better caster (#1480)

related to mps/rocm/cpu casting for fp16 and etc on clip

ldm_patched/modules/args_parser.py

@@ -102,7 +102,7 @@ vram_group.add_argument("--always-cpu", action="store_true")
 
 
 parser.add_argument("--always-offload-from-vram", action="store_true")
-
+parser.add_argument("--pytorch-deterministic", action="store_true")
 
 parser.add_argument("--disable-server-log", action="store_true")
 parser.add_argument("--debug-mode", action="store_true")

ldm_patched/modules/model_management.py

@@ -28,6 +28,10 @@ total_vram = 0
 lowvram_available = True
 xpu_available = False
 
+if args.pytorch_deterministic:
+    print("Using deterministic algorithms for pytorch")
+    torch.use_deterministic_algorithms(True, warn_only=True)
+
 directml_enabled = False
 if args.directml is not None:
     import torch_directml