[NVIDIA] chore: B200 single node DeepSeek v4 SGLang MTP

.github/configs/nvidia-master.yaml

@@ -1669,6 +1669,24 @@ dsr1-fp4-b200-sglang:
     - { tp: 4, ep: 4, conc-start: 4, conc-end: 128 }
     - { tp: 8, ep: 8, conc-start: 4, conc-end: 16 }
 
+dsv4-fp4-b200-sglang:
+  image: lmsysorg/sglang:deepseek-v4-blackwell
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b200
+  precision: fp4
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 1024, spec-decoding: "mtp" }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 512, spec-decoding: "mtp" }
+
 # NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/DeepSeek/DeepSeek-R1
 # does not have a B300-specific recipe, so this config reuses the existing DSR1 FP4
 # B200 SGLang recipe as-is until B300-specific tuning is available.

benchmarks/single_node/dsv4_fp4_b200.sh

@@ -0,0 +1,92 @@
+#!/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 $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+nvidia-smi
+
+export SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
+
+# The deepseek-v4-blackwell image bakes CUDA_VISIBLE_DEVICES=4,5,6,7 into its 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 lmsysorg/sglang:deepseek-v4-blackwell image installs sglang
+# editable at /workspace/sglang/python; prior sglang tags used /sgl-workspace/sglang.
+# The runner mounts our repo at a non-/workspace path for this image 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"
+
+set -x
+PYTHONNOUSERSITE=1 \
+SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2=1 \
+SGLANG_OPT_USE_TOPK_V2=1 \
+SGLANG_JIT_DEEPGEMM_PRECOMPILE=0 \
+sglang serve \
+  --trust-remote-code \
+  --model-path $MODEL \
+  --tp 8 \
+  --moe-runner-backend flashinfer_mxfp4 \
+  --speculative-algo EAGLE \
+  --speculative-num-steps 3 \
+  --speculative-eagle-topk 1 \
+  --speculative-num-draft-tokens 4 \
+  --chunked-prefill-size 4096 \
+  --disable-flashinfer-autotune \
+  --mem-fraction-static 0.82 \
+  --host 0.0.0.0 \
+  --port $PORT > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+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

benchmarks/single_node/dsv4_fp4_b200_mtp.sh

@@ -0,0 +1,93 @@
+#!/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 $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+nvidia-smi
+
+export SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
+
+# The deepseek-v4-blackwell image bakes CUDA_VISIBLE_DEVICES=4,5,6,7 into its 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 lmsysorg/sglang:deepseek-v4-blackwell image installs sglang
+# editable at /workspace/sglang/python; prior sglang tags used /sgl-workspace/sglang.
+# The runner mounts our repo at a non-/workspace path for this image 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"
+
+set -x
+PYTHONNOUSERSITE=1 \
+SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2=1 \
+SGLANG_OPT_USE_TOPK_V2=1 \
+SGLANG_JIT_DEEPGEMM_PRECOMPILE=0 \
+sglang serve \
+  --trust-remote-code \
+  --model-path $MODEL \
+  --tp 8 \
+  --moe-runner-backend flashinfer_mxfp4 \
+  --speculative-algo EAGLE \
+  --speculative-num-steps 3 \
+  --speculative-eagle-topk 1 \
+  --speculative-num-draft-tokens 4 \
+  --chunked-prefill-size 4096 \
+  --disable-flashinfer-autotune \
+  --mem-fraction-static 0.82 \
+  --host 0.0.0.0 \
+  --port $PORT > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+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/" \
+    --use-chat-template
+
+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

@@ -1,3 +1,13 @@
+- config-keys:
+    - dsv4-fp4-b200-sglang
+  description:
+    - "Add DeepSeek-V4-Pro single-node B200 SGLang benchmark (TP8, EP8, dp-attention)"
+    - "Container: lmsysorg/sglang:deepseek-v4-blackwell"
+    - "Recipe from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
+    - "Parallelism and sweep conc ranges match the dsv4-fp4-b200-vllm config"
+    - "Prefix caching and speculative decoding disabled for baseline numbers"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1131
+
 - config-keys:
     - dsr1-fp8-h100-dynamo-trt
     - dsr1-fp8-h100-dynamo-sglang

runners/launch_b200-dgxc-slurm.sh

@@ -249,13 +249,23 @@ 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 '')
     LOCK_FILE="${SQUASH_FILE}.lock"
 
+    # TODO(Cam): lmsysorg/sglang:deepseek-v4-blackwell installs sglang editable at
+    # /workspace/sglang/python (prior sglang tags used /sgl-workspace/sglang), so
+    # the default $GITHUB_WORKSPACE:/workspace/ bind-mount masks the install and
+    # breaks `import sglang`. Mount this one image at /ix instead; drop the
+    # conditional once the image stops installing editable under /workspace.
+    if [[ "$IMAGE" == *deepseek-v4-blackwell* ]]; then
+        CONTAINER_MOUNT_DIR=/ix
+    else
+        CONTAINER_MOUNT_DIR=/workspace
+    fi
+
     salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --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)
 
@@ -276,9 +286,9 @@ 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:$CONTAINER_MOUNT_DIR,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
         --no-container-mount-home \
-        --container-workdir=/workspace/ \
+        --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888 \
         bash benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
 fi

runners/launch_b200-nb.sh

@@ -7,14 +7,24 @@ SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
 
 UCX_NET_DEVICES=eth0
 
+# TODO(Cam): lmsysorg/sglang:deepseek-v4-blackwell installs sglang editable at
+# /workspace/sglang/python (prior sglang tags used /sgl-workspace/sglang), so
+# the default $GITHUB_WORKSPACE:/workspace/ bind-mount masks the install and
+# breaks `import sglang`. Mount this one image at /ix instead; drop the
+# conditional once the image stops installing editable under /workspace.
+if [[ "$IMAGE" == *deepseek-v4-blackwell* ]]; then
+    CONTAINER_MOUNT_DIR=/ix
+else
+    CONTAINER_MOUNT_DIR=/workspace
+fi
+
 set -x
 srun --partition=$PARTITION --gres=gpu:$TP --exclusive --job-name="$RUNNER_NAME" \
 --container-image=$IMAGE \
---container-name=$(echo "$IMAGE" | sed 's/[\/:@#]/_/g')-${USER} \
---container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+--container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
 --no-container-mount-home \
 --container-remap-root \
 --container-writable \
---container-workdir=/workspace/ \
+--container-workdir=$CONTAINER_MOUNT_DIR \
 --no-container-entrypoint --export=ALL,PORT=8888,UCX_NET_DEVICES=$UCX_NET_DEVICES \
 bash benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh