Merge 7658d7c155 into cd8c7a2306

This reverts commit ea6880b04b.

comfy/ldm/cosmos/predict2.py

@@ -839,6 +839,14 @@ class MiniTrainDIT(nn.Module):
         **kwargs,
     ):
         orig_shape = list(x.shape)
+
+        ref_latents = kwargs.get('ref_latents', None)
+        if ref_latents is not None:
+            for ref in ref_latents:
+                if ref.ndim == 4:
+                    ref = ref.unsqueeze(2)
+                x = torch.cat([x, ref.to(dtype=x.dtype, device=x.device)], dim=2)
+
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_temporal, self.patch_spatial, self.patch_spatial))
         x_B_C_T_H_W = x
         timesteps_B_T = timesteps

comfy/model_base.py

@@ -1212,6 +1212,7 @@ class CosmosPredict2(BaseModel):
 class Anima(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.anima.model.Anima)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1232,6 +1233,20 @@ class Anima(BaseModel):
                     out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights)
 
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = [1, 16, sum(math.prod(lat.size()[2:]) for lat in ref_latents)]
         return out
 
 class Lumina2(BaseModel):

comfy/model_patcher.py

@@ -26,6 +26,7 @@ import uuid
 from typing import Callable, Optional
 
 import torch
+import tqdm
 
 import comfy.float
 import comfy.hooks
@@ -1651,7 +1652,11 @@ class ModelPatcherDynamic(ModelPatcher):
                 self.model.model_loaded_weight_memory += casted_buf.numel() * casted_buf.element_size()
 
             force_load_stat = f" Force pre-loaded {len(self.backup)} weights: {self.model.model_loaded_weight_memory // 1024} KB." if len(self.backup) > 0 else ""
-            logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.{force_load_stat}")
+            log_key = (self.patches_uuid, allocated_size, num_patches, len(self.backup), self.model.model_loaded_weight_memory)
+            in_loop = bool(getattr(tqdm.tqdm, "_instances", None))
+            level = logging.DEBUG if in_loop and getattr(self, "_last_prepare_log_key", None) == log_key else logging.INFO
+            self._last_prepare_log_key = log_key
+            logging.log(level, f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.{force_load_stat}")
 
             self.model.device = device_to
             self.model.current_weight_patches_uuid = self.patches_uuid

server.py

@@ -560,7 +560,7 @@ class PromptServer():
                             buffer.seek(0)
 
                             return web.Response(body=buffer.read(), content_type=f'image/{image_format}',
-                                                headers={"Content-Disposition": f"attachment; filename=\"{filename}\""})
+                                                headers={"Content-Disposition": f"filename=\"{filename}\""})
 
                     if 'channel' not in request.rel_url.query:
                         channel = 'rgba'
@@ -580,7 +580,7 @@ class PromptServer():
                             buffer.seek(0)
 
                             return web.Response(body=buffer.read(), content_type='image/png',
-                                                headers={"Content-Disposition": f"attachment; filename=\"{filename}\""})
+                                                headers={"Content-Disposition": f"filename=\"{filename}\""})
 
                     elif channel == 'a':
                         with Image.open(file) as img:
@@ -597,7 +597,7 @@ class PromptServer():
                             alpha_buffer.seek(0)
 
                             return web.Response(body=alpha_buffer.read(), content_type='image/png',
-                                                headers={"Content-Disposition": f"attachment; filename=\"{filename}\""})
+                                                headers={"Content-Disposition": f"filename=\"{filename}\""})
                     else:
                         # Use the content type from asset resolution if available,
                         # otherwise guess from the filename.
@@ -614,7 +614,7 @@ class PromptServer():
                         return web.FileResponse(
                             file,
                             headers={
-                                "Content-Disposition": f"attachment; filename=\"{filename}\"",
+                                "Content-Disposition": f"filename=\"{filename}\"",
                                 "Content-Type": content_type
                             }
                         )