[Cherry-pick] Port MoE multi-stream fix to v0.11.0-dev (vllm-project#3753)

This PR moves the communication operation of shared experts out of extra
stream because I found that this might cause rtMemcpy related errors
when running shared experts multistream with aclgraph.

Furthermore, I utilize a global variable as extra stream object to avoid
allocating streams for each layer in full-graph mode.

Signed-off-by: whx-sjtu <[email protected]>

Author: whx-sjtu

tests/e2e/singlecard/test_multistream_overlap_shared_expert.py

@@ -28,7 +28,7 @@
 from tests.e2e.model_utils import check_outputs_equal
 
 MODELS = [
-    "Qwen/Qwen3-0.6B",
+    "vllm-ascend/DeepSeek-V2-Lite-W8A8",
 ]
 
 

vllm_ascend/ops/common_fused_moe.py

@@ -40,7 +40,8 @@
 from vllm_ascend.ops.moe.moe_comm_method import setup_moe_comm_method
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, enable_sp, is_310p,
                                is_enable_nz, npu_stream_switch,
-                               shared_expert_dp_enabled)
+                               shared_expert_dp_enabled,
+                               shared_experts_compute_stream)
 
 
 class AscendUnquantizedFusedMoEMethod(UnquantizedFusedMoEMethod):
@@ -419,8 +420,6 @@ def __init__(
         self.shared_expert_stream = None
         ascend_config = get_ascend_config()
         self.multistream_overlap_shared_expert = ascend_config.multistream_overlap_shared_expert
-        if self.multistream_overlap_shared_expert:
-            self.shared_expert_stream = torch.npu.Stream()
         if enable_sp():
             logger.info_once(
                 "Sequence parallelism is enabled, shared experts are replicated for best performance."
@@ -442,25 +441,28 @@ def forward_impl(self, hidden_states: torch.Tensor,
                      router_logits: torch.Tensor):
         # Make sure the shared experts stream begins after hidden_states are ready.
         if self.multistream_overlap_shared_expert:
-            self.shared_expert_stream.wait_stream(  # type: ignore
+            shared_experts_compute_stream().wait_stream(  # type: ignore
                 torch.npu.current_stream())
-        with npu_stream_switch(self.shared_expert_stream,
+        with npu_stream_switch(shared_experts_compute_stream(),
                                enabled=self.multistream_overlap_shared_expert):
             # Use a separate stream to run shared experts.
+            # Note that currently we only support calculations in separate streams with aclgraph.
+            # Communication operations in another stream might cause unknown errors.
             shared_out = self._shared_experts(hidden_states)
 
-            # NOTE: This is exactly the opposite of `maybe_all_reduce_tensor_model_parallel`
-            forward_context = get_forward_context()
-            moe_comm_type = forward_context.moe_comm_type
-            if moe_comm_type in {MoECommType.ALLTOALL, MoECommType.MC2} \
-                    and not shared_expert_dp_enabled():
-                shared_out = tensor_model_parallel_all_reduce(shared_out)
         fused_output = AscendFusedMoE.forward_impl(
             self,
             hidden_states=hidden_states,
             router_logits=router_logits,
         )
         # Make sure the default stream waits for the shared experts stream to finish.
         if self.multistream_overlap_shared_expert:
-            torch.npu.current_stream().wait_stream(self.shared_expert_stream)
+            torch.npu.current_stream().wait_stream(
+                shared_experts_compute_stream())
+        # NOTE: This is exactly the opposite of `maybe_all_reduce_tensor_model_parallel`
+        forward_context = get_forward_context()
+        moe_comm_type = forward_context.moe_comm_type
+        if moe_comm_type in {MoECommType.ALLTOALL, MoECommType.MC2} \
+                and not shared_expert_dp_enabled():
+            shared_out = tensor_model_parallel_all_reduce(shared_out)
         return shared_out, fused_output

vllm_ascend/utils.py

@@ -52,6 +52,7 @@
 _SLEEP_MODE_ENABLED = None
 _CURRENT_STREAM = None
 _PREFETCH_STREAM = None
+_SHARED_EXPERTS_COMPUTE_STREAM = None
 _ASCEND_CUSTOMOP_IS_REIGISTERED = False
 _DEFAULT_BUFFER_SIZE = 200
 _MIN_DP_BUFFER_SIZE = 50
@@ -259,6 +260,15 @@ def prefetch_stream() -> torch.npu.Stream:
     return _PREFETCH_STREAM
 
 
+def shared_experts_compute_stream() -> torch.npu.Stream:
+    global _SHARED_EXPERTS_COMPUTE_STREAM
+    if _SHARED_EXPERTS_COMPUTE_STREAM is None:
+        # when this function is called before any stream is set,
+        # we return the default stream.
+        _SHARED_EXPERTS_COMPUTE_STREAM = torch_npu.npu.Stream()
+    return _SHARED_EXPERTS_COMPUTE_STREAM
+
+
 def adapt_patch(is_global_patch: bool = False):
     if is_global_patch:
         from vllm_ascend.patch import platform  # noqa: F401