diff --git a/.github/configs/runners.yaml b/.github/configs/runners.yaml
index 245b74762..98080efea 100644
--- a/.github/configs/runners.yaml
+++ b/.github/configs/runners.yaml
@@ -24,6 +24,8 @@ b200:
 - 'b200-nvd_1'
 - 'b200-nvd_2'
 - 'b200-nvd_3'
+- 'b200-dgxc_1'
+- 'b200-dgxc_2'
 mi300x:
 - 'mi300x-amd_0'
 - 'mi300x-amd_1'
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
new file mode 100644
index 000000000..4d8ec0aed
--- /dev/null
+++ b/runners/launch_b200-dgxc.sh
@@ -0,0 +1,65 @@
+#!/usr/bin/bash
+
+HF_HUB_CACHE_MOUNT="/raid/hf_hub_cache/"
+FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
+PORT=8888
+
+# Create unique cache directory based on model parameters
+MODEL_NAME=$(basename "$MODEL")
+
+server_name="bmk-server"
+
+nvidia-smi
+
+# GPUs must be idle
+if nvidia-smi --query-compute-apps=pid --format=csv,noheader | grep -q '[0-9]'; then
+  echo "[ERROR] GPU busy from previous run"; nvidia-smi; exit 1
+fi
+
+set -x
+# Use --init flag to run an init process (PID 1) inside container for better signal handling and zombie process cleanup
+# Ref: https://www.paolomainardi.com/posts/docker-run-init/
+
+# NCCL_GRAPH_REGISTER tries to automatically enable user buffer registration with CUDA Graphs. 
+# Disabling it can reduce perf but will improve CI stability. i.e. we won't see vLLM/Sglang crashes.
+# Ref: https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html#nccl-graph-register
+
+if [[ "$MODEL" == "nvidia/DeepSeek-R1-0528-FP4" || "$MODEL" == "deepseek-ai/DeepSeek-R1-0528" ]]; then
+  if [[ "$OSL" == "8192" ]]; then
+    export NUM_PROMPTS=$(( CONC * 20 ))
+  else
+    export NUM_PROMPTS=$(( CONC * 50 ))
+  fi
+else
+  export NUM_PROMPTS=$(( CONC * 10 ))
+fi
+
+docker run --rm --init --network host --name $server_name \
+--runtime nvidia --gpus all --ipc host --privileged --shm-size=16g --ulimit memlock=-1 --ulimit stack=67108864 \
+-v $HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+-v $GITHUB_WORKSPACE:/workspace/ -w /workspace/ \
+-e HF_TOKEN -e HF_HUB_CACHE -e MODEL -e TP -e CONC -e MAX_MODEL_LEN -e ISL -e OSL -e PORT=$PORT -e EP_SIZE -e DP_ATTENTION \
+-e NCCL_GRAPH_REGISTER=0 \
+-e TORCH_CUDA_ARCH_LIST="10.0" -e CUDA_DEVICE_ORDER=PCI_BUS_ID -e CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" \
+-e PYTHONPYCACHEPREFIX=/tmp/pycache/ -e RESULT_FILENAME -e RANDOM_RANGE_RATIO -e NUM_PROMPTS \
+--entrypoint=/bin/bash \
+$(echo "$IMAGE" | sed 's/#/\//') \
+benchmarks/"${EXP_NAME%%_*}_${PRECISION}_b200${FRAMEWORK_SUFFIX}_docker.sh"
+
+# Try graceful first
+docker stop -t 90 "$server_name" || true
+# Wait until it's really dead
+docker wait "$server_name" >/dev/null 2>&1 || true
+# Force remove if anything lingers
+docker rm -f "$server_name" >/dev/null 2>&1 || true
+
+# Give a moment for GPU processes to fully terminate
+sleep 2
+# Verify GPUs are now idle; if not, print diag and (optionally) reset
+if nvidia-smi --query-compute-apps=pid --format=csv,noheader | grep -q '[0-9]'; then
+  echo "[WARN] After stop, GPU still busy:"; nvidia-smi
+  # Last resort if driver allows and GPUs appear idle otherwise:
+  #nvidia-smi --gpu-reset -i 0,1,2,3,4,5,6,7 2>/dev/null || true
+fi
+
+nvidia-smi
diff --git a/runners/launch_b200-nvd.sh b/runners/launch_b200-nvd.sh
index c5216b006..ebfa67458 100644
--- a/runners/launch_b200-nvd.sh
+++ b/runners/launch_b200-nvd.sh
@@ -39,7 +39,7 @@ docker run --rm --init --network host --name $server_name \
 --runtime nvidia --gpus all --ipc host --privileged --shm-size=16g --ulimit memlock=-1 --ulimit stack=67108864 \
 -v $HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
 -v $GITHUB_WORKSPACE:/workspace/ -w /workspace/ \
--e HF_TOKEN -e HF_HUB_CACHE -e MODEL -e TP -e CONC -e MAX_MODEL_LEN -e ISL -e OSL -e PORT=$PORT -e EP_SIZE \
+-e HF_TOKEN -e HF_HUB_CACHE -e MODEL -e TP -e CONC -e MAX_MODEL_LEN -e ISL -e OSL -e PORT=$PORT -e EP_SIZE -e DP_ATTENTION \
 -e NCCL_GRAPH_REGISTER=0 \
 -e TORCH_CUDA_ARCH_LIST="10.0" -e CUDA_DEVICE_ORDER=PCI_BUS_ID -e CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" \
 -e PYTHONPYCACHEPREFIX=/tmp/pycache/ -e RESULT_FILENAME -e RANDOM_RANGE_RATIO -e NUM_PROMPTS \