diff --git a/experiments/dsv4_h200_vllm_dspark_vs_default/results/20260708-142121/ttft_analysis.md b/experiments/dsv4_h200_vllm_dspark_vs_default/results/20260708-142121/ttft_analysis.md
new file mode 100644
index 0000000..be6e009
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_dspark_vs_default/results/20260708-142121/ttft_analysis.md
@@ -0,0 +1,85 @@
+# DSpark TTFT 长于 vLLM baseline 的根因分析
+
+## 数据回顾
+
+| Scenario | Backend | Conc | Input | Output | Mean TTFT(ms) | P95 TTFT(ms) | P99 TTFT(ms) | Mean TPOT(ms) | Req/s |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| c1_i512_o256 | vllm-dspark | 1 | 512 | 256 | 86.51 | 185.00 | 188.81 | 2.09 | 2.56 |
+| c1_i512_o256 | vllm-default | 1 | 512 | 256 | 49.50 | 68.50 | 71.09 | 6.66 | 0.99 |
+| c32_i512_o256 | vllm-dspark | 32 | 512 | 256 | 262.56 | 527.21 | 532.11 | 12.06 | 16.53 |
+| c32_i512_o256 | vllm-default | 32 | 512 | 256 | 122.62 | 272.00 | 304.80 | 15.77 | 12.98 |
+| c128_i512_o256 | vllm-dspark | 128 | 512 | 256 | 778.36 | 1071.85 | 1075.35 | 28.95 | 23.94 |
+| c128_i512_o256 | vllm-default | 128 | 512 | 256 | 617.72 | 767.18 | 770.89 | 18.31 | 29.99 |
+| c1_i4000_o512 | vllm-dspark | 1 | 4000 | 512 | 123.58 | 176.29 | 190.52 | 1.66 | 1.73 |
+| c1_i4000_o512 | vllm-default | 1 | 4000 | 512 | 131.50 | 191.53 | 268.69 | 6.68 | 0.50 |
+| c32_i4000_o512 | vllm-dspark | 32 | 4000 | 512 | 298.42 | 434.24 | 522.34 | 9.68 | 11.49 |
+| c32_i4000_o512 | vllm-default | 32 | 4000 | 512 | 382.74 | 681.98 | 684.75 | 16.43 | 5.38 |
+
+## 核心结论
+
+**DSpark 在短输入（i=512）场景下 TTFT 明显长于 vLLM baseline，是投机解码机制本身的正常 trade-off，但在当前配置下 overhead 偏大，有优化空间。**
+
+## 为什么短输入场景 TTFT 会变差？
+
+### 1. 投机解码的本质 overhead
+
+DSpark 在生成第一个 token 时，除了执行一次目标模型（target model）的 prefill，还要：
+- 加载并执行 draft model；
+- 让 draft model 连续预测最多 `spec-tokens=5` 个 token；
+- 用 target model 对这 5 个 token 做并行验证（verification）；
+- 根据验证结果回退或接受。
+
+对于 **短输入（512 tokens）**，baseline 的 prefill 本身非常快（P50 仅 44-65ms），draft model 的额外 forward + verification 的固定 overhead 占了 TTFT 的很大比例，导致 TTFT 反而变长。
+
+### 2. 长输入场景反而受益
+
+对于 **长输入（4000 tokens）**，prefill 是主要瓶颈。DSpark 通过 draft model 投机生成后续 token，能显著减少 target model 的 forward 次数，因此：
+- c1_i4000_o512：TTFT 与 baseline 接近（123ms vs 131ms），但 TPOT 从 6.68ms 降到 1.66ms；
+- c32_i4000_o512：TTFT 明显优于 baseline（298ms vs 383ms，-22%），TPOT 从 16.43ms 降到 9.68ms。
+
+这符合投机解码的设计目标：**用 draft model 的额外计算换取 target model 更少的完整 forward，从而在长序列或高输出长度上获得端到端收益。**
+
+### 3. 高并发短输入场景被放大
+
+在 c128_i512_o256 下：
+- DSpark 的 Req/s 低于 baseline（23.94 vs 29.99），说明单位时间内处理的请求更少；
+- 同时 DSpark 的 mean TPOT 高于 baseline（28.95ms vs 18.31ms）。
+
+这说明在并发较高、输入较短时，DSpark 的 draft-verification 流水线没有跑赢 baseline 的简单 decode，导致队列等待时间增加，进一步放大了 TTFT。
+
+## 是代码缺陷还是正常现象？
+
+### 属于正常现象的部分
+
+1. **短输入下投机解码增加 TTFT 是预期行为**：只要 draft model 不是零成本，第一个 token 必然比 baseline 慢。
+2. **DSpark 的收益主要体现在 TPOT/E2E/吞吐量**：所有场景下 DSpark 的 TPOT 都显著低于 baseline，长输入场景 E2E 也大幅缩短。
+
+### 可能存在优化空间的部分
+
+1. **`spec-tokens=5` 对短输入可能过大**
+   - 当前 DSpark 使用 `--spec-tokens 5`。
+   - 对于 512 token 的短输入，prefill 后只需要 256 个输出 token，一次 spec 5 个 token 的 verification overhead 相对较高。
+   - **建议**：对短输入场景尝试 `spec-tokens=1` 或 `spec-tokens=2`，看是否能降低 TTFT overhead 同时保持 TPOT 收益。
+
+2. **DSpark 与 default 使用不同 checkpoint**
+   - DSpark 模型路径是 `/data/models/DeepSeek-V4-Flash-DSpark`，而 baseline 是 `/data/models/DeepSeek-V4-Flash`。
+   - 如果 DSpark checkpoint 的 prefill 路径没有被充分优化（例如 MLA/KV cache 格式差异），可能会额外增加 prefill 时间。
+   - **建议**：确认两个 checkpoint 在 prefill 阶段是否使用相同的 kernel/量化路径，排除模型差异本身带来的 TTFT 偏移。
+
+3. **draft model 与 target model 的调度/重叠**
+   - 如果 draft model 的 forward 和 target model 的 verification 没有充分 overlap，TTFT 会进一步恶化。
+   - **建议**：通过 vLLM 的 profile/profiling 工具查看 prefill 阶段 draft model 是否串行执行，以及是否可以通过 `--spec-max-model-len` 或调度参数优化。
+
+4. **高并发下的 batching 策略**
+   - c128_i512_o256 下 DSpark 吞吐量低于 baseline，可能是 speculative decoding 在高并发下的 batching 效率下降。
+   - **建议**：测试不同 `max-num-seqs` 或 `--max-num-batched-tokens` 配置，观察并发吞吐和 TTFT 的变化。
+
+## 建议的下一步
+
+1. **参数调优实验**：固定模型和硬件，扫描 `spec-tokens ∈ {1, 2, 3, 5}`，重点观察短输入场景的 TTFT 和 TPOT 变化。
+2. **profiling**：在 c1_i512_o256 和 c32_i512_o256 下，用 vLLM 的 `--profile` 或 Nsight 抓取 prefill 时间线，确认 draft model 的 overhead 占比。
+3. **与 MTP 对比**：当前正在跑的 vLLM+MTP 实验会提供另一组投机解码数据，可以对比 DSpark 和官方 MTP 在 TTFT/TPOT 上的 trade-off 差异。
+
+## 总结
+
+DSpark 在短输入场景 TTFT 变差**不是代码 bug，而是投机解码机制的正常开销**。但当前 `spec-tokens=5` 的配置在短输入下 overhead 偏大，建议通过调参和 profiling 进一步优化，尤其是在 S1 实时交互层（输入短、延迟敏感）场景。
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/compare.py b/experiments/dsv4_h200_vllm_mtp_vs_default/compare.py
new file mode 100755
index 0000000..05db496
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/compare.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+"""Generate a side-by-side comparison of vLLM+MTP speculative decoding and vLLM default.
+
+Usage:
+    python3 compare.py --mtp <mtp_result_root> --default <default_result_root> \
+        [--output comparison.md]
+"""
+import argparse
+import json
+from collections import defaultdict
+from pathlib import Path
+
+
+def load_result(result_root: Path) -> dict:
+    path = result_root / "results.json"
+    with open(path, "r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--mtp", type=Path, required=True)
+    parser.add_argument("--default", type=Path, required=True)
+    parser.add_argument("-o", "--output", type=Path, default=Path("comparison.md"))
+    args = parser.parse_args()
+
+    mtp_data = load_result(args.mtp)
+    default_data = load_result(args.default)
+
+    by_scenario = defaultdict(dict)
+    for data in (mtp_data, default_data):
+        backend = data["metadata"]["engine"]
+        for s in data.get("scenarios", []):
+            key = s["name"]
+            by_scenario[key][backend] = s
+
+    with open(args.output, "w", encoding="utf-8") as f:
+        f.write("# vLLM+MTP speculative decoding vs vLLM default on DeepSeek-V4-Flash (H200, TP=8)\n\n")
+        f.write("## Summary\n\n")
+        f.write("- Model: `/data/models/DeepSeek-V4-Flash`\n")
+        f.write("- Hardware: 8x NVIDIA H200 143GB\n")
+        f.write("- Tensor Parallelism: 8\n")
+        f.write("- Benchmark client: `sglang.bench_serving --backend vllm`\n")
+        f.write("- Default: no speculative decoding\n")
+        f.write("- MTP: `--speculative-config '{\\\"method\\\":\\\"mtp\\\",\\\"num_speculative_tokens\\\":1}'` (official DeepSeek-V4 recipe)\n\n")
+
+        f.write("## Side-by-side results\n\n")
+        f.write("| Scenario | Backend | Conc | Input | Output | Req/s | OutTok/s | Mean TTFT(ms) | P95 TTFT(ms) | P99 TTFT(ms) | Mean TPOT(ms) | P99 TPOT(ms) | Mean E2E(ms) |\n")
+        f.write("|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|\n")
+
+        for scenario_name in sorted(by_scenario.keys()):
+            for backend in ("vllm-mtp", "vllm-default"):
+                s = by_scenario[scenario_name].get(backend)
+                if s is None:
+                    continue
+                cfg = s["config"]
+                m = s["metrics"]
+                f.write(
+                    f"| {scenario_name} | {backend} | {cfg['concurrency']} | {cfg['input_len']} | {cfg['output_len']} | "
+                    f"{m['request_throughput']:.2f} | {m['output_token_throughput']:.2f} | "
+                    f"{m['ttft_ms']['mean']:.2f} | {m['ttft_ms']['p95']:.2f} | {m['ttft_ms']['p99']:.2f} | "
+                    f"{m['tpot_ms']['mean']:.2f} | {m['tpot_ms']['p99']:.2f} | "
+                    f"{m['e2e_ms']['mean']:.2f} |\n"
+                )
+
+        f.write("\n## Notes\n\n")
+        f.write("- TTFT mean/P95/P99 are the main focus for verifying MTP's impact on time-to-first-token.\n")
+        f.write("- Mean TPOT and E2E are included to check whether speculative decoding pays back after first token.\n")
+
+    print(f"Wrote comparison to {args.output}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/config.env b/experiments/dsv4_h200_vllm_mtp_vs_default/config.env
new file mode 100644
index 0000000..a6ee34f
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/config.env
@@ -0,0 +1,32 @@
+# Configuration for vLLM MTP speculative decoding vs vLLM default comparison on H200.
