Enable Allgather/ReduceScatter backend for NaiveAllToAll (vllm-project#23964)

Signed-off-by: Shu Wang. <shuw@nvidia.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Shu Wang <shuw@nvidia.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>

Author: 3 people

vllm/distributed/device_communicators/all2all.py

@@ -5,6 +5,7 @@
 import torch
 import torch.distributed as dist
 
+from vllm.distributed import get_dp_group
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.utils import has_deep_ep, has_pplx
@@ -69,6 +70,44 @@ def destroy(self):
         pass
 
 
+class AgRsAll2AllManager(All2AllManagerBase):
+    """
+    An implementation of all2all communication based on
+    all-gather (dispatch) and reduce-scatter (combine).
+    """
+
+    def __init__(self, cpu_group):
+        super().__init__(cpu_group)
+
+    def dispatch(self, hidden_states: torch.Tensor,
+                 router_logits: torch.Tensor):
+        """
+        Gather hidden_states and router_logits from all dp ranks.
+        """
+        sizes = get_forward_context(
+        ).dp_metadata.get_chunk_sizes_across_dp_rank()
+        hidden_states, router_logits = get_dp_group().all_gatherv(
+            [hidden_states, router_logits],
+            dim=0,
+            sizes=sizes,
+        )
+        return hidden_states, router_logits
+
+    def combine(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """
+        Reduce-scatter hidden_states across all dp ranks.
+        """
+        sizes = get_forward_context(
+        ).dp_metadata.get_chunk_sizes_across_dp_rank()
+        hidden_states = get_dp_group().reduce_scatterv(hidden_states,
+                                                       dim=0,
+                                                       sizes=sizes)
+        return hidden_states
+
+    def destroy(self):
+        pass
+
+
 class PPLXAll2AllManager(All2AllManagerBase):
     """
     All2All communication based on PPLX kernels.

vllm/distributed/device_communicators/cuda_communicator.py

@@ -87,6 +87,10 @@ def __init__(self,
                 from .all2all import NaiveAll2AllManager
                 self.all2all_manager = NaiveAll2AllManager(self.cpu_group)
                 logger.info("Using naive all2all manager.")
+            elif all2all_backend == "allgather_reducescatter":
+                from .all2all import AgRsAll2AllManager
+                self.all2all_manager = AgRsAll2AllManager(self.cpu_group)
+                logger.info("Using AllGather-ReduceScatter all2all manager.")
             elif all2all_backend == "pplx":
                 from .all2all import PPLXAll2AllManager
                 self.all2all_manager = PPLXAll2AllManager(self.cpu_group)

vllm/envs.py

@@ -149,8 +149,11 @@
     VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
     VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
     VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
-    VLLM_ALL2ALL_BACKEND: Literal["naive", "pplx", "deepep_high_throughput",
-                                  "deepep_low_latency"] = "naive"
+    VLLM_ALL2ALL_BACKEND: Literal["naive", "pplx",
+                                  "deepep_high_throughput",
+                                  "deepep_low_latency",
+                                  "allgather_reducescatter"] = \
+                                  "allgather_reducescatter"
     VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
     VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
     VLLM_SLEEP_WHEN_IDLE: bool = False
@@ -1124,14 +1127,18 @@ def get_vllm_port() -> Optional[int]:
 
     # all2all backend for vllm's expert parallel communication
     # Available options:
-    # - "naive": naive all2all implementation using all-reduce
+    # - "naive": naive all2all implementation using broadcasts
+    # - "allgather_reducescatter": all2all implementation based on allgather and
+    #  reducescatter
     # - "pplx": use pplx kernels
     # - "deepep_high_throughput", use deepep high-throughput kernels
     # - "deepep_low_latency", use deepep low-latency kernels
     "VLLM_ALL2ALL_BACKEND":
-    env_with_choices("VLLM_ALL2ALL_BACKEND", "naive",
+    env_with_choices("VLLM_ALL2ALL_BACKEND", "allgather_reducescatter",
                      ["naive", "pplx",
-                     "deepep_high_throughput", "deepep_low_latency"]),
+                     "deepep_high_throughput",
+                     "deepep_low_latency",
+                     "allgather_reducescatter"]),
 
     # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
     # Both require compute capability 10.0 or above.