Add a --force-fp32 argument to force fp32 for debugging.

comfy/cli_args.py

@@ -9,6 +9,7 @@ parser.add_argument("--extra-model-paths-config", type=str, default=None, metava
 parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory.")
 parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
 parser.add_argument("--dont-upcast-attention", action="store_true", help="Disable upcasting of attention. Can boost speed but increase the chances of black images.")
+parser.add_argument("--force-fp32", action="store_true", help="Force fp32 (If this makes your GPU work better please report it).")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization instead of the sub-quadratic one. Ignored when xformers is used.")

comfy/model_management.py

@@ -69,6 +69,11 @@ elif args.novram:
 elif args.highvram:
     vram_state = VRAMState.HIGH_VRAM
 
+FORCE_FP32 = False
+if args.force_fp32:
+    print("Forcing FP32, if this improves things please report it.")
+    FORCE_FP32 = True
+
 
 if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
     try:
@@ -273,6 +278,9 @@ def mps_mode():
 
 def should_use_fp16():
     global xpu_available
+    if FORCE_FP32:
+        return False
+
     if cpu_mode() or mps_mode() or xpu_available:
         return False #TODO ?