+# The default-side results are reused from dsv4_h200_vllm_dspark_vs_default/results/20260708-142121/default.
+
+EXPERIMENT="dsv4_h200_vllm_mtp_vs_default"
+MODEL_NAME="DeepSeek-V4-Flash"
+MODEL_PATH="/data/models/DeepSeek-V4-Flash"
+SERVED_MODEL_NAME="deepseek-v4-flash"
+
+MTP_PORT="${MTP_PORT:-30009}"
+DEFAULT_RESULT_ROOT="${DEFAULT_RESULT_ROOT:-/data/user1/yy/experiments/dsv4_h200_vllm_dspark_vs_default/results/20260708-142121/default}"
+
+VENV_VLLM="${VENV_VLLM:-/data/user1/yy/envs/vllm}"
+VENV_CLIENT="${VENV_CLIENT:-/data/user1/yy/envs/sglang}"
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+TP=8
+
+MAX_MODEL_LEN=32768
+MAX_NUM_SEQS=256
+SPEC_METHOD="${SPEC_METHOD:-mtp}"
+SPEC_TOKENS="${SPEC_TOKENS:-1}"
+
+# Scenarios must match the ones in the reused default result set.
+declare -a SCENARIOS=(
+  "1   512   256  32"
+  "32  512   256  128"
+  "128 512   256  128"
+  "1   4000  512  32"
+  "32  4000  512  64"
+)
+
+MTP_START_SCRIPT="${SCRIPT_DIR:-.}/start_mtp.sh"
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/comparison.md b/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/comparison.md
new file mode 100644
index 0000000..0968e44
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/comparison.md
@@ -0,0 +1,30 @@
+# vLLM+MTP speculative decoding vs vLLM default on DeepSeek-V4-Flash (H200, TP=8)
+
+## Summary
+
+- Model: `/data/models/DeepSeek-V4-Flash`
+- Hardware: 8x NVIDIA H200 143GB
+- Tensor Parallelism: 8
+- Benchmark client: `sglang.bench_serving --backend vllm`
+- Default: no speculative decoding
+- MTP: `--speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":1}'` (official DeepSeek-V4 recipe)
+
+## Side-by-side results
+
+| Scenario | Backend | Conc | Input | Output | Req/s | OutTok/s | Mean TTFT(ms) | P95 TTFT(ms) | P99 TTFT(ms) | Mean TPOT(ms) | P99 TPOT(ms) | Mean E2E(ms) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| c128_i512_o256 | vllm-mtp | 128 | 512 | 256 | 35.03 | 4625.47 | 771.59 | 1110.39 | 1114.20 | 18.16 | 111.99 | 2506.29 |
+| c128_i512_o256 | vllm-default | 128 | 512 | 256 | 29.99 | 3959.89 | 617.72 | 767.18 | 770.89 | 18.31 | 65.44 | 2695.91 |
+| c1_i4000_o512 | vllm-mtp | 1 | 4000 | 512 | 0.83 | 234.90 | 161.27 | 216.12 | 289.08 | 3.70 | 3.89 | 1197.52 |
+| c1_i4000_o512 | vllm-default | 1 | 4000 | 512 | 0.50 | 140.13 | 131.50 | 191.53 | 268.69 | 6.68 | 6.73 | 2008.09 |
+| c1_i512_o256 | vllm-mtp | 1 | 512 | 256 | 1.64 | 239.37 | 60.63 | 70.69 | 71.97 | 3.80 | 4.33 | 607.04 |
+| c1_i512_o256 | vllm-default | 1 | 512 | 256 | 0.99 | 143.82 | 49.50 | 68.50 | 71.09 | 6.66 | 6.71 | 1010.98 |
+| c32_i4000_o512 | vllm-mtp | 32 | 4000 | 512 | 6.32 | 1785.09 | 409.17 | 667.77 | 673.18 | 14.18 | 19.84 | 4431.73 |
+| c32_i4000_o512 | vllm-default | 32 | 4000 | 512 | 5.38 | 1518.49 | 382.74 | 681.98 | 684.75 | 16.43 | 25.49 | 4927.71 |
+| c32_i512_o256 | vllm-mtp | 32 | 512 | 256 | 12.56 | 1657.98 | 279.91 | 761.67 | 766.29 | 16.73 | 37.56 | 2332.50 |
+| c32_i512_o256 | vllm-default | 32 | 512 | 256 | 12.98 | 1713.20 | 122.62 | 272.00 | 304.80 | 15.77 | 23.26 | 2154.65 |
+
+## Notes
+
+- TTFT mean/P95/P99 are the main focus for verifying MTP's impact on time-to-first-token.
