Install DeepEP from release wheels (#33932)

2026-08-07 15:38:44 -07:00
parent 115cd7bde1
commit eb3cc879e0
31 changed files with 141 additions and 353 deletions
@@ -40,7 +40,6 @@ env:
  SGLANG_CUDA_COREDUMP: "1"
  SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
  SKIP_PR_TEST_HEALTH_CHECK: ${{ inputs.skip_pr_test_health_check && 'true' || 'false' }}
-  FORCE_REBUILD_DEEPEP: '1'
  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
  USE_VENV: false

@@ -75,7 +75,6 @@ env:
  SGLANG_CUDA_COREDUMP: "1"
  SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
  SKIP_PR_TEST_HEALTH_CHECK: ${{ (fromJson(inputs.caller_inputs).skip_pr_test_health_check || fromJson(inputs.caller_inputs).test_parallel_dispatch || fromJson(inputs.caller_inputs).run_all_tests) && 'true' || 'false' }}
-  FORCE_REBUILD_DEEPEP: '1'
  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
  USE_VENV: false

@@ -136,8 +135,6 @@ jobs:

      - name: Install dependencies
        timeout-minutes: ${{ fromJson(steps.rc.outputs.install_timeout) }}
-        env:
-          GRACE_BLACKWELL: ${{ steps.rc.outputs.grace_blackwell || '0' }}
        run: |
          CUSTOM_BUILD_SGL_KERNEL=${{ fromJson(inputs.check_changes).sgl_kernel }} bash ${{ steps.rc.outputs.install }}

@@ -486,10 +486,8 @@ jobs:
      - uses: ./.github/actions/check-maintenance

      - name: Install dependencies
-        env:
-          GRACE_BLACKWELL: "1"
        run: |
-          bash scripts/ci/cuda/ci_install_deepep.sh
+          bash scripts/ci/cuda/ci_install_dependency.sh

      - name: Run test
        timeout-minutes: 600
@@ -64,7 +64,6 @@ env:
  SGLANG_CUDA_COREDUMP: "1"
  SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
  SKIP_PR_TEST_HEALTH_CHECK: ${{ (inputs.skip_pr_test_health_check == true || inputs.run_all_tests == true) && 'true' || 'false' }}
-  FORCE_REBUILD_DEEPEP: '1'
  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
  USE_VENV: false

@@ -238,49 +237,6 @@ jobs:
      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
    secrets: inherit

-  extra-b-test-deepep-4-gpu-h100:
-    needs: [check-changes, call-gate, sgl-kernel-build-wheels, rust-ext-build]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && (needs.call-gate.result == 'success' || needs.call-gate.result == 'skipped') }}
-    uses: ./.github/workflows/_pr-test-stage.yml
-    with:
-      self_name: extra-b-test-deepep-4-gpu-h100
-      runner_config: deepep-4-gpu-h100
-      check_changes: ${{ toJson(needs.check-changes.outputs) }}
-      caller_inputs: ${{ toJson(inputs) }}
-      partitions: ${{ needs.check-changes.outputs.partitions }}
-      run_timeout_minutes: '60'
-      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-    secrets: inherit
-
-  extra-b-test-deepep-4-gpu-b200:
-    needs: [check-changes, call-gate, sgl-kernel-build-wheels, rust-ext-build]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && (needs.call-gate.result == 'success' || needs.call-gate.result == 'skipped') }}
-    uses: ./.github/workflows/_pr-test-stage.yml
-    with:
-      self_name: extra-b-test-deepep-4-gpu-b200
-      runner_config: deepep-4-gpu-b200
-      check_changes: ${{ toJson(needs.check-changes.outputs) }}
-      caller_inputs: ${{ toJson(inputs) }}
-      partitions: ${{ needs.check-changes.outputs.partitions }}
-      run_timeout_minutes: '60'
-      timeout_per_file: '1800'
-      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-    secrets: inherit
-
-  extra-b-test-deepep-8-gpu-h200:
-    needs: [check-changes, call-gate, sgl-kernel-build-wheels, rust-ext-build]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && (needs.call-gate.result == 'success' || needs.call-gate.result == 'skipped') }}
-    uses: ./.github/workflows/_pr-test-stage.yml
-    with:
-      self_name: extra-b-test-deepep-8-gpu-h200
-      runner_config: deepep-8-gpu-h200
-      check_changes: ${{ toJson(needs.check-changes.outputs) }}
-      caller_inputs: ${{ toJson(inputs) }}
-      partitions: ${{ needs.check-changes.outputs.partitions }}
-      run_timeout_minutes: '60'
-      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-    secrets: inherit
-
  # =============================================== aggregator ====================================================
  # Mirrors pr-test.yml's `pr-test-finish` so notify-pr-states below only
  # depends on one job rather than re-listing every stage. Fails if any
@@ -298,9 +254,6 @@ jobs:
        extra-b-test-4-gpu-h100,
        extra-b-test-4-gpu-b200,
        extra-b-test-8-gpu-h200,
-        extra-b-test-deepep-4-gpu-h100,
-        extra-b-test-deepep-4-gpu-b200,
-        extra-b-test-deepep-8-gpu-h200,
      ]
    if: always()
    runs-on: ubuntu-latest
@@ -59,8 +59,6 @@ env:
  SGLANG_CUDA_COREDUMP: "1"
  SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
  SKIP_PR_TEST_HEALTH_CHECK: ${{ (inputs.skip_pr_test_health_check == true || inputs.test_parallel_dispatch == true || inputs.run_all_tests == true) && 'true' || 'false' }}
-  # TEMP: force DeepEP rebuilds during the initial Torch 2.13 rollout so every CI runner replaces its Torch 2.11 build; remove after the runner fleet has cycled.
-  FORCE_REBUILD_DEEPEP: '1'
  # Schedule / main-branch dispatch / workflow_call from main use refs/heads/main; PR events use refs/pull/*/merge
  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
  USE_VENV: false
@@ -400,6 +398,8 @@ jobs:
      caller_inputs: ${{ toJson(inputs) }}
      partitions: ${{ needs.check-changes.outputs.partitions }}
      run_timeout_minutes: '30'
+      warmup_deep_gemm_models: 'lmsys/sglang-ci-dsv3-test:4'
+      warmup_server_models: 'lmsys/sglang-ci-dsv3-test:4'
      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
    secrets: inherit

@@ -414,6 +414,7 @@ jobs:
      caller_inputs: ${{ toJson(inputs) }}
      partitions: ${{ needs.check-changes.outputs.partitions }}
      run_timeout_minutes: '30'
+      timeout_per_file: '1800'
      # Per-model TP must match the test's launch in test/registered/ -- see
      # FALLBACK_ARGS in scripts/ci/cuda/warmup_deep_gemm.py for extra dp/ep
      # flags. Only models that actually invoke DeepGEMM kernels at runtime
@@ -438,52 +439,6 @@ jobs:
      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
    secrets: inherit

-  base-c-test-deepep-4-gpu-h100:
-    needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels, rust-ext-build]
-    if: ${{ !failure() && !cancelled() }}
-    uses: ./.github/workflows/_pr-test-stage.yml
-    with:
-      self_name: base-c-test-deepep-4-gpu-h100
-      runner_config: deepep-4-gpu-h100
-      check_changes: ${{ toJson(needs.check-changes.outputs) }}
-      caller_inputs: ${{ toJson(inputs) }}
-      partitions: ${{ needs.check-changes.outputs.partitions }}
-      run_timeout_minutes: '30'
-      warmup_deep_gemm_models: 'lmsys/sglang-ci-dsv3-test:4'
-      warmup_server_models: 'lmsys/sglang-ci-dsv3-test:4'
-      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-    secrets: inherit
-
-  base-c-test-deepep-4-gpu-b200:
-    needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels, rust-ext-build]
-    if: ${{ !failure() && !cancelled() }}
-    uses: ./.github/workflows/_pr-test-stage.yml
-    with:
-      self_name: base-c-test-deepep-4-gpu-b200
-      runner_config: deepep-4-gpu-b200
-      check_changes: ${{ toJson(needs.check-changes.outputs) }}
-      caller_inputs: ${{ toJson(inputs) }}
-      partitions: ${{ needs.check-changes.outputs.partitions }}
-      run_timeout_minutes: '30'
-      timeout_per_file: '1800'
-      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-    secrets: inherit
-
-  base-c-test-deepep-8-gpu-h200:
-    needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels, rust-ext-build]
-    if: ${{ !failure() && !cancelled() }}
-    uses: ./.github/workflows/_pr-test-stage.yml
-    with:
-      self_name: base-c-test-deepep-8-gpu-h200
-      runner_config: deepep-8-gpu-h200
-      check_changes: ${{ toJson(needs.check-changes.outputs) }}
-      caller_inputs: ${{ toJson(inputs) }}
-      partitions: ${{ needs.check-changes.outputs.partitions }}
-      run_timeout_minutes: '30'
-      timeout_per_file: '1800'
-      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-    secrets: inherit
-
  base-c-test-4-gpu-b200:
    needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels, rust-ext-build]
    if: ${{ !failure() && !cancelled() }}
@@ -561,9 +516,6 @@ jobs:
        base-c-test-4-gpu-h100,
        base-c-test-8-gpu-h20,
        base-c-test-8-gpu-h200,
-        base-c-test-deepep-4-gpu-h100,
-        base-c-test-deepep-4-gpu-b200,
-        base-c-test-deepep-8-gpu-h200,
        base-c-test-4-gpu-b200,
        base-c-test-4-gpu-gb300,
        base-c-test-8-gpu-b300,
@@ -31,11 +31,6 @@ on:
        required: false
        type: string
        default: "20"
-      grace_blackwell:
-        description: "Set GRACE_BLACKWELL for the install step (cuda only)"
-        required: false
-        type: string
-        default: "0"
      rdma_devices:
        description: "SGLANG_CI_RDMA_ALL_DEVICES csv (cuda only; empty = unset)"
        required: false
@@ -62,8 +57,6 @@ env:
  SGLANG_ENABLE_ASYNC_ASSERT: true
  SGLANG_CUDA_COREDUMP: "1"
  SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
-  # TEMP: force DeepEP rebuilds during the initial Torch 2.13 rollout so every CI runner replaces its Torch 2.11 build; remove after the runner fleet has cycled.
-  FORCE_REBUILD_DEEPEP: '1'

 permissions:
  actions: write
@@ -109,8 +102,6 @@ jobs:

      - name: Install dependencies
        timeout-minutes: ${{ fromJson(inputs.install_timeout) }}
-        env:
-          GRACE_BLACKWELL: ${{ inputs.grace_blackwell }}
        run: |
          if [[ "${{ inputs.runs_on }}" == "1-gpu-5090" ]]; then
            source /etc/profile.d/sglang-ci.sh
@@ -120,7 +111,7 @@ jobs:
          # then dies with `ModuleNotFoundError: No module named 'sglang'`.
          # /rerun-test resolves install_script from runner_configs.yml; a manual
          # workflow_dispatch must pass it explicitly (it can't be derived from
-          # runs_on, which is shared across configs, e.g. default vs deepep).
+          # runs_on, which can be shared across runner configs).
          if [ -z "${{ inputs.install_script }}" ]; then
            echo "::error::install_script is required for cuda mode (empty would silently skip installing sglang). Pass e.g. scripts/ci/cuda/ci_install_dependency.sh"
            exit 1
@@ -67,6 +67,7 @@ dependencies = [
  "scipy",
  "sentencepiece",
  "setproctitle",
+  "sgl-deep-ep==0.1.0",
  "sgl-deep-gemm==0.1.5.post2",
  "sglang-kernel==0.4.6.post1",
  "smg-grpc-servicer>=0.5.0",
@@ -1,177 +0,0 @@
-#!/bin/bash
-# Install the dependency in CI.
-set -euxo pipefail
-
-# Source (not bash) so that venv activation, $PIP_CMD, $CU_VERSION, $NVCC_VER, and
-# $PIP_INSTALL_SUFFIX all propagate into this shell. Without sourcing, the subshell
-# exits and this script would fall back to system Python.
-#
-# Note: any `exit N` or `set -e` trip inside the sourced script terminates *this*
-# script too (bash runs sourced commands in the current shell, so `exit` is not
-# caught by `if`/`||`). The real error message appears upstream in the log.
-# shellcheck disable=SC1091
-source scripts/ci/cuda/ci_install_dependency.sh
-
-# In venv mode, PIP_CMD must be set by the sourced script. If it isn't, the
-# source chain is broken and we'd silently fall back to system `pip` below —
-# exactly the split-install bug the migration is meant to prevent.
-if [ -z "${PIP_CMD:-}" ]; then
-    echo "FATAL:PIP_CMD is unset after sourcing ci_install_dependency.sh"
-    exit 1
-fi
-
-export GDRCOPY_HOME=/usr/src/gdrdrv-2.5.1/
-export CUDA_HOME=/usr/local/cuda
-
-GRACE_BLACKWELL=${GRACE_BLACKWELL:-0}
-# Detect architecture
-ARCH=$(uname -m)
-if [ "$ARCH" != "x86_64" ] && [ "$ARCH" != "aarch64" ]; then
-    echo "Unsupported architecture: $ARCH"
-    exit 1
-fi
-
-if [ "${FORCE_REBUILD_DEEPEP:-0}" = "1" ]; then
-    echo "FORCE_REBUILD_DEEPEP=1; uninstalling any cached deep_ep before rebuild."
-    ${PIP_UNINSTALL_CMD:-pip uninstall -y} deep_ep ${PIP_UNINSTALL_SUFFIX:-} || true
-elif python3 -c "import deep_ep" >/dev/null 2>&1; then
-    echo "deep_ep is already installed or importable. Skipping installation."
-    exit 0
-fi
-
-# Install system dependencies
-# Use fallback logic in case apt fails due to unrelated broken packages on the runner
-DEEPEP_SYSTEM_DEPS="curl wget git sudo rdma-core infiniband-diags openssh-server perftest libibumad3 libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils libnl-3-200 libnl-route-3-200 librdmacm1 build-essential cmake"
-apt-get install -y --no-install-recommends $DEEPEP_SYSTEM_DEPS || {
-    echo "Warning: apt-get install failed, checking if required packages are available..."
-    for pkg in $DEEPEP_SYSTEM_DEPS; do
-        if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
-            echo "ERROR: Required package $pkg is not installed and apt-get failed"
-            exit 1
-        fi
-    done
-    echo "All required packages are already installed, continuing..."
-}
-
-# Install GDRCopy
-rm -rf /opt/gdrcopy && mkdir -p /opt/gdrcopy
-cd /opt/gdrcopy
-git clone https://github.com/NVIDIA/gdrcopy.git .
-git checkout v2.5.1
-apt-get update || true  # May fail due to unrelated broken packages
-GDRCOPY_DEPS_1="nvidia-dkms-580"
-GDRCOPY_DEPS_2="build-essential devscripts debhelper fakeroot pkg-config dkms"
-GDRCOPY_DEPS_3="check libsubunit0 libsubunit-dev python3-venv"
-for deps_group in "$GDRCOPY_DEPS_1" "$GDRCOPY_DEPS_2" "$GDRCOPY_DEPS_3"; do
-    apt-get install -y --no-install-recommends $deps_group || {
-        echo "Warning: apt-get install failed for '$deps_group', checking if packages are available..."
-        for pkg in $deps_group; do
-            if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
-                echo "ERROR: Required package $pkg is not installed and apt-get failed"
-                exit 1
-            fi
-        done
-        echo "All required packages from '$deps_group' are already installed, continuing..."
-    }
-done
-cd packages
-CUDA=/usr/local/cuda ./build-deb-packages.sh
-dpkg -i gdrdrv-dkms_*.deb
-dpkg -i libgdrapi_*.deb
-dpkg -i gdrcopy-tests_*.deb
-dpkg -i gdrcopy_*.deb
-
-# Set up library paths based on architecture
-LIB_PATH="/usr/lib/$ARCH-linux-gnu"
-if [ ! -e "$LIB_PATH/libmlx5.so" ]; then
-    ln -s $LIB_PATH/libmlx5.so.1 $LIB_PATH/libmlx5.so
-fi
-apt-get update || true
-apt-get install -y --no-install-recommends libfabric-dev || {
-    if ! dpkg -l libfabric-dev 2>/dev/null | grep -q "^ii"; then
-        echo "ERROR: Required package libfabric-dev is not installed and apt-get failed"
-        exit 1
-    fi
-    echo "libfabric-dev is already installed, continuing..."
-}
-
-# Install DeepEP
-DEEPEP_DIR=/root/.cache/deepep
-rm -rf ${DEEPEP_DIR}
-if [ "$GRACE_BLACKWELL" = "1" ]; then
-    GRACE_BLACKWELL_DEEPEP_BRANCH=hybrid-ep
-    git clone https://github.com/deepseek-ai/DeepEP.git -b ${GRACE_BLACKWELL_DEEPEP_BRANCH} ${DEEPEP_DIR} && \
-    pushd ${DEEPEP_DIR} && \
-    git checkout d28bd676c2120573c9f1425f0c16c39faa4117e6 && \
-    sed -i 's/#define NUM_CPU_TIMEOUT_SECS 100/#define NUM_CPU_TIMEOUT_SECS 1000/' csrc/kernels/configs.cuh && \
-    popd
-else
-    git clone https://github.com/deepseek-ai/DeepEP.git ${DEEPEP_DIR} && \
-    pushd ${DEEPEP_DIR} && \
-    git checkout 9af0e0d0e74f3577af1979c9b9e1ac2cad0104ee && \
-    popd
-fi
-
-cd ${DEEPEP_DIR}
-if [ "$GRACE_BLACKWELL" = "1" ]; then
-    # Resolve the toolkit CUDA version. Preference order:
-    #   1. $NVCC_VER inherited from the sourced ci_install_dependency.sh
-    #      (both scripts agree on the detected value, no re-detection cost).
-    #   2. Local `nvcc --version` (authoritative — container toolkit).
-    #   3. `nvidia-smi` (host driver; last resort).
-    if [ -n "${NVCC_VER:-}" ]; then
-        CUDA_VERSION="$NVCC_VER"
-    elif command -v nvcc >/dev/null 2>&1; then
-        CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[0-9]+\.[0-9]+')
-    else
-        CUDA_VERSION=$(nvidia-smi | grep "CUDA Version" | head -n1 | awk '{print $9}' || true)
-    fi
-    if [ -z "${CUDA_VERSION:-}" ]; then
-        echo "FATAL: could not determine CUDA toolkit version (NVCC_VER unset, nvcc missing, nvidia-smi empty)"
-        exit 1
-    fi
-    if [ "$CUDA_VERSION" = "12.8" ]; then
-        CHOSEN_TORCH_CUDA_ARCH_LIST='10.0'
-    elif awk -v ver="$CUDA_VERSION" 'BEGIN {exit !(ver > 12.8)}'; then
-        # CUDA > 12.8 supports sm_103 (Blackwell)
-        CHOSEN_TORCH_CUDA_ARCH_LIST='10.0;10.3'
-    else
-        echo "Unsupported CUDA version for Grace Blackwell: $CUDA_VERSION" && exit 1
-    fi && \
-    if [ "${CUDA_VERSION%%.*}" = "13" ]; then \
-        sed -i "/^    include_dirs = \['csrc\/'\]/a\    include_dirs.append('${CUDA_HOME}/include/cccl')" setup.py; \
-    fi
-    TORCH_CUDA_ARCH_LIST="${CHOSEN_TORCH_CUDA_ARCH_LIST}" ${PIP_CMD:-pip} install --no-build-isolation . ${PIP_INSTALL_SUFFIX:-}
-else
-    # CUDA 13.0 puts CCCL headers in /usr/local/cuda/include/cccl/ but nvshmem
-    # includes them as <cuda/__cccl_config> expecting /usr/local/cuda/include/cuda/.
-    # Add the cccl path to setup.py include_dirs so the compiler finds them.
-    NVCC_MAJOR=$(nvcc --version 2>/dev/null | grep -oP 'release \K[0-9]+' || echo "0")
-    if [ "$NVCC_MAJOR" = "13" ]; then
-        sed -i "/^    include_dirs = \['csrc\/'\]/a\    include_dirs.append('${CUDA_HOME:-/usr/local/cuda}/include/cccl')" setup.py
-    fi
-
-    # Build for both Hopper (sm_90) and Blackwell (sm_100) so the same wheel
-    # runs on H200 and B200 runners. Mirrors the CUDA-version-keyed list in
-    # docker/Dockerfile's DeepEP build stage.
-    if [ -n "${NVCC_VER:-}" ]; then
-        CUDA_VERSION="$NVCC_VER"
-    elif command -v nvcc >/dev/null 2>&1; then
-        CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[0-9]+\.[0-9]+')
-    else
-        CUDA_VERSION=$(nvidia-smi | grep "CUDA Version" | head -n1 | awk '{print $9}' || true)
-    fi
-    if [ -z "${CUDA_VERSION:-}" ]; then
-        echo "FATAL: could not determine CUDA toolkit version (NVCC_VER unset, nvcc missing, nvidia-smi empty)"
-        exit 1
-    fi
-    if [ "$CUDA_VERSION" = "12.8" ]; then
-        CHOSEN_TORCH_CUDA_ARCH_LIST='9.0;10.0'
-    elif awk -v ver="$CUDA_VERSION" 'BEGIN {exit !(ver > 12.8)}'; then
-        # CUDA > 12.8 supports sm_103 (Blackwell)
-        CHOSEN_TORCH_CUDA_ARCH_LIST='9.0;10.0;10.3'
-    else
-        CHOSEN_TORCH_CUDA_ARCH_LIST='9.0'
-    fi
-    TORCH_CUDA_ARCH_LIST="${CHOSEN_TORCH_CUDA_ARCH_LIST}" python3 setup.py install
-fi
@@ -136,7 +136,9 @@ cleanup_stale_shm() {
 install_apt_packages() {
    CI_APT_PACKAGES=(
        python3 python3-pip python3-venv python3-dev git libnuma-dev libssl-dev pkg-config
+        build-essential cmake rdma-core infiniband-diags perftest libibumad3
        libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils
+        libfabric-dev libnl-3-200 libnl-route-3-200 librdmacm1
        ffmpeg libavcodec-dev libavformat-dev libavutil-dev libswscale-dev
    )

@@ -164,6 +166,68 @@ install_apt_packages() {
    mark_step_done "${FUNCNAME[0]}"
 }

+install_gdrcopy() {
+    # DeepEP tests only run on 4+ GPU hosts. Keep GDRCopy in the shared CUDA
+    # bootstrap while avoiding a DKMS/package build on the 1- and 2-GPU jobs.
+    local gpu_count=0
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        gpu_count=$(
+            (nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null || true) |
+                awk 'NF {count++} END {print count + 0}'
+        )
+    fi
+    if [ "${gpu_count}" -lt 4 ]; then
+        echo "Skipping GDRCopy install on ${gpu_count}-GPU runner"
+        mark_step_done "${FUNCNAME[0]}"
+        return
+    fi
+
+    if ldconfig -p 2>/dev/null | grep 'libgdrapi\.so' >/dev/null; then
+        echo "GDRCopy userspace library is already installed"
+        mark_step_done "${FUNCNAME[0]}"
+        return
+    fi
+
+    local gdrcopy_root=/opt/gdrcopy
+    local gdrcopy_version=2.5.1
+    local -a gdrcopy_packages=(
+        nvidia-dkms-580 devscripts debhelper fakeroot dkms
+        check libsubunit0 libsubunit-dev python3-venv
+    )
+
+    apt-get update || true
+    apt-get install -y --no-install-recommends "${gdrcopy_packages[@]}" || {
+        echo "Warning: apt-get failed while installing GDRCopy build dependencies; checking installed packages"
+        local package
+        for package in "${gdrcopy_packages[@]}"; do
+            if ! dpkg -l "${package}" 2>/dev/null | grep -q '^ii'; then
+                echo "ERROR: Required GDRCopy package ${package} is unavailable"
+                exit 1
+            fi
+        done
+    }
+
+    rm -rf "${gdrcopy_root}"
+    git clone --branch "v${gdrcopy_version}" --depth 1 \
+        https://github.com/NVIDIA/gdrcopy.git "${gdrcopy_root}"
+    (
+        cd "${gdrcopy_root}/packages"
+        CUDA=/usr/local/cuda ./build-deb-packages.sh
+        dpkg -i gdrdrv-dkms_*.deb
+        dpkg -i libgdrapi_*.deb
+        dpkg -i gdrcopy-tests_*.deb
+        dpkg -i gdrcopy_*.deb
+    )
+
+    local lib_path="/usr/lib/${ARCH}-linux-gnu"
+    if [ ! -e "${lib_path}/libmlx5.so" ] && [ -e "${lib_path}/libmlx5.so.1" ]; then
+        ln -s "${lib_path}/libmlx5.so.1" "${lib_path}/libmlx5.so"
+    fi
+    ldconfig
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
 clean_site_packages() {
    # Clear torch compilation cache from every location it can be in; sglang
    # is not installed yet, so it cannot be asked which one is in use.
@@ -260,6 +324,11 @@ setup_pip_toolchain() {
    PIP_UNINSTALL_CMD="uv pip uninstall"
    PIP_UNINSTALL_SUFFIX=""

+    # Remove both the legacy source distribution and the SGLang wheel before
+    # resolving the pyproject pin. They own the same deep_ep module files, so
+    # leaving either installed can make pip preserve a mixed installation.
+    $PIP_UNINSTALL_CMD deep-ep sgl-deep-ep $PIP_UNINSTALL_SUFFIX || true
+
    # sglang-kernel stays: install_sglang_kernel version-gates and reinstalls it.
    $PIP_UNINSTALL_CMD sgl-kernel sglang sgl-fa4 flash-attn-4 $PIP_UNINSTALL_SUFFIX || true

@@ -368,6 +437,30 @@ install_pytorch_stack() {
    mark_step_done "${FUNCNAME[0]}"
 }

+install_cuda12_deepep_wheel() {
+    if [ "$CU_MAJOR" = "13" ]; then
+        echo "CUDA 13 uses the public sgl-deep-ep wheel declared in python/pyproject.toml"
+        mark_step_done "${FUNCNAME[0]}"
+        return
+    fi
+
+    local version
+    version=$(grep -Po -m1 '"sgl-deep-ep==\K[^"]+' python/pyproject.toml || true)
+    if [ -z "$version" ]; then
+        echo "ERROR: python/pyproject.toml must pin sgl-deep-ep"
+        exit 1
+    fi
+
+    # CUDA 12 wheels intentionally live only on the SGLang wheel index. Their
+    # local version satisfies the public-version pyproject pin, so the later
+    # editable SGLang install keeps this CUDA-matched wheel.
+    $PIP_CMD install "sgl-deep-ep==${version}+${CU_VERSION}" \
+        --index-url "https://docs.sglang.ai/whl/${CU_VERSION}/" \
+        --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
 require_prebuilt_rust_exts() {
    # Stages whose download succeeded set this to none. Runs before
    # setup_pip_toolchain uninstalls sglang, so clearing it here still reaches
@@ -708,6 +801,8 @@ verify_imports() {
    SGLANG_EXPECTED_INIT="${REPO_ROOT}/python/sglang/__init__.py" python3 -c '
 import torch
 print(torch.version.cuda)
+import deep_ep
+print(f"deep_ep loads from {deep_ep.__file__}")
 import cutlass
 import cutlass.cute

@@ -752,6 +847,7 @@ main() {
    kill_existing_processes
    cleanup_stale_shm
    install_apt_packages
+    install_gdrcopy
    clean_site_packages
    setup_cargo_cache
    require_prebuilt_rust_exts
@@ -759,6 +855,7 @@ main() {
    remove_stale_cuda12_nvidia_wheels
    uninstall_stale_flashinfer
    install_pytorch_stack
+    install_cuda12_deepep_wheel
    install_sglang
    # 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.
@@ -28,8 +28,8 @@ How `file -> models` is resolved (best effort, recall-favoring)
 How `runner label -> models` is aggregated
    Registration/prewarm decisions are made per GH runner *label* (a runner's
    `runs-on` tag), not per suite. Each suite's runner_config maps to a label
-    via scripts/ci/runner_configs.yml (several configs can share one label,
-    e.g. `4-gpu-h100` and `deepep-4-gpu-h100`), so `runner_labels` carries the
+    via scripts/ci/runner_configs.yml (runner configs may share a label), so
+    `runner_labels` carries the
    per-label UNION -- the set a runner registered under that label must have
    cached before it takes jobs. Suites without a mappable runner_config are
    listed in `unmapped_suites`.
@@ -367,9 +367,8 @@ def load_runner_labels(path: str) -> Dict[str, str]:
    """Parse ``{runner_config: runs_on label}`` out of runner_configs.yml.

    The mapping is what turns per-suite model sets into per-runner-LABEL sets:
-    a runner is registered under a `runs_on` label (several runner_configs can
-    share one, e.g. `4-gpu-h100` and `deepep-4-gpu-h100` both run on
-    `4-gpu-h100`), so a runner's cache must cover the union of every suite
+    a runner is registered under a `runs_on` label (runner configs may share a
+    label), so a runner's cache must cover the union of every suite
    that can land on its label. Raises ValueError on an entry without
    `runs_on` or a file with no entries at all -- a format drift must fail
    the workflow loudly, not silently empty the label aggregation.
@@ -8,8 +8,6 @@
 #   - artifact_version: actions/download-artifact major version
 #   - install_timeout: install-step wall-clock cap (minutes), enforced via
 #     `timeout-minutes:` on the install step in _pr-test-stage.yml
-#   - grace_blackwell (optional): exported as GRACE_BLACKWELL for the install
-#     step. Used by GB300 DeePEP setup.
 #   - runs_on: GHA runner label for the stage's `runs-on:`. The literal
 #     `$b200_runner` is substituted at workflow-load time with the dynamic
 #     b200 runner tag from check-changes (see runner_configs.py --map).
@@ -18,7 +16,6 @@

 _anchors:
  default_install: &default scripts/ci/cuda/ci_install_dependency.sh
-  deepep_install:  &deepep  scripts/ci/cuda/ci_install_deepep.sh
  kimi_k3_install: &kimi_k3 scripts/ci/cuda/ci_install_kimi_k3.sh

 runner_configs:
@@ -26,12 +23,9 @@ runner_configs:
  1-gpu-large:       { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 1-gpu-h100 }
  2-gpu-large:       { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 2-gpu-h100 }
  4-gpu-b200:        { install: *default, artifact_version: v6, install_timeout: "20", runs_on: $b200_runner }
-  4-gpu-gb300:       { install: *deepep,  artifact_version: v6, install_timeout: "20", grace_blackwell: "1", runs_on: 4-gpu-gb300 }
+  4-gpu-gb300:       { install: *default, artifact_version: v6, install_timeout: "20", runs_on: 4-gpu-gb300 }
  4-gpu-h100:        { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 4-gpu-h100 }
  8-gpu-h200:        { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200 }
  8-gpu-b200:        { install: *default, artifact_version: v6, install_timeout: "20", runs_on: 8-gpu-b200 }
  8-gpu-b300:        { install: *kimi_k3, artifact_version: v6, install_timeout: "20", runs_on: 8-gpu-b300 }
-  8-gpu-h20:         { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h20, rdma_devices: "mlx5_1,mlx5_2,mlx5_3,mlx5_4" }
-  deepep-4-gpu-h100: { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 4-gpu-h100 }
-  deepep-4-gpu-b200: { install: *deepep,  artifact_version: v6, install_timeout: "20", runs_on: $b200_runner }
-  deepep-8-gpu-h200: { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200 }
+  8-gpu-h20:         { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h20, rdma_devices: "mlx5_1,mlx5_2,mlx5_3,mlx5_4" }
@@ -412,7 +412,7 @@ _anchors:

 runner_configs:
  1-gpu:        { install: *default, artifact_version: v4, runs_on: 1-gpu-h100 }
-  deepep-1-gpu: { install: *default, artifact_version: v4, runs_on: 1-gpu-h100 }
+  alternate-1-gpu: { install: *default, artifact_version: v4, runs_on: 1-gpu-h100 }
  4-gpu-b200:   { install: *default, artifact_version: v6, runs_on: $b200_runner }
 """

@@ -431,7 +431,7 @@ class LoadRunnerLabels(unittest.TestCase):
            labels,
            {
                "1-gpu": "1-gpu-h100",
-                "deepep-1-gpu": "1-gpu-h100",
+                "alternate-1-gpu": "1-gpu-h100",
                "4-gpu-b200": lsm.B200_SENTINEL,
            },
        )
@@ -451,9 +451,7 @@ class LoadRunnerLabels(unittest.TestCase):
        labels = lsm.load_runner_labels(
            os.path.join(_REPO_ROOT, "scripts", "ci", "runner_configs.yml")
        )
-        # Two configs sharing a label is the reason the aggregation exists.
        self.assertEqual(labels["4-gpu-h100"], "4-gpu-h100")
-        self.assertEqual(labels["deepep-4-gpu-h100"], "4-gpu-h100")
        self.assertEqual(labels["4-gpu-b200"], lsm.B200_SENTINEL)
        self.assertGreaterEqual(len(labels), 10)

@@ -482,7 +480,7 @@ class RunnerLabelAggregation(unittest.TestCase):
                "b/test_b.py": self.REG.format(
                    calls=(
                        'register_cuda_ci(est_time=1, stage="base-y", '
-                        'runner_config="deepep-1-gpu")'
+                        'runner_config="alternate-1-gpu")'
                    ),
                    model="Qwen/Qwen3-8B",
                ),
@@ -520,7 +518,7 @@ class RunnerLabelAggregation(unittest.TestCase):
            )
            self.assertEqual(
                shared["suites"],
-                ["base-x-test-1-gpu", "base-y-test-deepep-1-gpu"],
+                ["base-x-test-1-gpu", "base-y-test-alternate-1-gpu"],
            )
            # Sentinel stays literal without --b200-runner.
            self.assertIn(lsm.B200_SENTINEL, inv["runner_labels"])
@@ -723,7 +723,7 @@ def _extract_legacy_suites(content):
 # matches the runner the nightly/weekly pipeline actually uses (see
 # .github/workflows/{nightly,weekly}-test-nvidia.yml), so /rerun-test can still
 # dispatch a single nightly/weekly test. The runner label, install script,
-# timeout, grace_blackwell, and rdma_devices are then resolved from
+# timeout and rdma_devices are then resolved from
 # runner_configs.yml as usual, keeping that file the single source of truth for
 # runner details.
 #
@@ -756,7 +756,6 @@ def _dispatch_err(suite, msg):
        "runner_label": None,
        "install_script": "",
        "install_timeout": "",
-        "grace_blackwell": "0",
        "rdma_devices": "",
        "is_cpu": False,
        "error": msg,
@@ -797,7 +796,6 @@ def _resolve_runner_config(rc, full_path, suite):
        "runner_label": runs_on,
        "install_script": install_script,
        "install_timeout": str(cfg["install_timeout"]),
-        "grace_blackwell": str(cfg.get("grace_blackwell", "0")),
        "rdma_devices": cfg.get("rdma_devices", ""),
        "is_cpu": False,
        "error": None,
@@ -811,8 +809,8 @@ def detect_suite(file_path_from_test):

    A CUDA file can carry multiple `register_cuda_ci(...)` calls — one per
    pool it should run on — so this returns a *list* of dispatch dicts, one
-    per registration. Runner label, install script, timeout, grace_blackwell,
-    and rdma_devices are all resolved from scripts/ci/runner_configs.yml — the
+    per registration. Runner label, install script, timeout, and rdma_devices
+    are all resolved from scripts/ci/runner_configs.yml — the
    same single source of truth that drives the main PR test pipeline.

    Legacy nightly/weekly CUDA suites (single-string `suite=`) are dispatchable
@@ -824,7 +822,7 @@ def detect_suite(file_path_from_test):
    `error` set.

    Each dict has keys: suite, runner_label, install_script,
-    install_timeout, grace_blackwell, rdma_devices, is_cpu, error.
+    install_timeout, rdma_devices, is_cpu, error.
    """
    full_path = f"test/{file_path_from_test}"
    with open(full_path, "r") as f:
@@ -860,7 +858,6 @@ def detect_suite(file_path_from_test):
                "runner_label": "ubuntu-latest",
                "install_script": "",
                "install_timeout": "",
-                "grace_blackwell": "0",
                "rdma_devices": "",
                "is_cpu": True,
                "error": None,
@@ -935,7 +932,6 @@ def _resolve_test_spec(test_spec):
                "runs_on": runner_label,
                "install_script": "",
                "install_timeout": "",
-                "grace_blackwell": "0",
                "rdma_devices": "",
                "error": None,
            }
@@ -954,7 +950,7 @@ def _resolve_test_spec(test_spec):
        print(
            f"Resolved: file={resolved_path}, selector={test_selector}, "
            f"suite={info['suite']}, mode={mode}, runs_on={info['runner_label']}, "
-            f"install={info['install_script']}, grace_blackwell={info['grace_blackwell']}, "
+            f"install={info['install_script']}, "
            f"rdma={info['rdma_devices']}, "
            f"command='{test_command}'"
        )
@@ -966,7 +962,6 @@ def _resolve_test_spec(test_spec):
                "runs_on": info["runner_label"],
                "install_script": info["install_script"],
                "install_timeout": info["install_timeout"],
-                "grace_blackwell": info["grace_blackwell"],
                "rdma_devices": info["rdma_devices"],
                "error": None,
            }
@@ -978,7 +973,7 @@ def _dispatch_batch(gh_repo, pr, batch, token, reply_comment_id="", reply_marker
    """
    Dispatch a single workflow run for a batch of resolved test specs that
    share the same dispatch shape (mode + runs_on + install_script +
-    install_timeout + grace_blackwell + rdma_devices).
+    install_timeout + rdma_devices).

    Returns a dict with keys: specs, success, test_commands, runner_label, run_url, error.
    """
@@ -987,7 +982,6 @@ def _dispatch_batch(gh_repo, pr, batch, token, reply_comment_id="", reply_marker
    runs_on = batch[0]["runs_on"]
    install_script = batch[0]["install_script"]
    install_timeout = batch[0]["install_timeout"]
-    grace_blackwell = batch[0]["grace_blackwell"]
    rdma_devices = batch[0]["rdma_devices"]

    # Join multiple commands with newlines for the workflow to iterate over
@@ -1020,7 +1014,6 @@ def _dispatch_batch(gh_repo, pr, batch, token, reply_comment_id="", reply_marker
            "runs_on": runs_on or "",
            "install_script": install_script,
            "install_timeout": install_timeout or "20",
-            "grace_blackwell": grace_blackwell or "0",
            "rdma_devices": rdma_devices,
            "reply_comment_id": str(reply_comment_id) if reply_comment_id else "",
            "reply_marker": reply_marker,
@@ -1128,7 +1121,7 @@ def handle_rerun_test(
    """
    Handles the /rerun-test command. Resolves all test specs, groups them by
    dispatch shape (mode + runs_on + install_script + install_timeout +
-    grace_blackwell + rdma_devices), and dispatches one workflow per group.
+    rdma_devices), and dispatches one workflow per group.
    """
    if not skip_permission_check and not _check_rerun_test_permissions(
        gh_repo, pr, comment, user_perms, "rerun-test"
@@ -1223,7 +1216,6 @@ def handle_rerun_test(
            r["runs_on"],
            r["install_script"],
            r["install_timeout"],
-            r["grace_blackwell"],
            r["rdma_devices"],
        )
        groups.setdefault(key, []).append(r)
@@ -102,7 +102,7 @@ class TestTboAttnDenseAttentionBackendCorrectness(CustomTestCase):
        only by the in_capture=True branch (via ``_bind_metadata_buffers``).
        If TBO short-circuits its capture to its own replay path, those dicts
        are empty and replay raises ``KeyError: bs``. Reproduces the
-        deepep-4-gpu-h100 failure where
+        4-gpu-h100 failure where
        ``flashattention_backend.target_verify_metadata[bs]`` lookup blew up
        during ``init_device_graphs``.

