docs: consolidate project docs (dedup, relocate, expand 910C client guide)

Project-level documentation was scattered and duplicated across README.md,
BENCHMARK_WORKFLOW.md, and docs/EXPERIMENT_GUIDE.md (directory layout +
scripts/common component table repeated 3x). Reorganize into a clear
single-source-of-truth structure.

Changes:
- README.md: drop the 6 stale changelog entries at the top (latest was
  07-21; history lives in git log). Replace the duplicated directory-
  layout + scripts/common sections with a one-line link to
  docs/EXPERIMENT_GUIDE.md. (151 -> 99 lines)
- BENCHMARK_WORKFLOW.md -> docs/BENCHMARK_WORKFLOW.md: relocate into docs/.
  Replace its duplicated Directory Layout and Quick Start/Adding sections
  with links to EXPERIMENT_GUIDE / README / NEW_PLATFORM_GUIDE; keep the
  unique parts (Rules, Naming Conventions, Final JSON Schema, Checklist).
  (394 -> 224 lines)
- docs/EXPERIMENT_GUIDE.md: now the single authority for directory layout
  + component table + experiment conventions. Add a cross-link from the
  results.json field list to BENCHMARK_WORKFLOW's full JSON Schema and
  Naming Conventions.
- docs/H200_QUICKSTART.md: deleted (outdated, repeatedly references
  removed legacy scripts; H200 usage is covered by ADAPTIVE_CONCURRENCY_USAGE
  and experiment READMEs).
- docs/DSV4_INFERENCE_COMPARISON_REPORT.md -> experiments/h200/
  dsv4_h200_vllm_mtp_vs_default/results/20260708-160349/: this is an
  experiment report, not a project doc; relocate next to its sibling
  report.md.
- envs/ASCEND_910C_ENV_SETUP.md §8: expand the vague "pip install sglang"
  note into a full sglang client image build guide -- pin sglang 0.5.2
  (not latest; >=0.5.16 deprecates bench_serving and breaks the parser),
  --no-deps minimal install loop, docker commit to a local image, with
  the exact commands used to build local/vllm-ascend:0.23-a3-dsv4-sglang.
- experiments/h200/dsv4_h200_vllm_tp2_custom_bench/README.md: fix the
  now-broken link to BENCHMARK_WORKFLOW.md (../../ -> ../../../docs/).
- .gitignore: ignore *.bak.glm52orig scratch backups.

Also includes the add16 adaptive_results produced by the dsv4 TP=4/DP=2
runs on 910c.1.

.gitignore

@@ -90,3 +90,6 @@ envs/*
 
 # Tooling artifacts
 skills-lock.json
+
+# 临时备份文件
+*.bak.glm52orig

README.md

@@ -1,32 +1,6 @@
 # sskj — 多平台大模型推理性能基准测试项目
 
-> **更新（2026-07-21 14:12:30 +0800）**
-> - RTX 6000D / DeepSeek-V4-Flash 的 SGLang 全量与 tiny 脚本切换到 `lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45`，MoE backend 保持 `auto`，由 nightly 自动选择可用的 FlashInfer MoE 实现；同时使用独立 nightly CUDA 13 cache，避免复用旧 SM120 修补镜像的编译缓存。
-
-> **更新（2026-07-21 14:00:30 +0800）**
-> - adaptive 并发搜索新增可配置的 `C=1 OOM` 配置级短路：同一 TP/DP 在最小并发仍 OOM 时，直接跳过该组更长输入，避免把已经确定不可部署的长度继续压测；6000D DSV4-Flash SGLang tiny 已启用，并支持从已有结果恢复时识别此前的该边界。
-
-> **更新（2026-07-21 11:54:30 +0800）**
-> - 新增 RTX 6000D / DeepSeek-V4-Flash tiny 自适应并发实验：输出固定 1K，输入遍历 1K、2K、4K、8K、16K、32K、64K、128K，并发采用 `C=16` 起、每次加 16、首点失败时按 `16 -> 8 -> 1` 回退。
-> - vLLM 覆盖 TP2/DP4、TP4/DP2、TP8/DP1，部署固定为 128K context、128 活跃请求、0.9 显存比例；SGLang 覆盖 TP4/DP2、TP8/DP1，固定为 128K、64、0.9，TP2/DP4 因 SM120 Marlin 权重加载 OOM 明确排除。
-
-> **更新（2026-07-20 17:55:20 +0800）**
-> - H20 SGLang TPxDP matrix 默认传入 `--cuda-graph-max-bs-decode 128`，使高并发 decode 在 batch 不超过 128 时持续使用 CUDA Graph；Docker 与 native 启动入口均已覆盖。
-
-> **更新（2026-07-20 17:25:40 +0800）**
-> - RTX 6000D 的 SGLang DSV4-Flash 默认传入 `--cuda-graph-max-bs-decode 16`，避免 SM120 sparse MLA 在大 batch CUDA Graph 预捕获时 OOM。该参数仅限制预捕获图的 decode batch 上限，不限制服务的 `max-running-requests` 或 benchmark 并发；其他平台默认不受影响。
-> - H20、RTX 6000D、P800 的配置不再硬编码 sskj 仓库绝对路径：仓库内的环境、数据集、缓存和 P800 patch 均通过共享的 `${ROOT_DIR}` 推导，仓库迁移后无需逐个修改路径。
-
-> **更新（2026-07-20，PR 自动化测试）**
-> - 仅用于验证 Gitee 受保护分支的自动 Pull Request 创建与评审流程；不修改实验代码、配置或结果口径。
-
-> **更新（2026-07-20，`903084c`）**
-> - RTX 6000D 的 SGLang DSV4-Flash 默认传入 `--cuda-graph-max-bs-decode 16`，避免 SM120 sparse MLA 在大 batch CUDA Graph 预捕获时 OOM。该参数仅限制预捕获图的 decode batch 上限，不限制服务的 `max-running-requests` 或 benchmark 并发；其他平台默认不受影响。
-> - H20 / RTX 6000D 的 TPxDP matrix baseline 不再显式限制模型上下文或服务内并发：vLLM 使用默认 `max-model-len` / `max-num-seqs`，SGLang 使用默认 `context-length` / `max-running-requests`。
-> - SGLang 固定使用 `--moe-runner-backend marlin`；vLLM 保持框架默认 MoE backend。
-> - `run_adaptive_concurrency_add16.sh` 在首点 OOM 时会重启服务并按 `C=16 -> 8 -> 1` 回退；仅 H20 / RTX 6000D matrix 默认开启，H200 / P800 保持原行为。
-> - RTX 6000D 旧 SGLang 结果使用 262K context cap，超过该范围或 C>32 的点不可与新 baseline 混用；请使用新的 `RUN_ID` 重跑受影响场景。
-
+> 历史更新见 `git log`。项目目的与工作流见下方。
 > **项目目的**：当新显卡（GPU/NPU）到货时，用最短时间完成大模型在该平台上的推理性能评估与部署配置选型。
 > 当前模型：DeepSeek-V4-Flash（FP8 / INT8）；后续接入 GLM5.2，**完全复用**本项目的实验与报告流程。
 > 新平台接入 SOP：[`docs/NEW_PLATFORM_GUIDE.md`](docs/NEW_PLATFORM_GUIDE.md)。
@@ -55,34 +29,7 @@
 
 ## 目录结构
 
-```
-platforms/                # 平台配置（每平台一个 .env）+ 平台补丁（patches/）
-scripts/common/           # 跨实验复用的编排组件（见下）
-experiments/              # 实验目录（三层结构）
-├── TEMPLATE/             # 老式固定场景对比实验模板（sglang vs vllm）
-├── h20/                  # NVIDIA H20 96GB
-├── h200/                 # NVIDIA H200 143GB
-├── p800/                 # Kunlun P800 XPU
-└── pro6000/              # NVIDIA RTX 6000D
-docs/                     # 规范与指南（SLO、实验规范、新卡接入）
-envs/                     # 环境搭建文档（具体 venv 不提交，见 .gitignore）
-datasets/                 # benchmark 数据集（不提交）
-BENCHMARK_WORKFLOW.md     # 结果目录与命名规范、results.json schema
-```
-
-### scripts/common/ 组件
-
-| 文件 | 职责 |
-|---|---|
-| `lib.sh` | 日志、health check、metadata JSON 生成 |
-| `platform.sh` | 平台自动探测并加载 `platforms/<chip>.env` |
-| `adaptive_bench_lib.sh` | **自适应并发搜索核心库**（引擎无关，靠 5 个 `engine_*` 回调注入平台逻辑） |
-| `adaptive_concurrency.py` | 上者的 Python 助手（shapes 生成、结果解析、汇总） |
-| `parse_backend.py` | 固定并发实验：raw jsonl → results.json + report.md |
-| `compare.py` | SGLang vs vLLM 横向对比表（老式实验用） |
-| `warmup.py` | 服务预热 |
-| `server_docker.sh` / `bench_client_docker.sh` | P800 专用 Docker 服务端/客户端生命周期 |
-| `adaptive_heartbeat.sh` | 长跑实验的心跳监控（手动独立运行） |
+仓库的目录布局、`scripts/common/` 组件职责、结果命名规范与 `results.json` schema，统一见 [`docs/EXPERIMENT_GUIDE.md`](docs/EXPERIMENT_GUIDE.md)（单一权威来源）与 [`docs/BENCHMARK_WORKFLOW.md`](docs/BENCHMARK_WORKFLOW.md)。
 
 ## 实验索引
 

docs/BENCHMARK_WORKFLOW.md

@@ -2,37 +2,7 @@
 
 ## Directory Layout
 
-```
-<repo-root>/                  # 本仓库根目录（如 /data1/yy/sskj，按机器调整）
-├── platforms/                # chip/accelerator platform configs
-│   ├── kunlun_p800.env
-│   ├── nvidia_h20.env
-│   ├── nvidia_h200.env
-│   ├── nvidia_rtx6000d.env
-│   └── patches/kunlun_p800/  # runtime patches required by some images
-├── scripts/
-│   └── common/               # reusable components (lib.sh, platform.sh, parse_backend.py, adaptive_bench_lib.sh, ...)
-├── docs/
-│   └── SLO_STANDARDS.md      # 推理服务 SLO 标准
-├── experiments/              # experiment-centric directories (preferred): experiments/<platform>/<experiment>/
-│   ├── TEMPLATE/             # template for fixed-scenario sglang-vs-vllm experiments
-│   └── <platform>/           # h20 / h200 / p800 / pro6000
-│       └── dsv4_p800_sglang/
-│           ├── README.md
-│           ├── config.env    # experiment-level configuration
-│           ├── start_server.sh
-│           ├── run_bench.sh
-│           ├── parse_results.py
-│           └── results/
-│               └── 20260708-XXXXXX/
-│                   ├── report.md
-│                   ├── results.json
-│                   └── logs/
-├── datasets/                 # benchmark datasets
-└── envs/                     # Python virtual environments
-```
-
-注：legacy 的 `scripts/benchmark_dspark_0707/` 已删除；顶层 `bench_results/` 与 `logs/` 已不在仓库中（历史结果已归档迁移至仓库外）。
+仓库目录结构、`scripts/common/` 组件职责见 [`EXPERIMENT_GUIDE.md`](EXPERIMENT_GUIDE.md) §1（单一权威来源，避免多处维护漂移）。
 
 ## Rules
 
@@ -230,152 +200,12 @@ The JSON file inside each experiment run directory (`experiments/<platform>/<exp
   - Keep at least one of `accelerator` or `chip`, and at least one of `engine` or `backend`, populated in every run.
 - If a metric is not applicable (e.g. `accept_length` for non-speculative decoding), set it to `null` rather than omitting the key.
 
-## Quick Start
+## Quick Start / Adding Experiments
 
-### Run P800 SGLang benchmark (Kunlun P800, Docker)
-
-```bash
-bash experiments/p800/dsv4_p800_sglang/run_bench.sh
-```
-
-### Run H200 DSpark benchmark (NVIDIA H200, native venv)
-
-```bash
-bash experiments/h200/dsv4_h200_dspark/run_bench.sh
-```
-
-Note: the wrapper's default `LEGACY_GRID_SCRIPT` / `SERVER_START_SCRIPT` point at the removed legacy scripts; override them before running (see `docs/H200_QUICKSTART.md`).
-
-### Legacy DSpark scripts (removed)
-
-The legacy suite `scripts/benchmark_dspark_0707/` (grid benchmark, spec-tokens comparison, etc.) has been removed from this repo.
-
-### Parse results
-
-```bash
-# Experiment-centric layout
-python3 experiments/p800/dsv4_p800_sglang/parse_results.py \
-  experiments/p800/dsv4_p800_sglang/results/<run_id>
-```
-
-## Adding a New Platform or Experiment
-
-### Quick Start: New Experiment on an Existing Platform (e.g. a new P800 box)
-
-Use this when the platform already exists (`platforms/<chip>.env` present) and you only want to test a new model/config on a machine.
-
-**0. Machine setup (one-time, only on a brand-new machine)**
-
-```bash
-git clone https://git.meta-stone.net/qqtang/sskj.git /data1/yy/sskj
-cd /data1/yy/sskj
-git config credential.helper store          # push once to save the token
-# Verify platform auto-detection:
-bash -c 'source scripts/common/platform.sh && echo "$PLATFORM | $HARDWARE | $ENGINE"'
-# Ensure platform deps exist (see platforms/<chip>.env): docker image pulled, MODEL_ROOT has the model, datasets/ present.
-```
-
-If the platform is not auto-detected, run with `PLATFORM=<chip> ...`. Platform-wide paths/images live in `platforms/<chip>.env` - edit that file (not the experiment) if a path differs on this machine.
-
-**1. Create the experiment from TEMPLATE**
-
-```bash
-cd /data1/yy/sskj
-cp -r experiments/TEMPLATE experiments/p800/<your_experiment_name>   # e.g. qwen3_8b_p800_sglang_tp4
-cd experiments/p800/<your_experiment_name>
-```
-
-**2. Edit `config.env`** - the single source of experiment params (model, TP, ports, scenarios). All values use `${VAR:-default}`, so they can be overridden by env vars at run time.
-
-**3. Edit `start_sglang.sh` / `start_server.sh`** - the deployment script (`docker run` + launch flags). This is where you change deployment params (TP, mem-fraction, attention backend, cuda graph, ...). Reference `config.env` variables (`$TP`, `$MEM_FRACTION_STATIC`, ...) rather than hardcoding.
-
-**4. Run**
-
-```bash
-bash run_bench.sh                                  # auto-detects platform
-# or with overrides / explicit RUN_ID:
-PLATFORM=kunlun_p800 RUN_ID=tp4_20260722 bash run_bench.sh
-# quick config probe without editing files:
-TP=4 MEM_FRACTION_STATIC=0.85 MAX_RUNNING=32 bash run_bench.sh
-```
-
-Results land in `results/<RUN_ID>/` (`report.md`, `results.json`; `raw_outputs/` and `logs/` are gitignored). Watch `logs/orchestrator.log` and `logs/server.log`.
-
-**5. Commit code + final artifacts, push, open PR**
-
-```bash
-cd /data1/yy/sskj
-git checkout -b feat/p800-<your_experiment_name>
-git add experiments/p800/<your_experiment_name>/config.env \
-        experiments/p800/<your_experiment_name>/start_sglang.sh \
-        experiments/p800/<your_experiment_name>/run_bench.sh \
-        experiments/p800/<your_experiment_name>/README.md \
-        experiments/p800/<your_experiment_name>/results/<RUN_ID>/report.md \
-        experiments/p800/<your_experiment_name>/results/<RUN_ID>/results.json
-git commit -m "feat(p800): add <your_experiment_name> experiment"
-git push -u origin feat/p800-<your_experiment_name>
-```
-
-Do NOT `git add` `raw_outputs/` or `logs/` (already ignored). Run through the [Checklist](#checklist-before-committing--archiving) below before pushing.
-
-> Adding a brand-new chip (not just a new experiment on an existing platform)? Follow steps 1-4 below instead.
-
-
-### 1. Add or update a platform config
-
-Create `platforms/<chip>.env` with identity and platform-wide paths:
-
-```bash
-CHIP="my_chip"
-ACCELERATOR="My Accelerator"
-HARDWARE="8x My Accelerator"
-ENGINE="vllm-myengine"
-DEFAULT_PORT="30000"
-MODEL_ROOT="/data/models"
-```
-
-For Docker-based platforms, also set `DOCKER_IMAGE`, `CONTAINER_NAME`,
-`CONTAINER_PYTHON`, and `PATCH_ROOT` (see `platforms/kunlun_p800.env`).
-For host-native platforms, set the relevant venv paths (see
-`platforms/nvidia_h200.env`).
-
-### 2. Create an experiment directory
-
-```
-experiments/<platform>/<name>/
-├── README.md          # Purpose and usage
-├── config.env         # Model, port, scenarios, engine overrides
-├── start_server.sh    # (optional) platform-specific server launch
-├── run_bench.sh       # Orchestrator
-└── parse_results.py   # Convert raw outputs to results.json + report.md
-```
-
-At minimum, `run_bench.sh` should:
-
-1. Source `scripts/common/lib.sh` and `scripts/common/platform.sh`.
-2. Read `config.env`.
-3. Create `experiments/<platform>/<name>/results/<RUN_ID>/`.
-4. Call `write_metadata_json` to create `results.json`.
-5. Record the exact server launch command/args in `results.json` `config.server_args` (or `phaseN_server_args`).
-6. Run the benchmark scenarios.
-7. Call `parse_results.py` to generate `report.md`.
-
-### 3. Reuse shared helpers
-
-- `scripts/common/lib.sh`: logging, health checks, metadata JSON.
-- `scripts/common/platform.sh`: platform auto-detection and env loading.
-- `scripts/common/warmup.py`: server 预热。
-- `scripts/common/parse_backend.py`: raw jsonl -> `results.json` + `report.md`。
-- `scripts/common/compare.py`: SGLang vs vLLM 横向对比表。
-
-Docker/XPU 平台可以在 `scripts/common/` 下新增自己的 helper，但不要在实验目录复制通用逻辑。
-
-### 4. Example experiments
-
-- Docker / XPU: `experiments/p800/dsv4_p800_sglang/`
-- Native / H200: `experiments/h200/dsv4_h200_dspark/`
-
-See `docs/H200_QUICKSTART.md` for a concrete H200 porting walkthrough.
+快速复现、新增实验、新平台接入的步骤见：
+- [`../README.md`](../README.md) §快速复现（命令示例）
+- [`EXPERIMENT_GUIDE.md`](EXPERIMENT_GUIDE.md) §2 新增实验（config.env 必备字段）
+- [`NEW_PLATFORM_GUIDE.md`](NEW_PLATFORM_GUIDE.md)（新平台接入 SOP）
 
 ## Checklist Before Committing / Archiving
 

docs/EXPERIMENT_GUIDE.md

@@ -93,6 +93,8 @@ VLLM_START_SCRIPT="${SCRIPT_DIR:-.}/start_vllm.sh"
 
 这些通过 `scripts/common/lib.sh` 中的 `write_metadata_json` 和 `jq` 注入。
 
+> 完整的 `results.json` schema（含 scenarios/latencies/slo_status 等字段）见 [`BENCHMARK_WORKFLOW.md`](BENCHMARK_WORKFLOW.md) §Final JSON Schema；结果目录命名规范见同文 §Naming Conventions。
+
 ## 3. 平台与硬件适配
 
 ### 3.1 芯片级默认配置

docs/H200_QUICKSTART.md

@@ -1,99 +0,0 @@
-# H200 Quick Start Guide
-
-This guide covers how to use this benchmark repository on an NVIDIA H200 machine.
-
-## Current state
-
-The older H200 experiments (e.g. `experiments/h200/dsv4_h200_dspark/`) are based on **native host virtual environments** (not Docker):
-
-- Server engine: `vllm-dspark`
-- Server env: `envs/vllm-dspark`
-- Benchmark client env: `envs/sglang` (uses `sglang.bench_serving --backend vllm`)
-- Default model: `/data/models/DeepSeek-V4-Flash-DSpark`
-- Default port: `30004`
-
-Note: the current mainstream H200 experiments — the TP/DP matrix adaptive-concurrency suites `experiments/h200/dsv4_h200_vllm_tp_dp_matrix/` and `experiments/h200/dsv4_h200_sglang_tp_dp_matrix/` — are Docker-based (`USE_DOCKER=1` for the server, benchmark client via `DOCKER_CLIENT_IMAGE`); see `experiments/ADAPTIVE_CONCURRENCY_USAGE.md`.
-
-The legacy benchmark suite `scripts/benchmark_dspark_0707/` has been removed from this repo. The migration wrapper experiment `experiments/h200/dsv4_h200_dspark/` still exists and produces the `experiments/h200/<name>/results/<RUN_ID>/` layout, but its default `LEGACY_GRID_SCRIPT` / `SERVER_START_SCRIPT` point at the removed scripts and must be overridden to run.
-
-## 1. Pull and verify
-
-```bash
-git clone <repo-url> sskj   # clone path: adjust to your machine
-cd sskj
-
-# Platform should auto-detect as nvidia_h200
-source scripts/common/platform.sh
-```
-
-If auto-detection fails, set it explicitly:
-
-```bash
-PLATFORM=nvidia_h200 source scripts/common/platform.sh
-```
-
-## 2. Legacy DSpark benchmark grid (removed)
-
-The legacy suite `scripts/benchmark_dspark_0707/` has been removed from this repo, along with its `bench_results/dspark_grid_<RUN_ID>/` output location (historical results were archived elsewhere). Use the wrapper experiment below, or the current tp_dp matrix experiments (`experiments/h200/dsv4_h200_{vllm,sglang}_tp_dp_matrix/`), instead.
-
-## 3. Run the migration wrapper experiment
-
-This produces results in the experiment-centric layout:
-
-```bash
-bash experiments/h200/dsv4_h200_dspark/run_bench.sh
-```
-
-Results land in `experiments/h200/dsv4_h200_dspark/results/<RUN_ID>/`.
-
-Caveat: the wrapper's default `LEGACY_GRID_SCRIPT` and `SERVER_START_SCRIPT` refer to the removed `scripts/benchmark_dspark_0707/` suite and `scripts/start_dsv4_dspark_8card.sh`; set both explicitly (or restore equivalent scripts) before running.
-
-## 4. Create a new H200 experiment
-
-To add a new H200 benchmark (for example a different model or engine), create:
-
-```
-experiments/h200/<your_name>/
-├── README.md          # What this experiment measures
-├── config.env         # Model, port, scenarios, venv paths
-├── start_server.sh    # (optional) native server launch
-├── run_bench.sh       # Orchestrator: server → benchmark → stop
-└── parse_results.py   # Generate results.json + report.md
-```
-
-Minimum `config.env`:
-
-```bash
-EXPERIMENT="${EXPERIMENT:-<your_name>}"
-MODEL_NAME="${MODEL_NAME:-DeepSeek-V4-Flash-DSpark}"
-MODEL_PATH="${MODEL_PATH:-/data/models/DeepSeek-V4-Flash-DSpark}"
-PORT="${PORT:-30004}"
-BACKEND="${BACKEND:-vllm}"
-ENGINE="${ENGINE:-vllm-dspark}"
-
-# Native venvs (paths are per-machine; adjust to yours)
-VENV_SERVER="${VENV_SERVER:-/path/to/envs/vllm-dspark}"
-VENV_CLIENT="${VENV_CLIENT:-/path/to/envs/sglang}"
-
-SCENARIOS=(
-  "32 512 256"
-)
-```
-
-Then source the platform loader and shared helpers at the top of `run_bench.sh`:
-
-```bash
-source "${SCRIPT_DIR}/../../../scripts/common/lib.sh"
-source "${SCRIPT_DIR}/../../../scripts/common/platform.sh"
-```
-
-For a concrete example, see `experiments/h200/dsv4_h200_dspark/`. For the current mainstream form (TP/DP matrix + adaptive concurrency search, Docker-based), copy `experiments/h200/dsv4_h200_vllm_tp_dp_matrix/` instead; usage in `experiments/ADAPTIVE_CONCURRENCY_USAGE.md`.
-
-## 5. Cross-platform comparison
-
-跨平台对比脚本目前未统一提供。可分别读取各实验 `results/<run_id>/results.json` 中的结构化数据，按 scenario 聚合后生成对比表。
-
-## Notes
-
-- `platforms/nvidia_h200.env` still carries venv defaults under `/data/user1/yy` (paths from another machine). Override `VENV_VLLM_DSPARK`, `VENV_SGLANG`, `MODEL_ROOT`, or `SERVER_START_SCRIPT` to match your machine.
-- Native server management helpers are not as mature as the Docker helpers in `scripts/common/server_docker.sh`. The wrapper's default `SERVER_START_SCRIPT` (`scripts/start_dsv4_dspark_8card.sh`) no longer exists in this repo and must be overridden.

envs/ASCEND_910C_ENV_SETUP.md

@@ -127,18 +127,63 @@ bash run_adaptive_concurrency_add16.sh
 
 ## 8. Benchmark Client 说明 ⚠️
 
-sskj 的压测客户端是 `sglang.bench_serving`，但 **vllm-ascend 官方镜像不含 sglang**。910C 上有两个选择：
+sskj 的压测客户端是 `sglang.bench_serving`，但 **vllm-ascend 官方镜像不含 sglang**。两个选择：
 
-1. **(推荐) 容器内装 sglang**：进容器 `pip install sglang`（或 `sglang[all]`），之后 `docker exec` 跑客户端。需要把 sglang 装进镜像或每次启动后手动装。
-2. **外部 sglang 镜像**：设 `USE_DOCKER_CLIENT=1`，提供 `DOCKER_CLIENT_IMAGE=lmsysorg/sglang:xxx`，用独立容器通过 host 网络打 vLLM 的 OpenAI API。但 sglang 官方镜像多为 x86 + CUDA，aarch64 NPU 节点上可能拉不到对应架构镜像。
+1. **(推荐) 构建带 sglang 的 vllm-ascend 镜像**：见 §8.1，一次性构建后所有实验复用。
+2. **外部 sglang 镜像**：设 `USE_DOCKER_CLIENT=1`，提供 `DOCKER_CLIENT_IMAGE=lmsysorg/sglang:xxx`，用独立容器通过 host 网络打 vLLM 的 OpenAI API。但 sglang 官方镜像多为 x86 + CUDA，aarch64 NPU 节点上拉不到对应架构镜像，不推荐。
 
-建议冒烟前先确认客户端方案，否则 adaptive 搜索会在 `engine_run_bench` 阶段失败。验证命令：
+建议冒烟前先确认客户端方案，否则 adaptive 搜索会在 `engine_run_bench` 阶段失败。
+
+### 8.1 构建带 sglang 客户端的 vllm-ascend 镜像
+
+> **版本要求**：必须装 **sglang 0.5.2**，不是最新版。
+> - sglang ≥ 0.5.16 已废弃 `sglang.bench_serving`（改为 `sglang.benchmark.serving`），输出格式变了，与 `scripts/common/adaptive_concurrency.py` / `parse_backend.py` 的解析逻辑不兼容。
+> - glm52 实验的解析器修复（commit 98cdb67）针对的就是 0.5.2 的输出格式。dsv4/glm52 实验的 `config.env` 默认镜像 `local/vllm-ascend:0.23-a3-*-sglang` 即基于 0.5.2 构建。
+
+**关键原则**：用 `pip install --no-deps` 只装 sglang 本体 + bench_serving 的轻量依赖，**不要装完整 sglang**（会拉 torch/transformers 等重依赖，破坏容器内 vllm 环境）。
+
+构建步骤（在 910C 节点上，需 sudo docker）：
 
 ```bash
-# 进容器看是否有 sglang
-docker run --rm <vllm-ascend-image> python -c "import sglang; print(sglang.__version__)"
+# 1. 启动一个临时容器（基础镜像 = 实验用的 vllm-ascend 镜像）
+BASE=quay.io/ascend/vllm-ascend:v0.23.0rc1-a3-openeuler
+sudo docker rm -f sglang-build 2>/dev/null
+sudo docker run -d --name sglang-build \
+    --device /dev/davinci0 --device /dev/davinci_manager \
+    --device /dev/devmm_svm --device /dev/hisi_hdc \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
+    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
+    -v /etc/ascend_install.info:/etc/ascend_install.info \
+    $BASE sleep infinity
+
+# 2. 装 sglang 0.5.2 本体（--no-deps，不碰 vllm 环境）
+sudo docker exec sglang-build pip install --no-deps "sglang==0.5.2"
+
+# 3. 循环补齐 bench_serving 缺失的轻量依赖（自动检测 ModuleNotFoundError 并安装）
+#    会装约 8 个包：ipython/traitlets/stack_data/executing/asttokens/pure_eval/prompt_toolkit/wcwidth
+#    （都是 IPython 依赖链，bench_serving 通过 sglang.utils 间接引入）
+sudo docker exec sglang-build bash -c '
+for i in $(seq 1 50); do
+  OUT=$(python -c "import sglang.bench_serving" 2>&1)
+  [ -z "$OUT" ] && { echo "IMPORT_OK"; break; }
+  MOD=$(echo "$OUT" | grep -oE "No module name .[a-zA-Z0-9_]+." | head -1 | sed "s/No module name //;s/.//")
+  [ -z "$MOD" ] && { echo "NON_MODULE_ERR: $OUT"; break; }
+  echo "iter $i: $MOD"
+  pip install --no-deps "$MOD" 2>&1 | grep -i Successfully
+done
+'
+
+# 4. 验证
+sudo docker exec sglang-build python -m sglang.bench_serving --help | head -3
+
+# 5. commit 成新镜像（tag 按实验命名，如 dsv4 / glm52）
+sudo docker commit sglang-build local/vllm-ascend:0.23-a3-dsv4-sglang
+sudo docker rm -f sglang-build
 ```
 