+- Mean TPOT and E2E are included to check whether speculative decoding pays back after first token.
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/report.md b/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/report.md
new file mode 100644
index 0000000..4f6b775
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/report.md
@@ -0,0 +1,19 @@
+# 8x NVIDIA H200 143GB VLLM Benchmark Report
+
+- Result root: `/data/user1/yy/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp`
+- Model: `/data/models/DeepSeek-V4-Flash`
+- Backend: VLLM
+- Benchmark client: `sglang.bench_serving --backend vllm`
+
+## Results
+
+| Scenario | Phase | Concurrency | Input | Output | Duration(s) | Success | Req/s | In tok/s | Out tok/s | Total tok/s | Mean TTFT(ms) | P95 TTFT(ms) | P99 TTFT(ms) | Mean TPOT(ms) | P95 TPOT(ms) | P99 TPOT(ms) | Mean E2E(ms) | P95 E2E(ms) | P99 E2E(ms) | SLO |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| c128_i512_o256 | mtp | 128 | 512 | 256 | 3.65 | 128 | 35.03 | 9647.26 | 4625.47 | 14272.73 | 771.59 | 1110.39 | 1114.20 | 18.16 | 32.29 | 111.99 | 2506.29 | 3456.20 | 3560.22 | ✅ |
+| c1_i4000_o512 | mtp | 1 | 4000 | 512 | 38.35 | 32 | 0.83 | 1694.86 | 234.90 | 1929.76 | 161.27 | 216.12 | 289.08 | 3.70 | 3.82 | 3.89 | 1197.52 | 2034.69 | 2078.68 | ✅ |
+| c1_i512_o256 | mtp | 1 | 512 | 256 | 19.46 | 32 | 1.64 | 419.94 | 239.37 | 659.31 | 60.63 | 70.69 | 71.97 | 3.80 | 4.11 | 4.33 | 607.04 | 1022.42 | 1023.52 | ✅ |
+| c32_i4000_o512 | mtp | 32 | 4000 | 512 | 10.12 | 64 | 6.32 | 13280.42 | 1785.09 | 15065.50 | 409.17 | 667.77 | 673.18 | 14.18 | 17.48 | 19.84 | 4431.73 | 8157.51 | 8520.71 | ✅ |
+| c32_i512_o256 | mtp | 32 | 512 | 256 | 10.19 | 128 | 12.56 | 3458.01 | 1657.98 | 5115.99 | 279.91 | 761.67 | 766.29 | 16.73 | 26.24 | 37.56 | 2332.50 | 4812.90 | 5073.77 | ✅ |
+
+SLO: S2 tier — TTFT P95 < 3000ms, TPOT mean < 50ms. ✅ pass, ⚠️ partial, ❌ fail.
+
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/results.json b/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/results.json
new file mode 100644
index 0000000..940635a
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/results.json
@@ -0,0 +1,311 @@
+{
+  "metadata": {
+    "experiment": "dsv4_h200_vllm_mtp_vs_default_mtp",
+    "run_id": "20260708-160349",
+    "timestamp": "2026-07-08T16:03:51+00:00",
+    "model": "/data/models/DeepSeek-V4-Flash",
+    "backend": "vllm",
+    "engine": "vllm-mtp",
+    "hardware": "8x NVIDIA H200 143GB",
+    "accelerator": "NVIDIA H200",
+    "chip": "nvidia_h200",
+    "script": "experiments/dsv4_h200_vllm_mtp_vs_default/run_bench.sh",
+    "env": "/data/user1/yy/envs/vllm",
+    "git_commit": "ca98fe0",
+    "git_dirty": "dirty",
+    "description": "H200 vLLM+MTP TP=8 benchmark for DeepSeek-V4-Flash"
+  },
+  "config": {
+    "tp": 8,
+    "cuda_visible_devices": "0,1,2,3,4,5,6,7",
+    "max_model_len": 32768,
+    "max_num_seqs": 256,
+    "spec_method": "mtp",
+    "spec_tokens": 1,
+    "backend": "vllm",
+    "server_start_script": "experiments/dsv4_h200_vllm_mtp_vs_default/start_mtp.sh",
+    "server_args": "vllm serve /data/models/DeepSeek-V4-Flash --trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-model-len 32768 --max-num-seqs 256 --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --speculative-config {\\\"method\\\":\\\"mtp\\\",\\\"num_speculative_tokens\\\":1} --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port 30009"
+  },
+  "scenarios": [
+    {
+      "name": "c128_i512_o256",
+      "config": {
+        "phase": "mtp",
+        "concurrency": 128,
+        "input_len": 512,
+        "output_len": 256,
+        "dataset": "random",
+        "num_prompts": 128
+      },
+      "metrics": {
+        "success": 128,
+        "failed": 0,
+        "duration_s": 3.6536797620065045,
+        "request_throughput": 35.03317431676217,
+        "input_token_throughput": 9647.260377478384,
+        "output_token_throughput": 4625.473796510006,
+        "total_token_throughput": 14272.734173988389,
+        "total_input_tokens": 35248,
+        "total_output_tokens": 16900,
+        "e2e_ms": {
+          "mean": 2506.2933596329913,
+          "p50": 2625.210165999306,
+          "p90": 3417.3656311962986,
+          "p95": 3456.20117305225,
