diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index ec9cbc11e..ae3c85285 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1796,6 +1796,30 @@ dsr1-fp8-b300-sglang:
     - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 }
     - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 }
 
+# NOTE: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+# lists H200 (not B300) as the FP8 target. Only the Flash-FP8 checkpoint
+# is live -- Pro-FP8 is still pending upload in the cookbook. This config
+# runs the H200 Flash Max-Throughput recipe (DP + DeepEP, no MTP) on B300
+# using the Blackwell image with SGLANG_DSV4_FP4_EXPERTS=0 to swap the
+# experts to FP8. Prefix caching is disabled.
+dsv4-fp8-b300-sglang:
+  image: lmsysorg/sglang:deepseek-v4-blackwell
+  model: sgl-project/DeepSeek-V4-Flash-FP8
+  model-prefix: dsv4
+  runner: b300
+  precision: fp8
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 4, ep: 4, dp-attn: true, conc-start: 4, conc-end: 1024 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 4, ep: 4, dp-attn: true, conc-start: 4, conc-end: 512 }
+
 qwen3.5-bf16-b200-sglang:
   image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e
   model: Qwen/Qwen3.5-397B-A17B
diff --git a/benchmarks/single_node/dsv4_fp8_b300.sh b/benchmarks/single_node/dsv4_fp8_b300.sh
new file mode 100755
index 000000000..e6e34b599
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp8_b300.sh
@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+
+# NOTE: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+# lists H200 (not B300) as the FP8 target, and only the Flash-FP8
+# checkpoint is live (Pro-FP8 is still pending upload). This script
+# reuses the H200 Flash Max-Throughput recipe (DP + DeepEP, no MTP) on
+# B300 using the Blackwell image with SGLANG_DSV4_FP4_EXPERTS=0 to swap
+# the experts to FP8. Prefix caching is disabled.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    EP_SIZE \
+    DP_ATTENTION
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+nvidia-smi
+
+export SGLANG_DSV4_FP4_EXPERTS=0
+export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256
+export SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+echo "TP: $TP, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+DP_ATTN_ARGS=""
+if [ "$DP_ATTENTION" = "true" ]; then
+    DP_ATTN_ARGS="--data-parallel-size $TP --enable-dp-attention"
+fi
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor
+
+set -x
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0.0.0.0 --port=$PORT \
+--trust-remote-code \
+--tensor-parallel-size=$TP --ep-size $EP_SIZE $DP_ATTN_ARGS \
+--moe-a2a-backend deepep \
+--deepep-config '{"normal_dispatch":{"num_sms":96},"normal_combine":{"num_sms":96}}' \
+--cuda-graph-max-bs 128 \
+--max-running-requests 256 \
+--disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ddc6409c2..63ee58f74 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1733,3 +1733,14 @@
     - "TP=2 and TP=4, concurrency 4-256 for 1k1k and 8k1k sequence lengths"
     - "Add --max-num-seqs and --gpu-memory-utilization 0.9 to server launch"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1043
+
+- config-keys:
+    - dsv4-fp8-b300-sglang
+  description:
+    - "Add DeepSeek-V4-Flash FP8 B300 SGLang benchmark"
+    - "Image: lmsysorg/sglang:deepseek-v4-blackwell (with SGLANG_DSV4_FP4_EXPERTS=0 to swap experts to FP8)"
+    - "Model: sgl-project/DeepSeek-V4-Flash-FP8"
+    - "Reuses the H200 Flash Max-Throughput recipe from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4 on B300; Pro-FP8 checkpoint is still pending upload in the cookbook"
+    - "DP + DeepEP, prefix caching disabled, no speculative decoding"
+    - "TP=4/EP=4/dp-attn=true, concurrency 4-1024 for 1k1k and 4-512 for 8k1k"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index b49391a3c..ecb24b1a1 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -248,27 +248,32 @@ find . -name '.nfs*' -delete 2>/dev/null || true
 
 else
 
-    HF_HUB_CACHE_MOUNT="/scratch/models"
-    # Qwen3.5-397B-A17B-FP8 is pre-staged under /scratch/models on the B300 cluster,
-    # so point MODEL at the local copy. Other models fall through and use `hf download`
-    # against the mounted cache from their benchmark script.
-    if [[ "$MODEL" == "Qwen/Qwen3.5-397B-A17B-FP8" ]]; then
-        export MODEL="/scratch/models/${MODEL#*/}"
-    fi
-    SQUASH_FILE="/data/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    HF_HUB_CACHE_MOUNT="/data/home/sa-shared/gharunners/hf-hub-cache"
+    SQUASH_FILE="/data/home/sa-shared/gharunners/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+    LOCK_FILE="${SQUASH_FILE}.lock"
 
     # Pin to one of the known-good B300 nodes; others have hardware/network
     # issues that cause benchmarks to hang or fail to start.
     salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --nodelist=b300-[001-006,008-012,017-020] -N 1 --gres=gpu:$TP --exclusive --time=180 --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
-    srun --jobid=$JOB_ID bash -c "enroot import -o $SQUASH_FILE docker://$IMAGE"
+    # Use flock to serialize concurrent imports to the same squash file
+    srun --jobid=$JOB_ID bash -c "
+        exec 9>\"$LOCK_FILE\"
+        flock -w 600 9 || { echo 'Failed to acquire lock for $SQUASH_FILE'; exit 1; }
+        if unsquashfs -l \"$SQUASH_FILE\" > /dev/null 2>&1; then
+            echo 'Squash file already exists and is valid, skipping import'
+        else
+            rm -f \"$SQUASH_FILE\"
+            enroot import -o \"$SQUASH_FILE\" docker://$IMAGE
+        fi
+    "
 
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \
-        --container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE_MOUNT \
+        --container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
         --no-container-mount-home \
         --container-workdir=/workspace/ \
         --no-container-entrypoint --export=ALL,PORT=8888 \