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

.github/configs/nvidia-master.yaml

@@ -1705,6 +1705,40 @@ dsv4-fp4-b200-sglang:
     # max-throughput
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 512 }
 
+dsv4-fp4-b200-sglang-mtp:
+  image: lmsysorg/sglang:deepseek-v4-blackwell
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b200-dsv4
+  precision: fp4
+  framework: sglang
+  multinode: false
+  # Mirrors the dsv4-fp4-b200-sglang baseline split (low-latency / balanced /
+  # max-throughput selected inside benchmarks/single_node/dsv4_fp4_b200_mtp.sh
+  # by CONC) so result filenames (ep=, dpa=) reflect the recipe. EAGLE/MTP is
+  # tuned per recipe in the script per
+  # https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+  # (max-throughput omits MTP since the verify cost exceeds savings at saturation).
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    # low-latency
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 32, spec-decoding: "mtp" }
+    # balanced
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: "mtp" }
+    # max-throughput
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024, spec-decoding: "mtp" }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    # low-latency
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 32, spec-decoding: "mtp" }
+    # balanced
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: "mtp" }
+    # max-throughput
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, 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_mtp.sh

@@ -0,0 +1,137 @@
+#!/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
+export SGLANG_ENABLE_SPEC_V2=1
+export SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2=1
+export SGLANG_OPT_USE_TOPK_V2=1
+
+# 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"
+
+# Three recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+# with EAGLE / MTP speculative decoding tuned per recipe:
+#   - low-latency    (CONC <= 32):        TP-only, MTP num-steps=3 / draft-tokens=4
+#   - balanced       (32 < CONC <= 128):  + DP-attn, MTP num-steps=1 / draft-tokens=2
+#   - max-throughput (CONC > 128):        + DP-attn, MTP disabled (verify cost > savings at saturation)
+DEEPEP_CONFIG='{"normal_dispatch":{"num_sms":96},"normal_combine":{"num_sms":96}}'
+
+SPEC_FLAGS=()
+
+if [[ $CONC -le 32 ]]; then
+    RECIPE=low-latency
+    RECIPE_FLAGS=(
+        --moe-runner-backend flashinfer_mxfp4
+        --chunked-prefill-size 4096
+        --disable-flashinfer-autotune
+        --mem-fraction-static 0.82
+    )
+    SPEC_FLAGS=(
+        --speculative-algo EAGLE
+        --speculative-num-steps 3
+        --speculative-eagle-topk 1
+        --speculative-num-draft-tokens 4
+    )
+elif [[ $CONC -le 128 ]]; then
+    RECIPE=balanced
+    export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256
+    RECIPE_FLAGS=(
+        --dp-size "$TP"
+        --enable-dp-attention
+        --moe-a2a-backend deepep
+        --deepep-config "$DEEPEP_CONFIG"
+        --mem-fraction-static 0.82
+        --cuda-graph-max-bs 64
+        --max-running-requests 128
+    )
+    SPEC_FLAGS=(
+        --speculative-algo EAGLE
+        --speculative-num-steps 1
+        --speculative-eagle-topk 1
+        --speculative-num-draft-tokens 2
+    )
+else
+    RECIPE=max-throughput
+    export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256
+    RECIPE_FLAGS=(
+        --dp-size "$TP"
+        --enable-dp-attention
+        --moe-a2a-backend deepep
+        --deepep-config "$DEEPEP_CONFIG"
+        --mem-fraction-static 0.82
+        --cuda-graph-max-bs 64
+        --max-running-requests 256
+    )
+fi
+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[@]}" "${SPEC_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/" \
+    --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

@@ -8,6 +8,14 @@
     - "Prefix caching and speculative decoding disabled for baseline numbers"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1131
 
+- config-keys:
+    - dsv4-fp4-b200-sglang-mtp
+  description:
+    - "Add DeepSeek-V4-Pro single-node B200 SGLang benchmark with EAGLE / MTP speculative decoding"
+    - "Mirrors the dsv4-fp4-b200-sglang baseline recipes (low-latency / balanced / max-throughput)"
+    - "MTP tuned per https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4: num-steps=3/4 (low-latency), 1/2 (balanced), disabled at max-throughput"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1145
+
 - config-keys:
     - dsr1-fp8-h100-dynamo-trt
     - dsr1-fp8-h100-dynamo-sglang