Merge branch 'master' into worksplit-multigpu

comfy/model_management.py

@@ -321,11 +321,16 @@ except:
 
 try:
     if is_amd():
+        try:
+            rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
+        except:
+            rocm_version = (6, -1)
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         logging.info("AMD arch: {}".format(arch))
+        logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if torch_version_numeric[0] >= 2 and torch_version_numeric[1] >= 7:  # works on 2.6 but doesn't actually seem to improve much
-                if any((a in arch) for a in ["gfx1100", "gfx1101"]):  # TODO: more arches
+                if any((a in arch) for a in ["gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches
                     ENABLE_PYTORCH_ATTENTION = True
 except:
     pass
@@ -723,7 +728,7 @@ def unet_inital_load_device(parameters, dtype):
         return torch_dev
 
     cpu_dev = torch.device("cpu")
-    if DISABLE_SMART_MEMORY:
+    if DISABLE_SMART_MEMORY or vram_state == VRAMState.NO_VRAM:
         return cpu_dev
 
     model_size = dtype_size(dtype) * parameters
@@ -1285,6 +1290,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     return False
 
 def supports_fp8_compute(device=None):
+    if args.supports_fp8_compute:
+        return True
+
     if not is_nvidia():
         return False