Register allgather/reducescatter buffers with symm memory (#12572)

Author: nvcastet

python/sglang/srt/distributed/device_communicators/pynccl_allocator.py

@@ -1,11 +1,11 @@
 import os
 import tempfile
+from contextlib import nullcontext
 
 import torch
 from torch.cuda.memory import CUDAPluggableAllocator
 
 from sglang.srt.distributed.parallel_state import GroupCoordinator
-from sglang.srt.server_args import get_global_server_args
 
 nccl_allocator_source = """
 
@@ -60,6 +60,9 @@
 
 
 def is_symmetric_memory_enabled():
+    # Import here to avoid circular import
+    from sglang.srt.server_args import get_global_server_args
+
     return get_global_server_args().enable_symm_mem
 
 
@@ -92,33 +95,25 @@ def get_nccl_mem_pool():
     return _mem_pool
 
 
-class use_symmetric_memory:
+class SymmetricMemoryContext:
     """
     Context manager for using symmetric memory with pynccl.
 
     To Utilize the symmetric memory feature in NCCL, the buffers need to be allocated
     by `ncclMemAlloc` and registered by `ncclCommWindowRegister`. Due to this, we introduce
     this context manager. All tensors created under this context will be correctly
     allocated and registered with a custom allocator.
-
-    In addition, developers need to manually tag the tensors that will be used as the input/output
-    of NCCL collectives with `tag(tensor)`.
     """
 
-    def __init__(self, group_coordinator: GroupCoordinator):
-        self.enabled = is_symmetric_memory_enabled()
-
-        if not self.enabled:
-            return
-
+    def __init__(
+        self,
+        group_coordinator: GroupCoordinator,
+    ):
         self.group_coordinator = group_coordinator
         self._mem_pool_ctx = torch.cuda.use_mem_pool(get_nccl_mem_pool())
         self.is_graph_capture = torch.cuda.is_current_stream_capturing()
 
     def __enter__(self):
-        if not self.enabled:
-            return self
-
         assert (
             self.group_coordinator.pynccl_comm is not None
         ), f"Symmetric memory requires pynccl to be enabled in group '{self.group_coordinator.group_name}'"
@@ -139,16 +134,16 @@ def __enter__(self):
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        if not self.enabled:
-            return
-
         self._mem_pool_ctx.__exit__(exc_type, exc_val, exc_tb)
 
         if self.is_graph_capture:
             torch._C._cuda_beginAllocateCurrentThreadToPool(_cur_device, _graph_pool_id)
 
-    def tag(self, tensor: torch.Tensor):
-        if not self.enabled:
-            return
 
-        tensor.symmetric_memory = True
+def use_symmetric_memory(group_coordinator: GroupCoordinator, disabled: bool = False):
+    disabled = (
+        not is_symmetric_memory_enabled()
+        or disabled
+        or group_coordinator.world_size == 1
+    )
+    return SymmetricMemoryContext(group_coordinator) if not disabled else nullcontext()

python/sglang/srt/distributed/parallel_state.py

@@ -188,6 +188,27 @@ def reg_all_gather_into_tensor_fake(
         fake_impl=reg_all_gather_into_tensor_fake,
     )
 
+    def reg_reduce_scatter_tensor(
+        output: torch.Tensor, input: torch.Tensor, group_name: str
+    ) -> None:
+        assert group_name in _groups, f"Group {group_name} is not found."
+        group = _groups[group_name]()
+        if group is None:
+            raise ValueError(f"Group {group_name} is destroyed.")
+        group._reduce_scatter_tensor(output, input)
+
+    def reg_reduce_scatter_tensor_fake(
+        output: torch.Tensor, input: torch.Tensor, group_name: str
+    ) -> None:
+        pass
+
+    direct_register_custom_op(
+        op_name="reg_reduce_scatter_tensor",
+        op_func=reg_reduce_scatter_tensor,
+        mutates_args=["output"],
+        fake_impl=reg_reduce_scatter_tensor_fake,
+    )
+
 
 class GroupCoordinator:
     """
@@ -314,10 +335,16 @@ def __init__(
         from sglang.srt.distributed.device_communicators.pynccl import (
             PyNcclCommunicator,
         )
+        from sglang.srt.distributed.device_communicators.pynccl_allocator import (
+            is_symmetric_memory_enabled,
+            use_symmetric_memory,
+        )
         from sglang.srt.distributed.device_communicators.torch_symm_mem import (
             TorchSymmMemCommunicator,
         )
 
+        self.is_symmetric_memory_enabled = is_symmetric_memory_enabled
+        self.use_symmetric_memory = use_symmetric_memory
         if is_hip():
             from sglang.srt.distributed.device_communicators.quick_all_reduce import (
                 QuickAllReduce,
@@ -552,7 +579,7 @@ def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
         if self.npu_communicator is not None and not self.npu_communicator.disabled:
             return self.npu_communicator.all_reduce(input_)
 
