Merge pull request #14597 from AUTOMATIC1111/improved-manual-cast

Improve the implementation of Manual Cast and IPEX support

1 files changed, 40 insertions, 16 deletions

diff --git a/modules/devices.py b/modules/devices.py
index ff279ac5..0321d12c 100644
--- a/modules/devices.py
+++ b/modules/devices.py
@@ -110,6 +110,7 @@ device_codeformer: torch.device = None
 dtype: torch.dtype = torch.float16
 dtype_vae: torch.dtype = torch.float16
 dtype_unet: torch.dtype = torch.float16
+dtype_inference: torch.dtype = torch.float16
 unet_needs_upcast = False
 
 
@@ -131,21 +132,44 @@ patch_module_list = [
 ]
 
 
-def manual_cast_forward(self, *args, **kwargs):
-    org_dtype = torch_utils.get_param(self).dtype
-    self.to(dtype)
-    args = [arg.to(dtype) if isinstance(arg, torch.Tensor) else arg for arg in args]
-    kwargs = {k: v.to(dtype) if isinstance(v, torch.Tensor) else v for k, v in kwargs.items()}
-    result = self.org_forward(*args, **kwargs)
-    self.to(org_dtype)
-    return result
+def manual_cast_forward(target_dtype):
+    def forward_wrapper(self, *args, **kwargs):
+        if any(
+            isinstance(arg, torch.Tensor) and arg.dtype != target_dtype
+            for arg in args
+        ):
+            args = [arg.to(target_dtype) if isinstance(arg, torch.Tensor) else arg for arg in args]
+            kwargs = {k: v.to(target_dtype) if isinstance(v, torch.Tensor) else v for k, v in kwargs.items()}
+
+        org_dtype = torch_utils.get_param(self).dtype
+        if org_dtype != target_dtype:
+            self.to(target_dtype)
+        result = self.org_forward(*args, **kwargs)
+        if org_dtype != target_dtype:
+            self.to(org_dtype)
+
+        if target_dtype != dtype_inference:
+            if isinstance(result, tuple):
+                result = tuple(
+                    i.to(dtype_inference)
+                    if isinstance(i, torch.Tensor)
+                    else i
+                    for i in result
+                )
+            elif isinstance(result, torch.Tensor):
+                result = result.to(dtype_inference)
+        return result
+    return forward_wrapper
 
 
 @contextlib.contextmanager
-def manual_cast():
+def manual_cast(target_dtype):
     for module_type in patch_module_list:
         org_forward = module_type.forward
-        module_type.forward = manual_cast_forward
+        if module_type == torch.nn.MultiheadAttention and has_xpu():
+            module_type.forward = manual_cast_forward(torch.float32)
+        else:
+            module_type.forward = manual_cast_forward(target_dtype)
         module_type.org_forward = org_forward
     try:
         yield None
@@ -161,15 +185,15 @@ def autocast(disable=False):
     if fp8 and device==cpu:
         return torch.autocast("cpu", dtype=torch.bfloat16, enabled=True)
 
-    if fp8 and (dtype == torch.float32 or shared.cmd_opts.precision == "full" or cuda_no_autocast()):
-        return manual_cast()
+    if fp8 and dtype_inference == torch.float32:
+        return manual_cast(dtype)
 
-    if has_mps() and shared.cmd_opts.precision != "full":
-        return manual_cast()
-
-    if dtype == torch.float32 or shared.cmd_opts.precision == "full":
+    if dtype == torch.float32 or dtype_inference == torch.float32:
         return contextlib.nullcontext()
 
+    if has_xpu() or has_mps() or cuda_no_autocast():
+        return manual_cast(dtype)
+
     return torch.autocast("cuda")