Merge branch 'master' into worksplit-multigpu

comfy/cli_args.py

@@ -79,6 +79,7 @@ fpte_group.add_argument("--fp8_e4m3fn-text-enc", action="store_true", help="Stor
 fpte_group.add_argument("--fp8_e5m2-text-enc", action="store_true", help="Store text encoder weights in fp8 (e5m2 variant).")
 fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text encoder weights in fp16.")
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
+fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 

comfy/clip_vision.py

@@ -110,9 +110,13 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
     elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd:
         json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json")
     elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
+        embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
         if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
-            json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
-        elif sd["vision_model.embeddings.position_embedding.weight"].shape[0] == 577:
+            if embed_shape == 729:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
+            elif embed_shape == 1024:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
+        elif embed_shape == 577:
             if "multi_modal_projector.linear_1.bias" in sd:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")
             else:

comfy/clip_vision_siglip_512.json

@@ -0,0 +1,13 @@
+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": 512,
+  "intermediate_size": 4304,
+  "model_type": "siglip_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 16,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/ldm/modules/attention.py

@@ -847,6 +847,7 @@ class SpatialTransformer(nn.Module):
         if not isinstance(context, list):
             context = [context] * len(self.transformer_blocks)
         b, c, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
         x_in = x
         x = self.norm(x)
         if not self.use_linear:
@@ -962,6 +963,7 @@ class SpatialVideoTransformer(SpatialTransformer):
         transformer_options={}
     ) -> torch.Tensor:
         _, _, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
         x_in = x
         spatial_context = None
         if exists(context):

comfy/lora_convert.py

@@ -1,4 +1,5 @@
 import torch
+import comfy.utils
 
 
 def convert_lora_bfl_control(sd): #BFL loras for Flux
@@ -11,7 +12,13 @@ def convert_lora_bfl_control(sd): #BFL loras for Flux
     return sd_out
 
 
+def convert_lora_wan_fun(sd): #Wan Fun loras
+    return comfy.utils.state_dict_prefix_replace(sd, {"lora_unet__": "lora_unet_"})
+
+
 def convert_lora(sd):
     if "img_in.lora_A.weight" in sd and "single_blocks.0.norm.key_norm.scale" in sd:
         return convert_lora_bfl_control(sd)
+    if "lora_unet__blocks_0_cross_attn_k.lora_down.weight" in sd:
+        return convert_lora_wan_fun(sd)
     return sd

comfy/model_base.py

@@ -1000,7 +1000,6 @@ class WAN21(BaseModel):
         device = kwargs["device"]
 
         if image is None:
-            image = torch.zeros_like(noise)
             shape_image = list(noise.shape)
             shape_image[1] = extra_channels
             image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
@@ -1013,6 +1012,9 @@ class WAN21(BaseModel):
         if not self.image_to_video or extra_channels == image.shape[1]:
             return image
 
+        if image.shape[1] > (extra_channels - 4):
+            image = image[:, :(extra_channels - 4)]
+
         mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
         if mask is None:
             mask = torch.zeros_like(noise)[:, :4]

comfy/model_management.py

@@ -851,6 +851,8 @@ def text_encoder_dtype(device=None):
         return torch.float8_e5m2
     elif args.fp16_text_enc:
         return torch.float16
+    elif args.bf16_text_enc:
+        return torch.bfloat16
     elif args.fp32_text_enc:
         return torch.float32
 
@@ -1263,6 +1265,8 @@ def soft_empty_cache(force=False):
         torch.xpu.empty_cache()
     elif is_ascend_npu():
         torch.npu.empty_cache()
+    elif is_mlu():
+        torch.mlu.empty_cache()
     elif torch.cuda.is_available():
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()

comfy/sd.py

@@ -265,6 +265,7 @@ class VAE:
         self.process_input = lambda image: image * 2.0 - 1.0
         self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)
         self.working_dtypes = [torch.bfloat16, torch.float32]
+        self.disable_offload = False
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
@@ -337,6 +338,7 @@ class VAE:
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+                self.disable_offload = True
             elif "blocks.2.blocks.3.stack.5.weight" in sd or "decoder.blocks.2.blocks.3.stack.5.weight" in sd or "layers.4.layers.1.attn_block.attn.qkv.weight" in sd or "encoder.layers.4.layers.1.attn_block.attn.qkv.weight" in sd: #genmo mochi vae
                 if "blocks.2.blocks.3.stack.5.weight" in sd:
                     sd = comfy.utils.state_dict_prefix_replace(sd, {"": "decoder."})
@@ -515,7 +517,7 @@ class VAE:
         pixel_samples = None
         try:
             memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
-            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
@@ -544,7 +546,7 @@ class VAE:
     def decode_tiled(self, samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
         self.throw_exception_if_invalid()
         memory_used = self.memory_used_decode(samples.shape, self.vae_dtype) #TODO: calculate mem required for tile
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
         dims = samples.ndim - 2
         args = {}
         if tile_x is not None:
@@ -578,7 +580,7 @@ class VAE:
             pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
         try:
             memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
-            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / max(1, memory_used))
             batch_number = max(1, batch_number)
@@ -612,7 +614,7 @@ class VAE:
             pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
 
         memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)  # TODO: calculate mem required for tile
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
 
         args = {}
         if tile_x is not None: