Merge branch 'master' into dr-support-pip-cm

comfy/cli_args.py

@@ -88,6 +88,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
+parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/ldm/chroma/model.py

@@ -163,7 +163,7 @@ class Chroma(nn.Module):
         distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype)
 
         # get all modulation index
-        modulation_index = timestep_embedding(torch.arange(mod_index_length), 32).to(img.device, img.dtype)
+        modulation_index = timestep_embedding(torch.arange(mod_index_length, device=img.device), 32).to(img.device, img.dtype)
         # we need to broadcast the modulation index here so each batch has all of the index
         modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype)
         # and we need to broadcast timestep and guidance along too

comfy/ldm/modules/attention.py

@@ -20,8 +20,11 @@ if model_management.xformers_enabled():
 if model_management.sage_attention_enabled():
     try:
         from sageattention import sageattn
-    except ModuleNotFoundError:
-        logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
+    except ModuleNotFoundError as e:
+        if e.name == "sageattention":
+            logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
+        else:
+            raise e
         exit(-1)
 
 if model_management.flash_attention_enabled():

comfy/ldm/wan/model.py

@@ -635,7 +635,7 @@ class VaceWanModel(WanModel):
         t,
         context,
         vace_context,
-        vace_strength=1.0,
+        vace_strength,
         clip_fea=None,
         freqs=None,
         transformer_options={},
@@ -661,8 +661,11 @@ class VaceWanModel(WanModel):
                 context = torch.concat([context_clip, context], dim=1)
             context_img_len = clip_fea.shape[-2]
 
+        orig_shape = list(vace_context.shape)
+        vace_context = vace_context.movedim(0, 1).reshape([-1] + orig_shape[2:])
         c = self.vace_patch_embedding(vace_context.float()).to(vace_context.dtype)
         c = c.flatten(2).transpose(1, 2)
+        c = list(c.split(orig_shape[0], dim=0))
 
         # arguments
         x_orig = x
@@ -682,8 +685,9 @@ class VaceWanModel(WanModel):
 
             ii = self.vace_layers_mapping.get(i, None)
             if ii is not None:
-                c_skip, c = self.vace_blocks[ii](c, x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
-                x += c_skip * vace_strength
+                for iii in range(len(c)):
+                    c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                    x += c_skip * vace_strength[iii]
                 del c_skip
         # head
         x = self.head(x, e)

comfy/model_base.py

@@ -1062,20 +1062,25 @@ class WAN21_Vace(WAN21):
         vace_frames = kwargs.get("vace_frames", None)
         if vace_frames is None:
             noise_shape[1] = 32
-            vace_frames = torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)
-
-        for i in range(0, vace_frames.shape[1], 16):
-            vace_frames = vace_frames.clone()
-            vace_frames[:, i:i + 16] = self.process_latent_in(vace_frames[:, i:i + 16])
+            vace_frames = [torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)]
 
         mask = kwargs.get("vace_mask", None)
         if mask is None:
             noise_shape[1] = 64
-            mask = torch.ones(noise_shape, device=noise.device, dtype=noise.dtype)
+            mask = [torch.ones(noise_shape, device=noise.device, dtype=noise.dtype)] * len(vace_frames)
 
-        out['vace_context'] = comfy.conds.CONDRegular(torch.cat([vace_frames.to(noise), mask.to(noise)], dim=1))
+        vace_frames_out = []
+        for j in range(len(vace_frames)):
+            vf = vace_frames[j].clone()
+            for i in range(0, vf.shape[1], 16):
+                vf[:, i:i + 16] = self.process_latent_in(vf[:, i:i + 16])
+            vf = torch.cat([vf, mask[j]], dim=1)
+            vace_frames_out.append(vf)
 
-        vace_strength = kwargs.get("vace_strength", 1.0)
+        vace_frames = torch.stack(vace_frames_out, dim=1)
+        out['vace_context'] = comfy.conds.CONDRegular(vace_frames)
+
+        vace_strength = kwargs.get("vace_strength", [1.0] * len(vace_frames_out))
         out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
         return out
 

comfy/model_detection.py

@@ -620,6 +620,9 @@ def convert_config(unet_config):
 
 
 def unet_config_from_diffusers_unet(state_dict, dtype=None):
+    if "conv_in.weight" not in state_dict:
+        return None
+
     match = {}
     transformer_depth = []
 

comfy/model_management.py

@@ -1257,6 +1257,9 @@ 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