Update app.py
Browse files
app.py
CHANGED
|
@@ -1107,7 +1107,7 @@ def _generate_impl(
|
|
| 1107 |
# ↓↓↓ 原 generate() 的完整函数体从这里开始,一行都不用改 ↓↓↓
|
| 1108 |
_current_generation_task_id.set(task_id)
|
| 1109 |
|
| 1110 |
-
|
| 1111 |
# _warm_wait_in_gpu(unet_choice)
|
| 1112 |
|
| 1113 |
# global _previous_arch_for_gc
|
|
@@ -1605,7 +1605,7 @@ class WarmupManager:
|
|
| 1605 |
with self._lock:
|
| 1606 |
plan = self._plan
|
| 1607 |
self._plan = None
|
| 1608 |
-
# 疑似bug
|
| 1609 |
# if plan is None and self._resume:
|
| 1610 |
# plan, self._resume = self._resume, None
|
| 1611 |
# self._seq += 1
|
|
@@ -1850,7 +1850,8 @@ def _boot_warmup():
|
|
| 1850 |
break
|
| 1851 |
time.sleep(0.2)
|
| 1852 |
|
| 1853 |
-
# 疑似bug
|
|
|
|
| 1854 |
# warmup.request_sweep(reason="boot_sweep")
|
| 1855 |
|
| 1856 |
# ==========================================================================
|
|
|
|
| 1107 |
# ↓↓↓ 原 generate() 的完整函数体从这里开始,一行都不用改 ↓↓↓
|
| 1108 |
_current_generation_task_id.set(task_id)
|
| 1109 |
|
| 1110 |
+
# 疑似bug 1:拿到 ZeroGPU 显卡后,还要先看一眼状态文件再决定要不要干活,结果因为状态文件被后台串台覆写,硬生生白等满 5 秒。
|
| 1111 |
# _warm_wait_in_gpu(unet_choice)
|
| 1112 |
|
| 1113 |
# global _previous_arch_for_gc
|
|
|
|
| 1605 |
with self._lock:
|
| 1606 |
plan = self._plan
|
| 1607 |
self._plan = None
|
| 1608 |
+
# 疑似bug 2:选了 Krea2,后台暂停了 MOSS;一旦 Krea2 读完,后台立刻又把 MOSS 捞出来继续读,直接把刚进内存的 Krea2 又挤出去了。
|
| 1609 |
# if plan is None and self._resume:
|
| 1610 |
# plan, self._resume = self._resume, None
|
| 1611 |
# self._seq += 1
|
|
|
|
| 1850 |
break
|
| 1851 |
time.sleep(0.2)
|
| 1852 |
|
| 1853 |
+
# 疑似bug:服务器一启动,不管你用不用,后台就会把 MOSS、Flux、Krea2 等所有模型按顺序全读一遍。
|
| 1854 |
+
# 注释掉后,启动时只预热你上次用过的那 1 个模型(_boot_warmup 的第一段阶梯预热依然保留)。
|
| 1855 |
# warmup.request_sweep(reason="boot_sweep")
|
| 1856 |
|
| 1857 |
# ==========================================================================
|