AscendGYM / docs /BRIDGE_USAGE.md
SLAI-AITP's picture
Fix evaluation guides and connect the Hugging Face leaderboard link
961d5d4 verified
|
Raw History Blame Contribute Delete
6.47 kB

ascend_gym_bridge 使用手册

编译执行环境可选择官方镜像或兼容自建服务。连接地址为服务基础 URL,op_manager.py run 的 --sandbox-url 优先于 SANDBOX_URL,均未设置时默认 http://127.0.0.1:9000。源码和冻结数据通过 HTTP 上传;客户端无需服务源码。请求与返回格式见服务协议。

本文说明 bridge 的完整用法。各 harness 的算法、配置与注意事项见对应 README:

1. 范围

bridge 是评测统一入口,负责:

  • 解析算子与输入包;
  • 生成 run 配置;
  • 启动 harness;
  • 批量调度、监控、停止与重测;
  • 将 harness 产物归一化为统一输出目录。

支持列表:native、bare。

2. 目录

bridge/
├── op_manager.py          # 单算子入口
├── start_all_runs.py      # 批量启动 / 状态 / 停止
├── sandbox_evaluator.py   # native 评测器
├── claude_agent_wrapper.py
├── bare_bench/            # bare 适配
├── configs/               # op_catalog 配置
└── tools/                 # 复测、协议代理、监控、任务队列

3. 前置条件

export ASCEND_GYM_WORKSPACE=/path/to/workspace
export ASCEND_GYM_DATA_ROOT=/path/to/ascend_gym_data
export SANDBOX_URL=http://<sandbox-host>:<port>

安装依赖:

pip install -r $ASCEND_GYM_WORKSPACE/requirements.txt

native 需要对应 agent CLI。详见各 harness README。

4. 命令

所有命令在 bridge/ 目录下执行。

4.1 解析与构建

python3 op_manager.py scan
python3 op_manager.py build --all
  • scan:解析 ops_src,写出 configs/op_catalog.resolved.json。
  • build:生成 outputs.inputs_dir 下的算子输入包。

4.2 校验

python3 op_manager.py validate --op swi_glu --harness native

bare 没有离线 preflight,使用 run --dry-run 检查配置。

4.3 单算子运行

python3 op_manager.py run \
  --op <OP> \
  --harness <native|bare> \
  --agent <AGENT> \
  --model <MODEL_KEY> \
  --candidate-evaluations-max <B> \
  --final-reprofile-count <N> \
  --sandbox-url $SANDBOX_URL \
  --tag <TAG>

常用选项:

选项 说明
--candidate-evaluations-max 候选评测硬上限
--final-reprofile-count 最终复测次数,不占候选预算
--candidate-solutions-target native 软目标(可选)
--wall-clock-seconds 搜索墙钟;仅 bare 使用,0 表示不限
--dry-run 打印命令与环境,不实际执行

4.4 参数优先级

结构化 CLI 参数覆盖 configs/op_catalog.json 中的默认值;bare 的 profiling 协议仍以 catalog 为准。

5. 批量运行

python3 start_all_runs.py \
  --harness <native|bare> \
  --agent <AGENT> \
  --ops all \
  --model <MODEL_KEY> \
  --candidate-evaluations-max <B> \
  --final-reprofile-count <N> \
  --parallel <N> \
  --tag <TAG>

--parallel 控制本机同时运行的 run 数量,不控制远端沙盒 worker 数。

查看状态与停止:

python3 start_all_runs.py --harness <H> --agent <A> --ops all --tag <TAG> --status
python3 start_all_runs.py --harness <H> --agent <A> --ops all --tag <TAG> --stop

5.1 final verification 固定卡失败

如果批次已结束,但某个 run 的 result.json.success=false,且 final_verification 的失败在固定同一张 NPU 上反复出现,而同一最佳版本在另一张 NPU 上能通过,不要直接重跑整个候选预算。先用 reprofile_best.py 对最佳版本发起复测;复测通过后它会调用 apply_final_recheck.py 写回:

# 1) 检查输入并预览;native 需要 SANDBOX_URL,bare 可省略
python3 tools/reprofile_best.py \
  --run-dir "$ASCEND_GYM_WORKSPACE/runs/<TAG>/<RUN_ID>" \
  --count 5 --npu node-0:npu0 --sandbox-url "$SANDBOX_URL" --dry-run

# 2) 复测 + 写回(native/bare 通用)
python3 tools/reprofile_best.py \
  --run-dir "$ASCEND_GYM_WORKSPACE/runs/<TAG>/<RUN_ID>" \
  --count 5 --npu node-0:npu0 --sandbox-url "$SANDBOX_URL" --apply

--count 不传时沿用原 run 的 final_verification.requested。如果只想复测、不写回,去掉 --apply,再按需执行:

python3 tools/apply_final_recheck.py \
  --run-dir "$ASCEND_GYM_WORKSPACE/runs/<TAG>/<RUN_ID>" \
  --npu node-0:npu0 --yes

写回会保留原失败到 recheck_history、替换 final_verification、更新 events.jsonl 并备份原 result.json。复测次数由 recheck.requested 决定,不硬编码 5 次。完整参数、写入内容和安全边界见 BRIDGE_TOOLS.md §2 与 §3。

6. 监控与队列

工具位于 tools/,说明见 BRIDGE_TOOLS.md。

python3 tools/monitor/native_monitor.py --tag <TAG> --single
python3 tools/task_dispatcher.py list --queue <QUEUE>

7. 产物

$ASCEND_GYM_WORKSPACE/runs/<TAG>/<op>-<agent>-<TAG>/

outputs.runs_dir 为 CWD 相对路径 ../runs;在 bridge/ 下执行时产物位于包根 runs/。

常见文件:

result.json
evaluations.jsonl
events.jsonl
usage.json
manifest.json
agent.stdout.txt / agent.stderr.txt
_client/                # bare 原始产物

各 harness 的结果字段见对应 README。

8. 配置

完整配置说明见 BRIDGE_CONFIG.md,字段参考见 CATALOG_REFERENCE.md。

9. 常见问题

现象 处理
operator not resolved 检查 source_root / search_roots 与实际源码
输入包缺失 执行 python3 op_manager.py build --all
沙盒连接失败 检查 SANDBOX_URL、/health、/status
模型 401 检查 models.<key>.api_key 与端点
不支持 harness 使用 native 或 bare

10. 文档索引