-        if self.pynccl_comm is not None and getattr(input_, "symmetric_memory", False):
+        if self.pynccl_comm is not None and self.is_symmetric_memory_enabled():
             with self.pynccl_comm.change_state(
                 enable=True, stream=get_current_device_stream_fast()
             ):
@@ -627,15 +654,33 @@ def _all_reduce_in_place(self, input_: torch.Tensor) -> None:
         else:
             torch.distributed.all_reduce(input_, group=self.device_group)
 
-    def reduce_scatter_tensor(
+    def _reduce_scatter_tensor(
         self,
         output: torch.Tensor,
         input: torch.Tensor,
-    ) -> None:
-        # TODO(ch-wan): support other backends
-        torch.distributed.reduce_scatter_tensor(output, input, group=self.device_group)
+    ) -> torch.Tensor:
+        pynccl_comm = self.pynccl_comm
+        if pynccl_comm is not None and (
+            not pynccl_comm.disabled or self.is_symmetric_memory_enabled()
+        ):
+            with pynccl_comm.change_state(
+                enable=True, stream=get_current_device_stream_fast()
+            ):
+                pynccl_comm.reduce_scatter(output, input)
+        else:
+            torch.distributed.reduce_scatter_tensor(
+                output, input, group=self.device_group
+            )
         return output
 
+    def reduce_scatter_tensor(self, output: torch.Tensor, input: torch.Tensor):
+        if _is_npu or not supports_custom_op():
+            self._reduce_scatter_tensor(output, input)
+        else:
+            torch.ops.sglang.reg_reduce_scatter_tensor(
+                output, input, group_name=self.unique_name
+            )
+
     def reduce_scatter(
         self,
         output: torch.Tensor,
@@ -682,8 +727,13 @@ def reduce_scatterv(
 
     def _all_gather_into_tensor(self, output: torch.Tensor, input: torch.Tensor):
         pynccl_comm = self.pynccl_comm
-        if pynccl_comm is not None and not pynccl_comm.disabled:
-            pynccl_comm.all_gather(output, input)
+        if pynccl_comm is not None and (
+            not pynccl_comm.disabled or self.is_symmetric_memory_enabled()
+        ):
+            with pynccl_comm.change_state(
+                enable=True, stream=get_current_device_stream_fast()
+            ):
+                pynccl_comm.all_gather(output, input)
         else:
             torch.distributed.all_gather_into_tensor(
                 output, input, group=self.device_group
@@ -745,9 +795,10 @@ def all_gather(
         # torch.compile . see https://github.com/pytorch/pytorch/issues/138795
         output_size = (input_size[0] * world_size,) + input_size[1:]
         # Allocate output tensor.
-        output_tensor = torch.empty(
-            output_size, dtype=input_.dtype, device=input_.device
-        )
+        with self.use_symmetric_memory(self):
+            output_tensor = torch.empty(
+                output_size, dtype=input_.dtype, device=input_.device
+            )
 
         # All-gather.
         if input_.is_cpu:
@@ -787,7 +838,7 @@ def all_gatherv(
                 pynccl_comm is not None and not pynccl_comm.disabled
             ), "pynccl is required for all_gatherv"
 
-            def _all_gather_single(
+            def _all_gather_allocate_output(
                 input_: torch.Tensor, sizes: Optional[List[int]] = None
             ):
                 input_size = input_.size()
@@ -801,19 +852,25 @@ def _all_gather_single(
                 else:
                     output_size = (input_size[0] * world_size,) + input_size[1:]
                 # Allocate output tensor.
-                output_tensor = torch.empty(
-                    output_size, dtype=input_.dtype, device=input_.device
-                )
-                pynccl_comm.all_gather(output_tensor, input_, sizes=sizes)
-                return output_tensor
+                with self.use_symmetric_memory(self, disabled=sizes is not None):
+                    output_tensor = torch.empty(
+                        output_size, dtype=input_.dtype, device=input_.device
+                    )
+                return output_tensor, sizes
 
             if isinstance(input_, torch.Tensor):
-                return _all_gather_single(input_, sizes)
+                input_ = [input_]
 
             output_list = []
-            pynccl_comm.group_start()
+            size_list = []
             for inp in input_:
-                output_list.append(_all_gather_single(inp, sizes=sizes))
+                output_tensor, s = _all_gather_allocate_output(inp, sizes=sizes)
+                output_list.append(output_tensor)
+                size_list.append(s)
+
+            pynccl_comm.group_start()
+            for i, inp in enumerate(input_):
+                pynccl_comm.all_gather(output_list[i], inp, sizes=size_list[i])
             pynccl_comm.group_end()
 
             return output_list

python/sglang/srt/layers/communicator.py

@@ -21,8 +21,12 @@
 
 from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
+    get_tp_group,
     tensor_model_parallel_all_reduce,
 )
+from sglang.srt.distributed.device_communicators.pynccl_allocator import (
+    use_symmetric_memory,
+)
 from sglang.srt.layers.dp_attention import (
     attn_tp_all_gather_into_tensor,
     attn_tp_reduce_scatter_tensor,
@@ -34,6 +38,7 @@
     get_attention_tp_size,
     get_global_dp_buffer,
     get_local_dp_buffer,
+    is_allocation_symmetric,
     is_dp_attention_enabled,
 )
 from sglang.srt.layers.moe import (
@@ -540,7 +545,12 @@ def _gather_hidden_states_and_residual(
             use_layer_norm_before_gather = context.attn_tp_size == 1
             if use_layer_norm_before_gather and hidden_states.shape[0] != 0:
                 residual = hidden_states
-                hidden_states = layernorm(hidden_states)
+                with use_symmetric_memory(
+                    get_tp_group(),
+                    disabled=not is_allocation_symmetric(),
+                ):
+                    hidden_states = layernorm(hidden_states)
+
             hidden_states, local_hidden_states = (
                 get_global_dp_buffer(),
                 hidden_states,

python/sglang/srt/layers/dp_attention.py

@@ -17,6 +17,9 @@
     get_tp_group,
     tensor_model_parallel_all_reduce,
 )
+from sglang.srt.distributed.device_communicators.pynccl_allocator import (
+    use_symmetric_memory,
+)
 from sglang.srt.utils import get_bool_env_var, is_hip
 
 if TYPE_CHECKING:
@@ -86,6 +89,7 @@ class _DpGatheredBufferWrapper:
     _device: torch.device
     _global_dp_buffer_len: int
     _local_dp_buffer_len: int
+    _dp_max_padding: bool
     _global_num_tokens: Optional[List[int]]
     _is_extend_in_batch: bool
 
@@ -100,27 +104,33 @@ def set_dp_buffer_len(
         cls,
         global_dp_buffer_len: int,
         local_dp_buffer_len: int,
+        dp_max_padding: bool,
         global_num_tokens: Optional[List[int]] = None,
     ):
         cls._global_dp_buffer_len = global_dp_buffer_len
         cls._local_dp_buffer_len = local_dp_buffer_len
+        cls._dp_max_padding = dp_max_padding
         cls._global_num_tokens = global_num_tokens
 
     @classmethod
     def get_global_dp_buffer(cls) -> torch.Tensor:
-        return torch.empty(
-            (cls._global_dp_buffer_len, cls._hidden_size),
-            dtype=cls._dtype,
-            device=cls._device,
-        )
+        with use_symmetric_memory(get_tp_group()):
+            buffer = torch.empty(
+                (cls._global_dp_buffer_len, cls._hidden_size),
+                dtype=cls._dtype,
+                device=cls._device,
+            )
+        return buffer
 
     @classmethod
     def get_local_dp_buffer(cls) -> torch.Tensor:
-        return torch.empty(
-            (cls._local_dp_buffer_len, cls._hidden_size),
-            dtype=cls._dtype,
-            device=cls._device,
-        )
+        with use_symmetric_memory(get_tp_group(), disabled=not cls._dp_max_padding):
+            buffer = torch.empty(
+                (cls._local_dp_buffer_len, cls._hidden_size),
+                dtype=cls._dtype,
+                device=cls._device,
+            )
+        return buffer
 
     @classmethod
     def get_global_dp_buffer_len(cls) -> int:
@@ -154,14 +164,19 @@ def set_is_extend_in_batch(cls, is_extend_in_batch: bool):
     def get_is_extend_in_batch(cls) -> bool:
         return cls._is_extend_in_batch
 
+    @classmethod
+    def is_dp_max_padding(cls) -> bool:
+        return cls._dp_max_padding
+
 
 def set_dp_buffer_len(
     global_dp_buffer_len: int,
     local_dp_buffer_len: int,
+    dp_max_padding: bool,
     global_num_tokens: Optional[List[int]] = None,
 ):
     _DpGatheredBufferWrapper.set_dp_buffer_len(
-        global_dp_buffer_len, local_dp_buffer_len, global_num_tokens
+        global_dp_buffer_len, local_dp_buffer_len, dp_max_padding, global_num_tokens
     )
 
 
@@ -205,6 +220,10 @@ def get_is_extend_in_batch() -> bool:
     return _DpGatheredBufferWrapper.get_is_extend_in_batch()
 
 
+def is_dp_max_padding() -> bool:
+    return _DpGatheredBufferWrapper.is_dp_max_padding()
+
+
 def compute_dp_attention_world_info(enable_dp_attention, tp_rank, tp_size, dp_size):
     if not enable_dp_attention:
         return tp_rank, tp_size, 0
@@ -298,6 +317,10 @@ def is_dp_attention_enabled() -> bool:
     return _ENABLE_DP_ATTENTION_FLAG
 
 
+def is_allocation_symmetric() -> bool:
+    return not is_dp_attention_enabled() or is_dp_max_padding()
+
+
 def get_attention_tp_group() -> GroupCoordinator:
     assert _ATTN_TP_GROUP is not None, "dp attention not initialized!"
     return _ATTN_TP_GROUP