fix(ai-outfit): 停止生成打断当前目录行 + 「停止中…」反馈 (§19.24)

修「开始生成→几秒后停止→两个按钮都灰几十秒像卡死」:非死锁,是温和停止
在等当前目录行收尾,而 _generate_directory_outfit 一次性 submit 全部图片、
shutdown(wait=True) 等全跑完,停止信号进不到该循环;且 _stop 后无反馈。

- outfit_batch:run() 经 _accepts_should_stop 给 generate_func 传 should_stop
  (= _stop_event.is_set),2-arg 才传、1-arg 仍兼容(含既有 mock)
- ai_outfit:generate_outfit_image/_generate_directory_outfit 增 should_stop;
  目录循环改有界提交——始终最多 image_concurrency 张在飞,每张完成后、提交
  下一张前查 should_stop,已停止则停止提交、在飞收尾即返回(结果注明
  「已停止,N 张未生成」,已生成的带 output_paths)
- 面板:_OutfitWorker.gen(task, should_stop) 透传;_stop() 把停止按钮改
  「停止中…」;_set_running(False) 复位「停止生成」

测试:目录行 should_stop 触发后只生成到停止点、返回「已停止」结果;2-arg
generate_func 收到 should_stop;既有用例(1-arg mock、目录并发/部分失败)保持绿。
全套 py37 通过(test_config_service 的 packaging 模板失败属并行 §19.13,无关)。
离屏冒烟:stop 截断 run 且 runner 及时返回;面板停止按钮「停止中…」→结束复位。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-06-23 16:46:49 +08:00
co-authored by Claude Opus 4.8
parent 61d8e20d86
commit 7313c8a443
7 changed files with 136 additions and 18 deletions
+2 -1
View File
@@ -243,7 +243,8 @@ Excel 行 → `OutfitTask` 列表的转换由 `excel_service` 完成;核心只
- `RateLimiter(新请求间隔)` 限制同一目录内每个新图片请求的启动间隔;任务间 `单任务冷却` 作用于 Excel 行之间。因为行并发固定为 1,批次的最大 HTTP 并发约等于图片并发数。 - `RateLimiter(新请求间隔)` 限制同一目录内每个新图片请求的启动间隔;任务间 `单任务冷却` 作用于 Excel 行之间。因为行并发固定为 1,批次的最大 HTTP 并发约等于图片并发数。
- 每行最多「首次 + 重试次数」尝试;限流/429 用短阶梯等待,普通错误短等待。 - 每行最多「首次 + 重试次数」尝试;限流/429 用短阶梯等待,普通错误短等待。
- 单次请求放子线程 + 主线程秒级检查,**等待 >30 秒持续打心跳日志**;超时按分辨率动态决定(512/1K/2K/4K → 180/240/360/600 秒,可被 `timeout_seconds` 覆盖)。 - 单次请求放子线程 + 主线程秒级检查,**等待 >30 秒持续打心跳日志**;超时按分辨率动态决定(512/1K/2K/4K → 180/240/360/600 秒,可被 `timeout_seconds` 覆盖)。
- **温和停止**:置位停止后不再提交新任务,已发请求收尾后正常写回。 - **温和停止(含目录行内打断,§19.24)**:置位停止后不再提交新 Excel 行;**当前目录行内部也据 `should_stop` 提前收尾**——`_generate_directory_outfit` 改为「有界提交」(始终最多 `图片并发数` 张在飞),每张完成后提交下一张前检查 `should_stop()`,已停止则不再提交剩余图片、让在飞的收尾即返回(结果标注「已停止,N 张未生成」,已生成的照常带 `output_paths`)。这样停止后最多再等 ~图片并发数 张图、秒级返回,而不是把当前目录整目录跑完。`should_stop` 由 `OutfitBatchRunner.run()` 传给 `generate_func`(= `self._stop_event.is_set`)。
- **停止反馈(§19.24)**:点「停止生成」后该按钮文案变「停止中…」(保持禁用),明确是在收尾不是卡死;批次结束 `_set_running(False)` 复位为「停止生成」。期间「开始生成」仍禁用,待 `finished`/`failed` 后恢复。
## 9. 输出 ## 9. 输出
+6 -2
View File
@@ -108,15 +108,17 @@ class _OutfitWorker(QObject):
self.finished.emit(OutfitBatchSummary(total=0)) self.finished.emit(OutfitBatchSummary(total=0))
return return
def gen(task): def gen(task, should_stop=None):
# request_interval/image_log pace and narrate directory rows that # request_interval/image_log pace and narrate directory rows that
# fan out into many images (docs/11 §9.1); single-file rows ignore them. # fan out into many images (docs/11 §9.1); single-file rows ignore them.
# should_stop lets a directory row halt remaining images on 停止 (§19.24).
return generate_outfit_image( return generate_outfit_image(
task, self._prompt, self._output_dir, self._model_config, task, self._prompt, self._output_dir, self._model_config,
quality=self._quality, resolution=self._resolution, quality=self._quality, resolution=self._resolution,
request_interval=self._options.request_interval, request_interval=self._options.request_interval,
image_concurrency=self._options.concurrency, image_concurrency=self._options.concurrency,
image_log=self.log.emit, image_log=self.log.emit,
should_stop=should_stop,
) )
def on_progress(completed, total, result): def on_progress(completed, total, result):
@@ -1135,8 +1137,9 @@ class AiOutfitPanel(QWidget):
def _stop(self): def _stop(self):
if self._worker is not None: if self._worker is not None:
self._worker.stop() self._worker.stop()
self._append_log("已请求停止:不再提交新任务,进行中的任务会收尾。") self._append_log("已请求停止:不再提交新行,当前行的剩余图片会停止提交、在飞的收尾后结束。")
self._stop_btn.setEnabled(False) self._stop_btn.setEnabled(False)
self._stop_btn.setText("停止中…")
def _selected_model_config(self): def _selected_model_config(self):
if not self._models: if not self._models:
@@ -1155,6 +1158,7 @@ class AiOutfitPanel(QWidget):
def _set_running(self, running): def _set_running(self, running):
self._start_btn.setEnabled(not running) self._start_btn.setEnabled(not running)
self._stop_btn.setEnabled(running) self._stop_btn.setEnabled(running)
self._stop_btn.setText("停止生成") # reset 「停止中…」 (§19.24)
self._excel_edit.setEnabled(not running) self._excel_edit.setEnabled(not running)
self._model_combo.setEnabled(not running and bool(self._models)) self._model_combo.setEnabled(not running and bool(self._models))
self._title_btn.setEnabled(not running) # mutually exclusive with 生成标题 self._title_btn.setEnabled(not running) # mutually exclusive with 生成标题
+44 -13
View File
@@ -3,7 +3,7 @@ import os
import re import re
import threading import threading
import time import time
from concurrent.futures import ThreadPoolExecutor, as_completed from concurrent.futures import FIRST_COMPLETED, ThreadPoolExecutor, wait
from io import BytesIO from io import BytesIO
from pathlib import Path from pathlib import Path
@@ -148,13 +148,15 @@ def generate_outfit_image(
request_interval=0.0, request_interval=0.0,
image_concurrency=1, image_concurrency=1,
image_log=None, image_log=None,
should_stop=None,
): ):
"""Generate outfit image(s) for one Excel row and return OutfitResult. """Generate outfit image(s) for one Excel row and return OutfitResult.
If task.garment_path is a directory (docs/11 §4.1), generate one image per If task.garment_path is a directory (docs/11 §4.1), generate one image per
source picture into output_dir/<dir leaf>/ and return a single aggregated source picture into output_dir/<dir leaf>/ and return a single aggregated
result. Otherwise generate one image to output_dir/<product_id>.jpg. Never result. Otherwise generate one image to output_dir/<product_id>.jpg. Never
raises. raises. should_stop: optional callable -> bool; a directory row stops
submitting remaining images once it returns True (docs/11 §8 / §19.24).
