Add model_options for text encoder.

comfy/sd.py

@@ -62,7 +62,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
 
 
 class CLIP:
-    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0):
+    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}):
         if no_init:
             return
         params = target.params.copy()
@@ -71,9 +71,14 @@ class CLIP:
 
         load_device = model_management.text_encoder_device()
         offload_device = model_management.text_encoder_offload_device()
-        dtype = model_management.text_encoder_dtype(load_device)
+        dtype = model_options.get("dtype", None)
+        if dtype is None:
+            dtype = model_management.text_encoder_dtype(load_device)
+
         params['dtype'] = dtype
         params['device'] = model_management.text_encoder_initial_device(load_device, offload_device, parameters * model_management.dtype_size(dtype))
+        params['model_options'] = model_options
+
         self.cond_stage_model = clip(**(params))
 
         for dt in self.cond_stage_model.dtypes:
@@ -394,7 +399,7 @@ class CLIPType(Enum):
     HUNYUAN_DIT = 5
     FLUX = 6
 
-def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION):
+def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
     clip_data = []
     for p in ckpt_paths:
         clip_data.append(comfy.utils.load_torch_file(p, safe_load=True))
@@ -464,7 +469,7 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
     for c in clip_data:
         parameters += comfy.utils.calculate_parameters(c)
 
-    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters)
+    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, model_options=model_options)
     for c in clip_data:
         m, u = clip.load_sd(c)
         if len(m) > 0:
@@ -506,14 +511,14 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
 
     return (model, clip, vae)
 
-def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}):
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     sd = comfy.utils.load_torch_file(ckpt_path)
-    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options)
+    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options)
     if out is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
     return out
 
-def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}):
+def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     clip = None
     clipvision = None
     vae = None
@@ -563,7 +568,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
             clip_sd = model_config.process_clip_state_dict(sd)
             if len(clip_sd) > 0:
                 parameters = comfy.utils.calculate_parameters(clip_sd)
-                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters)
+                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options)
                 m, u = clip.load_sd(clip_sd, full_model=True)
                 if len(m) > 0:
                     m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))

comfy/sd1_clip.py

@@ -84,7 +84,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
                  freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
                  special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False,
-                 return_projected_pooled=True, return_attention_masks=False):  # clip-vit-base-patch32
+                 return_projected_pooled=True, return_attention_masks=False, model_options={}):  # clip-vit-base-patch32
         super().__init__()
         assert layer in self.LAYERS
 
@@ -94,7 +94,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         with open(textmodel_json_config) as f:
             config = json.load(f)
 
-        self.operations = comfy.ops.manual_cast
+        operations = model_options.get("custom_operations", None)
+        if operations is None:
+            operations = comfy.ops.manual_cast
+
+        self.operations = operations
         self.transformer = model_class(config, dtype, device, self.operations)
         self.num_layers = self.transformer.num_layers
 
@@ -553,7 +557,7 @@ class SD1Tokenizer:
         return {}
 
 class SD1ClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=SDClipModel, name=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, clip_name="l", clip_model=SDClipModel, name=None, **kwargs):
         super().__init__()
 
         if name is not None:
@@ -563,7 +567,7 @@ class SD1ClipModel(torch.nn.Module):
             self.clip_name = clip_name
             self.clip = "clip_{}".format(self.clip_name)
 
-        setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
+        setattr(self, self.clip, clip_model(device=device, dtype=dtype, model_options=model_options, **kwargs))
 
         self.dtypes = set()
         if dtype is not None:

comfy/sdxl_clip.py

@@ -3,14 +3,14 @@ import torch
 import os
 
 class SDXLClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
@@ -38,10 +38,10 @@ class SDXLTokenizer:
         return {}
 
 class SDXLClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
-        self.clip_g = SDXLClipG(device=device, dtype=dtype)
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
+        self.clip_g = SDXLClipG(device=device, dtype=dtype, model_options=model_options)
         self.dtypes = set([dtype])
 
     def set_clip_options(self, options):
@@ -66,8 +66,8 @@ class SDXLClipModel(torch.nn.Module):
             return self.clip_l.load_sd(sd)
 
 class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG)
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG, model_options=model_options)
 
 
 class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer):
@@ -79,14 +79,14 @@ class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="g", tokenizer=StableCascadeClipGTokenizer)
 
 class StableCascadeClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
 
 class StableCascadeClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG)
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG, model_options=model_options)

comfy/text_encoders/aura_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class PT5XlModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_pile_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True, model_options=model_options)
 
 class PT5XlTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class AuraT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="pile_t5xl", tokenizer=PT5XlTokenizer)
 
 class AuraT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)

comfy/text_encoders/flux.py

@@ -6,9 +6,9 @@ import torch
 import os
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -35,11 +35,11 @@ class FluxTokenizer:
 
 
 class FluxClipModel(torch.nn.Module):
-    def __init__(self, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False)
-        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
+        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
         self.dtypes = set([dtype, dtype_t5])
 
     def set_clip_options(self, options):
@@ -66,6 +66,6 @@ class FluxClipModel(torch.nn.Module):
 
 def flux_clip(dtype_t5=None):
     class FluxClipModel_(FluxClipModel):
-        def __init__(self, device="cpu", dtype=None):
-            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype)
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_

comfy/text_encoders/hydit.py

@@ -7,9 +7,9 @@ import os
 import torch
 
 class HyditBertModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class HyditBertTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,9 +18,9 @@ class HyditBertTokenizer(sd1_clip.SDTokenizer):
 
 
 class MT5XLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class MT5XLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -50,10 +50,10 @@ class HyditTokenizer:
         return {"mt5xl.spiece_model": self.mt5xl.state_dict()["spiece_model"]}
 
 class HyditModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.hydit_clip = HyditBertModel(dtype=dtype)
-        self.mt5xl = MT5XLModel(dtype=dtype)
+        self.hydit_clip = HyditBertModel(dtype=dtype, model_options=model_options)
+        self.mt5xl = MT5XLModel(dtype=dtype, model_options=model_options)
 
         self.dtypes = set()
         if dtype is not None:

comfy/text_encoders/sa_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class T5BaseModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_base.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, model_options=model_options, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
 
 class T5BaseTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class SAT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5base", tokenizer=T5BaseTokenizer)
 
 class SAT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="t5base", clip_model=T5BaseModel, **kwargs)
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="t5base", clip_model=T5BaseModel, **kwargs)

comfy/text_encoders/sd2_clip.py

@@ -2,13 +2,13 @@ from comfy import sd1_clip
 import os
 
 class SD2ClipHModel(sd1_clip.SDClipModel):
-    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
-        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0}, model_options=model_options)
 
 class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
@@ -19,5 +19,5 @@ class SD2Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="h", tokenizer=SD2ClipHTokenizer)
 
 class SD2ClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
-        super().__init__(device=device, dtype=dtype, clip_name="h", clip_model=SD2ClipHModel, **kwargs)
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_name="h", clip_model=SD2ClipHModel, **kwargs)

comfy/text_encoders/sd3_clip.py

@@ -8,9 +8,9 @@ import comfy.model_management
 import logging
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -38,24 +38,24 @@ class SD3Tokenizer:
         return {}
 
 class SD3ClipModel(torch.nn.Module):
-    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         self.dtypes = set()
         if clip_l:
-            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False)
+            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_l = None
 
         if clip_g:
-            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype)
+            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_g = None
 
         if t5:
             dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
             self.dtypes.add(dtype_t5)
         else:
             self.t5xxl = None
@@ -132,6 +132,6 @@ class SD3ClipModel(torch.nn.Module):
 
 def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None):
     class SD3ClipModel_(SD3ClipModel):
-        def __init__(self, device="cpu", dtype=None):
-            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype)
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return SD3ClipModel_