+          "p99": 3560.217050112697
+        },
+        "ttft_ms": {
+          "mean": 771.5936086792681,
+          "p50": 743.3261279948056,
+          "p90": 964.7660584974801,
+          "p95": 1110.3887211014808,
+          "p99": 1114.1990597829863
+        },
+        "tpot_ms": {
+          "mean": 18.15787222271614,
+          "p50": 13.890511986549402,
+          "p90": 25.0715809341214,
+          "p95": 32.29210961239336,
+          "p99": 111.98906169440306
+        },
+        "itl_ms": {
+          "mean": 25.75249892379593,
+          "p50": 21.063708009023685,
+          "p90": 26.390794408507645,
+          "p95": 28.396654246898834,
+          "p99": 255.2068907531792
+        }
+      },
+      "slo_status": {
+        "ttft_p95_ok": true,
+        "tpot_mean_ok": true,
+        "overall": "✅"
+      },
+      "raw_file": "/data/user1/yy/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/raw_outputs/vllm_mtp_0708_128_512_256.jsonl"
+    },
+    {
+      "name": "c1_i4000_o512",
+      "config": {
+        "phase": "mtp",
+        "concurrency": 1,
+        "input_len": 4000,
+        "output_len": 512,
+        "dataset": "random",
+        "num_prompts": 32
+      },
+      "metrics": {
+        "success": 32,
+        "failed": 0,
+        "duration_s": 38.35236829900532,
+        "request_throughput": 0.8343682911709505,
+        "input_token_throughput": 1694.8627394591913,
+        "output_token_throughput": 234.90074797372165,
+        "total_token_throughput": 1929.7634874329128,
+        "total_input_tokens": 65002,
+        "total_output_tokens": 9009,
+        "e2e_ms": {
+          "mean": 1197.5197937804296,
+          "p50": 1222.8141454979777,
+          "p90": 1945.1575131970458,
+          "p95": 2034.6854075956799,
+          "p99": 2078.678046991845
+        },
+        "ttft_ms": {
+          "mean": 161.27064584452455,
+          "p50": 187.0977380021941,
+          "p90": 207.2812115002307,
+          "p95": 216.12448994783335,
+          "p99": 289.0834562029342
+        },
+        "tpot_ms": {
+          "mean": 3.698815812859111,
+          "p50": 3.6894009238415832,
+          "p90": 3.761323209835616,
+          "p95": 3.8155798700943198,
+          "p99": 3.8925407732209094
+        },
+        "itl_ms": {
+          "mean": 7.345683533679442,
+          "p50": 7.343719502387103,
+          "p90": 7.546291207836475,
+          "p95": 7.646059650869574,
+          "p99": 7.946166486799484
+        }
+      },
+      "slo_status": {
+        "ttft_p95_ok": true,
+        "tpot_mean_ok": true,
+        "overall": "✅"
+      },
+      "raw_file": "/data/user1/yy/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/raw_outputs/vllm_mtp_0708_1_4000_512.jsonl"
+    },
+    {
+      "name": "c1_i512_o256",
+      "config": {
+        "phase": "mtp",
+        "concurrency": 1,
+        "input_len": 512,
+        "output_len": 256,
+        "dataset": "random",
+        "num_prompts": 32
+      },
+      "metrics": {
+        "success": 32,
+        "failed": 0,
+        "duration_s": 19.455183718004264,
+        "request_throughput": 1.644805850401016,
+        "input_token_throughput": 419.9394936805094,
+        "output_token_throughput": 239.37065141617282,
+        "total_token_throughput": 659.3101450966822,
+        "total_input_tokens": 8170,
+        "total_output_tokens": 4657,
+        "e2e_ms": {
+          "mean": 607.0424630620437,
+          "p50": 709.1372035065433,
+          "p90": 974.411873005738,
+          "p95": 1022.4235178509844,
+          "p99": 1023.5160801192977
+        },
+        "ttft_ms": {
+          "mean": 60.633899186541385,
+          "p50": 67.31531800323864,
+          "p90": 70.1838248045533,
+          "p95": 70.6854243442649,
+          "p99": 71.96889203012688
+        },
+        "tpot_ms": {
+          "mean": 3.7982499249353343,
+          "p50": 3.7510736315107254,
+          "p90": 3.9678993170763492,
+          "p95": 4.107961743017731,
+          "p99": 4.327418892014148
+        },
+        "itl_ms": {
+          "mean": 7.296974936146524,
+          "p50": 7.262905994139146,
+          "p90": 7.464885500667151,
+          "p95": 7.645853242138401,
+          "p99": 9.469499189435757
+        }
