[TRTLLM-6589][feat] Support CUDA graph for DeepEP (NVIDIA#7514)

Signed-off-by: Yifei Zhang <[email protected]>

Author: yifeizhang-c

cpp/tensorrt_llm/deep_ep/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(DEEP_EP_COMMIT be2582ffe69b5e7d61c3bc9bf7a5316bc48261f9)
+set(DEEP_EP_COMMIT 5be51b228a7c82dbdb213ea58e77bffd12b38af8)
 set(NVSHMEM_URL_HASH
     SHA256=eb2c8fb3b7084c2db86bd9fd905387909f1dfd483e7b45f7b3c3d5fcf5374b5a)
 

tensorrt_llm/_torch/modules/fused_moe/deep_ep_utils.py

@@ -59,7 +59,7 @@ def reserve(self, hidden_size: int, hidden_dtype: torch.dtype):
 
     def dispatch(self, x: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
                  topk_idx: torch.Tensor, topk_weights: torch.Tensor,
-                 num_experts: int, global_expert_id_offset: int) -> \
+                 num_experts: int, global_expert_id_offset: int, all_rank_max_num_tokens: int, ep_size: int, use_cuda_graph: bool) -> \
             Tuple[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]], torch.Tensor, torch.Tensor, List, Tuple]:
         # NOTES: an optional `previous_event` means a CUDA event captured that you want to make it as a dependency
         # of the dispatch kernel, it may be useful with communication-computation overlap. For more information, please
@@ -71,13 +71,16 @@ def dispatch(self, x: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
         assert event.event is None
 
         # Do MoE dispatch
-        # NOTES: the CPU will wait for GPU's signal to arrive, so this is not compatible with CUDA graph
         # For more advanced usages, please refer to the docs of the `dispatch` function
+        if use_cuda_graph:
+            num_worst_tokens = all_rank_max_num_tokens * ep_size
+        else:
+            num_worst_tokens = 0
         recv_x, recv_topk_idx, recv_topk_weights, num_recv_tokens_per_expert_list, handle, event = \
             self.buffer.dispatch(x, topk_idx=topk_idx, topk_weights=topk_weights,
                                  num_tokens_per_rank=num_tokens_per_rank, num_tokens_per_rdma_rank=num_tokens_per_rdma_rank,
                                  is_token_in_rank=is_token_in_rank, num_tokens_per_expert=num_tokens_per_expert,
-                                 global_expert_id_offset=global_expert_id_offset)
+                                 global_expert_id_offset=global_expert_id_offset, num_worst_tokens=num_worst_tokens)
         assert event.event is None
 
         # For event management, please refer to the docs of the `EventOverlap` class

tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py

@@ -32,7 +32,7 @@ class AlltoallMethodType(IntEnum):
     NotEnabled = 0
     # MNNVL
     MNNVL = 1
-    # DeepEP intranode or internode: no CUDA Graphs support, IBGDA is required by internode
+    # DeepEP intranode or internode: CUDA Graphs are supported, IBGDA is required by internode
     DeepEP = 2
     # DeepEP low latency: CUDA Graphs are supported, IBGDA is required
     DeepEPLowLatency = 3
@@ -101,6 +101,8 @@ def __init__(
         self.repeat_idx = 0
         self.repeat_count = 1
 
+        self.use_cuda_graph = model_config.use_cuda_graph
+
         moe_load_balancer_config = model_config.moe_load_balancer
         init_expert_size_per_partition = moe_load_balancer_config.num_local_slots if moe_load_balancer_config else self.num_experts // self.ep_size
         self.initial_global_assignments = [
@@ -212,6 +214,9 @@ def __init__(
                         str(
                             min(model_config.max_num_tokens,
                                 self.moe_max_num_tokens))))
+                # Set nvshmem queue pair depth larger than the number of on-flight WRs (ref: https://github.com/deepseek-ai/DeepEP/issues/427)
+                os.environ['NVSHMEM_QP_DEPTH'] = str(
+                    2 * (self.deep_ep_max_num_tokens + 1))
                 self.deep_ep_buffer = buffer_pool.get_low_latency_buffer(
                     model_config.mapping)
                 self.deep_ep_buffer.reserve(self.deep_ep_max_num_tokens,
@@ -255,6 +260,25 @@ def _check_configs(self):
     def select_alltoall_method_type(mapping: Mapping, top_k: int,
                                     dtype: torch.dtype,
                                     use_cuda_graph: bool) -> AlltoallMethodType:
+
+        # Check if DeepEP is feasible for the given number of ranks
+        # DeepEP supports two modes:
+        # 1. Intranode: Single node with 2, 4, or 8 ranks
+        # 2. Internode: 2, 4, 8, or 16 nodes with 8 ranks per node
+        def is_deepep_feasible(num_ranks: int) -> bool:
+            NUM_INTRANODE_SUPPORTED_RANKS = {2, 4, 8}
+            REQUIRED_LOCAL_MPI_SIZE = 8
+            NUM_INTERNODE_SUPPORTED_RDMA_RANKS = {2, 4, 8, 16}
+            mpi_size = local_mpi_size()
+            # Intranode cases
+            if num_ranks == mpi_size and num_ranks in NUM_INTRANODE_SUPPORTED_RANKS:
+                return True
+            # Internode cases
+            if mpi_size != REQUIRED_LOCAL_MPI_SIZE:
+                return False
+            num_rdma_nodes = num_ranks // mpi_size
+            return num_rdma_nodes in NUM_INTERNODE_SUPPORTED_RDMA_RANKS
+
         all2all_method_type = os.environ.get("TRTLLM_FORCE_ALLTOALL_METHOD")
         if all2all_method_type is not None:
             return AlltoallMethodType[all2all_method_type]
@@ -276,12 +300,10 @@ def select_alltoall_method_type(mapping: Mapping, top_k: int,
 
         if os.environ.get("TRTLLM_CAN_USE_DEEP_EP", "0") == "1":
             if deep_ep_installed and dtype == torch.bfloat16:
-                if use_cuda_graph:
-                    # Here we can only choose DeepEPLowLatency since only this method supports CUDA Graphs.
-                    return AlltoallMethodType.DeepEPLowLatency
-                else:
-                    # Here we can choose DeepEP or DeepEPLowLatency if both are available. Now DeepEP is faster.
+                # Choose DeepEP if feasible
+                if is_deepep_feasible(mapping.moe_ep_size):
                     return AlltoallMethodType.DeepEP
+                return AlltoallMethodType.DeepEPLowLatency
 
         return AlltoallMethodType.NotEnabled
 
@@ -548,7 +570,7 @@ def forward_chunk(
                 if not use_postquant_alltoall:
                     x, recv_topk_idx, token_final_scales, num_recv_tokens_per_expert_list, deep_ep_handle = \
                         self.deep_ep_buffer.dispatch(x, token_selected_slots, token_final_scales, self.num_slots,
-                        self.expert_size_per_partition * self.mapping.moe_ep_rank)
+                        self.expert_size_per_partition * self.mapping.moe_ep_rank, all_rank_max_num_tokens, self.ep_size, self.use_cuda_graph)
                     padded, x, _, token_selected_slots, token_final_scales = self.pad_empty_recv_tensors(
                         x, None, recv_topk_idx, token_final_scales)
             elif self.alltoall_method_type == AlltoallMethodType.DeepEPLowLatency:
@@ -636,7 +658,7 @@ def forward_chunk(
                     x_sf = x_sf.view(torch.float32)
                 (x, x_sf), recv_topk_idx, token_final_scales, num_recv_tokens_per_expert_list, deep_ep_handle = \
                     self.deep_ep_buffer.dispatch((x, x_sf), token_selected_slots, token_final_scales, self.num_slots,
-                    self.expert_size_per_partition * self.mapping.moe_ep_rank)
+                    self.expert_size_per_partition * self.mapping.moe_ep_rank, all_rank_max_num_tokens, self.ep_size, self.use_cuda_graph)
                 padded, x, x_sf, token_selected_slots, token_final_scales = self.pad_empty_recv_tensors(
                     x, x_sf, recv_topk_idx, token_final_scales)
                 if x_sf is not None: