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

bare_harness

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。

1. 定位

bare harness 评测裸模型本身的算子优化能力。模型只能使用 Backend 提供的工具池(读取文件、回填代码、提交编译/精度/profiling、结束任务),不依赖 coding agent CLI,也不使用 Pi。

实现基于 k-lens-v2:

  • client_v2/bare_client/:单会话客户端;
  • client_v2/config.yaml + config_loader.py:客户端配置;
  • Backend/:k-lens-v2 Backend,包含 finish_task 工具。

2. 前置条件

  • Python 3.10+,依赖见根目录 requirements.txt。
  • 可访问的远端沙盒。
  • 一个 OpenAI 兼容的模型端点。

不需要 Node 或 claude/codex/opencode CLI。

3. 模型配置

模型登记在 bridge/configs/op_catalog.json → models,运行时用 --model <registry-key> 选择。

k-lens-v2 支持的 provider:

vllm / openrouter / openai / glm-api / glm-client

catalog 中的 local 会在 bare 路径映射为 vllm。本地 vLLM 示例:

"my-local-model": {
  "provider": "local",
  "name": "DeepSeek-V4-Flash-BF16",
  "openai_base": "http://127.0.0.1:8008/v1",
  "wire_api": "chat",
  "context": 1000000,
  "api_key": "any-value",
  "tool_format": "json"
}

4. 单算子运行

cd $ASCEND_GYM_WORKSPACE/bridge

python3 op_manager.py run \
  --op swi_glu \
  --harness bare \
  --agent bare \
  --model <MODEL_KEY> \
  --candidate-evaluations-max 8 \
  --final-reprofile-count 3 \
  --sandbox-url http://<sandbox-host>:<port> \
  --tag BARE-$(date +%Y%m%d-%H%M%S)

5. 批量运行

python3 start_all_runs.py \
  --harness bare --agent bare \
  --ops all \
  --model <MODEL_KEY> \
  --candidate-evaluations-max 8 --final-reprofile-count 3 \
  --parallel 2 \
  --sandbox-url http://<sandbox-host>:<port> \
  --tag BARE-BATCH-$(date +%Y%m%d-%H%M%S)

6. 预算与评测协议

  • --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:可选软目标,默认等于候选上限。
  • profiling 协议固定为 8 总轮 = 3 预热 + 5 测量,统计方式为 mean;与 bridge 的通用 --profiling-runs 口径不同。

7. 产物

输出目录:

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

主要文件:

文件 说明
result.json klens-bare-client-result-v1,包含 success、speedup、baseline/best、终止原因
evaluations.jsonl 每次候选评测
trajectory.jsonl 模型与工具交互轨迹
usage.json token 使用
_client/ bare_client 原始产物

8. 注意事项

  • bare 使用 k-lens-v2 Backend(含 finish_task),不要指向其他版本的 Backend。
  • bare 的 profiling 协议固定为 8 总轮 / 3 预热 / mean;与 native 的聚合口径不同,不应直接比较绝对 duration_ms。
  • 模型偶尔只返回文本而不发起工具调用时,可通过 catalog 模型条目的 protocol_correction_max 调整纠偏次数。

相关文档: