feat: large-scale EP(part 7: DeepEP integration)

.devcontainer/docker-compose.yml

@@ -1,7 +1,7 @@
 version: "3.9"
 services:
   tensorrt_llm-dev:
-    image: urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-x86_64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506021004-9420
+    image: urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-x86_64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506111045-4792
     network_mode: host
     ipc: host
 

docker/Dockerfile.multi

@@ -72,6 +72,10 @@ RUN bash ./install_pytorch.sh $TORCH_INSTALL_TYPE && rm install_pytorch.sh
 RUN pip3 uninstall -y opencv && rm -rf /usr/local/lib/python3*/dist-packages/cv2/
 RUN pip3 install opencv-python-headless --force-reinstall --no-deps --no-cache-dir
 
+# Install DeepEP
+COPY docker/common/install_deep_ep.sh install_deep_ep.sh
+RUN bash ./install_deep_ep.sh && rm install_deep_ep.sh
+
 # WARs against security issues inherited from pytorch:25.04
 # * https://github.com/advisories/GHSA-vqfr-h8mv-ghfj
 # * https://github.com/advisories/GHSA-7cx3-6m66-7c5m

docker/common/install_deep_ep.sh

@@ -0,0 +1,47 @@
+#!/bin/bash
+
+set -euxo pipefail
+
+GITHUB_URL=${GITHUB_MIRROR:-https://github.com}
+DEEP_EP_COMMIT=2b266cf6452134f993ab0fcb3ef2d5de7683c561
+
+if [ "$(. /etc/os-release && echo $ID)" == "rocky" ]; then
+    echo "Skipping DeepEP installation in the Rocky distribution."
+    exit 0
+fi
+libmlx5_dir=$(dirname $(ldconfig -p | grep libmlx5.so.1 | head -n1 | awk '{print $NF}'))
+
+export NVCC_APPEND_FLAGS="--threads 4"
+
+# Custom NVSHMEM
+curl -fsSL https://developer.download.nvidia.com/compute/redist/nvshmem/3.2.5/source/nvshmem_src_3.2.5-1.txz | tar xz
+pushd nvshmem_src
+curl -fsSL $GITHUB_URL/deepseek-ai/DeepEP/raw/$DEEP_EP_COMMIT/third-party/nvshmem.patch | patch -p1
+sed "s/TRANSPORT_VERSION_MAJOR 3/TRANSPORT_VERSION_MAJOR 103/" -i src/CMakeLists.txt
+ln -s libmlx5.so.1 "$libmlx5_dir/libmlx5.so"
+cmake -S . -B build \
+    -DCMAKE_INSTALL_PREFIX=/opt/custom_nvshmem \
+    -DGDRCOPY_HOME=/usr/include \
+    -DNVSHMEM_SHMEM_SUPPORT=0 \
+    -DNVSHMEM_UCX_SUPPORT=0 \
+    -DNVSHMEM_USE_NCCL=0 \
+    -DNVSHMEM_MPI_SUPPORT=0 \
+    -DNVSHMEM_IBGDA_SUPPORT=1 \
+    -DNVSHMEM_PMIX_SUPPORT=0 \
+    -DNVSHMEM_TIMEOUT_DEVICE_POLLING=0 \
+    -DNVSHMEM_USE_GDRCOPY=1 \
+    -DCMAKE_CUDA_ARCHITECTURES="90-real;100-real;120-real" \
+    -DNVSHMEM_BUILD_TESTS=0 \
+    -DNVSHMEM_BUILD_EXAMPLES=0
+cmake --build build -j`nproc`
+make -C build install
+popd
+
+# DeepEP
+curl -fsSL $GITHUB_URL/deepseek-ai/DeepEP/archive/$DEEP_EP_COMMIT.tar.gz | tar xz
+TORCH_CUDA_ARCH_LIST="9.0;10.0;12.0" NVSHMEM_DIR=/opt/custom_nvshmem pip install -v --no-cache-dir ./DeepEP-$DEEP_EP_COMMIT
+
+# Clean up
+rm -r nvshmem_src
+rm "$libmlx5_dir/libmlx5.so"
+rm -r DeepEP-$DEEP_EP_COMMIT

jenkins/L0_MergeRequest.groovy

@@ -28,10 +28,10 @@ UPLOAD_PATH = env.uploadPath ? env.uploadPath : "sw-tensorrt-generic/llm-artifac
 // Container configuration
 // available tags can be found in: https://urm.nvidia.com/artifactory/sw-tensorrt-docker/tensorrt-llm/
 // [base_image_name]-[arch]-[os](-[python_version])-[trt_version]-[torch_install_type]-[stage]-[date]-[mr_id]
-LLM_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-x86_64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506021004-9420"
-LLM_SBSA_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-aarch64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506021004-9420"
-LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-12.9.0-devel-rocky8-x86_64-rocky8-py310-trt10.10.0.31-skip-tritondevel-202506021004-9420"
-LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-12.9.0-devel-rocky8-x86_64-rocky8-py312-trt10.10.0.31-skip-tritondevel-202506021004-9420"
+LLM_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-x86_64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506111045-4792"
+LLM_SBSA_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-aarch64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506111045-4792"
+LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-12.9.0-devel-rocky8-x86_64-rocky8-py310-trt10.10.0.31-skip-tritondevel-202506111045-4792"
+LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-12.9.0-devel-rocky8-x86_64-rocky8-py312-trt10.10.0.31-skip-tritondevel-202506111045-4792"
 
 // TODO: Move common variables to an unified location
 BUILD_CORES_REQUEST = "8"

jenkins/controlCCache.groovy

@@ -1,7 +1,7 @@
 
 import java.lang.InterruptedException
 
-DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-x86_64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506021004-9420"
+DOCKER_IMAGE = "urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-25.04-py3-x86_64-ubuntu24.04-trt10.10.0.31-skip-tritondevel-202506111045-4792"
 
 def createKubernetesPodConfig(image, arch = "amd64")
 {

tensorrt_llm/_torch/model_config.py

@@ -84,6 +84,9 @@ class ModelConfig(Generic[TConfig]):
     # If true, enable min-latency mode. Currently only used for Llama4.
     enable_min_latency: bool = False
 
+    # Allow models to select op according to whether CUDA Graphs are used.
+    use_cuda_graph: bool = False
+
     extra_attrs: Dict = field(default_factory=dict, repr=False, init=False)
 
     _frozen: bool = field(default=False, init=False, repr=False)

tensorrt_llm/_torch/models/modeling_deepseekv3.py

@@ -38,7 +38,6 @@
 from tqdm import tqdm
 from transformers import PretrainedConfig
 
-from tensorrt_llm._mnnvl_utils import MnnvlMemory
 from tensorrt_llm.functional import PositionEmbeddingType
 from tensorrt_llm.llmapi.utils import enable_llm_debug
 from tensorrt_llm.models.modeling_utils import QuantConfig
@@ -351,10 +350,6 @@ def __init__(self,
         config = model_config.pretrained_config
         self.top_k = top_k
         self.use_dp = model_config.mapping.enable_attention_dp
-        self.enable_alltoall = Deepseekv3MoE.should_enable_alltoall(
-            model_config, top_k)
-        if self.enable_alltoall:
-            MnnvlMemory.initialize()
         self.gate = DeepseekV3Gate(
             hidden_size,
             num_experts,
@@ -377,7 +372,6 @@ def __init__(self,
             model_config=model_config,
             override_quant_config=override_quant_config,
             aux_stream=aux_stream_dict[AuxStreamType.MoeChunkingOverlap],
-            enable_alltoall=self.enable_alltoall,
             layer_idx=layer_idx)
 
         self.mapping = model_config.mapping
@@ -443,33 +437,14 @@ def _compute_shared_expert_tp_size(self, intermediate_size: int,
 
         return shared_tp_size, shared_output_scale
 
-    @staticmethod
-    def should_enable_alltoall(model_config: ModelConfig, top_k: int) -> bool:
-        if not model_config.mapping.enable_attention_dp:
-            return False
-
-        if model_config.mapping.tp_size == 1:
-            return False
-
-        if not MnnvlMemory.supports_mnnvl():
-            return False
-
-        if os.environ.get("TRTLLM_MOE_DISABLE_ALLTOALLV", "0") == "1":
-            return False
-
-        if model_config.mapping.moe_ep_size <= top_k:
-            return False
-
-        return True
-
     def compute_routed_output(self, hidden_states, hidden_states_fp4,
                               all_rank_num_tokens, do_finalize):
         # max-throughput
         use_dp_padding = False
         if self.use_dp and self.mapping.tp_size > 1:
             # FP4 all_gather moves this bf16 allgather in to after topk and fp4 quantization
             # to reduce allreduce BW
-            if disable_fp4_allgather() and not self.enable_alltoall:
+            if disable_fp4_allgather() and not self.experts.enable_alltoall:
                 hidden_states = allgather(hidden_states,
                                           self.mapping,
                                           dim=0,

tensorrt_llm/_torch/models/modeling_qwen3_moe.py

@@ -6,8 +6,6 @@
 from tqdm import tqdm
 from transformers import Qwen3MoeConfig
 
-from tensorrt_llm._mnnvl_utils import MnnvlMemory
-
 from ..attention_backend import AttentionMetadata
 from ..distributed import (AllReduce, AllReduceFusionOp, AllReduceParams,
                            allgather)
@@ -91,10 +89,6 @@ def __init__(
         self.mapping = model_config.mapping
         self.allreduce = AllReduce(mapping=model_config.mapping,
                                    strategy=model_config.allreduce_strategy)
-        self.enable_alltoall = Qwen3MoE.should_enable_alltoall(
-            model_config, self.top_k)
-        if self.enable_alltoall:
-            MnnvlMemory.initialize()
 
         self.gate = Qwen3Gate(
             hidden_size=self.hidden_dim,
@@ -117,25 +111,6 @@ def __init__(
             model_config=model_config,
         )
 
-    @staticmethod
-    def should_enable_alltoall(model_config: ModelConfig, top_k: int) -> bool:
-        if not model_config.mapping.enable_attention_dp:
-            return False
-
-        if model_config.mapping.tp_size == 1:
-            return False
-
-        if not MnnvlMemory.supports_mnnvl():
-            return False
-
-        if os.environ.get("TRTLLM_MOE_DISABLE_ALLTOALLV", "0") == "1":
-            return False
-
-        if model_config.mapping.moe_ep_size <= top_k:
-            return False
-
-        return True
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -151,7 +126,7 @@ def forward(
         if self.enable_attention_dp and self.mapping.tp_size > 1:
             # FP4 all_gather moves this bf16 allgather in to after topk and fp4 quantization
             # to reduce allreduce BW
-            if disable_fp4_allgather() and not self.enable_alltoall:
+            if disable_fp4_allgather() and not self.experts.enable_alltoall:
                 hidden_states = allgather(hidden_states,
                                           self.mapping,
                                           dim=0,

tensorrt_llm/_torch/modules/fused_moe/create_moe.py

@@ -52,7 +52,6 @@ def create_moe(
     aux_stream: Optional[torch.cuda.Stream] = None,
     weight_loading_mode: MoEWeightLoadingMode = MoEWeightLoadingMode.VANILLA,
     apply_router_weight_on_input: bool = False,
-    enable_alltoall: bool = False,
     layer_idx: Optional[int] = None,
 ) -> MoE:
     moe_cls = get_moe_cls(model_config, override_quant_config)
@@ -63,7 +62,6 @@ def create_moe(
 
     if moe_cls == TRTLLMGenFusedMoE:
         assert not apply_router_weight_on_input, "apply_router_weight_on_input is not supported in TRTLLMGenFusedMoE."
-        assert not enable_alltoall, "enable_alltoall is not supported in TRTLLMGenFusedMoE."
 
         return moe_cls(
             routing_method=routing_method,
@@ -88,12 +86,10 @@ def create_moe(
             aux_stream=aux_stream,
             weight_loading_mode=weight_loading_mode,
             apply_router_weight_on_input=apply_router_weight_on_input,
-            enable_alltoall=enable_alltoall,
             layer_idx=layer_idx,
         )
     elif moe_cls == VanillaMoE:
         assert not apply_router_weight_on_input, "apply_router_weight_on_input is not supported in VanillaMoE."
-        assert not enable_alltoall, "enable_alltoall is not supported in VanillaMoE."
 
         return moe_cls(
             routing_method=routing_method,