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    write_github_step_summary,
 )

-register_cuda_ci(est_time=500, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=500, stage="extra-b", runner_config="8-gpu-h200")

 DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"

@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=235, stage="extra-b", runner_config="deepep-4-gpu-b200")
+register_cuda_ci(est_time=235, stage="extra-b", runner_config="4-gpu-b200")

 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    write_github_step_summary,
 )

-register_cuda_ci(est_time=320, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=320, stage="extra-b", runner_config="8-gpu-h200")
 GLM52_MODEL_PATH = "zai-org/GLM-5.2-FP8"
 SERVER_LAUNCH_TIMEOUT = max(DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, 1800)

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
    write_github_step_summary,
 )

-register_cuda_ci(est_time=616, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=616, stage="extra-b", runner_config="8-gpu-h200")
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"


@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_cuda_ci(est_time=500, stage="extra-b", runner_config="deepep-4-gpu-h100")
+register_cuda_ci(est_time=500, stage="extra-b", runner_config="4-gpu-h100")

 GQA_MODEL_PATH = "Qwen/Qwen3-30B-A3B-FP8"

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=500, stage="base-c", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=500, stage="base-c", runner_config="8-gpu-h200")

 DSV4_FLASH_MODEL = "sgl-project/DeepSeek-V4-Flash-FP8"

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=1000, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=1000, stage="extra-b", runner_config="8-gpu-h200")

 DSV4_FLASH_MODEL = "sgl-project/DeepSeek-V4-Flash-FP8"
 DSV4_FLASH_LOADER_CONFIG = '{"enable_multithread_load": true, "num_threads": 64}'
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_cuda_ci(est_time=528, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=528, stage="extra-b", runner_config="8-gpu-h200")

 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_cuda_ci(est_time=478, stage="base-c", runner_config="deepep-4-gpu-h100")
+register_cuda_ci(est_time=478, stage="base-c", runner_config="4-gpu-h100")


 class TestPureDP(CustomTestCase):
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
 register_cuda_ci(
    est_time=189,
    stage="base-c",
-    runner_config="deepep-4-gpu-h100",
+    runner_config="4-gpu-h100",
    disabled="Temporarily disabled until the next Mooncake release includes the PyTorch 2.13 collective forwarding fix.",
 )

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_cuda_ci(est_time=900, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=900, stage="extra-b", runner_config="8-gpu-h200")

 DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"

@@ -4,7 +4,7 @@ Launches TP=4 with flashinfer_mxfp4 MoE runner + EAGLE speculative decoding.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.

-Registry: base-c-test-deepep-4-gpu-b200 (per-commit, 4x B200)
+Registry: base-c-test-4-gpu-b200 (per-commit, 4x B200)
 """

 import unittest
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=465, stage="base-c", runner_config="deepep-4-gpu-b200")
+register_cuda_ci(est_time=465, stage="base-c", runner_config="4-gpu-b200")

 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600
@@ -7,7 +7,7 @@ plus a GSM8K accuracy gate.
 Also covers SGLANG_DSV4_FP4_DEQUANT=1 (TP=8): FP4 experts dequantized to FP8
 during loading and served through the plain FP8 MoE path.

-Registry: base-c-test-deepep-8-gpu-h200 (per-commit, 8x H200)
+Registry: base-c-test-8-gpu-h200 (per-commit, 8x H200)
 """

 import unittest
@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=600, stage="base-c", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=600, stage="base-c", runner_config="8-gpu-h200")


 def _flashinfer_has_sm90_cutlass_mxfp4() -> bool:
@@ -4,7 +4,7 @@ Launches TP=4 with flashinfer_mxfp4 MoE runner + EAGLE speculative decoding.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.

-Registry: extra-b-test-deepep-4-gpu-b200 (label-gated, 4x B200)
+Registry: extra-b-test-4-gpu-b200 (label-gated, 4x B200)
 """

 import unittest
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=900, stage="extra-b", runner_config="deepep-4-gpu-b200")
+register_cuda_ci(est_time=900, stage="extra-b", runner_config="4-gpu-b200")

 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600
@@ -5,7 +5,7 @@ with FP4 experts disabled via SGLANG_DSV4_FP4_EXPERTS=0.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.

-Registry: extra-b-test-deepep-8-gpu-h200 (label-gated, 8x H200 — only 4 used by TP=4)
+Registry: extra-b-test-8-gpu-h200 (label-gated, 8x H200 — only 4 used by TP=4)
 """

 import unittest
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
    try_cached_model,
 )

-register_cuda_ci(est_time=560, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=560, stage="extra-b", runner_config="8-gpu-h200")

 MODEL_FP8 = "sgl-project/DeepSeek-V4-Flash-FP8"
 SERVER_LAUNCH_TIMEOUT = 3600
@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
    popen_launch_server,
 )

-register_cuda_ci(est_time=400, stage="extra-b", runner_config="deepep-4-gpu-h100")
+register_cuda_ci(est_time=400, stage="extra-b", runner_config="4-gpu-h100")

 # FP8 variant of Qwen3-30B-A3B: required because DeepEP normal/LL fast paths in
 # ep_moe/layer.py only run for {Fp8Config (via deep_gemm), W4AFp8Config, aiter,
@@ -83,9 +83,6 @@ PER_COMMIT_SUITES = {
        "base-c-test-8-gpu-h200",
        "base-c-test-8-gpu-b200",
        "base-c-test-8-gpu-b300",
-        "base-c-test-deepep-4-gpu-h100",
-        "base-c-test-deepep-4-gpu-b200",
-        "base-c-test-deepep-8-gpu-h200",
        # extra-a / extra-b: label-gated PR opt-in suites in pr-test-extra.yml
        # (tests still tagged per-commit but skipped on default PR runs).
        "extra-a-test-1-gpu-small",
@@ -94,9 +91,6 @@ PER_COMMIT_SUITES = {
        "extra-b-test-4-gpu-h100",
        "extra-b-test-4-gpu-b200",
        "extra-b-test-8-gpu-h200",
-        "extra-b-test-deepep-4-gpu-h100",
-        "extra-b-test-deepep-4-gpu-b200",
-        "extra-b-test-deepep-8-gpu-h200",
    ],
    HWBackend.NPU: [
        "base-a-test-1-gpu-small",