[NVIDIA] chore: B300 single node DeepSeek v4 SGLang LOW LATENCY ONLY

.github/configs/nvidia-master.yaml

@@ -1796,6 +1796,34 @@ 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: Low-latency fallback (TP=8, EP=1, no DP-attn, no DeepEP) while
+# the DeepEP FP8 weight-postprocess path is broken for DeepSeek-V4-Pro
+# on B300. Re-introduce balanced/max-throughput rows once fixed upstream.
+dsv4-fp4-b300-sglang:
+  image: lmsysorg/sglang:deepseek-v4-b300
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b300
+  precision: fp4
+  framework: sglang
+  multinode: false
+  # TODO(Cam): low-latency recipe only (TP-only, no DP-attn, no DeepEP)
+  # while the DeepEP FP8 weight-postprocess path is broken for this
+  # checkpoint on B300 (RuntimeError: Recipe must be a list/tuple of 3
+  # integers. raised from sglang.srt.layers.quantization.fp8
+  # .process_weights_after_loading_block_quant). Full concurrency sweep
+  # retained; revert to the recipe-per-CONC split on chore/dsv4-sgl-b300
+  # once sglang can load the checkpoint under --moe-a2a-backend deepep.
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 1024 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 512 }
+
 qwen3.5-bf16-b200-sglang:
   image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e
   model: Qwen/Qwen3.5-397B-A17B

benchmarks/single_node/dsv4_fp4_b300.sh

@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+# The B300 runner overrides MODEL to a pre-staged /data/models path, so skip
+# `hf download`. Only fetch when MODEL looks like a HF repo ID.
+if [[ "$MODEL" != /* ]]; then
+    hf download "$MODEL"
+fi
+
+nvidia-smi
+
+export SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
+
+# The deepseek-v4 sglang images (lmsysorg/sglang:deepseek-v4-blackwell and its
+# B300 forks) bake CUDA_VISIBLE_DEVICES=4,5,6,7 into their ENV, which masks half
+# of the 8 GPUs Slurm allocates us. Clear it so TP=8 can bind to all ranks.
+unset CUDA_VISIBLE_DEVICES
+
+# TODO(Cam): the deepseek-v4 sglang images install sglang editable at
+# /workspace/sglang/python; prior sglang tags used /sgl-workspace/sglang.
+# The runner mounts our repo at a non-/workspace path for these images so the
+# editable install stays visible. Paths in this script are $PWD-relative for
+# that reason. Drop the runner conditional once lmsys moves sglang back out of
+# /workspace.
+
+SERVER_LOG="$PWD/server.log"
+PORT=${PORT:-8888}
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor --output "$PWD/gpu_metrics.csv"
+
+# TODO(Cam): hardcoded to the low-latency recipe at every CONC until the
+# DeepEP FP8 weight-postprocess path is fixed for this checkpoint on B300
+# (RuntimeError: Recipe must be a list/tuple of 3 integers. raised from
+# sglang.srt.layers.quantization.fp8.process_weights_after_loading_block_quant).
+# Restore the CONC-based low-latency / balanced / max-throughput dispatch
+# on chore/dsv4-sgl-b300 once sglang can load the checkpoint under
+# --moe-a2a-backend deepep.
+RECIPE=low-latency
+RECIPE_FLAGS=(
+    --moe-runner-backend flashinfer_mxfp4
+    --chunked-prefill-size 4096
+    --disable-flashinfer-autotune
+    --mem-fraction-static 0.82
+)
+echo "Recipe: $RECIPE (CONC=$CONC)"
+
+set -x
+PYTHONNOUSERSITE=1 sglang serve \
+    --model-path $MODEL \
+    --host 0.0.0.0 \
+    --port $PORT \
+    --trust-remote-code \
+    --tp $TP \
+    --disable-radix-cache \
+    "${RECIPE_FLAGS[@]}" $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 "$RESULT_FILENAME" \
+    --result-dir "$PWD/"
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -1745,3 +1745,14 @@
     - "VLLM_ENGINE_READY_TIMEOUT_S=3600 to accommodate large weight loading"
     - "Configs: 1k1k conc 4-64, 8k1k conc 4-64"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1130
+
+- config-keys:
+    - dsv4-fp4-b300-sglang
+  description:
+    - "Add DeepSeek-V4-Pro FP4 B300 SGLang benchmark (low-latency fallback)"
+    - "Image: lmsysorg/sglang:deepseek-v4-b300"
+    - "Model: deepseek-ai/DeepSeek-V4-Pro"
+    - "Low-latency only (TP=8, EP=1, no DP-attn, no DeepEP) — DeepEP FP8 weight-postprocess path is broken for this checkpoint on B300"
+    - "Prefix caching disabled, no speculative decoding"
+    - "Configs: 1k1k conc 4-1024, 8k1k conc 4-512"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1143

runners/launch_b200-dgxc-slurm.sh

@@ -249,8 +249,7 @@ EOF
 
 else
 
-    HF_HUB_CACHE_MOUNT="/scratch/fsw/models"
-    export MODEL="$HF_HUB_CACHE_MOUNT/${MODEL#*/}"
+    HF_HUB_CACHE_MOUNT="/scratch/fsw/gharunners/hf-hub-cache"
     SQUASH_FILE="/home/sa-shared/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
@@ -276,7 +275,7 @@ else
 
     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 \

runners/launch_b300-nv.sh

@@ -248,29 +248,57 @@ 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.
+    # Pre-staged models on the B300 cluster live under /data/models. Point MODEL
+    # at the local copy so the benchmark skips `hf download` and reads from the
+    # mounted dir. Other models fall through and use `hf download` from their
+    # benchmark script.
+    HF_HUB_CACHE_MOUNT="/data/models"
     if [[ "$MODEL" == "Qwen/Qwen3.5-397B-A17B-FP8" ]]; then
-        export MODEL="/scratch/models/${MODEL#*/}"
+        export MODEL="$HF_HUB_CACHE_MOUNT/${MODEL#*/}"
+    elif [[ "$MODEL_PREFIX" == "dsv4" ]]; then
+        export MODEL="$HF_HUB_CACHE_MOUNT/dsv4-pro"
     fi
-    SQUASH_FILE="/data/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    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"
+
+    # TODO(Cam): the deepseek-v4 sglang images (lmsysorg/sglang:deepseek-v4-blackwell
+    # and its B300-recompiled forks like yhyang201/sglang-b300) install sglang
+    # editable at /workspace/sglang/python (prior sglang tags used /sgl-workspace/sglang),
+    # so the default $GITHUB_WORKSPACE:/workspace/ bind-mount masks the install
+    # and breaks `import sglang`. Mount these images at /ix instead; drop the
+    # conditional once the image stops installing editable under /workspace.
+    if [[ "$IMAGE" == *deepseek-v4-blackwell* || "$IMAGE" == *deepseek-v4-bw-ultra* || "$IMAGE" == *deepseek-v4-b300* || "$IMAGE" == *sglang-b300* ]]; then
+        CONTAINER_MOUNT_DIR=/ix
+    else
+        CONTAINER_MOUNT_DIR=/workspace
+    fi
+
+    # Import the squash file on the head node (outside any srun) under flock.
+    # Parallel GH jobs target the same shared squash path; flock serializes
+    # imports so only one job pulls and writes the file while the rest wait.
+    (
+        exec 9>"$LOCK_FILE"
+        flock -w 600 9 || { echo "Failed to acquire lock for $SQUASH_FILE" >&2; exit 1; }
+        if unsquashfs -l "$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
+    )
 
     # 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"
-
     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:$CONTAINER_MOUNT_DIR,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE_MOUNT \
         --no-container-mount-home \
-        --container-workdir=/workspace/ \
+        --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888 \
         bash benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b300${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh