Merge upstream/master, keep local README.md

comfy/ldm/chroma_radiance/model.py

@@ -37,7 +37,7 @@ class ChromaRadianceParams(ChromaParams):
     nerf_final_head_type: str
     # None means use the same dtype as the model.
     nerf_embedder_dtype: Optional[torch.dtype]
-
+    use_x0: bool
 
 class ChromaRadiance(Chroma):
     """
@@ -159,6 +159,9 @@ class ChromaRadiance(Chroma):
         self.skip_dit = []
         self.lite = False
 
+        if params.use_x0:
+            self.register_buffer("__x0__", torch.tensor([]))
+
     @property
     def _nerf_final_layer(self) -> nn.Module:
         if self.params.nerf_final_head_type == "linear":
@@ -276,6 +279,12 @@ class ChromaRadiance(Chroma):
         params_dict |= overrides
         return params.__class__(**params_dict)
 
+    def _apply_x0_residual(self, predicted, noisy, timesteps):
+
+        # non zero during training to prevent 0 div
+        eps = 0.0
+        return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps)
+
     def _forward(
         self,
         x: Tensor,
@@ -316,4 +325,11 @@ class ChromaRadiance(Chroma):
             transformer_options,
             attn_mask=kwargs.get("attention_mask", None),
         )
-        return self.forward_nerf(img, img_out, params)[:, :, :h, :w]
+
+        out = self.forward_nerf(img, img_out, params)[:, :, :h, :w]
+
+        # If x0 variant → v-pred, just return this instead
+        if hasattr(self, "__x0__"):
+            out = self._apply_x0_residual(out, img, timestep)
+        return out
+

comfy/ldm/hunyuan_video/model.py

@@ -43,6 +43,7 @@ class HunyuanVideoParams:
     meanflow: bool
     use_cond_type_embedding: bool
     vision_in_dim: int
+    meanflow_sum: bool
 
 
 class SelfAttentionRef(nn.Module):
@@ -317,7 +318,7 @@ class HunyuanVideo(nn.Module):
                 timesteps_r = transformer_options['sample_sigmas'][w[0] + 1]
                 timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype)
                 vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype))
-                vec = (vec + vec_r) / 2
+                vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2
 
         if ref_latent is not None:
             ref_latent_ids = self.img_ids(ref_latent)

comfy/model_detection.py

@@ -180,8 +180,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["use_cond_type_embedding"] = False
         if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys:
             dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0]
+            dit_config["meanflow_sum"] = True
         else:
             dit_config["vision_in_dim"] = None
+            dit_config["meanflow_sum"] = False
         return dit_config
 
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
@@ -257,6 +259,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_tile_size"] = 512
                 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
+            if "__x0__" in state_dict_keys: # x0 pred
+                dit_config["use_x0"] = True
         else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
             dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys

comfy/model_patcher.py

@@ -35,6 +35,7 @@ import comfy.model_management
 import comfy.patcher_extension
 import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
+from comfy.quant_ops import QuantizedTensor
 from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP
 
 
@@ -665,12 +666,18 @@ class ModelPatcher:
                 module_mem = comfy.model_management.module_size(m)
                 module_offload_mem = module_mem
                 if hasattr(m, "comfy_cast_weights"):
-                    weight_key = "{}.weight".format(n)
+                    def check_module_offload_mem(key):
-                    bias_key = "{}.bias".format(n)
+                        if key in self.patches:
-                    if weight_key in self.patches:
+                            return low_vram_patch_estimate_vram(self.model, key)
-                        module_offload_mem += low_vram_patch_estimate_vram(self.model, weight_key)
+                        model_dtype = getattr(self.model, "manual_cast_dtype", None)
-                    if bias_key in self.patches:
+                        weight, _, _ = get_key_weight(self.model, key)
-                        module_offload_mem += low_vram_patch_estimate_vram(self.model, bias_key)
+                        if model_dtype is None or weight is None:
+                            return 0
+                        if (weight.dtype != model_dtype or isinstance(weight, QuantizedTensor)):
+                            return weight.numel() * model_dtype.itemsize
+                        return 0
+                    module_offload_mem += check_module_offload_mem("{}.weight".format(n))
+                    module_offload_mem += check_module_offload_mem("{}.bias".format(n))
                 loading.append((module_offload_mem, module_mem, n, m, params))
         return loading
 
@@ -923,7 +930,7 @@ class ModelPatcher:
                                     patch_counter += 1
                             cast_weight = True
 
-                        if cast_weight:
+                        if cast_weight and hasattr(m, "comfy_cast_weights"):
                             m.prev_comfy_cast_weights = m.comfy_cast_weights
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False

comfy/ops.py

@@ -22,7 +22,6 @@ import comfy.model_management
 from comfy.cli_args import args, PerformanceFeature
 import comfy.float
 import comfy.rmsnorm
-import contextlib
 import json
 
 def run_every_op():
@@ -94,13 +93,6 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     else:
         offload_stream = None
 
-    if offload_stream is not None:
-        wf_context = offload_stream
-        if hasattr(wf_context, "as_context"):
-            wf_context = wf_context.as_context(offload_stream)
-    else:
-        wf_context = contextlib.nullcontext()
-
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
 
     weight_has_function = len(s.weight_function) > 0

comfy/quant_ops.py

@@ -399,7 +399,10 @@ class TensorCoreFP8Layout(QuantizedLayout):
         orig_dtype = tensor.dtype
 
         if isinstance(scale, str) and scale == "recalculate":
-            scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
+            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max
+            if tensor.dtype not in [torch.float32, torch.bfloat16]:  # Prevent scale from being too small
+                tensor_info = torch.finfo(tensor.dtype)
+                scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
 
         if scale is not None:
             if not isinstance(scale, torch.Tensor):

comfy/sd.py

@@ -127,6 +127,8 @@ class CLIP:
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
         self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        #Match torch.float32 hardcode upcast in TE implemention
+        self.patcher.set_model_compute_dtype(torch.float32)
         self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
         self.patcher.is_clip = True
         self.apply_hooks_to_conds = None

comfy/utils.py

@@ -803,12 +803,17 @@ def safetensors_header(safetensors_path, max_size=100*1024*1024):
             return None
         return f.read(length_of_header)
 
+ATTR_UNSET={}
+
 def set_attr(obj, attr, value):
     attrs = attr.split(".")
     for name in attrs[:-1]:
         obj = getattr(obj, name)
-    prev = getattr(obj, attrs[-1])
+    prev = getattr(obj, attrs[-1], ATTR_UNSET)
-    setattr(obj, attrs[-1], value)
+    if value is ATTR_UNSET:
+        delattr(obj, attrs[-1])
+    else:
+        setattr(obj, attrs[-1], value)
     return prev
 
 def set_attr_param(obj, attr, value):

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.76"
+__version__ = "0.4.0"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.76"
+version = "0.4.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.33.10
+comfyui-frontend-package==1.33.13
 comfyui-workflow-templates==0.7.54
 comfyui-embedded-docs==0.3.1
 torch