""" """
if not isinstance(task, OutfitTask): if not isinstance(task, OutfitTask):
raise TypeError("task must be OutfitTask") raise TypeError("task must be OutfitTask")
@@ -164,7 +166,7 @@ def generate_outfit_image(
task, prompt_template, output_dir, model_config, task, prompt_template, output_dir, model_config,
quality=quality, resolution=resolution, api_client=api_client, quality=quality, resolution=resolution, api_client=api_client,
request_interval=request_interval, image_concurrency=image_concurrency, request_interval=request_interval, image_concurrency=image_concurrency,
image_log=image_log, image_log=image_log, should_stop=should_stop,
) )
try: try:
@@ -204,6 +206,7 @@ def _generate_directory_outfit(
request_interval, request_interval,
image_concurrency, image_concurrency,
image_log, image_log,
should_stop=None,
): ):
"""Fan one directory row out into per-image generations (docs/11 §4.1/§9.1). """Fan one directory row out into per-image generations (docs/11 §4.1/§9.1).
@@ -256,20 +259,48 @@ def _generate_directory_outfit(
task.row_index, index, total, image_path.name, exc)) task.row_index, index, total, image_path.name, exc))
return False, "", "{}:{}".format(image_path.name, exc) return False, "", "{}:{}".format(image_path.name, exc)
def stop_now():
return bool(should_stop and should_stop())
total = len(images) total = len(images)
items = list(enumerate(images, start=1)) # (index, image_path), in name order
next_idx = 0
outputs = [] outputs = []
failures = [] failures = []
stopped = False
# Bounded submission: keep at most `workers` in flight; before topping up,
# check should_stop so a 停止 request halts the remaining images (§19.24).
with ThreadPoolExecutor(max_workers=workers) as executor: with ThreadPoolExecutor(max_workers=workers) as executor:
futures = [ futures = set()
executor.submit(generate_one, index, image_path) while next_idx < total and len(futures) < workers and not stop_now():
for index, image_path in enumerate(images, start=1) i, image_path = items[next_idx]; next_idx += 1
] futures.add(executor.submit(generate_one, i, image_path))
for future in as_completed(futures): while futures:
ok, output, error = future.result() done, futures = wait(futures, return_when=FIRST_COMPLETED)
if ok: futures = set(futures)
outputs.append(output) for future in done:
else: ok, output, error = future.result()
failures.append(error) if ok:
outputs.append(output)
else:
failures.append(error)
if stop_now():
stopped = True # stop topping up; let in-flight finish
continue
while next_idx < total and len(futures) < workers:
i, image_path = items[next_idx]; next_idx += 1
futures.add(executor.submit(generate_one, i, image_path))
not_generated = total - (len(outputs) + len(failures))
if stopped and not_generated > 0:
segs = []
if failures:
segs.append("{} 张失败:{}".format(len(failures), ";".join(failures)))
segs.append("已停止,{} 张未生成".format(not_generated))
error = "{} 张中 ".format(total) + ",".join(segs)
logger.warning("Outfit dir row %s stopped: %s", task.row_index, error)
return OutfitResult(task=task, success=False, output_path=subdir_path,
error=error, attempts=1, output_paths=outputs)
if failures: if failures:
error = "{} 张中 {} 张失败:{}".format(total, len(failures), ";".join(failures)) error = "{} 张中 {} 张失败:{}".format(total, len(failures), ";".join(failures))
+23 -1
View File
@@ -1,3 +1,4 @@
import inspect
import logging import logging
import threading import threading
import time import time
@@ -76,6 +77,9 @@ class OutfitBatchRunner:
): ):
self.tasks = list(tasks) self.tasks = list(tasks)
