Apply latest DeepEP branch (#34923)

docker/Dockerfile

@@ -9,6 +9,7 @@ ARG SGL_VERSION
 ARG SGL_DEEP_GEMM_VERSION=0.1.5.post2
 ARG USE_LATEST_SGLANG=0
 ARG GDRCOPY_VERSION=2.5.1
+ARG NCCL_VERSION=2.30.7
 ARG PIP_DEFAULT_INDEX
 ARG UBUNTU_MIRROR
 ARG GITHUB_ARTIFACTORY=github.com
@@ -173,6 +174,7 @@ ARG CUDA_VERSION
 ARG BUILD_TYPE
 ARG SGL_KERNEL_VERSION
 ARG GITHUB_ARTIFACTORY
+ARG NCCL_VERSION
 
 WORKDIR /sgl-workspace
 
@@ -248,6 +250,10 @@ RUN --mount=type=cache,target=/root/.cache/pip \
            sed -i 's/nvidia-cutlass-dsl\[cu13\]/nvidia-cutlass-dsl/' pyproject.toml; \
        fi \
     && python3 -m pip install --extra-index-url https://download.pytorch.org/whl/cu${CUINDEX} ".[${BUILD_TYPE}]" \
+    && if [ "${CUDA_VERSION%%.*}" = "13" ]; then \
+           python3 -m pip install --force-reinstall --no-deps \
+               "nvidia-nccl-cu13==${NCCL_VERSION}"; \
+       fi \
     && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
            pip list --format=freeze | awk -F'==' '/-cu13(==|$)/ {print $1}' \
                | xargs -r python3 -m pip uninstall -y && \

python/pyproject.toml

@@ -68,7 +68,7 @@ dependencies = [
   "scipy",
   "sentencepiece",
   "setproctitle",
-  "sgl-deep-ep==0.1.0",
+  "sgl-deep-ep==0.1.1",
   "sgl-deep-gemm==0.1.5.post3",
   "sglang-kernel==0.4.6.post1",
   "smg-grpc-servicer>=0.5.0",

python/sglang/srt/layers/moe/token_dispatcher/deepep.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import logging
+import os
 from contextlib import nullcontext
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, List, NamedTuple, Optional, Tuple, Union
@@ -68,6 +69,18 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and is_hip()
 
 logger = logging.getLogger(__name__)
 
+_NVSHMEM_QP_DEPTH_DEFAULT = 1024
+
+
+def _set_nvshmem_qp_depth(num_max_dispatch_tokens_per_rank: int) -> None:
+    min_qp_depth = 2 * (num_max_dispatch_tokens_per_rank + 1)
+    current_qp_depth = int(
+        os.environ.get("NVSHMEM_QP_DEPTH", _NVSHMEM_QP_DEPTH_DEFAULT)
+    )
+    os.environ["NVSHMEM_QP_DEPTH"] = str(
+        max(current_qp_depth, _NVSHMEM_QP_DEPTH_DEFAULT, min_qp_depth)
+    )
+
 
 def _is_mnnvl_fabric_supported() -> bool:
     if not is_flashinfer_available():
@@ -220,6 +233,8 @@ class DeepEPBuffer:
         if deepep_mode.enable_low_latency():
             assert num_max_dispatch_tokens_per_rank != -1
             assert num_experts != -1 and num_experts % group.size() == 0
+            if not _is_npu:
+                _set_nvshmem_qp_depth(num_max_dispatch_tokens_per_rank)
             num_rdma_bytes = max(
                 Buffer.get_low_latency_rdma_size_hint(
                     num_max_dispatch_tokens_per_rank,

scripts/ci/cuda/ci_install_dependency.sh

@@ -506,6 +506,17 @@ install_sglang() {
     mark_step_done "${FUNCNAME[0]}"
 }
 
+install_nccl() {
+    if [ "$CU_MAJOR" = "13" ]; then
+        $PIP_CMD install "nvidia-nccl-cu13==2.30.7" \
+            --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+    else
+        echo "CUDA ${CU_MAJOR} does not require the NCCL Gin wheel"
+    fi
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
 # Trust an installed wheel only if the version matches and every RECORD file is
 # on disk (dist-info can survive a partial install - cf. the cusparselt guard).
 # reject-local refuses wheels installed from a local file: a kernel-PR job
@@ -843,6 +854,7 @@ main() {
     install_pytorch_stack
     install_cuda12_deepep_wheel
     install_sglang
+    install_nccl
     # Diffusion B200 CI imports torch inside install_sglang_kernel after removing
     # stale CUDA 12 NVIDIA wheels, so opt into one early LD_LIBRARY_PATH refresh.
     if [ "${SGLANG_CI_EARLY_LD_LIBRARY_PATH:-0}" = "1" ]; then

test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py

@@ -93,6 +93,10 @@ class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):
         self.assertGreater(metrics["score"], 0.92)
 
 
+@unittest.skipIf(
+    True,
+    "Skipping since the current DeepEP doesn't accept the --enable-single-batch-overlap flag. To be fixed after https://github.com/sgl-project/DeepEP/pull/10",
+)
 class TestDummyWithSBO(CustomTestCase):
     @classmethod
     def setUpClass(cls):