Add dsv4-fp4-b300-vllm-mtp config (DSv4 vLLM B300 + MTP)

.github/configs/nvidia-master.yaml

@@ -2537,6 +2537,41 @@ dsv4-fp4-b300-vllm:
     - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 1024 }
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 2048 }
 
+# DeepSeek-V4-Pro on B300 with vLLM + MTP speculative decoding. The
+# deepseekv4-cu130 image registers DeepSeekV4MTPModel and exposes
+# --speculative-config '{"method":"deepseek_mtp",...}'; the script
+# defaults to num_speculative_tokens=1 (DSv4 publishes
+# num_nextn_predict_layers=1). Bench passes --dsv4 so prompts are
+# routed through encoding_dsv4.py -- required for honest MTP
+# acceptance numbers (random-token prompts silently regress).
+#
+# Two CONC bands (TP-only, mirrors dsv4-fp4-b300-sglang-mtp coverage):
+#   A: TP=8                 -- conc 1-8     low-conc, where MTP wins most
+#   B: TP=4                 -- conc 4-32    mid-batch
+# DP-attn (TP=4 ep=4 dp-attn / TP=8 ep=8 dp-attn) is wired in
+# benchmarks/single_node/dsv4_fp4_b300_vllm_mtp.sh via DP_ATTENTION=true
+# but is excluded from the initial sweep until the TP-only bands have
+# published numbers.
+dsv4-fp4-b300-vllm-mtp:
+  image: vllm/vllm-openai:deepseekv4-cu130
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b300
+  precision: fp4
+  framework: vllm
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, conc-start: 1, conc-end: 8, spec-decoding: mtp }
+    - { tp: 4, conc-start: 4, conc-end: 32, spec-decoding: mtp }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, conc-start: 1, conc-end: 8, spec-decoding: mtp }
+    - { tp: 4, conc-start: 4, conc-end: 32, spec-decoding: mtp }
+
 qwen3.5-fp8-h200-sglang:
   image: lmsysorg/sglang:v0.5.9-cu129-amd64
   model: Qwen/Qwen3.5-397B-A17B-FP8