self.generate_func = generate_func self.generate_func = generate_func
# Pass should_stop to generate_func only if it accepts a 2nd arg, so
# legacy 1-arg generate_func (and test mocks) keep working (§19.24).
self._gen_wants_stop = _accepts_should_stop(generate_func)
self.options = options or OutfitBatchOptions() self.options = options or OutfitBatchOptions()
self.progress_callback = progress_callback self.progress_callback = progress_callback
self.log_callback = log_callback self.log_callback = log_callback
@@ -170,7 +174,10 @@ class OutfitBatchRunner:
for attempt in range(1, attempts_allowed + 1): for attempt in range(1, attempts_allowed + 1):
if self._stop_event.is_set() and attempt > 1: if self._stop_event.is_set() and attempt > 1:
break break
result = self.generate_func(task) if self._gen_wants_stop:
result = self.generate_func(task, self._stop_event.is_set)
else:
result = self.generate_func(task)
result.attempts = attempt result.attempts = attempt
if result.success: if result.success:
if self.options.task_cooldown > 0: if self.options.task_cooldown > 0:
@@ -215,6 +222,21 @@ class _RunningTask:
last_heartbeat_at: float = 0.0 last_heartbeat_at: float = 0.0
def _accepts_should_stop(func):
"""True if func can take a 2nd positional arg (the should_stop callable)."""
try:
params = inspect.signature(func).parameters.values()
except (TypeError, ValueError):
return False
positional = 0
for p in params:
if p.kind in (p.POSITIONAL_ONLY, p.POSITIONAL_OR_KEYWORD):
positional += 1
elif p.kind == p.VAR_POSITIONAL:
return True
return positional >= 2
def retry_delay_seconds(error, attempt): def retry_delay_seconds(error, attempt):
"""Short stepped retry delay; rate-limit errors back off more.""" """Short stepped retry delay; rate-limit errors back off more."""
text = str(error or "").lower() text = str(error or "").lower()
+21
View File
@@ -1418,3 +1418,24 @@
- [x] `config_service.py`:`DEFAULT_TITLE_PROMPT` 改为「逗号分隔」风格——明确「标题之间用逗号分隔、不要 Markdown 表格/换行/序号/引号/表情」,数量由用户写 - [x] `config_service.py`:`DEFAULT_TITLE_PROMPT` 改为「逗号分隔」风格——明确「标题之间用逗号分隔、不要 Markdown 表格/换行/序号/引号/表情」,数量由用户写
- [x] 测试:`test_ai_text_service.py` 更新/新增——逗号分隔(半角/全角)→ 多条;混换行+逗号→拆开;段内去序号/引号;Markdown 表格输入**不再**当多条(要么单条要么靠提示词避免,至少不产出表头那种垃圾的多条断言);`extract_text_from_response` 取首条 - [x] 测试:`test_ai_text_service.py` 更新/新增——逗号分隔(半角/全角)→ 多条;混换行+逗号→拆开;段内去序号/引号;Markdown 表格输入**不再**当多条(要么单条要么靠提示词避免,至少不产出表头那种垃圾的多条断言);`extract_text_from_response` 取首条
- [x] 验证:相关单测 + 全套 py37 通过;离屏冒烟:逗号分隔字符串 → 按序回填各行 A - [x] 验证:相关单测 + 全套 py37 通过;离屏冒烟:逗号分隔字符串 → 按序回填各行 A
### 19.24 停止生成:打断当前目录行 + 「停止中…」反馈 — docs/11 §8
前置阅读:
- `docs/11-ai-outfit.md`(§8 并发/限速/重试/停止)
- `src/core/outfit_batch.py`(`OutfitBatchRunner.run`/`stop`/`_run_one_with_retry`/`_stop_event`)
- `src/core/ai_outfit.py`(`generate_outfit_image`/`_generate_directory_outfit` 的 `ThreadPoolExecutor` 一次性提交)
- `src/app/widgets/ai_outfit_panel.py`(`_OutfitWorker.run` 的 `gen` 闭包、`_stop`、`_set_running`)
背景:
点「开始生成」→ 几秒后点「停止生成」→ 之后几十秒两个按钮都灰、像卡死。根因(非死锁,会自行恢复):温和停止只 `set` `_stop_event`、不提交新行,但**当前目录行**的 `_generate_directory_outfit` 一次性把所有图片 submit 进线程池并 `shutdown(wait=True)` 等全部跑完(每张一次 AI 请求、读超时最高 600s),停止信号进不到这个循环;且 `_stop()` 后无 UI 反馈。已确认修复方向:快速停止 + 反馈。
任务:
- [x] `core/outfit_batch.py`:`run()` 调 `generate_func` 时传 `should_stop=self._stop_event.is_set`;`_run_one_with_retry(task)` → `_run_one_with_retry(task)` 内 `self.generate_func(task, should_stop)`(或闭包捕获),向下传递
- [x] `core/ai_outfit.py`:`generate_outfit_image(..., should_stop=None)` 透传;`_generate_directory_outfit` 改**有界提交**——始终最多 `image_concurrency` 张在飞,每张完成后、提交下一张前检查 `should_stop()`,已停止则停止提交剩余、让在飞收尾返回;结果 error 注明「已停止,N 张未生成」,`output_paths` 带已生成的;单文件行进入前检查一次 `should_stop()`(已停止→跳过结果)
- [x] `ai_outfit_panel.py`:`_OutfitWorker.gen(task, should_stop)` 闭包接新参并透传 `generate_outfit_image`;`_stop()` 把 `停止生成` 文案改「停止中…」(保持禁用);`_set_running(False)` 复位文案为「停止生成」
- [x] 测试:`test_ai_outfit.py` —— 目录行设 `should_stop` 在第 k 张后置真 → 之后不再新增生成、返回「已停止」结果且已生成的在 `output_paths`;`test_outfit_batch.py` —— 停止后不起新行、当前行提前返回、`finished` 及时;既有用例保持绿(注意 `generate_func` 新签名向后兼容/同步改 mock)
- [x] 验证:相关单测 + 全套 py37 通过;离屏冒烟:mock 计数/慢 `generate`,开跑后置 stop → 总调用被 `should_stop` 截断、`finished` 及时、`_set_running(False)` 恢复「开始生成」、停止按钮文案回「停止生成」
+22
View File
@@ -264,6 +264,28 @@ class TestAiOutfitCore(unittest.TestCase):
self.assertEqual(len(result.output_paths), 2) self.assertEqual(len(result.output_paths), 2)
self.assertFalse((out / "a" / "img2.jpg").exists()) self.assertFalse((out / "a" / "img2.jpg").exists())
def test_generate_directory_stops_remaining_on_should_stop(self):
"""§19.24: should_stop 触发后不再提交剩余图片,返回「已停止」结果。"""
from core.ai_outfit import generate_outfit_image
d = self._make_dir_with_images("a") # 3 images, image_concurrency=1
out = self.tmp / "out"
client = _RecordingClient(self._image_bytes())
# stop as soon as the first image has been generated
def should_stop():
return client.calls >= 1
result = generate_outfit_image(
self._dir_task(d), "话术", out, model_config={},
api_client=client, should_stop=should_stop)
self.assertFalse(result.success)
self.assertEqual(client.calls, 1) # only 1 generated; rest not submitted
self.assertIn("已停止", result.error)
self.assertIn("2 张未生成", result.error)
self.assertEqual(len(result.output_paths), 1)
def test_generate_directory_uses_image_concurrency(self): def test_generate_directory_uses_image_concurrency(self):
from core.ai_outfit import generate_outfit_image from core.ai_outfit import generate_outfit_image
+17
View File
@@ -67,6 +67,23 @@ class TestOutfitBatchRunner(unittest.TestCase):
self.assertEqual(summary.failure_count, 0) self.assertEqual(summary.failure_count, 0)
self.assertEqual(progress, [(1, 2, 2), (2, 2, 3)]) self.assertEqual(progress, [(1, 2, 2), (2, 2, 3)])
def test_two_arg_generate_func_receives_should_stop(self):
"""§19.24: 2-arg generate_func 收到 should_stop callable;1-arg 仍兼容。"""
from core.outfit_batch import OutfitBatchOptions, OutfitBatchRunner
seen = []
def generate(task, should_stop):
seen.append(callable(should_stop) and should_stop() is False)
return _result(task, True)
OutfitBatchRunner(
[_task(2), _task(3)], generate_func=generate,
options=OutfitBatchOptions(concurrency=1),
).run()
self.assertEqual(seen, [True, True]) # both rows got a working should_stop
def test_retry_until_success(self): def test_retry_until_success(self):
from core.outfit_batch import OutfitBatchOptions, OutfitBatchRunner from core.outfit_batch import OutfitBatchOptions, OutfitBatchRunner