Merge 7b7becc34c into 1265955b34

If the same weight is used multiple times within the same prefetch
window, it should only apply compute state mutations once. Mark the
weight as fully resident on the first pass accordingly.

comfy/model_management.py

@@ -721,13 +721,15 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
     else:
         minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
 
-    models_temp = set()
+    # Order-preserving dedup. A plain set() would randomize iteration order across runs
+    models_temp = {}
     for m in models:
-        models_temp.add(m)
+        models_temp[m] = None
         for mm in m.model_patches_models():
-            models_temp.add(mm)
+            models_temp[mm] = None
 
-    models = models_temp
+    models = list(models_temp)
+    models.reverse()
 
     models_to_load = []
 

comfy/model_prefetch.py

@@ -37,7 +37,8 @@ def prefetch_queue_pop(queue, device, module):
     consumed = queue.pop(0)
     if consumed is not None:
         offload_stream, prefetch_state = consumed
-        offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        if offload_stream is not None:
+            offload_stream.wait_stream(comfy.model_management.current_stream(device))
         _, comfy_modules = prefetch_state
         if comfy_modules is not None:
             cleanup_prefetched_modules(comfy_modules)

comfy/ops.py

@@ -253,6 +253,9 @@ def resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, w
     if bias is not None:
         bias = post_cast(s, "bias", bias, bias_dtype, prefetch["resident"], update_weight)
 
+    if prefetch["signature"] is not None:
+        prefetch["resident"] = True
+
     return weight, bias
 
 

comfy/sampler_helpers.py

@@ -89,7 +89,8 @@ def get_additional_models(conds, dtype):
         gligen += get_models_from_cond(conds[k], "gligen")
         add_models += get_models_from_cond(conds[k], "additional_models")
 
-    control_nets = set(cnets)
+    # Order-preserving dedup. A plain set() would randomize iteration order across runs
+    control_nets = list(dict.fromkeys(cnets))
 
     inference_memory = 0
     control_models = []

comfy_extras/frame_interpolation_models/film_net.py

@@ -199,6 +199,9 @@ class FILMNet(nn.Module):
     def get_dtype(self):
         return self.extract.extract_sublevels.convs[0][0].conv.weight.dtype
 
+    def memory_used_forward(self, shape, dtype):
+        return 1700 * shape[1] * shape[2] * dtype.itemsize
+
     def _build_warp_grids(self, H, W, device):
         """Pre-compute warp grids for all pyramid levels."""
         if (H, W) in self._warp_grids:

comfy_extras/frame_interpolation_models/ifnet.py

@@ -74,6 +74,9 @@ class IFNet(nn.Module):
     def get_dtype(self):
         return self.encode.cnn0.weight.dtype
 
+    def memory_used_forward(self, shape, dtype):
+        return 300 * shape[1] * shape[2] * dtype.itemsize
+
     def _build_warp_grids(self, H, W, device):
         if (H, W) in self._warp_grids:
             return

comfy_extras/nodes_frame_interpolation.py

@@ -37,7 +37,7 @@ class FrameInterpolationModelLoader(io.ComfyNode):
         model = cls._detect_and_load(sd)
         dtype = torch.float16 if model_management.should_use_fp16(model_management.get_torch_device()) else torch.float32
         model.eval().to(dtype)
-        patcher = comfy.model_patcher.ModelPatcher(
+        patcher = comfy.model_patcher.CoreModelPatcher(
             model,
             load_device=model_management.get_torch_device(),
             offload_device=model_management.unet_offload_device(),
@@ -98,16 +98,13 @@ class FrameInterpolate(io.ComfyNode):
         if num_frames < 2 or multiplier < 2:
             return io.NodeOutput(images)
 
-        model_management.load_model_gpu(interp_model)
         device = interp_model.load_device
         dtype = interp_model.model_dtype()
         inference_model = interp_model.model
-
-        # Free VRAM for inference activations (model weights + ~20x a single frame's worth)
-        H, W = images.shape[1], images.shape[2]
-        activation_mem = H * W * 3 * images.element_size() * 20
-        model_management.free_memory(activation_mem, device)
+        activation_mem = inference_model.memory_used_forward(images.shape, dtype)
+        model_management.load_models_gpu([interp_model], memory_required=activation_mem)
         align = getattr(inference_model, "pad_align", 1)
+        H, W = images.shape[1], images.shape[2]
 
         # Prepare a single padded frame on device for determining output dimensions
         def prepare_frame(idx):