Star0
5
代码
代码
Issues2
Pull Requests
Actions
项目讨论
Wiki
分析
Star0
5
已合并
chore(precision-test): 同步 test_selector 上游并适配 torch_npu 改为 pytorch (#18) #44
tzing_t创建于 5 天前
chore(precision-test): 同步 test_selector 上游并适配 torch_npu 改为 pytorch (#18) #44
已合并
tzing_t创建于 5 天前
从precision合入到master
共 21 个文件变更+595-180
Mprecision-test/README.md+30-30
@@ -1,12 +1,12 @@
1# precision-test1# precision-test
2 2 
3-<!-- 基于代码覆盖率变更的精准测试用例选择器——分析 PR 代码变更，按覆盖率数据以行、函数两种粒度筛选相关测试用例（文件级作为重命名/删除场景的内部兜底），避免每次 PR 跑全量用例。统一支持 vllm_ascend / sglang / torch_npu 三仓库。 -->3+<!-- 基于代码覆盖率变更的精准测试用例选择器——分析 PR 代码变更，按覆盖率数据以行、函数两种粒度筛选相关测试用例（文件级作为重命名/删除场景的内部兜底），避免每次 PR 跑全量用例。统一支持 vllm_ascend / sglang / pytorch 三仓库。 -->
4 4 
5## 核心特性 (Features)5## 核心特性 (Features)
6 6 
7-- **三仓库统一**：一套插件支持 `vllm_ascend` / `sglang` / `torch_npu`，通过 `repo-name` 选择仓库适配器7+- **三仓库统一**：一套插件支持 `vllm_ascend` / `sglang` / `pytorch`，通过 `repo-name` 选择仓库适配器
8- **多粒度匹配**：行级 + 函数级并行匹配，结果合并去重；文件级匹配作为重命名/删除场景的内部兜底（不可手动开关）8- **多粒度匹配**：行级 + 函数级并行匹配，结果合并去重；文件级匹配作为重命名/删除场景的内部兜底（不可手动开关）
9-- **PR 源自动适配**：vllm_ascend / sglang 走 GitHub（`github-pr`），torch_npu 走 GitCode（`gitcode-pr`），两者互斥9+- **PR 源自动适配**：三仓库均走 GitHub（`github-pr`，pytorch 检测上游 pytorch/pytorch）；`gitcode-pr` 作为 GitCode PR 通用入口，与 `github-pr` 互斥
10- **基于覆盖率选择**：利用覆盖率数据（SQLite 格式）精准识别受代码变更影响的测试用例10- **基于覆盖率选择**：利用覆盖率数据（SQLite 格式）精准识别受代码变更影响的测试用例
11- **安全加固**：内置 DoS 防护、子进程环境变量白名单、路径与 PR 标识校验11- **安全加固**：内置 DoS 防护、子进程环境变量白名单、路径与 PR 标识校验
12- **可配置**：匹配粒度、阈值、虚拟环境等均可通过 action 输入参数调整12- **可配置**：匹配粒度、阈值、虚拟环境等均可通过 action 输入参数调整
@@ -57,7 +57,7 @@ jobs:
57 fi57 fi
58```58```
59 59 
60-#### torch_npu（GitCode PR）60+#### pytorch（GitHub PR）
61 61 
62```yaml62```yaml
63name: 精准测试63name: 精准测试
@@ -79,8 +79,8 @@ jobs:
79 - name: 精准测试选择79 - name: 精准测试选择
80 uses: openlibing/test-action/precision-test@v2.0.080 uses: openlibing/test-action/precision-test@v2.0.0
81 with:81 with:
82- repo-name: torch_npu82+ repo-name: pytorch
83- gitcode-pr: ${{ atomgit.repository }}#${{ atomgit.event.pull_request.number }}83+ github-pr: pytorch/pytorch#${{ github.event.pull_request.number }}
84 source-dir: covstub84 source-dir: covstub
85 min-affected: 185 min-affected: 1
86 86 
@@ -110,23 +110,23 @@ jobs:
110 110 
111## 输入参数 (Inputs)111## 输入参数 (Inputs)
112 112 
113-| 参数名 | 说明 | 必填 | 默认值 |113+| 参数名 | 说明 | 必填 | 默认值 |
114-| ------------------------ | ----------------------------------------------------------------------------------------------------- | ------ | -------------------- |114+| ------------------------ | --------------------------------------------------------------------------------------------------------------- | ------ | -------------------- |
115-| `github-pr` | GitHub PR（vllm_ascend / sglang），格式 `owner/repo#pr_number` 或仅 `pr_number`；与 `gitcode-pr` 互斥 | 二选一 | - |115+| `github-pr` | GitHub PR（vllm_ascend / sglang / pytorch），格式 `owner/repo#pr_number` 或仅 `pr_number`；与 `gitcode-pr` 互斥 | 二选一 | - |
116-| `gitcode-pr` | GitCode PR（torch_npu），格式 `owner/repo#pr_number` 或仅 `pr_number`；与 `github-pr` 互斥 | 二选一 | - |116+| `gitcode-pr` | GitCode PR，格式 `owner/repo#pr_number` 或仅 `pr_number`；与 `github-pr` 互斥 | 二选一 | - |
117-| `repo-name` | 仓库适配器：`vllm_ascend` / `sglang` / `torch_npu` | 否 | `vllm_ascend` |117+| `repo-name` | 仓库适配器：`vllm_ascend` / `sglang` / `pytorch` | 否 | `vllm_ascend` |
118-| `source-dir` | 源代码目录（函数级匹配与噪音过滤需要） | 否 | `covstub` |118+| `source-dir` | 源代码目录（函数级匹配与噪音过滤需要） | 否 | `covstub` |
119-| `map-file` | 测试用例映射文件路径 | 否 | `test_case_map.json` |119+| `map-file` | 测试用例映射文件路径 | 否 | `test_case_map.json` |
120-| `coverage-dir` | 覆盖率数据目录；构建 map 时（`build-map: true` 或 map 文件不存在时）必填 | 否 | `coverage` |120+| `coverage-dir` | 覆盖率数据目录；构建 map 时（`build-map: true` 或 map 文件不存在时）必填 | 否 | `coverage` |
121-| `build-map` | 是否强制重建测试用例映射 | 否 | `false` |121+| `build-map` | 是否强制重建测试用例映射 | 否 | `false` |
122-| `min-affected` | 最小受影响行数阈值 | 否 | `1` |122+| `min-affected` | 最小受影响行数阈值 | 否 | `1` |
123-| `dedup` | 是否启用去重（相同覆盖行仅保留一个用例） | 否 | `false` |123+| `dedup` | 是否启用去重（相同覆盖行仅保留一个用例） | 否 | `false` |
124-| `enable-line-match` | 是否启用行级匹配 | 否 | `true` |124+| `enable-line-match` | 是否启用行级匹配 | 否 | `true` |
125-| `disable-line-match` | 关闭行级匹配（**优先级高于 `enable-line-match`**） | 否 | `false` |125+| `disable-line-match` | 关闭行级匹配（**优先级高于 `enable-line-match`**） | 否 | `false` |
126-| `enable-function-match` | 是否启用函数级匹配 | 否 | `true` |126+| `enable-function-match` | 是否启用函数级匹配 | 否 | `true` |
127-| `disable-function-match` | 关闭函数级匹配（**优先级高于 `enable-function-match`**） | 否 | `false` |127+| `disable-function-match` | 关闭函数级匹配（**优先级高于 `enable-function-match`**） | 否 | `false` |
128-| `skip-imports` | 函数级匹配时是否跳过 import 语句行 | 否 | `false` |128+| `skip-imports` | 函数级匹配时是否跳过 import 语句行 | 否 | `false` |
129-| `skip-venv` | 跳过虚拟环境创建，直接使用系统 `python3` / `pip3` | 否 | `true` |129+| `skip-venv` | 跳过虚拟环境创建，直接使用系统 `python3` / `pip3` | 否 | `true` |
130 130 
131**匹配粒度规则**：`disable-*` 优先于 `enable-*`。即 `disable-line-match: true` 时无论 `enable-line-match` 取值如何，行级匹配均关闭。文件级匹配仅作为重命名/删除场景的内部兜底，不对外暴露开关。131**匹配粒度规则**：`disable-*` 优先于 `enable-*`。即 `disable-line-match: true` 时无论 `enable-line-match` 取值如何，行级匹配均关闭。文件级匹配仅作为重命名/删除场景的内部兜底，不对外暴露开关。
132 132 
@@ -148,10 +148,10 @@ permissions:
148 148 
149## 仓库差异对照149## 仓库差异对照
150 150 
151-| 维度 | vllm_ascend | sglang | torch_npu |151+| 维度 | vllm_ascend | sglang | pytorch |
152| ------------ | ------------------------- | ------------------------------------ | ------------------------------- |152| ------------ | ------------------------- | ------------------------------------ | ------------------------------- |
153-| PR 源 | GitHub（`github-pr`） | GitHub（`github-pr`） | GitCode（`gitcode-pr`） |153+| PR 源 | GitHub（`github-pr`） | GitHub（`github-pr`） | GitHub（`github-pr`） |
154-| 产品代码前缀 | `vllm_ascend/` | `python/sglang/` | `torch_npu/` |154+| 产品代码前缀 | `vllm_ascend/` | `python/sglang/` | `torch/` |
155| 全量触发变更 | `csrc/` 目录（非 `.md`） | 无 | 无 |155| 全量触发变更 | `csrc/` 目录（非 `.md`） | 无 | 无 |
156| 测试目录识别 | `tests__` 前缀 / `cpu-ut` | `____w__sglang__sglang__test__` 前缀 | 目录名含 `__` 或以 `test_` 开头 |156| 测试目录识别 | `tests__` 前缀 / `cpu-ut` | `____w__sglang__sglang__test__` 前缀 | 目录名含 `__` 或以 `test_` 开头 |
157 157 
@@ -191,7 +191,7 @@ permissions:
191| --------------- | ------ | ------------------------------------------------------------------- |191| --------------- | ------ | ------------------------------------------------------------------- |
192| `GITHUB_TOKEN` | - | GitHub API token（vllm_ascend / sglang 拉 PR diff，可选，提升速率） |192| `GITHUB_TOKEN` | - | GitHub API token（vllm_ascend / sglang 拉 PR diff，可选，提升速率） |
193| `GH_TOKEN` | - | GitHub API token 备选（`GITHUB_TOKEN` 未设置时回退） |193| `GH_TOKEN` | - | GitHub API token 备选（`GITHUB_TOKEN` 未设置时回退） |
194-| `GITCODE_TOKEN` | - | GitCode API token（torch_npu 拉 PR diff，可选，提升速率） |194+| `GITCODE_TOKEN` | - | GitCode API token（`gitcode-pr` 拉 PR diff，可选，提升速率） |
195 195 
196> 插件通过子进程环境变量白名单（`SUBPROCESS_ENV_SAFE_NAMES`）向 Python 子进程显式传递 `PATH`、`HOME`、代理、SSL 证书、`GITHUB_TOKEN` / `GH_TOKEN` / `GITCODE_TOKEN` 等安全变量，屏蔽 `ACTIONS_*` / `ATOMGIT_*` / `INPUT_*` / `HUAWEICLOUD_*` 等平台敏感前缀，避免密钥泄露到 Python 进程。196> 插件通过子进程环境变量白名单（`SUBPROCESS_ENV_SAFE_NAMES`）向 Python 子进程显式传递 `PATH`、`HOME`、代理、SSL 证书、`GITHUB_TOKEN` / `GH_TOKEN` / `GITCODE_TOKEN` 等安全变量，屏蔽 `ACTIONS_*` / `ATOMGIT_*` / `INPUT_*` / `HUAWEICLOUD_*` 等平台敏感前缀，避免密钥泄露到 Python 进程。
197 197 
@@ -223,8 +223,8 @@ GitHub / GitCode API 均需直连外网，内网环境必须配置代理：
223 PYTHON_EXEC_TIMEOUT_SECONDS: 900223 PYTHON_EXEC_TIMEOUT_SECONDS: 900
224 GITCODE_TOKEN: ${{ secrets.GITCODE_TOKEN }}224 GITCODE_TOKEN: ${{ secrets.GITCODE_TOKEN }}
225 with:225 with:
226- repo-name: torch_npu226+ repo-name: pytorch
227- gitcode-pr: ${{ atomgit.repository }}#${{ atomgit.event.pull_request.number }}227+ github-pr: pytorch/pytorch#${{ github.event.pull_request.number }}
228```228```
229 229 
230## 数据格式230## 数据格式
@@ -234,7 +234,7 @@ GitHub / GitCode API 均需直连外网，内网环境必须配置代理：
234- **file**：文件路径信息234- **file**：文件路径信息
235- **arc**：覆盖弧数据，含 `fromno`、`tono` 字段235- **arc**：覆盖弧数据，含 `fromno`、`tono` 字段
236 236 
237-覆盖率目录布局自动探测：测试目录下存在 `covdata/` 子目录（vllm_ascend / torch_npu）则从 `covdata/` 读取 `coverage.*`；无 `covdata/` 子目录（sglang）则直接读取测试目录下的 `coverage.*`。237+覆盖率目录布局自动探测：测试目录下存在 `covdata/` 子目录（vllm_ascend / pytorch）则从 `covdata/` 读取 `coverage.*`；无 `covdata/` 子目录（sglang）则直接读取测试目录下的 `coverage.*`。
238 238 
239## 上下游串联239## 上下游串联
240 240 
Mprecision-test/action.yml+4-4
@@ -1,16 +1,16 @@
1name: "Precision Test Selector"1name: "Precision Test Selector"
2-description: "Coverage-based precision test selector with line, function, and file granularity (vllm_ascend / sglang / torch_npu)"2+description: "Coverage-based precision test selector with line, function, and file granularity (vllm_ascend / sglang / pytorch)"
3author: "openlibing"3author: "openlibing"
4 4 
5inputs:5inputs:
6 github-pr:6 github-pr:
7- description: "GitHub PR (vllm_ascend / sglang), format: owner/repo#pr_number or just pr_number"7+ description: "GitHub PR (vllm_ascend / sglang / pytorch), format: owner/repo#pr_number or just pr_number"
8 required: false8 required: false
9 gitcode-pr:9 gitcode-pr:
10- description: "GitCode PR (torch_npu), format: owner/repo#pr_number or just pr_number; mutually exclusive with github-pr"10+ description: "GitCode PR, format: owner/repo#pr_number or just pr_number; mutually exclusive with github-pr"
11 required: false11 required: false
12 repo-name:12 repo-name:
13- description: "Repository adapter: vllm_ascend / sglang / torch_npu (default: vllm_ascend)"13+ description: "Repository adapter: vllm_ascend / sglang / pytorch (default: vllm_ascend)"
14 required: false14 required: false
15 default: "vllm_ascend"15 default: "vllm_ascend"
16 source-dir:16 source-dir:
Mprecision-test/dist/index.js+1-1
@@ -25651,7 +25651,7 @@ const path = __nccwpck_require__(6928);
25651/**25651/**
25652 * Available repository adapters supported by the precision-test tool.25652 * Available repository adapters supported by the precision-test tool.
25653 */25653 */
25654-const AVAILABLE_REPOS = ["vllm_ascend", "sglang", "torch_npu"];25654+const AVAILABLE_REPOS = ["vllm_ascend", "sglang", "pytorch"];
25655 25655 
25656/**25656/**
25657 * Whitelist of environment variable names safe to pass to the Python subprocess.25657 * Whitelist of environment variable names safe to pass to the Python subprocess.
Mprecision-test/dist/test_selector/__init__.py+2-2
@@ -1,10 +1,10 @@
1-"""test_selector - 覆盖率驱动的精准测试选择器（vllm_ascend / sglang / torch_npu 统一架构）。1+"""test_selector - 覆盖率驱动的精准测试选择器（vllm_ascend / sglang / pytorch 统一架构）。
2 2 
3公共逻辑集中在 test_selector/ 包（diff 解析、噪音过滤、覆盖率选择、CLI 等），3公共逻辑集中在 test_selector/ 包（diff 解析、噪音过滤、覆盖率选择、CLI 等），
4仓库差异通过 test_selector.repos.RepoAdapter 注入，保证架构统一、差异按仓库归类。4仓库差异通过 test_selector.repos.RepoAdapter 注入，保证架构统一、差异按仓库归类。
5 5 
6入口：6入口：
7- python -m test_selector --repo <vllm_ascend|sglang|torch_npu> ...7+ python -m test_selector --repo <vllm_ascend|sglang|pytorch> ...
8"""8"""
9 9 
10from .repos import AVAILABLE_REPOS, RepoAdapter, get_adapter10from .repos import AVAILABLE_REPOS, RepoAdapter, get_adapter
Mprecision-test/dist/test_selector/change_detector.py+7-0
@@ -106,6 +106,13 @@ class CodeChangeDetector:
106 NamedTuple base) or module level, unless the annotation expression106 NamedTuple base) or module level, unless the annotation expression
107 may have runtime side effects (calls, Annotated metadata) and the107 may have runtime side effects (calls, Annotated metadata) and the
108 file has no ``from __future__ import annotations``.108 file has no ``from __future__ import annotations``.
109+ - First-time bindings are excluded for pure insertions (needs base
110+ content): every added line is an assignment binding simple names
111+ that occur nowhere in the base file, with inert right-hand sides;
112+ no meaningful line ends in a comma (such lines are call/signature
113+ arguments, not statements); class-scope insertions follow dataclass
114+ field-order rules (plain assigns are not fields, annotated fields
115+ must land at or after the last existing field).
109 - Isolated blank-line deletion (neighbours not deleted): treated as a116 - Isolated blank-line deletion (neighbours not deleted): treated as a
110 one-line insertion -> candidate pair (line above, line below).117 one-line insertion -> candidate pair (line above, line below).
111 - Pure insertions and blank-deletion pairs are classified via ast of118 - Pure insertions and blank-deletion pairs are classified via ast of
Mprecision-test/dist/test_selector/cli.py+2-2
@@ -10,11 +10,11 @@
10 10 
11用法：11用法：
12 python -m test_selector --repo vllm_ascend --github-pr owner/repo#12312 python -m test_selector --repo vllm_ascend --github-pr owner/repo#123
13- python -m test_selector --repo torch_npu --gitcode-pr Ascend/pytorch#12313+ python -m test_selector --repo torch_npu --github-pr pytorch/pytorch#123
14 python -m test_selector --repo sglang --build-map --coverage-dir coverage14 python -m test_selector --repo sglang --build-map --coverage-dir coverage
15 python vllm/test_selector.py --github-pr 123 # 薄入口（默认 vllm_ascend）15 python vllm/test_selector.py --github-pr 123 # 薄入口（默认 vllm_ascend）
16 python sglang/test_selector.py --github-pr 123 # 薄入口（默认 sglang）16 python sglang/test_selector.py --github-pr 123 # 薄入口（默认 sglang）
17- python PyTorch/test_selector.py --gitcode-pr 123 # 薄入口（默认 torch_npu）17+ python PyTorch/test_selector.py --github-pr 123 # 薄入口（默认 torch_npu，检测上游 pytorch/pytorch）
18"""18"""
19 19 
20import argparse20import argparse
Mprecision-test/dist/test_selector/coverage_selector.py+5-4
@@ -49,20 +49,21 @@ class CoverageSelector:
49 if not self.coverage_data_dir or not self.coverage_data_dir.exists():49 if not self.coverage_data_dir or not self.coverage_data_dir.exists():
50 print(f" Warning: Coverage data directory not found: {self.coverage_data_dir}")50 print(f" Warning: Coverage data directory not found: {self.coverage_data_dir}")
51 return test_cases51 return test_cases
52- for item in self.coverage_data_dir.iterdir():52+ for item in self.coverage_data_dir.rglob("*"):
53 if not item.is_dir():53 if not item.is_dir():
54 continue54 continue
55 name = item.name55 name = item.name
56 # 目录名规则由仓库适配器决定（vllm: tests__ 前缀 / cpu-ut；sglang: ____w__... 前缀）56 # 目录名规则由仓库适配器决定（vllm: tests__ 前缀 / cpu-ut；sglang: ____w__... 前缀）
57 if not self.adapter.is_test_case_dir(name):57 if not self.adapter.is_test_case_dir(name):
58 continue58 continue
59- # 布局探测：covdata/ 子目录（vllm/torch_npu）或目录下直接放置（sglang）59+ # 布局探测：covdata/ 子目录（vllm/pytorch）或目录下直接放置（sglang）
60 covdata_dir = item / "covdata"60 covdata_dir = item / "covdata"
61 has_cov_files = any(item.glob(self.adapter.coverage_file_glob))61 has_cov_files = any(item.glob(self.adapter.coverage_file_glob))
62 if covdata_dir.exists() and any(covdata_dir.glob(self.adapter.coverage_file_glob)):62 if covdata_dir.exists() and any(covdata_dir.glob(self.adapter.coverage_file_glob)):
63 has_cov_files = True63 has_cov_files = True
64 if has_cov_files:64 if has_cov_files:
65- test_cases.append(name)65+ rel = item.relative_to(self.coverage_data_dir).as_posix()
66+ test_cases.append(rel)
66 return sorted(test_cases)67 return sorted(test_cases)
67 68 
68 def normalize_test_name(self, test_name: str) -> str:69 def normalize_test_name(self, test_name: str) -> str:
@@ -142,7 +143,7 @@ class CoverageSelector:
142 143 
143 file_lines_map = defaultdict(set) # filepath -> set of lines144 file_lines_map = defaultdict(set) # filepath -> set of lines
144 145 
145- # 布局探测：vllm/torch_npu 将 coverage 放在 covdata/ 子目录，sglang 直接放在测试目录下146+ # 布局探测：vllm/pytroch 将 coverage 放在 covdata/ 子目录，sglang 直接放在测试目录下
146 cov_dirs = [covdata_dir] if covdata_dir.exists() else [test_case_dir]147 cov_dirs = [covdata_dir] if covdata_dir.exists() else [test_case_dir]
147 148 
148 for cov_dir in cov_dirs:149 for cov_dir in cov_dirs:
Mprecision-test/dist/test_selector/diff_parser.py+209-13
@@ -4,6 +4,8 @@
4代码变更的场景：4代码变更的场景：
5- 纯注释/docstring 变更5- 纯注释/docstring 变更
6- 纯类型注解变更（插入/替换/删除三个方向，可证明惰性时豁免）6- 纯类型注解变更（插入/替换/删除三个方向，可证明惰性时豁免）
7+- 首次新增的变量绑定（纯插入块全部为绑定全新名字的带值赋值，RHS 惰性，
8+ 无尾逗号行——尾逗号行是调用/签名参数而非语句）
7- 函数/类定义之间的空行插入9- 函数/类定义之间的空行插入
8- 新增函数/类整体10- 新增函数/类整体
9 11 
@@ -318,14 +320,27 @@ def _class_consumes_annotations(cls_node: ast.ClassDef, class_map: dict) -> bool
318 return False320 return False
319 321 
320 322 
321-def _collect_class_info(source: str) -> tuple[list[tuple], dict[str, ast.ClassDef], bool]:323+def _arg_names(args: ast.arguments) -> set[str]:
322- """Parse source and return (scopes, class_map, has_future):324+ """Names bound by a function/lambda parameter list."""
325+ names = {a.arg for a in args.posonlyargs + args.args + args.kwonlyargs}
326+ if args.vararg:
327+ names.add(args.vararg.arg)
328+ if args.kwarg:
329+ names.add(args.kwarg.arg)
330+ return names
323 331 
324- scopes : [(lineno, end_lineno, col_offset, node)] of every332+ 
325- function/class definition, for annotation scope resolution;333+def _collect_class_info(source: str) -> tuple[list[tuple], dict[str, ast.ClassDef], bool, set[str]]:
326- class_map : {class name -> ClassDef} for same-file base resolution334+ """Parse source and return (scopes, class_map, has_future, known_names):
327- (a later definition shadows an earlier one);335+ 
328- has_future: the file has ``from __future__ import annotations``.336+ scopes : [(lineno, end_lineno, col_offset, node)] of every
337+ function/class definition, for annotation scope resolution;
338+ class_map : {class name -> ClassDef} for same-file base resolution
339+ (a later definition shadows an earlier one);
340+ has_future : the file has ``from __future__ import annotations``.
341+ known_names : every name the base file already binds or reads (Name
342+ occurrences, def/class names, import aliases, parameters),
343+ for first-time-binding detection.