+构建完成后，把 `config.env` 的 `DOCKER_IMAGE` 指向新镜像即可。容器内 Python 路径是 `/usr/local/python3.12.13/bin/python3`（不是 `/usr/local/bin/python`，后者不存在），`config.env` 的 `CONTAINER_PYTHON` 已设为此值。
+
 ## 9. NPU 监控
 
 公共库 `adaptive_bench_lib.sh` 的 GPU 监控写死 `nvidia-smi`，910C 实验脚本已用 `npu-smi info` 重写 `adaptive_start_gpu_monitor` / `start_gpu_monitor`，输出与 nvidia-smi 相同的 CSV 列（timestamp, index, memory.used, memory.total, utilization.gpu），下游 `parse_backend.py` 无需改动。

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260728-165926/adaptive_summary.md

@@ -0,0 +1,6 @@
+# Adaptive concurrency search summary
+
+| Engine | TP | DP | ISL | OSL | Stop | Saturation C | Best TPS C | Best Total TPS | Max successful C |
+|---|---:|---:|---:|---:|---|---:|---:|---:|---:|
+
+`Saturation C` is the first point in the final low-gain streak. `Best TPS C` is the tested point with the highest observed Total TPS.

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260728-165926/run_manifest.json

@@ -0,0 +1,27 @@
+{
+  "experiment": "dsv4_910c_vllm_tp_dp_matrix",
+  "engine": "vllm",
+  "run_id": "adaptive_20260728-165926",
+  "model": "/mnt/models/DeepSeek-V4-Flash-w8a8-mtp",
+  "hardware": "8x Ascend910 (16 dies, 64GB HBM/die)",
+  "matrix": "/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/matrix.json",
+  "dataset": "random",
+  "tokenize_prompt": false,
+  "random_range_ratio": 1.0,
+  "search": {
+    "start_concurrency": 16,
+    "max_concurrency": 256,
+    "multiplier": 2,
+    "initial_backoff_concurrencies": [
+      8,
+      1
+    ],
+    "num_prompts_multiplier": 5,
+    "min_tps_gain_pct": 2.0,
+    "plateau_patience": 2,
+    "warmup_max_requests": 0,
+    "ttft_slo_ms": 4000.0,
+    "enable_ttft_slo_stop": 1,
+    "ttft_group_skip_ms": 8000.0
+  }
+}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260728-165926/shapes.tsv

@@ -0,0 +1,31 @@
+mark	input_len	output_len
+Y	1024	128
+Y	1024	256
+Y	1024	512
+Y	1024	1024
+Y	1024	2048
+Y	1024	4096
+Y	4096	128
+Y	4096	256
+Y	4096	512
+Y	4096	1024
+Y	4096	2048
+Y	4096	4096
+Y	8192	128
+Y	8192	256
+Y	8192	512
+Y	8192	1024
+Y	8192	2048
+Y	8192	4096
+Y	16384	128
+Y	16384	256
+Y	16384	512
+Y	16384	1024
+Y	16384	2048
+Y	32768	128
+Y	32768	256
+Y	32768	512
+Y	32768	1024
+Y	65536	128
+Y	65536	256
+Y	131072	128

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-110354/adaptive_summary.md

@@ -0,0 +1,6 @@
+# Adaptive concurrency search summary
+
+| Engine | TP | DP | ISL | OSL | Stop | Saturation C | Best TPS C | Best Total TPS | Max successful C |
+|---|---:|---:|---:|---:|---|---:|---:|---:|---:|
+
+`Saturation C` is the first point in the final low-gain streak. `Best TPS C` is the tested point with the highest observed Total TPS.

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-110354/run_manifest.json

@@ -0,0 +1,27 @@
+{
+  "experiment": "dsv4_910c_vllm_tp_dp_matrix",
+  "engine": "vllm",
+  "run_id": "adaptive_20260729-110354",
+  "model": "/mnt/models/DeepSeek-V4-Flash-w8a8-mtp",
+  "hardware": "8x Ascend910 (16 dies, 64GB HBM/die)",
+  "matrix": "/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/matrix.json",
+  "dataset": "random",
+  "tokenize_prompt": false,
+  "random_range_ratio": 1.0,
+  "search": {
+    "start_concurrency": 16,
+    "max_concurrency": 256,
+    "multiplier": 2,
+    "initial_backoff_concurrencies": [
+      8,
+      1
+    ],
+    "num_prompts_multiplier": 5,
+    "min_tps_gain_pct": 2.0,
+    "plateau_patience": 2,
+    "warmup_max_requests": 0,
+    "ttft_slo_ms": 4000.0,
+    "enable_ttft_slo_stop": 1,
+    "ttft_group_skip_ms": 8000.0
+  }
+}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-110354/shapes.tsv

@@ -0,0 +1,31 @@
+mark	input_len	output_len
+Y	1024	128
+Y	1024	256
+Y	1024	512
+Y	1024	1024
+Y	1024	2048
+Y	1024	4096
+Y	4096	128
+Y	4096	256
+Y	4096	512
+Y	4096	1024
+Y	4096	2048
+Y	4096	4096
+Y	8192	128
+Y	8192	256
+Y	8192	512
+Y	8192	1024
+Y	8192	2048
+Y	8192	4096
+Y	16384	128
+Y	16384	256
+Y	16384	512
+Y	16384	1024
+Y	16384	2048
+Y	32768	128
+Y	32768	256
+Y	32768	512
+Y	32768	1024
+Y	65536	128
+Y	65536	256
+Y	131072	128

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-112009/adaptive_summary.md

@@ -0,0 +1,6 @@
+# Adaptive concurrency search summary
+
+| Engine | TP | DP | ISL | OSL | Stop | Saturation C | Best TPS C | Best Total TPS | Max successful C |
+|---|---:|---:|---:|---:|---|---:|---:|---:|---:|
+
+`Saturation C` is the first point in the final low-gain streak. `Best TPS C` is the tested point with the highest observed Total TPS.

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-112009/run_manifest.json

@@ -0,0 +1,27 @@
+{
+  "experiment": "dsv4_910c_vllm_tp_dp_matrix",
+  "engine": "vllm",
+  "run_id": "adaptive_20260729-112009",
+  "model": "/mnt/models/DeepSeek-V4-Flash-w8a8-mtp",
+  "hardware": "8x Ascend910 (16 dies, 64GB HBM/die)",
+  "matrix": "/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/matrix.json",
+  "dataset": "random",
+  "tokenize_prompt": false,
+  "random_range_ratio": 1.0,
+  "search": {
+    "start_concurrency": 16,
+    "max_concurrency": 256,
+    "multiplier": 2,
+    "initial_backoff_concurrencies": [
+      8,
+      1
+    ],
+    "num_prompts_multiplier": 5,
+    "min_tps_gain_pct": 2.0,
+    "plateau_patience": 2,
+    "warmup_max_requests": 0,
+    "ttft_slo_ms": 4000.0,
+    "enable_ttft_slo_stop": 1,
+    "ttft_group_skip_ms": 8000.0
+  }
+}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-112009/shapes.tsv

@@ -0,0 +1,31 @@
+mark	input_len	output_len
+Y	1024	128
+Y	1024	256
+Y	1024	512
+Y	1024	1024
+Y	1024	2048
+Y	1024	4096
+Y	4096	128
+Y	4096	256
+Y	4096	512
+Y	4096	1024
+Y	4096	2048
+Y	4096	4096
+Y	8192	128
+Y	8192	256
+Y	8192	512
+Y	8192	1024
+Y	8192	2048
+Y	8192	4096
+Y	16384	128
+Y	16384	256
+Y	16384	512
+Y	16384	1024
+Y	16384	2048
+Y	32768	128
+Y	32768	256
+Y	32768	512
+Y	32768	1024
+Y	65536	128
+Y	65536	256
+Y	131072	128

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113439/adaptive_summary.md

@@ -0,0 +1,6 @@
+# Adaptive concurrency search summary
+
+| Engine | TP | DP | ISL | OSL | Stop | Saturation C | Best TPS C | Best Total TPS | Max successful C |
+|---|---:|---:|---:|---:|---|---:|---:|---:|---:|
+
+`Saturation C` is the first point in the final low-gain streak. `Best TPS C` is the tested point with the highest observed Total TPS.

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113439/run_manifest.json

@@ -0,0 +1,27 @@
+{
+  "experiment": "dsv4_910c_vllm_tp_dp_matrix",
+  "engine": "vllm",
+  "run_id": "adaptive_20260729-113439",
+  "model": "/mnt/models/DeepSeek-V4-Flash-w8a8-mtp",
+  "hardware": "8x Ascend910 (16 dies, 64GB HBM/die)",
+  "matrix": "/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/matrix.json",
+  "dataset": "random",
+  "tokenize_prompt": false,
+  "random_range_ratio": 1.0,
+  "search": {
+    "start_concurrency": 16,
+    "max_concurrency": 256,
+    "multiplier": 2,
+    "initial_backoff_concurrencies": [
+      8,
+      1
+    ],
+    "num_prompts_multiplier": 5,
+    "min_tps_gain_pct": 2.0,
+    "plateau_patience": 2,
+    "warmup_max_requests": 0,
+    "ttft_slo_ms": 4000.0,
+    "enable_ttft_slo_stop": 1,
+    "ttft_group_skip_ms": 8000.0
+  }
+}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113439/shapes.tsv

@@ -0,0 +1,31 @@
+mark	input_len	output_len
+Y	1024	128
+Y	1024	256
+Y	1024	512
+Y	1024	1024
+Y	1024	2048
+Y	1024	4096
+Y	4096	128
+Y	4096	256
+Y	4096	512
+Y	4096	1024
+Y	4096	2048
+Y	4096	4096
+Y	8192	128
+Y	8192	256
+Y	8192	512
+Y	8192	1024
+Y	8192	2048
+Y	8192	4096
+Y	16384	128
+Y	16384	256
+Y	16384	512
+Y	16384	1024
+Y	16384	2048
+Y	32768	128
+Y	32768	256
+Y	32768	512
+Y	32768	1024
+Y	65536	128
+Y	65536	256
+Y	131072	128

