agentic client: venv rebuilt on every run and caller's UV_CACHE_DIR overridden
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
调研方向
Start with infx/bench/agentic/venv.py and the existing expectations in infx/tests/bench/test_agentic_runtime.py:70-97; infx/bench/agentic/run.py:248-256 shows where bootstrap is called. Read the venv setup and tests first, then update the tests for cache selection and matching-venv reuse. Done when matching requirements avoid uv calls, changed requirements rebuild, and a caller-supplied UV_CACHE_DIR is honored.
由索引模型根据 Issue 内容生成。
描述
The agentic benchmark client rebuilds its venv on every run and ignores a caller-supplied UV_CACHE_DIR, so a caller that runs many benchmarks pays a full client download each time. Downstream: AMD-AGI/Magpie#128, AMD-AGI/Hyperloom#1919.
Current behaviour (main @ 288471b, read from code)
inferencex-e2e/benchmarks/srt_agentic.shrunspython3 -m infx.bench agentic, which callsvenv.bootstrapwhen not already inside the client venv (infx/bench/agentic/run.py:248-256).bootstrapalways deletes the venv (infx/bench/agentic/venv.py:66), with no check of whether the existing one already matches.- It always sets
UV_CACHE_DIR=<runtime root>/uv-cache(venv.py:67-69), overriding a value from the caller. The CI workflows set their ownUV_CACHE_DIR(e.g..github/workflows/benchmark-tmpl.yml:246), which this also overrides. - The runtime root is
AIPERF_RUNTIME_DIR, else a per-job temp directory (venv.py:28-34). When the root is new per run, as it is for Magpie's AgentX runs, the cache is always cold and every run downloads the full client requirements again, plusuvitself if it is not onPATH(infx/bench/uv.py:21-32). infx/tests/bench/test_agentic_runtime.py:70-97pins the rebuild-and-override behaviour.
The older bash path had the same shape: 408c015 inferencex-e2e/benchmarks/benchmark_lib.sh:3133 deletes the venv and :3140-3142 force the cache path; 3d55815 made AIPERF_UV_CACHE_DIR overridable (benchmarks/benchmark_lib.sh:1787-1790) but still deleted the venv (:1841). Main dropped that override.
Impact
For a caller such as Magpie, the client starts after the server is healthy, so the download runs while the server holds its GPUs and counts against the client timeout. On a slow or restricted network a run can fail for reasons unrelated to what is being benchmarked.
Proposed change (all in infx/bench/agentic/venv.py, plus tests)
- Honour the caller's cache: use
os.environ.get("UV_CACHE_DIR")when set, falling back to<root>/uv-cache. - Reuse a matching venv: key it on
PYTHON_VERSIONplus the bytes ofREQUIREMENTS, store the key in the venv after a successful install, and skip the rebuild when the key matches andagentxandhfare executable. Otherwise rebuild as today. - If the venv is meant to be shared across concurrent runs through a persistent
AIPERF_RUNTIME_DIR, take a file lock around the check-and-build. - Tests: a matching key makes no
uvcalls; changed requirements rebuild;UV_CACHE_DIRfrom the caller is used.
- 主要语言
- Python
- 星标
- 1.9k
- 派生
- 343
- 平均合并
- 2 天 40 分钟
- 30 天内合并 PR
- 431
环境准备
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
SemiAnalysisAI/InferenceX 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 68/100
SemiAnalysisAI/InferenceX#2125 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 68/100
SemiAnalysisAI/InferenceX#1587 ·
维护者通常 1 天内回复
-
[experimental chore/agentx-v0.2-aiperf-testing branch] agentic build_replay_cmd: missing --tokenizer-trust-remote-code breaks Kimi-K2.5 (and other custom-code tokenizer models) / agentic build_replay_cmd 缺少 --tokenizer-trust-remote-code 导致 Kimi-K2.5 等自定义 tokenizer 模型失败可能重新可做 @sahil-bhatti-ai 于 170 天前认领，目前没有进行中的 PR。 未关闭
难度 1/5 1 小时以内 新手友好度 78/100
SemiAnalysisAI/InferenceX#1369 · 3 条评论 ·
维护者通常 1 天内回复
-
难度 1/5 1-3 小时 新手友好度 76/100
SemiAnalysisAI/InferenceX#1359 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1 小时以内 新手友好度 18/100
SemiAnalysisAI/InferenceX#4655 · 1 条评论 ·
维护者通常 1 天内回复
查看 SemiAnalysisAI/InferenceX 的全部 Issue
相似的 Issue
-
bug
难度 2/5 1-3 小时 新手友好度 86/100
NousResearch/hermes-agent#147104 ·
维护者通常 1 天内回复
-
[oblt-aw][autodoc] Document new required `shared-org-key` input for `obs-aw-dependency-review`可能已有人在做 @github-actions 今天认领。 未关闭
难度 1/5 1 小时以内 新手友好度 90/100
维护者通常 1 天内回复
-
bot-triaged module: inductor oncall: pt2 topic: fuzzer triaged
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 1 天内回复
-
getchaintips leaves out the active tip once a header is staged on it, where Core always lists it未关闭
难度 2/5 1-3 小时 新手友好度 88/100
btclib-org/btclib-node#2158 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 75/100
patrick-hudson/put.gd-desktop#979 · 1 条评论 ·
维护者通常 1 天内回复