329 """344 """
330 tree = ast.parse(source)345 tree = ast.parse(source)
331 scopes = [346 scopes = [
@@ -344,7 +359,24 @@ def _collect_class_info(source: str) -> tuple[list[tuple], dict[str, ast.ClassDe
344 and any(alias.name == "annotations" for alias in n.names)359 and any(alias.name == "annotations" for alias in n.names)
345 for n in tree.body360 for n in tree.body
346 )361 )
347- return scopes, class_map, has_future362+ known_names = set()
363+ for node in ast.walk(tree):
364+ if isinstance(node, ast.Name):
365+ known_names.add(node.id)
366+ elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
367+ known_names.add(node.name)
368+ known_names.update(_arg_names(node.args))
369+ elif isinstance(node, ast.ClassDef):
370+ known_names.add(node.name)
371+ elif isinstance(node, ast.Lambda):
372+ known_names.update(_arg_names(node.args))
373+ elif isinstance(node, ast.ImportFrom):
374+ for alias in node.names:
375+ known_names.add(alias.asname or alias.name)
376+ elif isinstance(node, ast.Import):
377+ for alias in node.names:
378+ known_names.add(alias.asname or alias.name.split(".")[0])
379+ return scopes, class_map, has_future, known_names
348 380 
349 381 
350def _insertion_scope(scopes: list[tuple], a: int, b: int, add_indent: int | None):382def _insertion_scope(scopes: list[tuple], a: int, b: int, add_indent: int | None):
@@ -391,7 +423,7 @@ def _pure_annotation_insert_reason(
391 """423 """
392 if ann_ctx is None or not add_texts:424 if ann_ctx is None or not add_texts:
393 return None425 return None
394- scopes, class_map, has_future = ann_ctx426+ scopes, class_map, has_future, _known = ann_ctx
395 anns = _parse_pure_annotations(add_texts)427 anns = _parse_pure_annotations(add_texts)
396 if not anns:428 if not anns:
397 return None429 return None
@@ -422,7 +454,7 @@ def _pure_annotation_change_exempt(
422 """454 """
423 if ann_ctx is None:455 if ann_ctx is None:
424 return False456 return False
425- scopes, class_map, has_future = ann_ctx457+ scopes, class_map, has_future, _known = ann_ctx
426 del_anns = _parse_pure_annotations([t for _, t in del_lines])458 del_anns = _parse_pure_annotations([t for _, t in del_lines])
427 if not del_anns:459 if not del_anns:
428 return False460 return False
@@ -446,6 +478,161 @@ def _pure_annotation_change_exempt(
446 return not _class_consumes_annotations(node, class_map)478 return not _class_consumes_annotations(node, class_map)
447 479 
448 480 
481+def _parse_assignments(texts) -> list[tuple[ast.stmt, list[str]]] | None:
482+ """Parse diff lines as a block of value-carrying simple-name assignments.
483+ 
484+ Returns [(stmt, target_names), ...] when every meaningful (non-blank,
485+ non-comment) line forms assignments (``x = v`` or ``x: T = v``) whose
486+ targets are all plain names; None when the block is unparsable or
487+ contains anything else (bare annotations, attribute/subscript targets,
488+ tuple unpacking, any other statement).
489+ """
490+ lines = [t for t in texts if t.strip() and not t.strip().startswith("#")]
491+ if not lines:
492+ return None
493+ block = textwrap.dedent("\n".join(lines))
494+ try:
495+ tree = ast.parse(block)
496+ except (SyntaxError, ValueError):
497+ return None
498+ stmts = []
499+ for stmt in tree.body:
500+ if isinstance(stmt, ast.Assign):
501+ if not all(isinstance(t, ast.Name) for t in stmt.targets):
502+ return None
503+ stmts.append((stmt, [t.id for t in stmt.targets]))
504+ elif isinstance(stmt, ast.AnnAssign):
505+ if stmt.value is None or not isinstance(stmt.target, ast.Name):
506+ return None
507+ stmts.append((stmt, [stmt.target.id]))
508+ else:
509+ return None
510+ return stmts
511+ 
512+ 
513+_LAZY_RHS_NODES = (
514+ ast.Constant,
515+ ast.Name,
516+ ast.Attribute,
517+ ast.Tuple,
518+ ast.List,
519+ ast.Set,
520+ ast.Dict,
521+ ast.BinOp,
522+ ast.UnaryOp,
523+ ast.BoolOp,
524+ ast.operator,
525+ ast.unaryop,
526+ ast.boolop,
527+ ast.expr_context,
528+)
529+ 
530+ 
531+def _assign_rhs_lazy(value: ast.expr) -> bool:
532+ """True when the right-hand side is free of runtime side effects:
533+ literals, names, attributes and containers/arithmetic over those only
534+ (no calls, subscripts, comprehensions, lambdas, walrus...). Attribute
535+ access and arithmetic can still trigger property/__add__ hooks - an
536+ accepted risk; dict unpacking ({**d}) is not: it iterates the source
537+ mapping at runtime."""
538+ for node in ast.walk(value):
539+ if not isinstance(node, _LAZY_RHS_NODES):
540+ return False
541+ if isinstance(node, ast.Dict) and any(k is None for k in node.keys):
542+ return False
543+ return True
544+ 
545+ 
546+def _decorator_tail(dec: ast.expr) -> str | None:
547+ """Simple name of a decorator (call decorators use their callee):
548+ 'dataclass' for @dataclass / @dataclasses.dataclass / @dataclass(...)."""
549+ expr = dec.func if isinstance(dec, ast.Call) else dec
550+ if isinstance(expr, ast.Name):
551+ return expr.id
552+ if isinstance(expr, ast.Attribute):
553+ return expr.attr
554+ return None
555+ 
556+ 
557+def _class_dataclass_like(cls_node: ast.ClassDef, class_map: dict) -> bool:
558+ """True when the class or a same-file ancestor is decorated with
559+ @dataclass (annotated class-body statements then become __init__ fields
560+ whose order matters)."""
561+ stack, seen = [cls_node], set()
562+ while stack:
563+ node = stack.pop()
564+ if id(node) in seen:
565+ continue
566+ seen.add(id(node))
567+ if any(_decorator_tail(d) == "dataclass" for d in node.decorator_list):
568+ return True
569+ for base in node.bases:
570+ name = _base_simple_name(base)
571+ parent = class_map.get(name) if name else None
572+ if parent is not None and id(parent) not in seen:
573+ stack.append(parent)
574+ return False
575+ 
576+ 
577+def _last_field_line(cls_node: ast.ClassDef) -> int | None:
578+ """Last base line of the class's annotated fields (dataclass __init__
579+ parameter order follows field order); None when the class has none."""
580+ last = None
581+ for stmt in cls_node.body:
582+ if isinstance(stmt, ast.AnnAssign):
583+ end = stmt.end_lineno or stmt.lineno
584+ last = end if last is None else max(last, end)
585+ return last
586+ 
587+ 
588+def _new_binding_insert_reason(
589+ add_texts,
590+ a: int,
591+ b: int,
592+ add_indent: int | None,
593+ ann_ctx: tuple | None,
594+) -> str | None:
595+ """Skip reason when a pure insertion consists solely of assignments that
596+ bind provably-new names with inert right-hand sides; None to record.
597+ 
598+ Base code cannot read a name that does not exist yet, so every reader of
599+ a first-time binding is itself an added line carrying its own anchor -
600+ the anchor here is noise (an import-time fingerprint in module/class
601+ scopes). Gates: no meaningful line ends in a comma (such lines are call
602+ or signature arguments, not statements); the block is all simple-name
603+ assignments; no target is a
604+ dunder or already occurs anywhere in the base file (Name occurrences,
605+ def/class/import/parameter bindings); right-hand sides are inert; and in
606+ class scope the class must be plain or dataclass-like, where plain
607+ assigns are never fields while annotated fields are only exempt when
608+ inserted at or after the last existing field (a middle field insertion
609+ would silently rebind positional constructor calls).
610+ """
611+ if ann_ctx is None or not add_texts:
612+ return None
613+ effective = [t for t in add_texts if t.strip() and not t.strip().startswith("#")]
614+ if not effective or effective[-1].rstrip().endswith(","):
615+ return None
616+ scopes, class_map, has_future, known_names = ann_ctx
617+ stmts = _parse_assignments(add_texts)
618+ if not stmts:
619+ return None
620+ for stmt, names in stmts:
621+ if any(name.startswith("__") or name in known_names for name in names):
622+ return None
623+ if not _assign_rhs_lazy(stmt.value):
624+ return None
625+ scope = _insertion_scope(scopes, a, b, add_indent)
626+ if isinstance(scope, ast.ClassDef) and _class_consumes_annotations(scope, class_map):
627+ if not _class_dataclass_like(scope, class_map):
628+ return None
629+ if any(isinstance(stmt, ast.AnnAssign) for stmt, _ in stmts):
630+ last_field = _last_field_line(scope)
631+ if last_field is not None and a < last_field:
632+ return None
633+ return "first-time binding of new name(s)"
634+ 
635+ 
449def _classify_candidate_pairs(636def _classify_candidate_pairs(
450 affected: set[int],637 affected: set[int],
451 pairs: list[tuple],638 pairs: list[tuple],
@@ -460,10 +647,17 @@ def _classify_candidate_pairs(
460 comment/docstring line in the base file AND the added lines are comments or647 comment/docstring line in the base file AND the added lines are comments or
461 doc prose (not parseable Python), i.e. a pure comment/docstring change.648 doc prose (not parseable Python), i.e. a pure comment/docstring change.
462 Pure type annotation changes are dropped in every direction (insertion,649 Pure type annotation changes are dropped in every direction (insertion,
463- replacement, deletion) when provably inert: function-local annotations are650+ replacement, deletion) when provably inert: function-local annotations
464- never evaluated; class-level ones need a plain class (no decorators, no651+ are never evaluated; class-level ones need a plain class (no decorators,
465- TypedDict/Protocol/Enum/BaseModel/NamedTuple base) and side-effect-free652+ no TypedDict/Protocol/Enum/BaseModel/NamedTuple base) and side-effect-free
466 annotation expressions unless the file has future annotations.653 annotation expressions unless the file has future annotations.
654+ First-time bindings are dropped when a pure insertion (no meaningful
655+ line ends in a comma - those are call/signature arguments) consists
656+ solely of assignments binding simple names that occur nowhere in the
657+ base file with inert right-hand sides (base code cannot read a name that
658+ does not exist yet; readers added in the same PR carry their own anchors);
659+ in dataclass-like classes annotated fields are only dropped when inserted
660+ at or after the last existing field.
467 Candidate pairs: without base content (or non-parseable Python) both sides661 Candidate pairs: without base content (or non-parseable Python) both sides
468 of each pair are counted, bounded by the hunk.662 of each pair are counted, bounded by the hunk.