benchmarks/single_node/dsv4_fp4_b300_vllm_mtp.sh

@@ -0,0 +1,139 @@
+#!/usr/bin/env bash
+
+# DeepSeek-V4-Pro B300 single-node vLLM recipe with MTP (Multi-Token
+# Prediction) speculative decoding. Mirrors the dsv4_fp4_b300_vllm.sh
+# split: TP mode (dp-attn=false) runs without expert parallel; DP mode
+# (dp-attn=true) enables expert parallel (EP_SIZE=TP value = DP size).
+#
+# MTP plumbing: the deepseekv4-cu130 image registers DeepSeekV4MTPModel
+# (vllm/model_executor/models/deepseek_v4_mtp.py) and remaps
+# model_type=deepseek_v4 -> deepseek_mtp inside SpeculativeConfig
+# (vllm/config/speculative.py). DSv4-Pro publishes
+# num_nextn_predict_layers=1 so the MTP head is invoked once per
+# spec step; vLLM accepts num_speculative_tokens > n_predict as long as
+# it's divisible (the head is reused). We start with k=1 (matches the
+# DSv4 b200 dynamo sglang MTP entry); bump SPEC_TOKENS to sweep.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+hf download "$MODEL"
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+# DeepSeek-V4-Pro weights are large; engine startup can exceed the default
+# 600s. Give it an hour to load.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+EP_ARGS=()
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+if [ "${DP_ATTENTION}" = "true" ]; then
+    MAX_NUM_BATCHED_TOKENS=2048
+else
+    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
+fi
+
+BENCHMARK_MAX_MODEL_LEN="$MAX_MODEL_LEN"
+if [ "$ISL" -eq 1024 ] && [ "$OSL" -eq 1024 ]; then
+    BENCHMARK_MAX_MODEL_LEN=4096
+fi
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    EVAL_MAX_MODEL_LEN=$(compute_eval_context_length "$MODEL" "$BENCHMARK_MAX_MODEL_LEN")
+    export EVAL_MAX_MODEL_LEN
+    SERVE_MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+else
+    SERVE_MAX_MODEL_LEN="$BENCHMARK_MAX_MODEL_LEN"
+fi
+
+# MTP speculative decoding: DSv4 has one nextn predictor layer baked
+# into the checkpoint. Bump for sweeps; vLLM reuses the head when
+# SPEC_TOKENS > n_predict (must be divisible by n_predict=1, so any int
+# works).
+SPEC_TOKENS="${SPEC_TOKENS:-1}"
+SPECULATIVE_CONFIG="{\"method\":\"deepseek_mtp\",\"num_speculative_tokens\":${SPEC_TOKENS}}"
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    --pipeline-parallel-size 1 \
+    --kv-cache-dtype fp8 \
+    --trust-remote-code \
+    --block-size 256 \
+    --no-enable-prefix-caching \
+    "${EP_ARGS[@]}" \
+    --speculative-config "$SPECULATIVE_CONFIG" \
+    --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
+    --attention_config.use_fp4_indexer_cache True \
+    --tokenizer-mode deepseek_v4 \
+    --tool-call-parser deepseek_v4 \
+    --enable-auto-tool-choice \
+    --reasoning-parser deepseek_v4 \
+    --max-cudagraph-capture-size 2048 \
+    --max-model-len "$SERVE_MAX_MODEL_LEN" \
+    --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" > "$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"
+
+pip install -q datasets pandas
+
+# --dsv4 routes prompts through encoding_dsv4.py (PR #1153), which emits the
+# <bos><User>...<Assistant><think> framing DeepSeek-V4-Pro expects. The DSv4-Pro
+# tokenizer ships without a jinja chat_template, so plain --use-chat-template
+# would crash; --dsv4 sidesteps that and satisfies the AGENTS.md rule that all
+# MTP scripts must benchmark against chat-formatted inputs (MTP acceptance
+# silently regresses on raw random tokens).
+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/ \
+    --trust-remote-code \
+    --dsv4
+
+# 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

@@ -1918,3 +1918,15 @@
     - "Three CONC bands: A=TP8 (1-8), B=TP4 (16-128), C=DP4 dp-attn (64-512); B/C overlap at conc 64,128"
     - "Configs: 1k1k and 8k1k, no validation.py / launcher / yaml-field changes (knob-free)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1180
+
+- config-keys:
+    - dsv4-fp4-b300-vllm-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP4 B300 vLLM benchmark with MTP (Multi-Token Prediction) speculative decoding"
+    - "Image: vllm/vllm-openai:deepseekv4-cu130 (registers DeepSeekV4MTPModel; speculative.py remaps model_type=deepseek_v4 -> deepseek_mtp)"
+    - "Model: deepseek-ai/DeepSeek-V4-Pro"
+    - "MTP via --speculative-config '{\"method\":\"deepseek_mtp\",\"num_speculative_tokens\":1}' (DSv4-Pro publishes num_nextn_predict_layers=1; SPEC_TOKENS env var allows sweeping >1 since vLLM reuses the head when divisible by n_predict)"
+    - "Bench passes --dsv4 (auto-enables --use-chat-template) so MTP acceptance is measured against DSv4 chat-formatted prompts (encoding_dsv4.py from PR #1153) — random-token prompts silently regress acceptance"
+    - "Two CONC bands (TP-only, mirrors dsv4-fp4-b300-sglang-mtp coverage): TP8 conc 1-8, TP4 conc 4-32, for both 1k1k and 8k1k"
+    - "DP-attn (TP=4 ep=4 / TP=8 ep=8) supported in script via DP_ATTENTION=true but excluded from initial sweep until TP-only numbers land"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1203