+      },
+      "slo_status": {
+        "ttft_p95_ok": true,
+        "tpot_mean_ok": true,
+        "overall": "✅"
+      },
+      "raw_file": "/data/user1/yy/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/raw_outputs/vllm_mtp_0708_1_512_256.jsonl"
+    },
+    {
+      "name": "c32_i4000_o512",
+      "config": {
+        "phase": "mtp",
+        "concurrency": 32,
+        "input_len": 4000,
+        "output_len": 512,
+        "dataset": "random",
+        "num_prompts": 64
+      },
+      "metrics": {
+        "success": 64,
+        "failed": 0,
+        "duration_s": 10.122198707002099,
+        "request_throughput": 6.322736971733974,
+        "input_token_throughput": 13280.415045301297,
+        "output_token_throughput": 1785.086474097831,
+        "total_token_throughput": 15065.501519399128,
+        "total_input_tokens": 134427,
+        "total_output_tokens": 18069,
+        "e2e_ms": {
+          "mean": 4431.7270394060415,
+          "p50": 4118.452171998797,
+          "p90": 7965.46315389569,
+          "p95": 8157.508349696581,
+          "p99": 8520.712041212246
+        },
+        "ttft_ms": {
+          "mean": 409.17025550083963,
+          "p50": 412.25154899439076,
+          "p90": 663.4122133953497,
+          "p95": 667.7715254438226,
+          "p99": 673.1776816866477
+        },
+        "tpot_ms": {
+          "mean": 14.18404794198284,
+          "p50": 15.168117225958857,
+          "p90": 17.233481520907418,
+          "p95": 17.483692691057605,
+          "p99": 19.836913968959017
+        },
+        "itl_ms": {
+          "mean": 28.557059211652966,
+          "p50": 13.894067000364885,
+          "p90": 29.86279199831188,
+          "p95": 174.40551639592738,
+          "p99": 181.762090458069
+        }
+      },
+      "slo_status": {
+        "ttft_p95_ok": true,
+        "tpot_mean_ok": true,
+        "overall": "✅"
+      },
+      "raw_file": "/data/user1/yy/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/raw_outputs/vllm_mtp_0708_32_4000_512.jsonl"
+    },
+    {
+      "name": "c32_i512_o256",
+      "config": {
+        "phase": "mtp",
+        "concurrency": 32,
+        "input_len": 512,
+        "output_len": 256,
+        "dataset": "random",
+        "num_prompts": 128
+      },
+      "metrics": {
+        "success": 128,
+        "failed": 0,
+        "duration_s": 10.193132108004647,
+        "request_throughput": 12.55747484126904,
+        "input_token_throughput": 3458.0146344144614,
+        "output_token_throughput": 1657.9791001363028,
+        "total_token_throughput": 5115.993734550764,
+        "total_input_tokens": 35248,
+        "total_output_tokens": 16900,
+        "e2e_ms": {
+          "mean": 2332.5013259607204,
+          "p50": 2112.3746865050634,
+          "p90": 4070.417974996962,
+          "p95": 4812.899109590944,
+          "p99": 5073.772265354783
+        },
+        "ttft_ms": {
+          "mean": 279.9133442812263,
+          "p50": 206.422848496004,
+          "p90": 663.5394259006716,
+          "p95": 761.6650494928763,
+          "p99": 766.2871305434965
+        },
+        "tpot_ms": {
+          "mean": 16.7286719523913,
+          "p50": 16.23081167053897,
+          "p90": 23.97215435499593,
+          "p95": 26.244743178375007,
+          "p99": 37.558756268073765
+        },
+        "itl_ms": {
+          "mean": 30.42237174675025,
+          "p50": 14.531328000884969,
+          "p90": 61.01081449742196,
+          "p95": 163.01492899947334,
+          "p99": 182.57800715073245
+        }
+      },
+      "slo_status": {
+        "ttft_p95_ok": true,
+        "tpot_mean_ok": true,
+        "overall": "✅"
+      },
+      "raw_file": "/data/user1/yy/experiments/dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/mtp/raw_outputs/vllm_mtp_0708_32_512_256.jsonl"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/run_bench.sh b/experiments/dsv4_h200_vllm_mtp_vs_default/run_bench.sh
new file mode 100755
index 0000000..7715bb3
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/run_bench.sh
@@ -0,0 +1,216 @@
+#!/usr/bin/env bash
+# vLLM+MTP speculative decoding vs vLLM default comparison on H200.
+# Reuses the default-side results from dsv4_h200_vllm_dspark_vs_default.
+set -Eeuo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+EXPERIMENT_NAME="$(basename "$SCRIPT_DIR")"
+
+# shellcheck source=/dev/null
+source "${SCRIPT_DIR}/../../scripts/common/lib.sh"
+# shellcheck source=/dev/null
+source "${SCRIPT_DIR}/../../scripts/common/platform.sh"
+# shellcheck source=/dev/null
+source "${SCRIPT_DIR}/config.env"
+
+RUN_ID="${RUN_ID:-$(date '+%Y%m%d-%H%M%S')}"
+RESULT_BASE="${SCRIPT_DIR}/results"
+
+log_dir_global="${RESULT_BASE}/${RUN_ID}/logs"
+mkdir -p "$log_dir_global"
+log_init "${log_dir_global}/orchestrator.log"
+
+log "experiment=${EXPERIMENT_NAME}"
+log "run_id=${RUN_ID}"
+log "platform=${PLATFORM}"
+log "hardware=${HARDWARE}"
+log "model=${MODEL_PATH}"
+log "spec_method=${SPEC_METHOD}"
+log "spec_tokens=${SPEC_TOKENS}"
+log "default_result_root=${DEFAULT_RESULT_ROOT}"
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+is_server_healthy() {
+  local port="$1"
+  curl --fail --silent --show-error --max-time 5 "http://127.0.0.1:${port}/health" >/dev/null 2>&1
+}
+
+stop_server() {
+  local pid_file="/data/user1/yy/${EXPERIMENT_NAME}_mtp.pid"
+  if [[ -f "$pid_file" ]]; then
+    local pid
+    pid="$(cat "$pid_file")"
+    if kill -0 "$pid" 2>/dev/null; then
+      log "stopping mtp server pid=${pid}"
+      kill "$pid" 2>/dev/null || true
+      sleep 5
+      kill -9 "$pid" 2>/dev/null || true
+    fi
+    rm -f "$pid_file"
+  fi
+  pkill -9 -f "vllm serve.*DeepSeek-V4-Flash" 2>/dev/null || true
+  sleep 2
+}
+
+start_server() {
+  log "starting mtp server with ${MTP_START_SCRIPT}"
+  bash "${MTP_START_SCRIPT}" >> "${log_dir_global}/mtp.server.outer.log" 2>&1
+
+  if ! is_server_healthy "$MTP_PORT"; then
+    log "error: mtp server failed to become healthy"
+    return 1
+  fi
+  log "mtp server is healthy on port ${MTP_PORT}"
+}
+
+run_warmup() {
+  log "warming up mtp (input=4000, output=512, num=2)"
+  "${VENV_CLIENT}/bin/python" "${SCRIPT_DIR}/../../scripts/common/warmup.py" \
+    --backend vllm \
+    --port "$MTP_PORT" \
+    --input-len 4000 \
+    --output-len 512 \
+    --num 2 \
+    --env-python "${VENV_CLIENT}/bin/python" \
+    >> "${log_dir_global}/mtp.warmup.log" 2>&1
+  log "warmup for mtp completed"
+}
+
+scenario_already_completed() {
+  local output_file="$1"
+  local expected="$2"
+  [[ -s "$output_file" ]] || return 1
+  local completed
+  completed="$(${VENV_CLIENT}/bin/python -c "
+import json, sys
+path = sys.argv[1]
+try:
+    with open(path, 'r', encoding='utf-8') as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                data = json.loads(line)
+                print(data.get('completed', 0))
+                break
+except Exception:
+    print(0)
+" "$output_file")"
+  [[ "${completed:-0}" -ge "$expected" ]]
+}
+
+run_benchmark() {
+  local result_root="${RESULT_BASE}/${RUN_ID}/mtp"
+  local raw_dir="${result_root}/raw_outputs"
+  local bench_log_dir="${result_root}/logs"
+
+  mkdir -p "$raw_dir" "$bench_log_dir"
+
+  log "===== mtp BENCHMARK START ====="
+
+  stop_server
+  start_server
+  run_warmup
+
+  for scenario in "${SCENARIOS[@]}"; do
+    read -r concurrency input_len output_len num_prompts <<< "$scenario"
+    output_file="${raw_dir}/vllm_mtp_$(date '+%m%d')_${concurrency}_${input_len}_${output_len}.jsonl"
+    detail_log="${bench_log_dir}/vllm_mtp_c${concurrency}_i${input_len}_o${output_len}.log"
+
+    if scenario_already_completed "$output_file" "$num_prompts"; then
+      log "skipping already-completed mtp scenario: c=${concurrency} i=${input_len} o=${output_len}"
+      continue
+    fi
+
+    log "running mtp scenario: c=${concurrency} i=${input_len} o=${output_len} n=${num_prompts}"
+
+    "${VENV_CLIENT}/bin/python" -m sglang.bench_serving \
+      --backend vllm \
+      --host 127.0.0.1 \
+      --port "$MTP_PORT" \
+      --dataset-name random \
+      --random-input-len "$input_len" \
+      --random-output-len "$output_len" \
+      --num-prompts "$num_prompts" \
+      --max-concurrency "$concurrency" \
+      --request-rate 10000 \
+      --output-file "$output_file" \
+      --output-details \
+      > "$detail_log" 2>&1 || {
+        log "ERROR: mtp scenario c=${concurrency} i=${input_len} o=${output_len} failed; see ${detail_log}"
+        continue
+      }
+
+    log "finished mtp scenario: output=${output_file}"
+  done
+
+  stop_server
+  log "===== mtp BENCHMARK DONE ====="
+}
+
+parse_mtp() {
+  local result_root="${RESULT_BASE}/${RUN_ID}/mtp"
