward_batch_info.py

@@ -138,6 +138,7 @@ class ForwardBatch:
     # For DP attention
     global_num_tokens: Optional[List[int]] = None
     gathered_buffer: Optional[torch.Tensor] = None
+    can_run_dp_cuda_graph: bool = False
 
     def compute_mrope_positions(
         self, model_runner: ModelRunner, batch: ModelWorkerBatch
@@ -221,6 +222,7 @@ class ForwardBatch:
             return_logprob=batch.return_logprob,
             top_logprobs_nums=batch.top_logprobs_nums,
             global_num_tokens=batch.global_num_tokens,
+            can_run_dp_cuda_graph=batch.can_run_dp_cuda_graph,
             lora_paths=batch.lora_paths,
             sampling_info=batch.sampling_info,
         )

python/sglang/srt/model_executor/model_runner.py

@@ -592,6 +592,9 @@ class ModelRunner:
             )
 
     def forward_idle(self, forward_batch: ForwardBatch):
+        if self.cuda_graph_runner and self.cuda_graph_runner.can_run(forward_batch):
+            return self.cuda_graph_runner.replay(forward_batch)
+
         return self.model.forward(
             forward_batch.input_ids, forward_batch.positions, forward_batch
         )

python/sglang/srt/server_args.py

@@ -191,11 +191,12 @@ class ServerArgs:
         if self.enable_dp_attention:
             self.dp_size = self.tp_size
             self.chunked_prefill_size = self.chunked_prefill_size // 2
-            self.disable_cuda_graph = True
+            self.cuda_graph_max_bs = min(self.cuda_graph_max_bs, 96)
             self.enable_overlap_schedule = False
             logger.warning(
                 f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE workload issue. "
-                "The CUDA graph is disabled. Data parallel size is adjust to be the same as tensor parallel size."
+                f"The CUDA graph max batch size is adjusted to {self.cuda_graph_max_bs}. "
+                "Data parallel size is adjusted to be the same as tensor parallel size."
             )
 
         if self.enable_overlap_schedule:

scripts/playground/reference_hf.py

@@ -31,7 +31,7 @@ from transformers import AutoModelForCausalLM
 from sglang.srt.hf_transformers_utils import get_tokenizer
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def normal_text(args):
     t = get_tokenizer(args.model_path, trust_remote_code=True)
     m = AutoModelForCausalLM.from_pretrained(
@@ -69,7 +69,7 @@ def normal_text(args):
         print(output_str)
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def synthetic_tokens(args):
     m = AutoModelForCausalLM.from_pretrained(
         args.model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True