Merge 9ee905bc47 into 439bd807f8

comfy/model_base.py

@@ -1735,22 +1735,23 @@ class WAN21_HuMo(WAN21):
         if audio_embed is not None:
             out['audio_embed'] = comfy.conds.CONDRegular(audio_embed)
 
-        if "c_concat" not in out:  # 1.7B model
             reference_latents = kwargs.get("reference_latents", None)
-            if reference_latents is not None:
+
+            if "c_concat" not in out and reference_latents is not None and reference_latents[0].shape[1] == 16:  # 1.7B model
                 out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1]))
-        else:
-            noise_shape = list(noise.shape)
-            noise_shape[1] += 4
-            concat_latent = torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)
-            zero_vae_values_first = torch.tensor([0.8660, -0.4326, -0.0017, -0.4884, -0.5283, 0.9207, -0.9896, 0.4433, -0.5543, -0.0113, 0.5753, -0.6000, -0.8346, -0.3497, -0.1926, -0.6938]).view(1, 16, 1, 1, 1)
-            zero_vae_values_second = torch.tensor([1.0869, -1.2370, 0.0206, -0.4357, -0.6411, 2.0307, -1.5972, 1.2659, -0.8595, -0.4654, 0.9638, -1.6330, -1.4310, -0.1098, -0.3856, -1.4583]).view(1, 16, 1, 1, 1)
-            zero_vae_values = torch.tensor([0.8642, -1.8583, 0.1577, 0.1350, -0.3641, 2.5863, -1.9670, 1.6065, -1.0475, -0.8678, 1.1734, -1.8138, -1.5933, -0.7721, -0.3289, -1.3745]).view(1, 16, 1, 1, 1)
-            concat_latent[:, 4:] = zero_vae_values
-            concat_latent[:, 4:, :1] = zero_vae_values_first
-            concat_latent[:, 4:, 1:2] = zero_vae_values_second
-            out['c_concat'] = comfy.conds.CONDNoiseShape(concat_latent)
-            reference_latents = kwargs.get("reference_latents", None)
+            else:
+                concat_latent_image = kwargs.get("concat_latent_image", None)
+                if concat_latent_image is None:
+                    noise_shape = list(noise.shape)
+                    noise_shape[1] += 4
+                    concat_latent = torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)
+                    zero_vae_values_first = torch.tensor([0.8660, -0.4326, -0.0017, -0.4884, -0.5283, 0.9207, -0.9896, 0.4433, -0.5543, -0.0113, 0.5753, -0.6000, -0.8346, -0.3497, -0.1926, -0.6938]).view(1, 16, 1, 1, 1)
+                    zero_vae_values_second = torch.tensor([1.0869, -1.2370, 0.0206, -0.4357, -0.6411, 2.0307, -1.5972, 1.2659, -0.8595, -0.4654, 0.9638, -1.6330, -1.4310, -0.1098, -0.3856, -1.4583]).view(1, 16, 1, 1, 1)
+                    zero_vae_values = torch.tensor([0.8642, -1.8583, 0.1577, 0.1350, -0.3641, 2.5863, -1.9670, 1.6065, -1.0475, -0.8678, 1.1734, -1.8138, -1.5933, -0.7721, -0.3289, -1.3745]).view(1, 16, 1, 1, 1)
+                    concat_latent[:, 4:] = zero_vae_values
+                    concat_latent[:, 4:, :1] = zero_vae_values_first
+                    concat_latent[:, 4:, 1:2] = zero_vae_values_second
+                    out['c_concat'] = comfy.conds.CONDNoiseShape(concat_latent)
             if reference_latents is not None:
                 ref_latent = self.process_latent_in(reference_latents[-1])
                 ref_latent_shape = list(ref_latent.shape)

comfy/sd.py

@@ -468,6 +468,9 @@ class CLIP:
     def decode(self, token_ids, skip_special_tokens=True):
         return self.tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens)
 
+    def is_dynamic(self):
+        return self.patcher.is_dynamic()
+
 class VAE:
     def __init__(self, sd=None, device=None, config=None, dtype=None, metadata=None):
         if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
@@ -1251,6 +1254,8 @@ class VAE:
         except:
             return None
 
+    def is_dynamic(self):
+        return self.patcher.is_dynamic()
 
 class StyleModel:
     def __init__(self, model, device="cpu"):

comfy_execution/caching.py

@@ -503,6 +503,21 @@ RAM_CACHE_DEFAULT_RAM_USAGE = 0.05
 
 RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER = 1.3
 
+
+def all_outputs_dynamic(outputs):
+    if outputs is None:
+        return False
+
+    for output in outputs:
+        if isinstance(output, (list, tuple)):
+            if not all_outputs_dynamic(output):
+                return False
+        elif not hasattr(output, "is_dynamic") or not output.is_dynamic():
+            return False
+
+    return True
+
+
 class RAMPressureCache(LRUCache):
 
     def __init__(self, key_class, enable_providers=False):
@@ -533,7 +548,11 @@ class RAMPressureCache(LRUCache):
         for key, cache_entry in self.cache.items():
             if not free_active and self.used_generation[key] == self.generation:
                 continue
-            oom_score =  RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER ** (self.generation - self.used_generation[key])
+
+            if all_outputs_dynamic(cache_entry.outputs) and self.used_generation[key] == self.generation:
+                continue
+
+            oom_score = RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER ** (self.generation - self.used_generation[key])
 
             ram_usage = RAM_CACHE_DEFAULT_RAM_USAGE
             def scan_list_for_ram_usage(outputs):