+  log "parsing mtp results in ${result_root}"
+  "${VENV_CLIENT}/bin/python" "${SCRIPT_DIR}/../../scripts/common/parse_backend.py" "$result_root" --backend vllm \
+    >> "${result_root}/logs/parse.log" 2>&1 || {
+      log "WARNING: parser failed for mtp; see ${result_root}/logs/parse.log"
+    }
+}
+
+write_mtp_metadata() {
+  local result_root="${RESULT_BASE}/${RUN_ID}/mtp"
+  ensure_result_root "$result_root"
+  local meta_json="${result_root}/results.json"
+
+  local server_args
+  server_args="vllm serve $MODEL_PATH --trust-remote-code --tensor-parallel-size $TP --kv-cache-dtype fp8 --max-model-len $MAX_MODEL_LEN --max-num-seqs $MAX_NUM_SEQS --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --speculative-config {\\\"method\\\":\\\"mtp\\\",\\\"num_speculative_tokens\\\":$SPEC_TOKENS} --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port $MTP_PORT"
+
+  write_metadata_json \
+    "$meta_json" \
+    "${EXPERIMENT_NAME}_mtp" \
+    "$RUN_ID" \
+    "$MODEL_PATH" \
+    "vllm" \
+    "vllm-mtp" \
+    "$HARDWARE" \
+    "$ACCELERATOR" \
+    "$CHIP" \
+    "experiments/${EXPERIMENT_NAME}/run_bench.sh" \
+    "$VENV_VLLM" \
+    "H200 vLLM+MTP TP=8 benchmark for DeepSeek-V4-Flash"
+
+  jq --arg server_args "$server_args" \
+     '.config = {
+        "tp": 8,
+        "cuda_visible_devices": "0,1,2,3,4,5,6,7",
+        "max_model_len": 32768,
+        "max_num_seqs": 256,
+        "spec_method": "mtp",
+        "spec_tokens": 1,
+        "backend": "vllm",
+        "server_start_script": "experiments/dsv4_h200_vllm_mtp_vs_default/start_mtp.sh",
+        "server_args": $server_args
+      }' "$meta_json" > "${meta_json}.tmp" && mv "${meta_json}.tmp" "$meta_json"
+}
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+stop_server
+write_mtp_metadata
+run_benchmark
+parse_mtp
+
+log "generating comparison report"
+"${VENV_CLIENT}/bin/python" "${SCRIPT_DIR}/compare.py" \
+  --mtp "${RESULT_BASE}/${RUN_ID}/mtp" \
+  --default "$DEFAULT_RESULT_ROOT" \
+  --output "${RESULT_BASE}/${RUN_ID}/comparison.md" \
+  >> "${log_dir_global}/compare.log" 2>&1 || {
+    log "WARNING: comparison script failed; see ${log_dir_global}/compare.log"
+  }
+
+log "all results saved to ${RESULT_BASE}/${RUN_ID}"
diff --git a/experiments/dsv4_h200_vllm_mtp_vs_default/start_mtp.sh b/experiments/dsv4_h200_vllm_mtp_vs_default/start_mtp.sh
new file mode 100755
index 0000000..c41fab0
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_mtp_vs_default/start_mtp.sh
@@ -0,0 +1,66 @@
+#!/bin/bash
+# Start vLLM with MTP speculative decoding for the comparison experiment.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source=/dev/null
+source "${SCRIPT_DIR}/config.env"
+
+cd /data/user1/yy
+mkdir -p logs tmp
+
+VENV="${VENV_VLLM}"
+export PATH="$VENV/bin:$PATH"
+export PYTHONUNBUFFERED=1
+export TMPDIR=/data/user1/yy/tmp
+export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES}"
+
+LOG="/data/user1/yy/logs/dsv4_h200_vllm_mtp_vs_default_mtp_$(date +%Y%m%d_%H%M%S).log"
+PID_FILE="/data/user1/yy/dsv4_h200_vllm_mtp_vs_default_mtp.pid"
+
+rm -f "$PID_FILE"
+
+echo "=== Starting vLLM+MTP server (TP=$TP, spec_method=$SPEC_METHOD, spec_tokens=$SPEC_TOKENS) ==="
+echo "Model: $MODEL_PATH"
+echo "Port: $MTP_PORT"
+echo "Log: $LOG"
+
+nohup vllm serve "$MODEL_PATH" \
+    --trust-remote-code \
+    --tensor-parallel-size "$TP" \
+    --kv-cache-dtype fp8 \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --max-num-seqs "$MAX_NUM_SEQS" \
+    --block-size 256 \
+    --gpu-memory-utilization 0.90 \
+    --tokenizer-mode deepseek_v4 \
+    --reasoning-parser deepseek_v4 \
+    --speculative-config "{\"method\":\"mtp\",\"num_speculative_tokens\":$SPEC_TOKENS}" \
+    --no-disable-hybrid-kv-cache-manager \
+    --disable-uvicorn-access-log \
+    --port "$MTP_PORT" \
+    > "$LOG" 2>&1 &
+
+PID=$!
+echo $PID > "$PID_FILE"
+echo "PID: $PID"
+echo "Waiting for health..."
+
+for i in $(seq 1 240); do
+    if curl --fail --silent --show-error --max-time 5 "http://127.0.0.1:${MTP_PORT}/health" >/dev/null 2>&1; then
+        echo "vLLM+MTP server is ready at http://127.0.0.1:${MTP_PORT}"
+        echo "Log: $LOG"
+        exit 0
+    fi
+    if ! kill -0 $PID 2>/dev/null; then
+        echo "ERROR: vLLM+MTP server exited early"
+        tail -200 "$LOG"
+        exit 1
+    fi
+    echo "Waiting... ($i/240)"
+    sleep 5
+done
+
+echo "ERROR: vLLM+MTP server not healthy after 240 retries"
+tail -200 "$LOG"
+exit 1