[CI] Clean DeepSeek V4 tests and installation scripts (#26609)

.github/workflows/pr-test-extra.yml

@@ -214,6 +214,20 @@ jobs:
       run_timeout_minutes: '60'
     secrets: inherit
 
+  extra-b-test-deepep-4-gpu-b200:
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    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'
+    secrets: inherit
+
   extra-b-test-deepep-8-gpu-h200:
     needs: [check-changes, call-gate, sgl-kernel-build-wheels]
     if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && (needs.call-gate.result == 'success' || needs.call-gate.result == 'skipped') }}
@@ -243,6 +257,7 @@ 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-b200,
         extra-b-test-deepep-8-gpu-h200,
       ]
     if: always()

.github/workflows/pr-test.yml

@@ -462,13 +462,13 @@ jobs:
       warmup_server_models: 'lmsys/sglang-ci-dsv3-test:4'
     secrets: inherit
 
-  base-c-test-4-gpu-b200:
+  base-c-test-deepep-4-gpu-b200:
     needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels]
     if: ${{ !failure() && !cancelled() }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
-      self_name: base-c-test-4-gpu-b200
+      self_name: base-c-test-deepep-4-gpu-b200
-      runner_config: 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 }}
@@ -476,27 +476,27 @@ jobs:
       timeout_per_file: '1800'
     secrets: inherit
 
-  base-c-test-dsv4-4-gpu-b200:
+  base-c-test-deepep-8-gpu-h200:
     needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels]
     if: ${{ !failure() && !cancelled() }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
-      self_name: base-c-test-dsv4-4-gpu-b200
+      self_name: base-c-test-deepep-8-gpu-h200
-      runner_config: dsv4-4-gpu-b200
+      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: '45'
+      run_timeout_minutes: '30'
       timeout_per_file: '1800'
     secrets: inherit
 
-  base-c-test-dsv4-8-gpu-h200:
+  base-c-test-4-gpu-b200:
     needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels]
     if: ${{ !failure() && !cancelled() }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
-      self_name: base-c-test-dsv4-8-gpu-h200
+      self_name: base-c-test-4-gpu-b200
-      runner_config: dsv4-8-gpu-h200
+      runner_config: 4-gpu-b200
       check_changes: ${{ toJson(needs.check-changes.outputs) }}
       caller_inputs: ${{ toJson(inputs) }}
       partitions: ${{ needs.check-changes.outputs.partitions }}
@@ -531,9 +531,9 @@ jobs:
         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-dsv4-4-gpu-b200,
-        base-c-test-dsv4-8-gpu-h200,
         # base-c-test-4-gpu-gb200,  # Temporarily disabled — no GB200 runner
       ]
     if: always()

docker/Dockerfile

@@ -548,20 +548,12 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # the `nixl` import path) but unconditionally requires nixl-cu12, so we install
 # it with --no-deps and pair it with the matching nixl-cu12 / nixl-cu13 binary
 # to avoid shipping wrong-CUDA libs on cu13 images.
-# The upstream flash-mla packages are required for running deepseek-v4 models
 RUN --mount=type=cache,target=/root/.cache/pip if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
     python3 -m pip install nixl nixl-cu12 --no-deps ; \
     python3 -m pip install cuda-python==12.9 ; \
-    cd /sgl-workspace && git clone https://github.com/deepseek-ai/FlashMLA.git flash-mla \
-    && cd flash-mla && git submodule update --init --recursive \
-    && pip install --no-build-isolation -v . ; \
 elif [ "${CUDA_VERSION%%.*}" = "13" ]; then \
     python3 -m pip install nixl nixl-cu13 --no-deps ; \
     python3 -m pip install cuda-python==13.2.0 ; \
-    cd /sgl-workspace && git clone https://github.com/deepseek-ai/FlashMLA.git flash-mla \
-    && ln -s /usr/local/cuda/include/cccl/cuda /usr/local/cuda/include/cuda \
-    && cd flash-mla && git submodule update --init --recursive \
-    && pip install --no-build-isolation -v . ; \
 fi
 
 # Add yank script

scripts/ci/cuda/ci_install_deepep.sh

@@ -150,5 +150,28 @@ else
     if [ "$NVCC_MAJOR" = "13" ]; then
         sed -i "/^    include_dirs = \['csrc\/'\]/a\    include_dirs.append('${CUDA_HOME:-/usr/local/cuda}/include/cccl')" setup.py
     fi
-    python3 setup.py install
+
+    # Build for both Hopper (sm_90) and Blackwell (sm_100) so the same wheel
+    # runs on H200 and B200 runners. Mirrors the CUDA-version-keyed list in
+    # docker/Dockerfile's DeepEP build stage.
+    if [ -n "${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

scripts/ci/cuda/ci_install_dsv4_dep.sh

@@ -1,161 +0,0 @@
-#!/bin/bash
-set -euxo pipefail
-
-source scripts/ci/cuda/ci_install_dependency.sh
-
-if [ -z "${PIP_CMD:-}" ]; then
-    echo "FATAL:PIP_CMD is unset after sourcing ci_install_dependency.sh"
-    exit 1
-fi
-
-export GDRCOPY_HOME=/usr/src/gdrdrv-2.5.1/
-export CUDA_HOME=/usr/local/cuda
-
-# Detect architecture
-ARCH=$(uname -m)
-if [ "$ARCH" != "x86_64" ] && [ "$ARCH" != "aarch64" ]; then
-    echo "Unsupported architecture: $ARCH"
-    exit 1
-fi
-
-###############################################################################
-# Install FlashMLA
-###############################################################################
-INSTALL_FLASH_MLA=1
-if [ "${FORCE_REBUILD_FLASH_MLA:-0}" = "1" ]; then
-    echo "FORCE_REBUILD_FLASH_MLA=1; uninstalling any cached flash_mla before rebuild."
-    ${PIP_UNINSTALL_CMD:-pip uninstall -y} flash_mla ${PIP_UNINSTALL_SUFFIX:-} || true
-elif python3 -c "import flash_mla" >/dev/null 2>&1; then
-    echo "flash_mla is already installed or importable. Skipping installation."
-    INSTALL_FLASH_MLA=0
-fi
-
-if [ "$INSTALL_FLASH_MLA" = "1" ]; then
-    # CUDA 13.0 puts CCCL headers under /usr/local/cuda/include/cccl/cuda but
-    # FlashMLA's build expects them at /usr/local/cuda/include/cuda. Symlink so
-    # the compiler finds them. Idempotent: skip if the link/dir already exists.
-    if [ ! -e /usr/local/cuda/include/cuda ] && [ -d /usr/local/cuda/include/cccl/cuda ]; then
-        ln -s /usr/local/cuda/include/cccl/cuda /usr/local/cuda/include/cuda
-    fi
-
-    FLASH_MLA_DIR=/root/.cache/flash-mla
-    rm -rf ${FLASH_MLA_DIR}
-    git clone https://github.com/deepseek-ai/FlashMLA.git ${FLASH_MLA_DIR}
-    pushd ${FLASH_MLA_DIR}
-    git submodule update --init --recursive
-    ${PIP_CMD:-pip} install --no-build-isolation -v . ${PIP_INSTALL_SUFFIX:-}
-    popd
-fi
-
-###############################################################################
-# Install DeepEP
-###############################################################################
-# Default to a forced rebuild so changes to TORCH_CUDA_ARCH_LIST or any other
-# build-time input don't silently reuse a cached deep_ep from a prior run.
-INSTALL_DEEPEP=1
-if [ "${FORCE_REBUILD_DEEPEP:-1}" = "1" ]; then
-    echo "FORCE_REBUILD_DEEPEP=1; uninstalling any cached deep_ep before rebuild."
-    ${PIP_UNINSTALL_CMD:-pip uninstall -y} deep_ep ${PIP_UNINSTALL_SUFFIX:-} || true
-elif python3 -c "import deep_ep" >/dev/null 2>&1; then
-    echo "deep_ep is already installed or importable. Skipping installation."
-    INSTALL_DEEPEP=0
-fi
-
-if [ "$INSTALL_DEEPEP" = "1" ]; then
-    # 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}
-    git clone https://github.com/deepseek-ai/DeepEP.git ${DEEPEP_DIR}
-    pushd ${DEEPEP_DIR}
-    git checkout 9af0e0d0e74f3577af1979c9b9e1ac2cad0104ee
-    popd
-
-    cd ${DEEPEP_DIR}
-    # CUDA 13.0 puts CCCL headers in /usr/local/cuda/include/cccl/ but nvshmem
-    # includes them as <cuda/__cccl_config> expecting /usr/local/cuda/include/cuda/.
-    # Add the cccl path to setup.py include_dirs so the compiler finds them.
-    NVCC_MAJOR=$(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

scripts/ci/runner_configs.yml

@@ -17,7 +17,6 @@
 _anchors:
   default_install: &default scripts/ci/cuda/ci_install_dependency.sh
   deepep_install:  &deepep  scripts/ci/cuda/ci_install_deepep.sh
-  dsv4_install:    &dsv4    scripts/ci/cuda/ci_install_dsv4_dep.sh
 
 runner_configs:
   1-gpu-small:       { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 1-gpu-5090 }
@@ -28,6 +27,5 @@ runner_configs:
   8-gpu-h200:        { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200 }
   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-8-gpu-h200: { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200-deepep }
+  deepep-4-gpu-b200: { install: *deepep,  artifact_version: v6, install_timeout: "20", runs_on: $b200_runner }
-  dsv4-4-gpu-b200:   { install: *dsv4,    artifact_version: v6, install_timeout: "30", runs_on: $b200_runner }
+  deepep-8-gpu-h200: { install: *deepep,  artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200 }
-  dsv4-8-gpu-h200:   { install: *dsv4,    artifact_version: v4, install_timeout: "30", runs_on: 8-gpu-h200 }

test/registered/cp/test_deepseek_v4_flash_fp4_b200_cp.py

@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=235, stage="extra-b", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=235, stage="extra-b", runner_config="deepep-4-gpu-b200")
 
 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/disaggregation/test_disaggregation_dsv4.py

@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=250, stage="base-c", runner_config="dsv4-8-gpu-h200")
+register_cuda_ci(est_time=250, stage="base-c", runner_config="deepep-8-gpu-h200")
 
 DSV4_FLASH_MODEL = "sgl-project/DeepSeek-V4-Flash-FP8"
 

test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py

@@ -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-dsv4-4-gpu-b200 (per-commit, 4x B200)
+Registry: base-c-test-deepep-4-gpu-b200 (per-commit, 4x B200)
 """
 
 import unittest
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=465, stage="base-c", runner_config="dsv4-4-gpu-b200")
+register_cuda_ci(est_time=465, stage="base-c", runner_config="deepep-4-gpu-b200")
 
 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/models_e2e/test_deepseek_v4_flash_fp4_h200.py

@@ -4,7 +4,7 @@ Launches TP=4 with Marlin FP4 MoE runner + EAGLE speculative decoding.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.
 
-Registry: base-c-test-dsv4-8-gpu-h200 (per-commit, 8x H200 — only 4 used by TP=4)
+Registry: base-c-test-deepep-8-gpu-h200 (per-commit, 8x H200 — only 4 used by TP=4)
 """
 
 import unittest
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=370, stage="base-c", runner_config="dsv4-8-gpu-h200")
+register_cuda_ci(est_time=370, stage="base-c", runner_config="deepep-8-gpu-h200")
 
 
 def _flashinfer_has_sm90_cutlass_mxfp4() -> bool:

test/registered/models_e2e/test_deepseek_v4_flash_fp4_megamoe_b200.py

@@ -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-dsv4-4-gpu-b200 (per-commit, 4x B200)
+Registry: extra-b-test-deepep-4-gpu-b200 (label-gated, 4x B200)
 """
 
 import unittest
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=900, stage="base-c", runner_config="dsv4-4-gpu-b200")
+register_cuda_ci(est_time=900, stage="extra-b", runner_config="deepep-4-gpu-b200")
 
 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/models_e2e/test_deepseek_v4_flash_fp8_h200.py

@@ -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: base-c-test-dsv4-8-gpu-h200 (per-commit, 8x H200 — only 4 used by TP=4)
+Registry: extra-b-test-deepep-8-gpu-h200 (label-gated, 8x H200 — only 4 used by TP=4)
 """
 
 import unittest
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=280, stage="base-c", runner_config="dsv4-8-gpu-h200")
+register_cuda_ci(est_time=280, stage="extra-b", runner_config="deepep-8-gpu-h200")
 
 MODEL_FP8 = "sgl-project/DeepSeek-V4-Flash-FP8"
 SERVER_LAUNCH_TIMEOUT = 3600

test/run_suite.py

@@ -59,8 +59,8 @@ PER_COMMIT_SUITES = {
         "base-c-test-8-gpu-h200",
         "base-c-test-8-gpu-b200",
         "base-c-test-deepep-4-gpu-h100",
-        "base-c-test-dsv4-4-gpu-b200",
+        "base-c-test-deepep-4-gpu-b200",
-        "base-c-test-dsv4-8-gpu-h200",
+        "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",
@@ -69,6 +69,7 @@ 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-b200",
         "extra-b-test-deepep-8-gpu-h200",
     ],
     HWBackend.NPU: [