469 """663 """
@@ -532,6 +726,8 @@ def _classify_candidate_pairs(
532 reason = "belongs to a newly added function/class"726 reason = "belongs to a newly added function/class"
533 if reason is None:727 if reason is None:
534 reason = _pure_annotation_insert_reason(add_texts, a, b, add_indent, ann_ctx)728 reason = _pure_annotation_insert_reason(add_texts, a, b, add_indent, ann_ctx)
729+ if reason is None:
730+ reason = _new_binding_insert_reason(add_texts, a, b, add_indent, ann_ctx)
535 # kind == 'blank': isolated blank deletion == one-line insertion731 # kind == 'blank': isolated blank deletion == one-line insertion
536 elif _between_definitions(a, b, end_lines, start_lines, blanks):732 elif _between_definitions(a, b, end_lines, start_lines, blanks):
537 reason = "between function/class definitions"733 reason = "between function/class definitions"
Mprecision-test/dist/test_selector/repos/__init__.py+2-2
@@ -11,7 +11,7 @@ __all__ = ["RepoAdapter", "VllmAscendAdapter", "SglangAdapter", "TorchNpuAdapter
11_ADAPTERS = {11_ADAPTERS = {
12 "vllm_ascend": VllmAscendAdapter,12 "vllm_ascend": VllmAscendAdapter,
13 "sglang": SglangAdapter,13 "sglang": SglangAdapter,
14- "torch_npu": TorchNpuAdapter,14+ "pytorch": TorchNpuAdapter,
15}15}
16 16 
17AVAILABLE_REPOS = tuple(_ADAPTERS.keys())17AVAILABLE_REPOS = tuple(_ADAPTERS.keys())
@@ -21,7 +21,7 @@ def get_adapter(repo_name: str) -> RepoAdapter:
21 """按仓库名获取适配器实例。21 """按仓库名获取适配器实例。
22 22 
23 Args:23 Args:
24- repo_name: 仓库名（vllm_ascend / sglang / torch_npu）24+ repo_name: 仓库名（vllm_ascend / sglang / pytorch）
25 25 
26 Returns:26 Returns:
27 RepoAdapter 实例27 RepoAdapter 实例
Mprecision-test/dist/test_selector/repos/base.py+1-1
@@ -21,7 +21,7 @@ class RepoAdapter(ABC):
21 #: 覆盖率数据目录下的测试用例文件夹命名前缀；None 表示无固定前缀（vllm 用 tests__ 前缀 + cpu-ut）21 #: 覆盖率数据目录下的测试用例文件夹命名前缀；None 表示无固定前缀（vllm 用 tests__ 前缀 + cpu-ut）
22 test_case_dir_prefix: str | None = None22 test_case_dir_prefix: str | None = None
23 #: 覆盖率文件名 glob 模式（测试目录或其 covdata/ 子目录下）。23 #: 覆盖率文件名 glob 模式（测试目录或其 covdata/ 子目录下）。
24- #: 默认 "coverage*" 兼容 pytest-cov 裸文件（torch_npu: coverage）与原始格式24+ #: 默认 "coverage*" 兼容 pytest-cov 裸文件（pytorch: coverage）与原始格式
25 #: 带后缀文件（vllm/sglang: coverage.linux-...-workflow.xxx）。25 #: 带后缀文件（vllm/sglang: coverage.linux-...-workflow.xxx）。
26 coverage_file_glob: str = "coverage*"26 coverage_file_glob: str = "coverage*"
27 27 
Mprecision-test/dist/test_selector/repos/torch_npu.py+20-17
@@ -1,14 +1,16 @@
1-"""torch_npu 仓库适配器（PyTorch / GitCode 社区）。1+"""torch_npu 仓库适配器（检测上游 PyTorch：https://github.com/pytorch/pytorch）。
2 2 
3-- REPO_NAME = "torch_npu"，产品代码前缀 "torch_npu/"3+- REPO_NAME = "torch"，产品代码前缀 "torch/"，PR 源为 GitHub（--github-pr pytorch/pytorch#N）
4-- 测试用例目录：目录名含 "__" 或以 "test_" 开头（如 test__xxx 编码路径），4+- 测试用例目录：目录名含 "__" 或以 "test_" 开头（如 _inductor__test_add），
5 覆盖率文件位于 covdata/ 子目录5 覆盖率文件位于 covdata/ 子目录
6- 测试文件规则：test/ 下 test_*.py6- 测试文件规则：test/ 下 test_*.py
7- 测试名规范化：-- -> ::，__ -> /，文件级补 .py7- 测试名规范化：-- -> ::，__ -> /，文件级补 .py
8- 粒度默认值跟随合并版：line=True, function=True8- 粒度默认值跟随合并版：line=True, function=True
9-- 源文件解析使用基类默认候选（source_dir/torch_npu/<rel>，运行需 -s 指向 covstub）9+- 源文件解析使用基类默认候选（source_dir/torch/<rel>，运行需 -s 指向 covstub，
10+ covstub/torch/ 对应 pytorch/pytorch 仓库根目录）
10"""11"""
11 12 
13+import re
12from fnmatch import fnmatchcase14from fnmatch import fnmatchcase
13from pathlib import PurePosixPath15from pathlib import PurePosixPath
14 16 
@@ -18,10 +20,14 @@ from .base import RepoAdapter
18class TorchNpuAdapter(RepoAdapter):20class TorchNpuAdapter(RepoAdapter):
19 """torch_npu 仓库适配器。"""21 """torch_npu 仓库适配器。"""
20 22 
21- repo_name = "torch_npu"23+ repo_name = "torch"
22- product_code_prefix = "torch_npu/"24+ product_code_prefix = "torch/"
23 test_case_dir_prefix = None25 test_case_dir_prefix = None
24 26 
27+ #: 精确匹配路径中独立的 torch/ 段（前为路径起点或 /），
28+ #: 避免 pytorch/、torchvision/、test_torch.py 等含 torch 子串的路径被误剥
29+ _TORCH_PATH_RE = re.compile(r"(?:^|/)torch/(.+)")
30+ 
25 #: 可纳入 PR 检测的测试根目录31 #: 可纳入 PR 检测的测试根目录
26 TEST_ROOTS = ("test/",)32 TEST_ROOTS = ("test/",)
27 33 
@@ -30,7 +36,7 @@ class TorchNpuAdapter(RepoAdapter):
30 # ------------------------------------------------------------------36 # ------------------------------------------------------------------
31 def is_test_case_dir(self, name: str) -> bool:37 def is_test_case_dir(self, name: str) -> bool:
32 # PyTorch: "__" in name or name.startswith("test_")38 # PyTorch: "__" in name or name.startswith("test_")
33- return ("__" in name) or name.startswith("test_")39+ return name.endswith(".py") or ("__" in name) or name.startswith("test_")
34 40 
35 def normalize_test_name(self, test_name: str) -> str:41 def normalize_test_name(self, test_name: str) -> str:
36 """42 """
@@ -40,19 +46,16 @@ class TorchNpuAdapter(RepoAdapter):
40 """46 """
41 # 先转换 -- 为 ::（函数级测试标记）47 # 先转换 -- 为 ::（函数级测试标记）
42 result = test_name.replace("--", "::")48 result = test_name.replace("--", "::")
43- # 再转换 __ 为 /49+ if result.endswith(".py") or (".py::" in result):
44- result = result.replace("__", "/")50+ if not result.startswith("test/"):
45- # 无 ::（文件级测试）则补 .py 后缀51+ result = f"test/{result}"
46- if "::" not in result:52+ return result
47- result = result + ".py"
48 return result53 return result
49 54 
50 def covered_path_to_rel(self, path: str) -> str | None:55 def covered_path_to_rel(self, path: str) -> str | None:
51- """覆盖率库路径包含 torch_npu/ 时剥离前缀得到相对路径，否则返回 None。"""56+ """覆盖率库路径含独立的 torch/ 段时剥离前缀得到相对路径，否则返回 None。"""
52- if self.repo_name not in path:57+ match = self._TORCH_PATH_RE.search(path)
53- return None58+ return match.group(1) if match else None
54- marker = f"{self.repo_name}/"
55- return path.split(marker)[-1] if marker in path else path
56 59 
57 # ------------------------------------------------------------------60 # ------------------------------------------------------------------
58 # 测试文件检测（PR 内容检测）61 # 测试文件检测（PR 内容检测）
Mprecision-test/src/README.md+63-62
@@ -8,10 +8,10 @@
8 8 
9```bash9```bash
10python -m test_selector --repo vllm_ascend --github-pr "vllm-project/vllm-ascend#12379"10python -m test_selector --repo vllm_ascend --github-pr "vllm-project/vllm-ascend#12379"
11-python -m test_selector --repo torch_npu --gitcode-pr "Ascend/pytorch#46780"11+python -m test_selector --repo torch_npu --github-pr "pytorch/pytorch#150000"
12```12```
13 13 
14-> **PR 源选择**：vllm_ascend / sglang 使用 GitHub（`--github-pr`），torch_npu 使用 GitCode（`--gitcode-pr`）。14+> **PR 源选择**：三仓库均使用 GitHub（`--github-pr`）；torch_npu 检测上游仓库 [pytorch/pytorch](https://github.com/pytorch/pytorch)。
15 15 
16**核心工作流**（main 函数结构）：16**核心工作流**（main 函数结构）：
17 17 
@@ -27,7 +27,7 @@ python -m test_selector --repo torch_npu --gitcode-pr "Ascend/pytorch#46780"
27├─────────────────────────────────────────────────────────────────┤27├─────────────────────────────────────────────────────────────────┤
28│ 动作2: 全量触发变更检测（adapter.has_full_suite_changes） │28│ 动作2: 全量触发变更检测（adapter.has_full_suite_changes） │
29│ - vllm_ascend: csrc/ 目录变更（非 .md）→ 全量测试 │29│ - vllm_ascend: csrc/ 目录变更（非 .md）→ 全量测试 │
30-│ - sglang / torch_npu: 恒为 False（原生/子模块变更走精准匹配） │30+│ - sglang / torch_npu: 恒为 False（原生代码变更不解析进精准匹配） │
31├─────────────────────────────────────────────────────────────────┤31├─────────────────────────────────────────────────────────────────┤
32│ 动作3: 产品代码变更检测（change_detector） │32│ 动作3: 产品代码变更检测（change_detector） │
33│ - parse_pr_diff_file() → changed_files_with_lines, renames │33│ - parse_pr_diff_file() → changed_files_with_lines, renames │
@@ -49,7 +49,7 @@ python -m test_selector --repo torch_npu --gitcode-pr "Ascend/pytorch#46780"
49- **精准匹配**：无全量触发时，基于行/函数/文件级匹配选择测试49- **精准匹配**：无全量触发时，基于行/函数/文件级匹配选择测试
50- **文件重命名检测**：rename 新路径在覆盖率数据中不存在，用旧路径做文件级匹配召回50- **文件重命名检测**：rename 新路径在覆盖率数据中不存在，用旧路径做文件级匹配召回
51- **新增/删除处理**：新增测试文件直接加入；已删除测试文件从结果中移除51- **新增/删除处理**：新增测试文件直接加入；已删除测试文件从结果中移除
52-- **排除变更场景**：纯注释/docstring、纯类型注解、函数/类定义间空行插入、新增 def/class 整体、新文件等不构成代码变更的场景，在 diff 解析阶段剔除（见第五章）52+- **排除变更场景**：纯注释/docstring、纯类型注解、首次新增的变量绑定、函数/类定义间空行插入、新增 def/class 整体、新文件等不构成代码变更的场景，在 diff 解析阶段剔除（见第五章）
53 53 
54**代码结构**：54**代码结构**：
55 55 
@@ -59,7 +59,7 @@ test_selector/
59├── __main__.py # python -m test_selector 入口59├── __main__.py # python -m test_selector 入口
60├── github.py # PR 拉取（公共）60├── github.py # PR 拉取（公共）
61├── pr_detector.py # 检测 PR 内容：测试文件/删除文件/全量触发（规则委托 adapter）61├── pr_detector.py # 检测 PR 内容：测试文件/删除文件/全量触发（规则委托 adapter）
62-├── diff_parser.py # 排除变更场景：纯注释/docstring/纯注解/新文件跳过等62+├── diff_parser.py # 排除变更场景：纯注释/docstring/纯注解/首次绑定/新文件跳过等
63├── noise_filter.py # map 噪音处理：import/def/class/docstring/空行63├── noise_filter.py # map 噪音处理：import/def/class/docstring/空行
64├── function_parser.py # 函数区间解析（sglang 优化版，区间线性扫描）64├── function_parser.py # 函数区间解析（sglang 优化版，区间线性扫描）
65├── coverage_selector.py # 构建 测试用例→覆盖文件/行号 映射65├── coverage_selector.py # 构建 测试用例→覆盖文件/行号 映射
@@ -69,7 +69,7 @@ test_selector/
69 ├── base.py # RepoAdapter 抽象接口69 ├── base.py # RepoAdapter 抽象接口
70 ├── vllm_ascend.py # vllm_ascend 适配器70 ├── vllm_ascend.py # vllm_ascend 适配器
71 ├── sglang.py # sglang 适配器71 ├── sglang.py # sglang 适配器
72- ├── torch_npu.py # torch_npu（PyTorch / GitCode）适配器72+ ├── torch_npu.py # torch_npu（上游 PyTorch / GitHub）适配器
73 └── __init__.py # 适配器注册表 get_adapter()73 └── __init__.py # 适配器注册表 get_adapter()
74```74```
75 75 
@@ -97,22 +97,22 @@ test_selector/
97│ ├── <sglang 布局>/97│ ├── <sglang 布局>/
98│ │ └── ____w__sglang__sglang__test__registered__npu__...__test_xxx/98│ │ └── ____w__sglang__sglang__test__registered__npu__...__test_xxx/
99│ │ └── coverage.linux-* # sglang：覆盖率文件直接放在测试目录下99│ │ └── coverage.linux-* # sglang：覆盖率文件直接放在测试目录下
100-│ └── <torch_npu 布局>/100+│ └── <pytorch 布局>/
101-│ └── _inductor__test_add/ # torch_npu：目录名含 __ 或以 test_ 开头101+│ └── _inductor__test_add/ # pytorch：目录名含 __ 或以 test_ 开头
102-│ └── covdata/ # torch_npu：覆盖率文件在 covdata/ 子目录102+│ └── covdata/ # pytorch：覆盖率文件在 covdata/ 子目录
103│ └── coverage.*103│ └── coverage.*
104│104│
105├── covstub/ # 源码目录（函数级匹配需要）105├── covstub/ # 源码目录（函数级匹配需要）
106│ ├── vllm_ascend/ # vllm：对应仓库根目录106│ ├── vllm_ascend/ # vllm：对应仓库根目录
107│ ├── sglang/ # sglang：对应仓库 python/sglang/107│ ├── sglang/ # sglang：对应仓库 python/sglang/
108-│ └── torch_npu/ # torch_npu：对应 PyTorch 仓库根目录108+│ └── torch/ # torch_npu：对应 pytorch/pytorch 仓库根目录
109├── test_case_map.json # 自动生成的映射文件109├── test_case_map.json # 自动生成的映射文件
110└── recommended_pytest_paths.txt # 推荐测试用例列表（输出）110└── recommended_pytest_paths.txt # 推荐测试用例列表（输出）
111```111```
112 112 
113**覆盖率目录布局自动探测**（三仓库兼容）：113**覆盖率目录布局自动探测**（三仓库兼容）：
114 114 
115-- 测试目录下存在 `covdata/` 子目录（vllm / torch_npu 布局）→ 从 `covdata/` 读取 `coverage.*`115+- 测试目录下存在 `covdata/` 子目录（vllm / pytorch 布局）→ 从 `covdata/` 读取 `coverage.*`
116- 无 `covdata/` 子目录（sglang 布局）→ 直接读取测试目录下的 `coverage.*`116- 无 `covdata/` 子目录（sglang 布局）→ 直接读取测试目录下的 `coverage.*`
117 117 
118**测试用例目录识别规则**（adapter 提供）：118**测试用例目录识别规则**（adapter 提供）：
@@ -137,7 +137,7 @@ test_selector/
137| ----------- | ---------------- | ----------------------------------------------------------------- |137| ----------- | ---------------- | ----------------------------------------------------------------- |
138| vllm_ascend | `vllm_ascend/` | `vllm_ascend/core/worker.py` → `core/worker.py` |138| vllm_ascend | `vllm_ascend/` | `vllm_ascend/core/worker.py` → `core/worker.py` |
139| sglang | `python/sglang/` | `python/sglang/srt/models/qwen3_vl.py` → `srt/models/qwen3_vl.py` |139| sglang | `python/sglang/` | `python/sglang/srt/models/qwen3_vl.py` → `srt/models/qwen3_vl.py` |
140-| torch_npu | `torch_npu/` | `torch_npu/contrib/xxx.py` → `contrib/xxx.py` |140+| torch_npu | `torch/` | `torch/distributed/utils.py` → `distributed/utils.py` |
141 141 
142---142---
143 143 
@@ -189,9 +189,9 @@ python -m test_selector --repo sglang \
189 --github-pr "sgl-project/sglang#37043" \189 --github-pr "sgl-project/sglang#37043" \
190 --source-dir ./covstub190 --source-dir ./covstub
191 191 
192-# torch_npu（GitCode）192+# torch_npu（GitHub 上游 pytorch/pytorch）
193python -m test_selector --repo torch_npu \193python -m test_selector --repo torch_npu \
194- --gitcode-pr "Ascend/pytorch#46780" \194+ --github-pr "pytorch/pytorch#150000" \
195 --source-dir ./covstub195 --source-dir ./covstub
196```196```
197 197 
@@ -205,18 +205,18 @@ python -m test_selector --repo torch_npu \
205 205 
206### 公共参数206### 公共参数
207 207 
208-| 参数 | 必填 | 说明 |208+| 参数 | 必填 | 说明 |
209-| ----------------------- | ----------- | ------------------------------------------------------------------------------------------------------------ |209+| ----------------------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------- |
210-| `--repo` / `-r` | 否 | 仓库适配器：`vllm_ascend` / `sglang` / `torch_npu`（默认：`vllm_ascend`，薄入口默认各自仓库） |210+| `--repo` / `-r` | 否 | 仓库适配器：`vllm_ascend` / `sglang` / `torch_npu`（默认：`vllm_ascend`，薄入口默认各自仓库） |
211-| `--github-pr` / `-pr` | 二选一 | GitHub PR（vllm_ascend / sglang），格式：`owner/repo#pr_number` 或仅 `pr_number`（自动从 git remote 推导） |211+| `--github-pr` / `-pr` | 二选一 | GitHub PR（三仓库通用），格式：`owner/repo#pr_number` 或仅 `pr_number`（自动从 git remote 推导）；torch_npu 传上游 `pytorch/pytorch#N` |
212-| `--gitcode-pr` | 二选一 | GitCode PR（torch_npu），格式：`owner/repo#pr_number` 或仅 `pr_number`（自动从 git remote 推导）；两参数互斥 |212+| `--gitcode-pr` | 二选一 | GitCode PR（公共备用模块，当前无仓库默认使用），格式：`owner/repo#pr_number` 或仅 `pr_number`（自动从 git remote 推导）；两参数互斥 |
213-| `--source-dir` / `-s` | 建议 | 源码目录（默认：`covstub`；函数级匹配与噪音过滤需要） |213+| `--source-dir` / `-s` | 建议 | 源码目录（默认：`covstub`；函数级匹配与噪音过滤需要） |
214-| `--map-file` / `-m` | 否 | 映射文件（默认：`test_case_map.json`） |214+| `--map-file` / `-m` | 否 | 映射文件（默认：`test_case_map.json`） |
215-| `--coverage-dir` / `-c` | 构建 map 时 | 覆盖率数据目录（默认：`coverage`） |215+| `--coverage-dir` / `-c` | 构建 map 时 | 覆盖率数据目录（默认：`coverage`） |
216-| `--build-map` / `-b` | 否 | 强制重建测试用例映射；仅构建 map 时使用 |216+| `--build-map` / `-b` | 否 | 强制重建测试用例映射；仅构建 map 时使用 |
217-| `--min-affected` / `-a` | 否 | 最少受影响行数阈值（默认：1） |217+| `--min-affected` / `-a` | 否 | 最少受影响行数阈值（默认：1） |
218-| `--dedup` | 否 | 去重：相同覆盖行的测试只保留一个（默认关闭） |218+| `--dedup` | 否 | 去重：相同覆盖行的测试只保留一个（默认关闭） |
219-| `--skip-imports` | 否 | 函数级匹配时跳过 import 语句行（默认关闭） |219+| `--skip-imports` | 否 | 函数级匹配时跳过 import 语句行（默认关闭） |
220 220 
221### 匹配粒度开关221### 匹配粒度开关
222 222 
@@ -242,12 +242,12 @@ if args.disable_function_match:
242 242 
243| 维度 | vllm_ascend | sglang | torch_npu |243| 维度 | vllm_ascend | sglang | torch_npu |
244| -------------- | ----------------------------------------------------- | ----------------------------------------------------------------- | ----------------------------------- |244| -------------- | ----------------------------------------------------- | ----------------------------------------------------------------- | ----------------------------------- |
245-| PR 源 | GitHub | GitHub | GitCode |245+| PR 源 | GitHub | GitHub | GitHub（pytorch/pytorch） |
246-| 产品代码前缀 | `vllm_ascend/` | `python/sglang/` | `torch_npu/` |246+| 产品代码前缀 | `vllm_ascend/` | `python/sglang/` | `torch/` |
247| 测试目录识别 | `tests__` 前缀 / `cpu-ut` | `____w__sglang__sglang__test__` 前缀 | 目录名含 `__` 或以 `test_` 开头 |247| 测试目录识别 | `tests__` 前缀 / `cpu-ut` | `____w__sglang__sglang__test__` 前缀 | 目录名含 `__` 或以 `test_` 开头 |
248| 覆盖率文件位置 | `covdata/` 子目录 | 测试目录下直接放置（兼容探测 covdata） | `covdata/` 子目录 |248| 覆盖率文件位置 | `covdata/` 子目录 | 测试目录下直接放置（兼容探测 covdata） | `covdata/` 子目录 |
249| 测试文件规则 | `tests/e2e/pull_request/`、`tests/ut/` 下 `test_*.py` | `test/registered/`、`test/{unit,e2e,integration}/` 下 `test_*.py` | `test/` 下 `test_*.py` |249| 测试文件规则 | `tests/e2e/pull_request/`、`tests/ut/` 下 `test_*.py` | `test/registered/`、`test/{unit,e2e,integration}/` 下 `test_*.py` | `test/` 下 `test_*.py` |
250-| 全量触发变更 | csrc/ 目录（非 .md）→ 全量测试 | 无（csrc/rust 走精准匹配） | 无（submodule/原生变更走精准匹配） |250+| 全量触发变更 | csrc/ 目录（非 .md）→ 全量测试 | 无（csrc/rust 走精准匹配） | 无（原生变更走精准匹配） |
251| 测试名规范化 | `--`→`::`、`__`→`/`、文件级补 `.py` | 剥离编码前缀恢复 `test/`、`__`→`/`、`--`→`::` | `--`→`::`、`__`→`/`、文件级补 `.py` |251| 测试名规范化 | `--`→`::`、`__`→`/`、文件级补 `.py` | 剥离编码前缀恢复 `test/`、`__`→`/`、`--`→`::` | `--`→`::`、`__`→`/`、文件级补 `.py` |
252| 变更检测 | 本地哈希比对（默认）或 PR diff | 同左 | 同左 |252| 变更检测 | 本地哈希比对（默认）或 PR diff | 同左 | 同左 |
253 253 
@@ -270,21 +270,22 @@ if args.disable_function_match:
270 270 
271> sglang 不触发全量：`csrc/*.cu`、`rust/*.rs` 等原生代码变更不会被解析进精准匹配（仅保留 `.py`），推荐结果由 Python 产品代码匹配 + 新增/删除测试文件决定。271> sglang 不触发全量：`csrc/*.cu`、`rust/*.rs` 等原生代码变更不会被解析进精准匹配（仅保留 `.py`），推荐结果由 Python 产品代码匹配 + 新增/删除测试文件决定。
272>272>
273-> torch_npu 不触发全量：submodule 指针更新（如 `third_party/torchair/torchair` 的 commit 变更）与原生代码变更均不会被解析进精准匹配。**注意**：submodule 指针更新意味着子模块内部有真实代码变更，但工具无法穿透到子模块 diff，此时可能返回 0 推荐，存在漏测风险，建议人工评估（详见故障排除表）。273+> torch_npu 不触发全量：`torch/csrc/` 等原生代码（C++/CUDA）变更不会被解析进精准匹配（仅保留 `.py`）。**注意**：纯原生代码变更的 PR 可能返回 0 推荐，此时存在漏测风险，建议人工评估是否需跑相关用例（详见故障排除表）。
274 274 
275### 排除变更场景（diff_parser）275### 排除变更场景（diff_parser）
276 276 
277`_parse_diff_base_lines()` 解析统一 diff 文本为受影响的 base（变更前）行号，并在解析阶段/后续分类阶段排除以下**不构成代码变更**的场景：277`_parse_diff_base_lines()` 解析统一 diff 文本为受影响的 base（变更前）行号，并在解析阶段/后续分类阶段排除以下**不构成代码变更**的场景：
278 278 
279-| 场景 | 处理 |279+| 场景 | 处理 |
280-| ----------------------------------------- | ------------------------------------------------------------ |280+| ----------------------------------------- | -------------------------------------------------------------------------------- |
281-| 纯注释/docstring 变更 | 删除组全为注释/docstring 且新增为注释/doc 说明 → 剔除 |281+| 纯注释/docstring 变更 | 删除组全为注释/docstring 且新增为注释/doc 说明 → 剔除 |
282-| 纯类型注解变更（插入/替换/删除） | 可证明惰性的注解（函数体内、普通类、模块级，无副作用）→ 豁免 |282+| 纯类型注解变更（插入/替换/删除） | 可证明惰性的注解（函数体内、普通类、模块级，无副作用）→ 豁免 |
283-| 函数/类定义之间的空行插入 | 位于两个 def/class 之间 → 排除 |283+| 首次新增的变量绑定（纯插入） | 全部为绑定全新名字的带值赋值（名字在 base 中不存在、RHS 惰性、无尾逗号行）→ 豁免 |
284-| 新增 def/class 整体 | 插入文本属于新定义的函数/类 → 排除 |284+| 函数/类定义之间的空行插入 | 位于两个 def/class 之间 → 排除 |
285-| 新文件（`--- /dev/null` + `@@ -0,0 ...`） | 无 base 版本，跳过行级解析（避免无意义的 base 内容拉取） |285+| 新增 def/class 整体 | 插入文本属于新定义的函数/类 → 排除 |
286-| 函数体内纯插入 | 记录插入位置上方的行号 |286+| 新文件（`--- /dev/null` + `@@ -0,0 ...`） | 无 base 版本，跳过行级解析（避免无意义的 base 内容拉取） |
287-| 隔离空行删除 | 按单行插入处理为候选对（上、下两行） |287+| 函数体内纯插入 | 记录插入位置上方的行号 |
288+| 隔离空行删除 | 按单行插入处理为候选对（上、下两行） |
288 289 
289**候选对分类**（`_classify_candidate_pairs`，需要 base 文件内容）：通过 GitHub contents API / GitCode raw 接口单路拉取 base 内容后做 AST 分类；拉取失败重试 3 次后直接退出（不允许降级），仅 AST 解析失败等无法分类的场景回退为 hunk 范围内双侧计数。290**候选对分类**（`_classify_candidate_pairs`，需要 base 文件内容）：通过 GitHub contents API / GitCode raw 接口单路拉取 base 内容后做 AST 分类；拉取失败重试 3 次后直接退出（不允许降级），仅 AST 解析失败等无法分类的场景回退为 hunk 范围内双侧计数。
290 291 
@@ -387,9 +388,9 @@ for path, label in file_level_paths: # rename 旧路径 + 删除路径
387- **PR 模式**：`parse_pr_diff_file()` 解析 PR diff，产出变更行号 + 重命名映射 + 删除列表388- **PR 模式**：`parse_pr_diff_file()` 解析 PR diff，产出变更行号 + 重命名映射 + 删除列表
388- **本地模式**（默认）：`detect_changes_by_comparison()` 扫描 `--source-dir` 下全部 `.py`，与 `.file_hashes.json` 基线比对 MD5，变更文件保守返回全部行号（1~9999），首次运行生成基线389- **本地模式**（默认）：`detect_changes_by_comparison()` 扫描 `--source-dir` 下全部 `.py`，与 `.file_hashes.json` 基线比对 MD5，变更文件保守返回全部行号（1~9999），首次运行生成基线
389 390 
390-### GitCode PR 拉取（torch_npu）391+### GitCode PR 拉取（公共备用模块）
391 392 
392-torch_npu 的 PR 拉取通过公共模块 `test_selector.gitcode`（`--gitcode-pr`），与 GitHub 版（`test_selector.github`，`--github-pr`）的差异：393+GitCode PR 拉取通过公共模块 `test_selector.gitcode`（`--gitcode-pr`，当前无仓库默认使用；torch_npu 已切换到 GitHub 上游仓库 pytorch/pytorch），与 GitHub 版（`test_selector.github`，`--github-pr`）的差异：
393 394 
394| 维度 | github.py | gitcode.py |395| 维度 | github.py | gitcode.py |
395| ----------------- | -------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |396| ----------------- | -------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
@@ -454,7 +455,7 @@ torch_npu 的 PR 拉取通过公共模块 `test_selector.gitcode`（`--gitcode-p
454**注意**：455**注意**：
455 456 
456- map 的 key 为 normalize 后的测试路径（vllm：`tests/.../test_xxx.py` 或 `...::test_func`、`cpu-ut`；sglang：`test/.../test_xxx.py`；torch_npu：`test/.../test_xxx.py`、`_inductor/test_add.py` 等，文件级补 `.py`）457- map 的 key 为 normalize 后的测试路径（vllm：`tests/.../test_xxx.py` 或 `...::test_func`、`cpu-ut`；sglang：`test/.../test_xxx.py`；torch_npu：`test/.../test_xxx.py`、`_inductor/test_add.py` 等，文件级补 `.py`）
457-- files 的 key 为剥离产品代码前缀后的相对路径（vllm：`core/worker.py`；sglang：`srt/xxx.py`；torch_npu：`contrib/xxx.py`）458+- files 的 key 为剥离产品代码前缀后的相对路径（vllm：`core/worker.py`；sglang：`srt/xxx.py`；torch_npu：`distributed/xxx.py`）
458- 保存时强制使用 LF 换行（`newline="\n"`），避免 Windows CRLF 导致 JSON 体积膨胀，保证跨平台 git 差异对比一致459- 保存时强制使用 LF 换行（`newline="\n"`），避免 Windows CRLF 导致 JSON 体积膨胀，保证跨平台 git 差异对比一致
459 460 
460---461---
@@ -531,9 +532,9 @@ python -m test_selector --repo vllm_ascend \
531python -m test_selector --repo sglang \532python -m test_selector --repo sglang \
532 --github-pr "sgl-project/sglang#37043" -s ./covstub533 --github-pr "sgl-project/sglang#37043" -s ./covstub
533 534 
534-# GitCode PR（torch_npu）535+# GitHub PR（torch_npu，上游 pytorch/pytorch）
535python -m test_selector --repo torch_npu \536python -m test_selector --repo torch_npu \
536- --gitcode-pr "Ascend/pytorch#46780" -s ./covstub537+ --github-pr "pytorch/pytorch#150000" -s ./covstub
537```538```
538 539 
539### 场景2：更新覆盖率数据后重建映射540### 场景2：更新覆盖率数据后重建映射
@@ -592,23 +593,23 @@ selected, reason = ts.select_tests(changed, source_dir="covstub")
592 593 
593## 九、故障排除594## 九、故障排除
594 595 
595-| 问题 | 可能原因 | 解决方案 |596+| 问题 | 可能原因 | 解决方案 |
596-| ------------------------------------------------------------------- | --------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |597+| ------------------------------------------------------------------- | ------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
597-| `Error: --coverage-dir is required when building the test case map` | map 文件不存在且未传覆盖率目录 | 首次使用需传 `--coverage-dir`（或先 `--build-map`） |598+| `Error: --coverage-dir is required when building the test case map` | map 文件不存在且未传覆盖率目录 | 首次使用需传 `--coverage-dir`（或先 `--build-map`） |
598-| `Coverage data directory not found` | 覆盖率目录路径不对 | 确认 `--coverage-dir` 指向含测试用例目录的根目录 |599+| `Coverage data directory not found` | 覆盖率目录路径不对 | 确认 `--coverage-dir` 指向含测试用例目录的根目录 |
599-| `没有测试用例覆盖变更的代码行` | 变更文件无覆盖数据 | 确认覆盖率数据来自变更前的全量用例 |600+| `没有测试用例覆盖变更的代码行` | 变更文件无覆盖数据 | 确认覆盖率数据来自变更前的全量用例 |
600-| `解析 PR 失败` / GitHub API 超时 | 内网无法直连 GitHub | 设置代理环境变量后重试（见下方） |601+| `解析 PR 失败` / GitHub API 超时 | 内网无法直连 GitHub | 设置代理环境变量后重试（见下方） |
601-| `WinError 10060` 连接超时（GitCode） | 内网无法直连 GitCode API | 设置代理环境变量后重试（见下方）；git 全局已配置 `proxycn2.huawei.com:8080` 认证代理时，PowerShell 中执行 `$env:HTTPS_PROXY=...` 后重试 |602+| `WinError 10060` 连接超时（GitCode） | 内网无法直连 GitCode API | 设置代理环境变量后重试（见下方）；git 全局已配置 `proxycn2.huawei.com:8080` 认证代理时，PowerShell 中执行 `$env:HTTPS_PROXY=...` 后重试 |
602-| `函数级匹配失败` | 源码目录路径不对 | 确认 `--source-dir` 指向包含 `vllm_ascend/` / `sglang/` / `torch_npu/` 包的根目录 |603+| `函数级匹配失败` | 源码目录路径不对 | 确认 `--source-dir` 指向包含 `vllm_ascend/` / `sglang/` / `torch/` 包的根目录 |
603-| 推荐结果过多 | 变更函数被大量测试覆盖 | 启用 `--dedup` 或提高 `--min-affected` |604+| 推荐结果过多 | 变更函数被大量测试覆盖 | 启用 `--dedup` 或提高 `--min-affected` |
604-| 推荐结果为空 | PR 仅含原生代码变更且无新测试文件（sglang） | 正常，原生代码变更不触发全量 |605+| 推荐结果为空 | PR 仅含原生代码变更且无新测试文件（sglang） | 正常，原生代码变更不触发全量 |
605-| 推荐结果为空 | PR 仅含 submodule 指针更新（torch_npu，如 `third_party/torchair/torchair`） | 正常但**存在漏测风险**：submodule commit 变更意味着子模块内部有真实代码变更，工具无法穿透；建议人工评估是否需跑 torchchair 相关用例 |606+| 推荐结果为空 | PR 仅含原生代码变更（torch_npu，如 `torch/csrc/` 下 C++/CUDA 文件） | 正常但**存在漏测风险**：原生代码变更不解析进精准匹配；建议人工评估是否需跑相关用例 |
606-| `Full-Suite Changes Detected`（vllm） | diff 含 csrc/ 目录变更 | 正常行为，触发全量测试 |607+| `Full-Suite Changes Detected`（vllm） | diff 含 csrc/ 目录变更 | 正常行为，触发全量测试 |
607-| `New Test Files Added: X` | diff 中有新增测试文件 | 正常，直接加入推荐列表 |608+| `New Test Files Added: X` | diff 中有新增测试文件 | 正常，直接加入推荐列表 |
608-| `Deleted Test Files Removed: X` | diff 中有已删除测试文件 | 正常，从推荐列表移除 |609+| `Deleted Test Files Removed: X` | diff 中有已删除测试文件 | 正常，从推荐列表移除 |
609-| `Detected X Renamed File(s)` | diff 中包含文件重命名 | 正常，旧路径做文件级匹配 |610+| `Detected X Renamed File(s)` | diff 中包含文件重命名 | 正常，旧路径做文件级匹配 |
610-| 变更行数偏多、含 context 行 | diff 解析按 hunk 范围推算 | 已知限制，算法保守偏多报 |611+| 变更行数偏多、含 context 行 | diff 解析按 hunk 范围推算 | 已知限制，算法保守偏多报 |
611-| `Unknown repo 'xxx'` | `--repo` 拼写错误 | 使用 `vllm_ascend` / `sglang` / `torch_npu` |612+| `Unknown repo 'xxx'` | `--repo` 拼写错误 | 使用 `vllm_ascend` / `sglang` / `torch_npu` |
612 613 
613**代理配置示例（华为内网环境）**：614**代理配置示例（华为内网环境）**：
614 615 
@@ -618,7 +619,7 @@ set HTTP_PROXY=http://<user>:<pwd>@proxycn2.huawei.com:8080/
618set HTTPS_PROXY=http://<user>:<pwd>@proxycn2.huawei.com:8080/619set HTTPS_PROXY=http://<user>:<pwd>@proxycn2.huawei.com:8080/
619 620 
620python -m test_selector --repo vllm_ascend --github-pr "vllm-project/vllm-ascend#16104" -s ./covstub621python -m test_selector --repo vllm_ascend --github-pr "vllm-project/vllm-ascend#16104" -s ./covstub
621-python -m test_selector --repo torch_npu --gitcode-pr "Ascend/pytorch#46780" -s ./covstub622+python -m test_selector --repo torch_npu --github-pr "pytorch/pytorch#150000" -s ./covstub
622```623```
623 624 
624> **说明**：625> **说明**：
Mprecision-test/src/__init__.py+2-2
@@ -1,10 +1,10 @@
1-"""test_selector - 覆盖率驱动的精准测试选择器（vllm_ascend / sglang / torch_npu 统一架构）。1+"""test_selector - 覆盖率驱动的精准测试选择器（vllm_ascend / sglang / pytorch 统一架构）。
2 2 
3公共逻辑集中在 test_selector/ 包（diff 解析、噪音过滤、覆盖率选择、CLI 等），3公共逻辑集中在 test_selector/ 包（diff 解析、噪音过滤、覆盖率选择、CLI 等），
4仓库差异通过 test_selector.repos.RepoAdapter 注入，保证架构统一、差异按仓库归类。4仓库差异通过 test_selector.repos.RepoAdapter 注入，保证架构统一、差异按仓库归类。
5 5 
6入口：6入口：
7- python -m test_selector --repo <vllm_ascend|sglang|torch_npu> ...7+ python -m test_selector --repo <vllm_ascend|sglang|pytorch> ...
8"""8"""
9 9 
10from .repos import AVAILABLE_REPOS, RepoAdapter, get_adapter10from .repos import AVAILABLE_REPOS, RepoAdapter, get_adapter
Mprecision-test/src/change_detector.py+7-0
@@ -106,6 +106,13 @@ class CodeChangeDetector:
106 NamedTuple base) or module level, unless the annotation expression106 NamedTuple base) or module level, unless the annotation expression
107 may have runtime side effects (calls, Annotated metadata) and the107 may have runtime side effects (calls, Annotated metadata) and the
108 file has no ``from __future__ import annotations``.108 file has no ``from __future__ import annotations``.
109+ - First-time bindings are excluded for pure insertions (needs base
110+ content): every added line is an assignment binding simple names
111+ that occur nowhere in the base file, with inert right-hand sides;
112+ no meaningful line ends in a comma (such lines are call/signature
113+ arguments, not statements); class-scope insertions follow dataclass
114+ field-order rules (plain assigns are not fields, annotated fields
115+ must land at or after the last existing field).
109 - Isolated blank-line deletion (neighbours not deleted): treated as a116 - Isolated blank-line deletion (neighbours not deleted): treated as a
110 one-line insertion -> candidate pair (line above, line below).117 one-line insertion -> candidate pair (line above, line below).
111 - Pure insertions and blank-deletion pairs are classified via ast of118 - Pure insertions and blank-deletion pairs are classified via ast of
Mprecision-test/src/cli.py+2-2
@@ -10,11 +10,11 @@
10 10 
11用法：11用法：
12 python -m test_selector --repo vllm_ascend --github-pr owner/repo#12312 python -m test_selector --repo vllm_ascend --github-pr owner/repo#123
13- python -m test_selector --repo torch_npu --gitcode-pr Ascend/pytorch#12313+ python -m test_selector --repo torch_npu --github-pr pytorch/pytorch#123
14 python -m test_selector --repo sglang --build-map --coverage-dir coverage14 python -m test_selector --repo sglang --build-map --coverage-dir coverage
15 python vllm/test_selector.py --github-pr 123 # 薄入口（默认 vllm_ascend）15 python vllm/test_selector.py --github-pr 123 # 薄入口（默认 vllm_ascend）
16 python sglang/test_selector.py --github-pr 123 # 薄入口（默认 sglang）16 python sglang/test_selector.py --github-pr 123 # 薄入口（默认 sglang）
17- python PyTorch/test_selector.py --gitcode-pr 123 # 薄入口（默认 torch_npu）17+ python PyTorch/test_selector.py --github-pr 123 # 薄入口（默认 torch_npu，检测上游 pytorch/pytorch）
18"""18"""
19 19 
20import argparse20import argparse
Mprecision-test/src/coverage_selector.py+5-4
@@ -49,20 +49,21 @@ class CoverageSelector:
49 if not self.coverage_data_dir or not self.coverage_data_dir.exists():49 if not self.coverage_data_dir or not self.coverage_data_dir.exists():
50 print(f" Warning: Coverage data directory not found: {self.coverage_data_dir}")50 print(f" Warning: Coverage data directory not found: {self.coverage_data_dir}")
51 return test_cases51 return test_cases
52- for item in self.coverage_data_dir.iterdir():52+ for item in self.coverage_data_dir.rglob("*"):
53 if not item.is_dir():53 if not item.is_dir():
54 continue54 continue
55 name = item.name55 name = item.name
56 # 目录名规则由仓库适配器决定（vllm: tests__ 前缀 / cpu-ut；sglang: ____w__... 前缀）56 # 目录名规则由仓库适配器决定（vllm: tests__ 前缀 / cpu-ut；sglang: ____w__... 前缀）
57 if not self.adapter.is_test_case_dir(name):57 if not self.adapter.is_test_case_dir(name):
58 continue58 continue
59- # 布局探测：covdata/ 子目录（vllm/torch_npu）或目录下直接放置（sglang）59+ # 布局探测：covdata/ 子目录（vllm/pytorch）或目录下直接放置（sglang）
60 covdata_dir = item / "covdata"60 covdata_dir = item / "covdata"
61 has_cov_files = any(item.glob(self.adapter.coverage_file_glob))61 has_cov_files = any(item.glob(self.adapter.coverage_file_glob))
62 if covdata_dir.exists() and any(covdata_dir.glob(self.adapter.coverage_file_glob)):62 if covdata_dir.exists() and any(covdata_dir.glob(self.adapter.coverage_file_glob)):
63 has_cov_files = True63 has_cov_files = True
64 if has_cov_files:64 if has_cov_files:
65- test_cases.append(name)65+ rel = item.relative_to(self.coverage_data_dir).as_posix()
66+ test_cases.append(rel)
66 return sorted(test_cases)67 return sorted(test_cases)
67 68 
68 def normalize_test_name(self, test_name: str) -> str:69 def normalize_test_name(self, test_name: str) -> str:
@@ -142,7 +143,7 @@ class CoverageSelector:
142 143 
143 file_lines_map = defaultdict(set) # filepath -> set of lines144 file_lines_map = defaultdict(set) # filepath -> set of lines
144 145 
145- # 布局探测：vllm/torch_npu 将 coverage 放在 covdata/ 子目录，sglang 直接放在测试目录下146+ # 布局探测：vllm/pytroch 将 coverage 放在 covdata/ 子目录，sglang 直接放在测试目录下
146 cov_dirs = [covdata_dir] if covdata_dir.exists() else [test_case_dir]147 cov_dirs = [covdata_dir] if covdata_dir.exists() else [test_case_dir]
147 148 
148 for cov_dir in cov_dirs:149 for cov_dir in cov_dirs:
Mprecision-test/src/diff_parser.py+209-13
@@ -4,6 +4,8 @@
4代码变更的场景：4代码变更的场景：
5- 纯注释/docstring 变更5- 纯注释/docstring 变更
6- 纯类型注解变更（插入/替换/删除三个方向，可证明惰性时豁免）6- 纯类型注解变更（插入/替换/删除三个方向，可证明惰性时豁免）
7+- 首次新增的变量绑定（纯插入块全部为绑定全新名字的带值赋值，RHS 惰性，
8+ 无尾逗号行——尾逗号行是调用/签名参数而非语句）
7- 函数/类定义之间的空行插入9- 函数/类定义之间的空行插入
8- 新增函数/类整体10- 新增函数/类整体
9 11 
@@ -318,14 +320,27 @@ def _class_consumes_annotations(cls_node: ast.ClassDef, class_map: dict) -> bool
318 return False320 return False
319 321 
320 322 
321-def _collect_class_info(source: str) -> tuple[list[tuple], dict[str, ast.ClassDef], bool]:323+def _arg_names(args: ast.arguments) -> set[str]:
322- """Parse source and return (scopes, class_map, has_future):324+ """Names bound by a function/lambda parameter list."""
325+ names = {a.arg for a in args.posonlyargs + args.args + args.kwonlyargs}
326+ if args.vararg:
327+ names.add(args.vararg.arg)
328+ if args.kwarg:
329+ names.add(args.kwarg.arg)
330+ return names
323 331 
324- scopes : [(lineno, end_lineno, col_offset, node)] of every332+ 
325- function/class definition, for annotation scope resolution;333+def _collect_class_info(source: str) -> tuple[list[tuple], dict[str, ast.ClassDef], bool, set[str]]:
326- class_map : {class name -> ClassDef} for same-file base resolution334+ """Parse source and return (scopes, class_map, has_future, known_names):
327- (a later definition shadows an earlier one);335+ 
328- has_future: the file has ``from __future__ import annotations``.336+ scopes : [(lineno, end_lineno, col_offset, node)] of every
337+ function/class definition, for annotation scope resolution;
338+ class_map : {class name -> ClassDef} for same-file base resolution
339+ (a later definition shadows an earlier one);
340+ has_future : the file has ``from __future__ import annotations``.
341+ known_names : every name the base file already binds or reads (Name
342+ occurrences, def/class names, import aliases, parameters),
343+ for first-time-binding detection.
329 """344 """
330 tree = ast.parse(source)345 tree = ast.parse(source)
331 scopes = [346 scopes = [
@@ -344,7 +359,24 @@ def _collect_class_info(source: str) -> tuple[list[tuple], dict[str, ast.ClassDe
344 and any(alias.name == "annotations" for alias in n.names)359 and any(alias.name == "annotations" for alias in n.names)
345 for n in tree.body360 for n in tree.body
346 )361 )
347- return scopes, class_map, has_future362+ known_names = set()
363+ for node in ast.walk(tree):
364+ if isinstance(node, ast.Name):
365+ known_names.add(node.id)
366+ elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
367+ known_names.add(node.name)
368+ known_names.update(_arg_names(node.args))
369+ elif isinstance(node, ast.ClassDef):
370+ known_names.add(node.name)
371+ elif isinstance(node, ast.Lambda):
372+ known_names.update(_arg_names(node.args))
373+ elif isinstance(node, ast.ImportFrom):
374+ for alias in node.names:
375+ known_names.add(alias.asname or alias.name)
376+ elif isinstance(node, ast.Import):
377+ for alias in node.names:
378+ known_names.add(alias.asname or alias.name.split(".")[0])
379+ return scopes, class_map, has_future, known_names
348 380 
349 381 
350def _insertion_scope(scopes: list[tuple], a: int, b: int, add_indent: int | None):382def _insertion_scope(scopes: list[tuple], a: int, b: int, add_indent: int | None):
@@ -391,7 +423,7 @@ def _pure_annotation_insert_reason(
391 """423 """
392 if ann_ctx is None or not add_texts:424 if ann_ctx is None or not add_texts:
393 return None425 return None
394- scopes, class_map, has_future = ann_ctx426+ scopes, class_map, has_future, _known = ann_ctx
395 anns = _parse_pure_annotations(add_texts)427 anns = _parse_pure_annotations(add_texts)
396 if not anns:428 if not anns:
397 return None429 return None
@@ -422,7 +454,7 @@ def _pure_annotation_change_exempt(
422 """454 """
423 if ann_ctx is None:455 if ann_ctx is None:
424 return False456 return False
425- scopes, class_map, has_future = ann_ctx457+ scopes, class_map, has_future, _known = ann_ctx
426 del_anns = _parse_pure_annotations([t for _, t in del_lines])458 del_anns = _parse_pure_annotations([t for _, t in del_lines])
427 if not del_anns:459 if not del_anns:
428 return False460 return False
@@ -446,6 +478,161 @@ def _pure_annotation_change_exempt(
446 return not _class_consumes_annotations(node, class_map)478 return not _class_consumes_annotations(node, class_map)
447 479 
448 480 
481+def _parse_assignments(texts) -> list[tuple[ast.stmt, list[str]]] | None:
482+ """Parse diff lines as a block of value-carrying simple-name assignments.
483+ 
484+ Returns [(stmt, target_names), ...] when every meaningful (non-blank,
485+ non-comment) line forms assignments (``x = v`` or ``x: T = v``) whose
486+ targets are all plain names; None when the block is unparsable or
487+ contains anything else (bare annotations, attribute/subscript targets,
488+ tuple unpacking, any other statement).
489+ """
490+ lines = [t for t in texts if t.strip() and not t.strip().startswith("#")]
491+ if not lines:
492+ return None
493+ block = textwrap.dedent("\n".join(lines))
494+ try:
495+ tree = ast.parse(block)
496+ except (SyntaxError, ValueError):
497+ return None
498+ stmts = []
499+ for stmt in tree.body:
500+ if isinstance(stmt, ast.Assign):
501+ if not all(isinstance(t, ast.Name) for t in stmt.targets):
502+ return None
503+ stmts.append((stmt, [t.id for t in stmt.targets]))
504+ elif isinstance(stmt, ast.AnnAssign):
505+ if stmt.value is None or not isinstance(stmt.target, ast.Name):
506+ return None
507+ stmts.append((stmt, [stmt.target.id]))
508+ else:
509+ return None
510+ return stmts
511+ 
512+ 
513+_LAZY_RHS_NODES = (
514+ ast.Constant,
515+ ast.Name,
516+ ast.Attribute,
517+ ast.Tuple,
518+ ast.List,
519+ ast.Set,
520+ ast.Dict,
521+ ast.BinOp,
522+ ast.UnaryOp,
523+ ast.BoolOp,
524+ ast.operator,
525+ ast.unaryop,
526+ ast.boolop,
527+ ast.expr_context,
528+)
529+ 
530+ 
531+def _assign_rhs_lazy(value: ast.expr) -> bool:
532+ """True when the right-hand side is free of runtime side effects:
533+ literals, names, attributes and containers/arithmetic over those only
534+ (no calls, subscripts, comprehensions, lambdas, walrus...). Attribute
535+ access and arithmetic can still trigger property/__add__ hooks - an
536+ accepted risk; dict unpacking ({**d}) is not: it iterates the source
537+ mapping at runtime."""
538+ for node in ast.walk(value):
539+ if not isinstance(node, _LAZY_RHS_NODES):
540+ return False
541+ if isinstance(node, ast.Dict) and any(k is None for k in node.keys):
542+ return False
543+ return True
544+ 
545+ 
546+def _decorator_tail(dec: ast.expr) -> str | None:
547+ """Simple name of a decorator (call decorators use their callee):
548+ 'dataclass' for @dataclass / @dataclasses.dataclass / @dataclass(...)."""
549+ expr = dec.func if isinstance(dec, ast.Call) else dec
550+ if isinstance(expr, ast.Name):
551+ return expr.id
552+ if isinstance(expr, ast.Attribute):
553+ return expr.attr
554+ return None
555+ 
556+ 
557+def _class_dataclass_like(cls_node: ast.ClassDef, class_map: dict) -> bool:
558+ """True when the class or a same-file ancestor is decorated with
559+ @dataclass (annotated class-body statements then become __init__ fields
560+ whose order matters)."""
561+ stack, seen = [cls_node], set()
562+ while stack:
563+ node = stack.pop()
564+ if id(node) in seen:
565+ continue
566+ seen.add(id(node))
567+ if any(_decorator_tail(d) == "dataclass" for d in node.decorator_list):
568+ return True
569+ for base in node.bases:
570+ name = _base_simple_name(base)
571+ parent = class_map.get(name) if name else None
572+ if parent is not None and id(parent) not in seen:
573+ stack.append(parent)
574+ return False
575+ 
576+ 
577+def _last_field_line(cls_node: ast.ClassDef) -> int | None:
578+ """Last base line of the class's annotated fields (dataclass __init__
579+ parameter order follows field order); None when the class has none."""
