[Fix] fix allreduce bug in Piecewise Graph (#12106)

python/sglang/srt/compilation/backend.py

@@ -392,7 +392,7 @@ class SGLangBackend:
         self.configure_post_pass()
 
         self.split_gm, self.piecewise_graphs = split_graph(
-            graph, ["sglang.unified_attention_with_output"]
+            graph, ["sglang.unified_attention_with_output", "sglang.inplace_all_reduce"]
         )
 
         from torch._dynamo.utils import lazy_format_graph_code

python/sglang/srt/distributed/parallel_state.py

@@ -340,17 +340,10 @@ class GroupCoordinator:
         self.qr_comm: Optional[QuickAllReduce] = None
         if use_custom_allreduce and self.world_size > 1:
             # Initialize a custom fast all-reduce implementation.
-            if torch_compile is not None and torch_compile:
-                # For piecewise CUDA graph, the requirement for custom allreduce is larger to
-                # avoid illegal cuda memory access.
-                ca_max_size = 256 * 1024 * 1024
-            else:
-                ca_max_size = 8 * 1024 * 1024
             try:
                 self.ca_comm = CustomAllreduce(
                     group=self.cpu_group,
                     device=self.device,
-                    max_size=ca_max_size,
                 )
             except Exception as e:
                 logger.warning(

python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py

@@ -32,7 +32,6 @@ from sglang.srt.distributed import get_tensor_model_parallel_rank
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
     set_graph_pool_id,
 )
-from sglang.srt.distributed.parallel_state import graph_capture
 from sglang.srt.layers.dp_attention import (
     DpPaddingMode,
     get_attention_tp_rank,
@@ -281,10 +280,10 @@ class PiecewiseCudaGraphRunner:
         # Trigger CUDA graph capture for specific shapes.
         # Capture the large shapes first so that the smaller shapes
         # can reuse the memory pool allocated for the large shapes.
-        with freeze_gc(
+        with freeze_gc(self.model_runner.server_args.enable_cudagraph_gc):
-            self.model_runner.server_args.enable_cudagraph_gc
+            if self.model_runner.tp_group.ca_comm is not None:
-        ), graph_capture() as graph_capture_context:
+                old_ca_disable = self.model_runner.tp_group.ca_comm.disabled
-            self.stream = graph_capture_context.stream
+                self.model_runner.tp_group.ca_comm.disabled = True
             avail_mem = get_available_gpu_memory(
                 self.model_runner.device,
                 self.model_runner.gpu_id,
@@ -312,9 +311,10 @@ class PiecewiseCudaGraphRunner:
 
                 # Save gemlite cache after each capture
                 save_gemlite_cache()
+            if self.model_runner.tp_group.ca_comm is not None:
+                self.model_runner.tp_group.ca_comm.disabled = old_ca_disable
 
     def capture_one_batch_size(self, num_tokens: int):
-        stream = self.stream
         bs = 1
 
         # Graph inputs
@@ -479,6 +479,9 @@ class PiecewiseCudaGraphRunner:
         forward_batch: ForwardBatch,
         **kwargs,
     ) -> Union[LogitsProcessorOutput, PPProxyTensors]:
+        if self.model_runner.tp_group.ca_comm is not None:
+            old_ca_disable = self.model_runner.tp_group.ca_comm.disabled
+            self.model_runner.tp_group.ca_comm.disabled = True
         static_forward_batch = self.replay_prepare(forward_batch, **kwargs)
         # Replay
         with set_forward_context(static_forward_batch, self.attention_layers):
@@ -504,6 +507,8 @@ class PiecewiseCudaGraphRunner:
                 raise NotImplementedError(
                     "PPProxyTensors is not supported in PiecewiseCudaGraphRunner yet."
                 )
+        if self.model_runner.tp_group.ca_comm is not None:
+            self.model_runner.tp_group.ca_comm.disabled = old_ca_disable
 
     def get_spec_info(self, num_tokens: int):
         spec_info = None