Apply the correct behavior of precision='full'

modules/devices.py

@@ -185,11 +185,14 @@ def autocast(disable=False):
     if fp8 and device==cpu:
         return torch.autocast("cpu", dtype=torch.bfloat16, enabled=True)
 
-    if dtype == torch.float32:
-        return contextlib.nullcontext()
-
     if has_xpu() or has_mps() or cuda_no_autocast():
-        return manual_cast(dtype_inference)
+        return manual_cast(dtype)
+
+    if fp8 and dtype_inference == torch.float32:
+        return manual_cast(dtype)
+
+    if dtype == torch.float32 or dtype_inference == torch.float32:
+        return contextlib.nullcontext()
 
     return torch.autocast("cuda")