Fixed many autocast problems.

fooocus_version.py

@@ -1 +1 @@
-version = '2.1.734'
+version = '2.1.735'

modules/patch.py

@@ -1,3 +1,4 @@
+import contextlib
 import os
 import torch
 import time
@@ -463,16 +464,36 @@ def text_encoder_device_patched():
     return fcbh.model_management.get_torch_device()
 
 
-def patched_get_autocast_device(dev):
+def patched_autocast(device_type, dtype=None, enabled=True, cache_enabled=None):
     # https://github.com/lllyasviel/Fooocus/discussions/571
     # https://github.com/lllyasviel/Fooocus/issues/620
-    result = ''
-    if hasattr(dev, 'type'):
-        result = str(dev.type)
-    if 'cuda' in result:
-        return 'cuda'
-    else:
-        return 'cpu'
+    # https://github.com/lllyasviel/Fooocus/issues/759
+
+    supported = False
+
+    if device_type == 'cuda' and dtype == torch.float32 and enabled:
+        supported = True
+
+    if device_type == 'cuda' and dtype == torch.float16 and enabled:
+        supported = True
+
+    if device_type == 'cuda' and dtype == torch.bfloat16 and enabled:
+        supported = True
+
+    if not supported:
+        print(f'[Fooocus Autocast Warning] Requested unsupported torch autocast ['
+              f'device_type={str(device_type)}, '
+              f'dtype={str(dtype)}, '
+              f'enabled={str(enabled)}, '
+              f'cache_enabled={str(cache_enabled)}]. '
+              f'Fooocus fixed it automatically, feel free to report to Fooocus on GitHub if this may cause potential problems.')
+        return contextlib.nullcontext()
+
+    return torch.amp.autocast_mode.autocast_origin(
+        device_type=device_type,
+        dtype=dtype,
+        enabled=enabled,
+        cache_enabled=cache_enabled)
 
 
 def patched_load_models_gpu(*args, **kwargs):
@@ -535,8 +556,18 @@ def patch_all():
     if not hasattr(fcbh.model_management, 'load_models_gpu_origin'):
         fcbh.model_management.load_models_gpu_origin = fcbh.model_management.load_models_gpu
 
+    if not hasattr(torch.amp.autocast_mode, 'autocast_origin'):
+        torch.amp.autocast_mode.autocast_origin = torch.amp.autocast_mode.autocast
+
+    torch.amp.autocast_mode.autocast = patched_autocast
+    torch.amp.autocast = patched_autocast
+    torch.autocast = patched_autocast
+
+    # # Test if this will fail
+    # with torch.autocast(device_type='cpu', dtype=torch.float32):
+    #     print(torch.ones(10))
+
     fcbh.model_management.load_models_gpu = patched_load_models_gpu
-    fcbh.model_management.get_autocast_device = patched_get_autocast_device
     fcbh.model_management.text_encoder_device = text_encoder_device_patched
     fcbh.model_patcher.ModelPatcher.calculate_weight = calculate_weight_patched
     fcbh.cldm.cldm.ControlNet.forward = patched_cldm_forward

update_log.md

@@ -1,3 +1,7 @@
+# 2.1.735
+
+* Fixed many problems related to torch autocast.
+
 # 2.1.733
 
 * Increased allowed random seed range.