Update model_loading.py

model_loading.py

@@ -4,6 +4,8 @@ import torch.nn as nn
 import json
 import folder_paths
 import comfy.model_management as mm
+from .utils import check_diffusers_version, remove_specific_blocks, log
+check_diffusers_version()
 
 from diffusers.models import AutoencoderKLCogVideoX
 from diffusers.schedulers import CogVideoXDDIMScheduler
@@ -20,7 +22,6 @@ from .cogvideox_fun.pipeline_cogvideox_control import CogVideoX_Fun_Pipeline_Con
 
 from .videosys.cogvideox_transformer_3d import CogVideoXTransformer3DModel as CogVideoXTransformer3DModelPAB
 
-from .utils import check_diffusers_version, remove_specific_blocks, log
 from comfy.utils import load_torch_file
 
 script_directory = os.path.dirname(os.path.abspath(__file__))
@@ -32,7 +33,7 @@ class CogVideoLoraSelect:
             "required": {
                "lora": (folder_paths.get_filename_list("cogvideox_loras"), 
                 {"tooltip": "LORA models are expected to be in ComfyUI/models/CogVideo/loras with .safetensors extension"}),
-                "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01, "tooltip": "LORA strength, set to 0.0 to unmerge the LORA"}),
+                "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.0001, "tooltip": "LORA strength, set to 0.0 to unmerge the LORA"}),
             },
             "optional": {
                 "prev_lora":("COGLORA", {"default": None, "tooltip": "For loading multiple LoRAs"}),
@@ -182,16 +183,7 @@ class DownloadAndLoadCogVideoModel:
         if block_edit is not None:
             transformer = remove_specific_blocks(transformer, block_edit)
         
-        #fp8
-        if fp8_transformer == "enabled" or fp8_transformer == "fastmode":
-            for name, param in transformer.named_parameters():
-                params_to_keep = {"patch_embed", "lora", "pos_embedding"}
-                if not any(keyword in name for keyword in params_to_keep):
-                    param.data = param.data.to(torch.float8_e4m3fn)
         
-            if fp8_transformer == "fastmode":
-                from .fp8_optimization import convert_fp8_linear
-                convert_fp8_linear(transformer, dtype)
 
         with open(scheduler_path) as f:
             scheduler_config = json.load(f)
@@ -221,8 +213,7 @@ class DownloadAndLoadCogVideoModel:
                 adapter_list = []
                 adapter_weights = []
                 for l in lora:
-                    if l["fuse_lora"]:
-                        fuse = True
+                    fuse = True if l["fuse_lora"] else False
                     lora_sd = load_torch_file(l["path"])             
                     for key, val in lora_sd.items():
                         if "lora_B" in key:
@@ -239,9 +230,18 @@ class DownloadAndLoadCogVideoModel:
                 for l in lora:
                     pipe.set_adapters(adapter_list, adapter_weights=adapter_weights)
                 if fuse:
-                    pipe.fuse_lora(lora_scale=1 / lora_rank, components=["transformer"])
+                    pipe.fuse_lora(lora_scale=lora[-1]["strength"] / lora_rank, components=["transformer"])
 
+        #fp8
+        if fp8_transformer == "enabled" or fp8_transformer == "fastmode":
+            for name, param in pipe.transformer.named_parameters():
+                params_to_keep = {"patch_embed", "lora", "pos_embedding"}
+                if not any(keyword in name for keyword in params_to_keep):
+                    param.data = param.data.to(torch.float8_e4m3fn)
         
+            if fp8_transformer == "fastmode":
+                from .fp8_optimization import convert_fp8_linear
+                convert_fp8_linear(pipe.transformer, dtype)
 
         if enable_sequential_cpu_offload:
             pipe.enable_sequential_cpu_offload()
@@ -312,7 +312,6 @@ class DownloadAndLoadCogVideoGGUFModel:
                 "pab_config": ("PAB_CONFIG", {"default": None}),
                 "block_edit": ("TRANSFORMERBLOCKS", {"default": None}),
                 "compile": (["disabled","torch"], {"tooltip": "compile the model for faster inference, these are advanced options only available on Linux, see readme for more info"}),
-              
             }
         }
 

requirements.txt

@@ -1,5 +1,5 @@
 huggingface_hub
-diffusers>=0.30.3
+diffusers>=0.31.0
 accelerate>=0.33.0
 einops
 peft