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/adaptive_points.jsonl

@@ -0,0 +1,3 @@
+{"status":"COMPLETED","completed":80,"failed":0,"duration_s":38.5779427299276,"request_tps":2.073723851996349,"input_tps":2123.493224444261,"output_tps":265.43665305553264,"total_tps":2388.9298774997937,"mean_input_tokens":1024.0,"mean_output_tokens":128.0,"ttft_p50_ms":1218.7625300139189,"ttft_p95_ms":3272.40676897927,"ttft_p99_ms":4362.865626341663,"tpot_p50_ms":49.4835413781574,"tpot_p95_ms":124.33025477457703,"tpot_p99_ms":79.74750237895067,"e2e_p50_ms":7354.705074976664,"e2e_p95_ms":9499.298503511818,"e2e_p99_ms":11670.241948680481,"itl_p50_ms":49.363330006599426,"itl_p95_ms":408.277097926474,"itl_p99_ms":448.8152776751667,"validation_errors":[],"timestamp":"2026-07-29T11:48:16+08:00","engine":"vllm","tp":4,"dp":2,"mark":"Y","isl":1024,"osl":128,"concurrency":16,"num_prompts":80,"warmup_requests":16,"attempt":1,"gain_pct":null,"plateau_streak":0,"error_type":"","raw_file":"/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/tp4_dp2/raw_outputs/vllm_adaptive_c16_i1024_o128_a1.jsonl","detail_log":"/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/tp4_dp2/logs/vllm_c16_i1024_o128_a1.log"}
+{"status":"COMPLETED","completed":160,"failed":0,"duration_s":48.83359965996351,"request_tps":3.2764326429775124,"input_tps":3355.0670264089727,"output_tps":419.3833783011216,"total_tps":3774.450404710094,"mean_input_tokens":1024.0,"mean_output_tokens":128.0,"ttft_p50_ms":1306.9384199916385,"ttft_p95_ms":3467.1723649604246,"ttft_p99_ms":3896.632032970665,"tpot_p50_ms":61.83210448829428,"tpot_p95_ms":161.7796285074291,"tpot_p99_ms":98.65987394490887,"e2e_p50_ms":9634.722630027682,"e2e_p95_ms":11989.051085483516,"e2e_p99_ms":14206.873630409826,"itl_p50_ms":57.19536996912211,"itl_p95_ms":433.56416895403527,"itl_p99_ms":470.6485329172574,"validation_errors":[],"timestamp":"2026-07-29T11:49:30+08:00","engine":"vllm","tp":4,"dp":2,"mark":"Y","isl":1024,"osl":128,"concurrency":32,"num_prompts":160,"warmup_requests":32,"attempt":1,"gain_pct":57.997539,"plateau_streak":0,"error_type":"","raw_file":"/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/tp4_dp2/raw_outputs/vllm_adaptive_c32_i1024_o128_a1.jsonl","detail_log":"/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/tp4_dp2/logs/vllm_c32_i1024_o128_a1.log"}
+{"status":"COMPLETED","completed":240,"failed":0,"duration_s":54.53471646993421,"request_tps":4.4008663753173725,"input_tps":4506.487168324989,"output_tps":563.3108960406237,"total_tps":5069.798064365613,"mean_input_tokens":1024.0,"mean_output_tokens":128.0,"ttft_p50_ms":1688.9932350022718,"ttft_p95_ms":4221.1321365553895,"ttft_p99_ms":5367.613992423278,"tpot_p50_ms":68.93910724380031,"tpot_p95_ms":164.9316360476405,"tpot_p99_ms":92.13227955046051,"e2e_p50_ms":11271.64134499617,"e2e_p95_ms":12833.146476407997,"e2e_p99_ms":13655.718730769586,"itl_p50_ms":61.67635991005227,"itl_p95_ms":453.4679424541537,"itl_p99_ms":482.7886850689538,"validation_errors":[],"timestamp":"2026-07-29T11:50:47+08:00","engine":"vllm","tp":4,"dp":2,"mark":"Y","isl":1024,"osl":128,"concurrency":48,"num_prompts":240,"warmup_requests":48,"attempt":1,"gain_pct":34.318842,"plateau_streak":0,"error_type":"","raw_file":"/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/tp4_dp2/raw_outputs/vllm_adaptive_c48_i1024_o128_a1.jsonl","detail_log":"/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/tp4_dp2/logs/vllm_c48_i1024_o128_a1.log"}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/adaptive_shapes.jsonl

@@ -0,0 +1 @@
+{"timestamp":"2026-07-29T11:50:47+08:00","engine":"vllm","tp":4,"dp":2,"mark":"Y","isl":1024,"osl":128,"status":"TTFT_BOUNDARY","stop_reason":"TTFT_SLO_EXCEEDED","tested_points":3,"search_cap":256,"max_successful_concurrency":48,"saturation_concurrency":null,"stop_probe_concurrency":48,"best_tps_concurrency":48,"best_total_tps":5069.798064365613,"last_total_tps":5069.798064365613}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/run_manifest.json

@@ -0,0 +1,27 @@
+{
+  "experiment": "dsv4_910c_vllm_tp_dp_matrix",
+  "engine": "vllm",
+  "run_id": "adaptive_20260729-113709",
+  "model": "/mnt/models/DeepSeek-V4-Flash-w8a8-mtp",
+  "hardware": "8x Ascend910 (16 dies, 64GB HBM/die)",
+  "matrix": "/mnt/yy/sskj/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/matrix.json",
+  "dataset": "random",
+  "tokenize_prompt": false,
+  "random_range_ratio": 1.0,
+  "search": {
+    "start_concurrency": 16,
+    "max_concurrency": 256,
+    "multiplier": 2,
+    "initial_backoff_concurrencies": [
+      8,
+      1
+    ],
+    "num_prompts_multiplier": 5,
+    "min_tps_gain_pct": 2.0,
+    "plateau_patience": 2,
+    "warmup_max_requests": 0,
+    "ttft_slo_ms": 4000.0,
+    "enable_ttft_slo_stop": 1,
+    "ttft_group_skip_ms": 8000.0
+  }
+}

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/adaptive_results/adaptive_20260729-113709/shapes.tsv

@@ -0,0 +1,31 @@
+mark	input_len	output_len
+Y	1024	128
+Y	1024	256
+Y	1024	512
+Y	1024	1024
+Y	1024	2048
+Y	1024	4096
+Y	4096	128
+Y	4096	256
+Y	4096	512
+Y	4096	1024
+Y	4096	2048
+Y	4096	4096
+Y	8192	128
+Y	8192	256
+Y	8192	512
+Y	8192	1024
+Y	8192	2048
+Y	8192	4096
+Y	16384	128
+Y	16384	256
+Y	16384	512
+Y	16384	1024
+Y	16384	2048
+Y	32768	128
+Y	32768	256
+Y	32768	512
+Y	32768	1024
+Y	65536	128
+Y	65536	256
+Y	131072	128

experiments/h200/dsv4_h200_vllm_tp2_custom_bench/README.md

@@ -239,7 +239,7 @@ python3 bench_client.py --lb-strategy round_robin --ports 30005,30006,30007,3000
 
 ### 结构化结果 (`results.json`)
 
-遵循仓库统一的 [JSON Schema](../../BENCHMARK_WORKFLOW.md#final-json-schema)。
+遵循仓库统一的 [JSON Schema](../../../docs/BENCHMARK_WORKFLOW.md#final-json-schema)。
 
 ### 人类可读报告 (`report.md`)