580+ last = None
581+ for stmt in cls_node.body:
582+ if isinstance(stmt, ast.AnnAssign):
583+ end = stmt.end_lineno or stmt.lineno
584+ last = end if last is None else max(last, end)
585+ return last
586+ 
587+ 
588+def _new_binding_insert_reason(
589+ add_texts,
590+ a: int,
591+ b: int,
592+ add_indent: int | None,
593+ ann_ctx: tuple | None,
594+) -> str | None:
595+ """Skip reason when a pure insertion consists solely of assignments that
596+ bind provably-new names with inert right-hand sides; None to record.
597+ 
598+ Base code cannot read a name that does not exist yet, so every reader of
599+ a first-time binding is itself an added line carrying its own anchor -
600+ the anchor here is noise (an import-time fingerprint in module/class
601+ scopes). Gates: no meaningful line ends in a comma (such lines are call
602+ or signature arguments, not statements); the block is all simple-name
603+ assignments; no target is a
604+ dunder or already occurs anywhere in the base file (Name occurrences,
605+ def/class/import/parameter bindings); right-hand sides are inert; and in
606+ class scope the class must be plain or dataclass-like, where plain
607+ assigns are never fields while annotated fields are only exempt when
608+ inserted at or after the last existing field (a middle field insertion
609+ would silently rebind positional constructor calls).
610+ """
611+ if ann_ctx is None or not add_texts:
612+ return None
613+ effective = [t for t in add_texts if t.strip() and not t.strip().startswith("#")]
614+ if not effective or effective[-1].rstrip().endswith(","):
615+ return None
616+ scopes, class_map, has_future, known_names = ann_ctx
617+ stmts = _parse_assignments(add_texts)
618+ if not stmts:
619+ return None
620+ for stmt, names in stmts:
621+ if any(name.startswith("__") or name in known_names for name in names):
622+ return None
623+ if not _assign_rhs_lazy(stmt.value):
624+ return None
625+ scope = _insertion_scope(scopes, a, b, add_indent)
626+ if isinstance(scope, ast.ClassDef) and _class_consumes_annotations(scope, class_map):
627+ if not _class_dataclass_like(scope, class_map):
628+ return None
629+ if any(isinstance(stmt, ast.AnnAssign) for stmt, _ in stmts):
630+ last_field = _last_field_line(scope)
631+ if last_field is not None and a < last_field:
632+ return None
633+ return "first-time binding of new name(s)"
634+ 
635+ 
449def _classify_candidate_pairs(636def _classify_candidate_pairs(
450 affected: set[int],637 affected: set[int],
451 pairs: list[tuple],638 pairs: list[tuple],
@@ -460,10 +647,17 @@ def _classify_candidate_pairs(
460 comment/docstring line in the base file AND the added lines are comments or647 comment/docstring line in the base file AND the added lines are comments or
461 doc prose (not parseable Python), i.e. a pure comment/docstring change.648 doc prose (not parseable Python), i.e. a pure comment/docstring change.
462 Pure type annotation changes are dropped in every direction (insertion,649 Pure type annotation changes are dropped in every direction (insertion,
463- replacement, deletion) when provably inert: function-local annotations are650+ replacement, deletion) when provably inert: function-local annotations
464- never evaluated; class-level ones need a plain class (no decorators, no651+ are never evaluated; class-level ones need a plain class (no decorators,
465- TypedDict/Protocol/Enum/BaseModel/NamedTuple base) and side-effect-free652+ no TypedDict/Protocol/Enum/BaseModel/NamedTuple base) and side-effect-free
466 annotation expressions unless the file has future annotations.653 annotation expressions unless the file has future annotations.
654+ First-time bindings are dropped when a pure insertion (no meaningful
655+ line ends in a comma - those are call/signature arguments) consists
656+ solely of assignments binding simple names that occur nowhere in the
657+ base file with inert right-hand sides (base code cannot read a name that
658+ does not exist yet; readers added in the same PR carry their own anchors);
659+ in dataclass-like classes annotated fields are only dropped when inserted
660+ at or after the last existing field.
467 Candidate pairs: without base content (or non-parseable Python) both sides661 Candidate pairs: without base content (or non-parseable Python) both sides
468 of each pair are counted, bounded by the hunk.662 of each pair are counted, bounded by the hunk.
469 """663 """
@@ -532,6 +726,8 @@ def _classify_candidate_pairs(
532 reason = "belongs to a newly added function/class"726 reason = "belongs to a newly added function/class"
533 if reason is None:727 if reason is None:
534 reason = _pure_annotation_insert_reason(add_texts, a, b, add_indent, ann_ctx)728 reason = _pure_annotation_insert_reason(add_texts, a, b, add_indent, ann_ctx)
729+ if reason is None:
730+ reason = _new_binding_insert_reason(add_texts, a, b, add_indent, ann_ctx)
535 # kind == 'blank': isolated blank deletion == one-line insertion731 # kind == 'blank': isolated blank deletion == one-line insertion
536 elif _between_definitions(a, b, end_lines, start_lines, blanks):732 elif _between_definitions(a, b, end_lines, start_lines, blanks):
537 reason = "between function/class definitions"733 reason = "between function/class definitions"
Mprecision-test/src/repos/__init__.py+2-2
@@ -11,7 +11,7 @@ __all__ = ["RepoAdapter", "VllmAscendAdapter", "SglangAdapter", "TorchNpuAdapter
11_ADAPTERS = {11_ADAPTERS = {
12 "vllm_ascend": VllmAscendAdapter,12 "vllm_ascend": VllmAscendAdapter,
13 "sglang": SglangAdapter,13 "sglang": SglangAdapter,
14- "torch_npu": TorchNpuAdapter,14+ "pytorch": TorchNpuAdapter,
15}15}
16 16 
17AVAILABLE_REPOS = tuple(_ADAPTERS.keys())17AVAILABLE_REPOS = tuple(_ADAPTERS.keys())
@@ -21,7 +21,7 @@ def get_adapter(repo_name: str) -> RepoAdapter:
21 """按仓库名获取适配器实例。21 """按仓库名获取适配器实例。
22 22 
23 Args:23 Args:
24- repo_name: 仓库名（vllm_ascend / sglang / torch_npu）24+ repo_name: 仓库名（vllm_ascend / sglang / pytorch）
25 25 
26 Returns:26 Returns:
27 RepoAdapter 实例27 RepoAdapter 实例
Mprecision-test/src/repos/base.py+1-1
@@ -21,7 +21,7 @@ class RepoAdapter(ABC):
21 #: 覆盖率数据目录下的测试用例文件夹命名前缀；None 表示无固定前缀（vllm 用 tests__ 前缀 + cpu-ut）21 #: 覆盖率数据目录下的测试用例文件夹命名前缀；None 表示无固定前缀（vllm 用 tests__ 前缀 + cpu-ut）
22 test_case_dir_prefix: str | None = None22 test_case_dir_prefix: str | None = None
23 #: 覆盖率文件名 glob 模式（测试目录或其 covdata/ 子目录下）。23 #: 覆盖率文件名 glob 模式（测试目录或其 covdata/ 子目录下）。
24- #: 默认 "coverage*" 兼容 pytest-cov 裸文件（torch_npu: coverage）与原始格式24+ #: 默认 "coverage*" 兼容 pytest-cov 裸文件（pytorch: coverage）与原始格式
25 #: 带后缀文件（vllm/sglang: coverage.linux-...-workflow.xxx）。25 #: 带后缀文件（vllm/sglang: coverage.linux-...-workflow.xxx）。
26 coverage_file_glob: str = "coverage*"26 coverage_file_glob: str = "coverage*"
27 27 
Mprecision-test/src/repos/torch_npu.py+20-17
@@ -1,14 +1,16 @@
1-"""torch_npu 仓库适配器（PyTorch / GitCode 社区）。1+"""torch_npu 仓库适配器（检测上游 PyTorch：https://github.com/pytorch/pytorch）。
2 2 
3-- REPO_NAME = "torch_npu"，产品代码前缀 "torch_npu/"3+- REPO_NAME = "torch"，产品代码前缀 "torch/"，PR 源为 GitHub（--github-pr pytorch/pytorch#N）
4-- 测试用例目录：目录名含 "__" 或以 "test_" 开头（如 test__xxx 编码路径），4+- 测试用例目录：目录名含 "__" 或以 "test_" 开头（如 _inductor__test_add），
5 覆盖率文件位于 covdata/ 子目录5 覆盖率文件位于 covdata/ 子目录
6- 测试文件规则：test/ 下 test_*.py6- 测试文件规则：test/ 下 test_*.py
7- 测试名规范化：-- -> ::，__ -> /，文件级补 .py7- 测试名规范化：-- -> ::，__ -> /，文件级补 .py
8- 粒度默认值跟随合并版：line=True, function=True8- 粒度默认值跟随合并版：line=True, function=True
9-- 源文件解析使用基类默认候选（source_dir/torch_npu/<rel>，运行需 -s 指向 covstub）9+- 源文件解析使用基类默认候选（source_dir/torch/<rel>，运行需 -s 指向 covstub，
10+ covstub/torch/ 对应 pytorch/pytorch 仓库根目录）
10"""11"""
11 12 
13+import re
12from fnmatch import fnmatchcase14from fnmatch import fnmatchcase
13from pathlib import PurePosixPath15from pathlib import PurePosixPath
14 16 
@@ -18,10 +20,14 @@ from .base import RepoAdapter
18class TorchNpuAdapter(RepoAdapter):20class TorchNpuAdapter(RepoAdapter):
19 """torch_npu 仓库适配器。"""21 """torch_npu 仓库适配器。"""
20 22 
21- repo_name = "torch_npu"23+ repo_name = "torch"
22- product_code_prefix = "torch_npu/"24+ product_code_prefix = "torch/"
23 test_case_dir_prefix = None25 test_case_dir_prefix = None
24 26 
27+ #: 精确匹配路径中独立的 torch/ 段（前为路径起点或 /），
28+ #: 避免 pytorch/、torchvision/、test_torch.py 等含 torch 子串的路径被误剥
29+ _TORCH_PATH_RE = re.compile(r"(?:^|/)torch/(.+)")
30+ 
25 #: 可纳入 PR 检测的测试根目录31 #: 可纳入 PR 检测的测试根目录
26 TEST_ROOTS = ("test/",)32 TEST_ROOTS = ("test/",)
27 33 
@@ -30,7 +36,7 @@ class TorchNpuAdapter(RepoAdapter):
30 # ------------------------------------------------------------------36 # ------------------------------------------------------------------
31 def is_test_case_dir(self, name: str) -> bool:37 def is_test_case_dir(self, name: str) -> bool:
32 # PyTorch: "__" in name or name.startswith("test_")38 # PyTorch: "__" in name or name.startswith("test_")
33- return ("__" in name) or name.startswith("test_")39+ return name.endswith(".py") or ("__" in name) or name.startswith("test_")
34 40 
35 def normalize_test_name(self, test_name: str) -> str:41 def normalize_test_name(self, test_name: str) -> str:
36 """42 """
@@ -40,19 +46,16 @@ class TorchNpuAdapter(RepoAdapter):
40 """46 """
41 # 先转换 -- 为 ::（函数级测试标记）47 # 先转换 -- 为 ::（函数级测试标记）
42 result = test_name.replace("--", "::")48 result = test_name.replace("--", "::")
43- # 再转换 __ 为 /49+ if result.endswith(".py") or (".py::" in result):
44- result = result.replace("__", "/")50+ if not result.startswith("test/"):
45- # 无 ::（文件级测试）则补 .py 后缀51+ result = f"test/{result}"
46- if "::" not in result:52+ return result
47- result = result + ".py"
48 return result53 return result
49 54 
50 def covered_path_to_rel(self, path: str) -> str | None:55 def covered_path_to_rel(self, path: str) -> str | None:
51- """覆盖率库路径包含 torch_npu/ 时剥离前缀得到相对路径，否则返回 None。"""56+ """覆盖率库路径含独立的 torch/ 段时剥离前缀得到相对路径，否则返回 None。"""
52- if self.repo_name not in path:57+ match = self._TORCH_PATH_RE.search(path)
53- return None58+ return match.group(1) if match else None
54- marker = f"{self.repo_name}/"
55- return path.split(marker)[-1] if marker in path else path
56 59 
57 # ------------------------------------------------------------------60 # ------------------------------------------------------------------
58 # 测试文件检测（PR 内容检测）61 # 测试文件检测（PR 内容检测）
Mprecision-test/validators.js+1-1
@@ -3,7 +3,7 @@ const path = require("path");
3/**3/**
4 * Available repository adapters supported by the precision-test tool.4 * Available repository adapters supported by the precision-test tool.
5 */5 */
6-const AVAILABLE_REPOS = ["vllm_ascend", "sglang", "torch_npu"];6+const AVAILABLE_REPOS = ["vllm_ascend", "sglang", "pytorch"];
7 7 
8/**8/**
9 * Whitelist of environment variable names safe to pass to the Python subprocess.9 * Whitelist of environment variable names safe to pass to the Python subprocess.