Add qwen3.5-fp4-b200-trt-mtp single-node TensorRT-LLM benchmark

.github/configs/nvidia-master.yaml

@@ -12242,6 +12242,32 @@ qwen3.5-fp4-b200-trt:
       - { tp: 4, ep: 4, dp-attn: true, conc-list: [1024] }
       - { tp: 8, ep: 8, dp-attn: true, conc-list: [256, 512, 1024] }
 
+qwen3.5-fp4-b200-trt-mtp:
+  image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc18
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model-prefix: qwen3.5
+  runner: b200
+  precision: fp4
+  framework: trt
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 2, ep: 1, spec-decoding: "mtp", conc-list: [8] }
+      - { tp: 2, ep: 2, spec-decoding: "mtp", conc-list: [4] }
+      - { tp: 8, ep: 8, spec-decoding: "mtp", conc-list: [4] }
+      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: "mtp", conc-list: [64, 128, 256, 512, 1024] }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 2, ep: 1, spec-decoding: "mtp", conc-list: [4] }
+      - { tp: 2, ep: 2, spec-decoding: "mtp", conc-list: [8, 16] }
+      - { tp: 4, ep: 4, spec-decoding: "mtp", conc-list: [4] }
+      - { tp: 8, ep: 8, spec-decoding: "mtp", conc-list: [4] }
+      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: "mtp", conc-list: [128, 256, 1024] }
+
 # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3).
 # 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint
 # (MiniMaxAI/MiniMax-M3-MXFP8, ~444 GB) quantized by NVIDIA — native MX tensor

benchmarks/single_node/fixed_seq_len/qwen3.5_fp4_b200_trt_mtp.sh

@@ -0,0 +1,170 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    DP_ATTENTION \
+    EP_SIZE
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
+
+# MTP (multi-token prediction) speculative decode requires the FlashInfer GDN
+# prefill path to be disabled.
+export TLLM_USE_FLASHINFER_GDN_PREFILL="0"
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+nvidia-smi
+
+SERVER_LOG=/workspace/server.log
+EXTRA_CONFIG_FILE="qwen3.5-fp4-trt-mtp.yml"
+NUM_NEXTN_PREDICT_LAYERS=3
+
+# Attention-DP layouts run CUTEDSL MoE; everything else runs the TRTLLM backend.
+# With MTP the served batch is much smaller than raw concurrency: attention-DP
+# runs at CONC/8, everything else at CONC. The KV-cache memory fraction is tuned
+# per layout (there is no single derivable rule).
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    MAX_BATCH_SIZE=$(( CONC / 8 ))
+    MOE_BACKEND="CUTEDSL"
+    # attention-DP: 0.9 up to conc 512, backed off to 0.8 at conc 1024.
+    if (( CONC >= 1024 )); then KV_MEMORY_FRACTION=0.8; else KV_MEMORY_FRACTION=0.9; fi
+    MODE_CONFIG="enable_attention_dp: true
+attention_dp_config:
+    enable_balance: true
+    batching_wait_iters: 10
+    timeout_iters: 500"
+else
+    MAX_BATCH_SIZE="$CONC"
+    MOE_BACKEND="TRTLLM"
+    # non-attention-DP fraction, tuned per (ISL, TP, EP) layout.
+    case "${ISL}_tp${TP}_ep${EP_SIZE}" in
+        1024_tp2_ep1) KV_MEMORY_FRACTION=0.6 ;;
+        1024_tp2_ep2) KV_MEMORY_FRACTION=0.75 ;;
+        1024_tp8_ep8) KV_MEMORY_FRACTION=0.8 ;;
+        8192_tp2_ep1) KV_MEMORY_FRACTION=0.7 ;;
+        8192_tp2_ep2) KV_MEMORY_FRACTION=0.6 ;;
+        8192_tp4_ep4) KV_MEMORY_FRACTION=0.75 ;;
+        8192_tp8_ep8) KV_MEMORY_FRACTION=0.8 ;;
+        *)            KV_MEMORY_FRACTION=0.8 ;;
+    esac
+    # Short-context runs hold less in flight, so they wait on a tighter token
+    # ratio before flushing a batch.
+    case "$ISL" in
+        1024) BATCH_WAIT_MAX_TOKENS_RATIO=0.0625 ;;
+        *)    BATCH_WAIT_MAX_TOKENS_RATIO=0.45 ;;
+    esac
+    MODE_CONFIG="batch_wait_timeout_iters: 50
+batch_wait_max_tokens_ratio: $BATCH_WAIT_MAX_TOKENS_RATIO"
+fi
+
+cat > "$EXTRA_CONFIG_FILE" << EOF
+backend: pytorch
+print_iter_log: true
+enable_layerwise_nvtx_marker: false
+disable_overlap_scheduler: false
+enable_iter_perf_stats: true
+enable_chunked_prefill: false
+stream_interval: 20
+num_postprocess_workers: 4
+scheduler_config:
+    capacity_scheduler_policy: MAX_UTILIZATION
+    context_chunking_policy: FIRST_COME_FIRST_SERVED
+kv_cache_config:
+    free_gpu_memory_fraction: $KV_MEMORY_FRACTION
+    enable_block_reuse: false
+    dtype: fp8
+cuda_graph_config:
+    enable_padding: true
+    batch_sizes:
+    - 1
+    - 2
+    - 4
+    - 8
+    - 16
+    - 32
+    - 64
+    - 128
+moe_config:
+    backend: $MOE_BACKEND
+    use_low_precision_moe_combine: true
+speculative_config:
+    decoding_type: MTP
+    num_nextn_predict_layers: $NUM_NEXTN_PREDICT_LAYERS
+$MODE_CONFIG
+EOF
+
+echo "Generated config file contents:"
+cat "$EXTRA_CONFIG_FILE"
+
+MAX_MODEL_LEN=$(( MAX_MODEL_LEN > 8192 ? MAX_MODEL_LEN : 8192 ))
+
+case "${ISL}_${OSL}" in
+    8192_1024) MAX_NUM_TOKENS=32768 ;;
+    1024_1024) MAX_NUM_TOKENS=16384 ;;
+    *)
+        MAX_NUM_TOKENS=$(( ISL + OSL + 256 ))
+        MAX_NUM_TOKENS=$(( MAX_NUM_TOKENS > 8192 ? MAX_NUM_TOKENS : 8192 ))
+        ;;
+esac
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+    MAX_NUM_TOKENS="$EVAL_MAX_MODEL_LEN"
+fi
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+mpirun -n 1 --oversubscribe --allow-run-as-root \
+    trtllm-serve "$MODEL" --port="$PORT" \
+    --trust_remote_code \
+    --backend=pytorch \
+    --max_batch_size "$MAX_BATCH_SIZE" \
+    --max_seq_len="$MAX_MODEL_LEN" \
+    --max_num_tokens="$MAX_NUM_TOKENS" \
+    --tp_size="$TP" --ep_size="$EP_SIZE" \
+    --extra_llm_api_options="$EXTRA_CONFIG_FILE" \
+    > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend openai \
+    --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/ \
+    --use-chat-template
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -4123,6 +4123,13 @@
     - "Allocate FlashInfer MNNVL workspace for one-shot TP8 all-reduce during CUDA graph capture"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1897
 
+- config-keys:
+    - qwen3.5-fp4-b200-trt-mtp
+  description:
+    - "Add Qwen3.5-397B-A17B-NVFP4 B200 single-node TensorRT-LLM benchmark with MTP speculative decode (1k/1k and 8k/1k) across a TP/TEP/DEP parallelism sweep"
+    - "Image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1894
+
 - config-keys:
     - dsr1-fp4-b200-sglang
   description: