Merge 9ee905bc47 into fc0cb10bcb

.github/workflows/stable-release.yml

@@ -117,7 +117,7 @@ jobs:
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
 
-          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
           ./python.exe -s -m pip install -r requirements_comfyui.txt
           rm requirements_comfyui.txt
 

comfy/model_patcher.py

@@ -718,7 +718,6 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
-                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -791,12 +790,11 @@ class ModelPatcher:
                 for param in params:
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
-            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)

comfy/ops.py

@@ -654,29 +654,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 run_every_op()
 
                 input_shape = input.shape
-                reshaped_3d = False
+                tensor_3d = input.ndim == 3
+
+                if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
+                    return self.forward_comfy_cast_weights(input, *args, **kwargs)
 
                 if (getattr(self, 'layout_type', None) is not None and
-                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
-                    not getattr(self, 'comfy_force_cast_weights', False) and
-                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
+                    not isinstance(input, QuantizedTensor)):
 
                     # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
-                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
+                    if tensor_3d:
+                        input = input.reshape(-1, input_shape[2])
 
-                    # Fall back to non-quantized for non-2D tensors
-                    if input_reshaped.ndim == 2:
-                        reshaped_3d = input.ndim == 3
-                        # dtype is now implicit in the layout class
-                        scale = getattr(self, 'input_scale', None)
-                        if scale is not None:
-                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
-                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
+                    if input.ndim != 2:
+                        # Fall back to comfy_cast_weights for non-2D tensors
+                        return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs)
 
-                output = self.forward_comfy_cast_weights(input)
+                    # dtype is now implicit in the layout class
+                    input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None))
+
+                output = self._forward(input, self.weight, self.bias)
 
                 # Reshape output back to 3D if input was 3D
-                if reshaped_3d:
+                if tensor_3d:
                     output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
 
                 return output

comfy/quant_ops.py

@@ -19,7 +19,6 @@ try:
         cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
         if cuda_version < (13,):
             ck.registry.disable("cuda")
-            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
     ck.registry.disable("triton")
     for k, v in ck.list_backends().items():

comfy/sd.py

@@ -218,7 +218,7 @@ class CLIP:
             if unprojected:
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-            self.load_model(tokens)
+            self.load_model()
             self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
             all_hooks.reset()
             self.patcher.patch_hooks(None)
@@ -266,7 +266,7 @@ class CLIP:
         if return_pooled == "unprojected":
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-        self.load_model(tokens)
+        self.load_model()
         self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
         o = self.cond_stage_model.encode_token_weights(tokens)
         cond, pooled = o[:2]
@@ -299,11 +299,8 @@ class CLIP:
             sd_clip[k] = sd_tokenizer[k]
         return sd_clip
 
-    def load_model(self, tokens={}):
-        memory_used = 0
-        if hasattr(self.cond_stage_model, "memory_estimation_function"):
-            memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device)
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+    def load_model(self):
+        model_management.load_model_gpu(self.patcher)
         return self.patcher
 
     def get_key_patches(self):

comfy/supported_models.py

@@ -845,7 +845,7 @@ class LTXAV(LTXV):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = 0.061  # TODO
+        self.memory_usage_factor = 0.055  # TODO
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.LTXAV(self, device=device)

comfy/text_encoders/lt.py

@@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
-        if llama_quantization_metadata is not None:
+        llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None)
+        if llama_scaled_fp8 is not None:
             model_options = model_options.copy()
-            model_options["quantization_metadata"] = llama_quantization_metadata
+            model_options["scaled_fp8"] = llama_scaled_fp8
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
@@ -98,13 +98,10 @@ class LTXAVTEModel(torch.nn.Module):
 
         out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
         out_device = out.device
-        if comfy.model_management.should_use_bf16(self.execution_device):
-            out = out.to(device=self.execution_device, dtype=torch.bfloat16)
         out = out.movedim(1, -1).to(self.execution_device)
         out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
         out = out.reshape((out.shape[0], out.shape[1], -1))
         out = self.text_embedding_projection(out)
-        out = out.float()
         out_vid = self.video_embeddings_connector(out)[0]
         out_audio = self.audio_embeddings_connector(out)[0]
         out = torch.concat((out_vid, out_audio), dim=-1)
@@ -121,21 +118,13 @@ class LTXAVTEModel(torch.nn.Module):
 
             return self.load_state_dict(sdo, strict=False)
 
-    def memory_estimation_function(self, token_weight_pairs, device=None):
-        constant = 6.0
-        if comfy.model_management.should_use_bf16(device):
-            constant /= 2.0
 
-        token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
-        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
-        return num_tokens * constant * 1024 * 1024
-
-def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
+def ltxav_te(dtype_llama=None, llama_scaled_fp8=None):
     class LTXAVTEModel_(LTXAVTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
+            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+                model_options["llama_scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)

comfy_extras/nodes_lt_audio.py

@@ -185,10 +185,6 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
                 io.Combo.Input(
                     "ckpt_name",
                     options=folder_paths.get_filename_list("checkpoints"),
-                ),
-                io.Combo.Input(
-                    "device",
-                    options=["default", "cpu"],
                 )
             ],
             outputs=[io.Clip.Output()],
@@ -201,11 +197,7 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
         clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder)
         clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)
 
-        model_options = {}
-        if device == "cpu":
-            model_options["load_device"] = model_options["offload_device"] = torch.device("cpu")
-
-        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type, model_options=model_options)
+        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type)
         return io.NodeOutput(clip)
 
 

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.35.9
-comfyui-workflow-templates==0.7.69
+comfyui-workflow-templates==0.7.67
 comfyui-embedded-docs==0.3.1
 torch
 torchsde
@@ -21,7 +21,7 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.5
+comfy-kitchen>=0.2.3
 
 #non essential dependencies:
 kornia>=0.7.1