Merge e8f3bc5ab7 into 22ad513c72

* add dependency aware cache that removed a cached node as soon as all of its decendents have executed. This allows users with lower RAM to run workflows they would otherwise not be able to run. The downside is that every workflow will fully run each time even if no nodes have changed.

* remove test code

* tidy code

comfy/cli_args.py

@@ -101,6 +101,7 @@ parser.add_argument("--preview-size", type=int, default=512, help="Sets the maxi
 cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")

comfy/lora.py

@@ -20,6 +20,7 @@ from __future__ import annotations
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
+import comfy.weight_adapter as weight_adapter
 import logging
 import torch
 
@@ -49,139 +50,12 @@ def load_lora(lora, to_load, log_missing=True):
             dora_scale = lora[dora_scale_name]
             loaded_keys.add(dora_scale_name)
 
-        reshape_name = "{}.reshape_weight".format(x)
+        for adapter_cls in weight_adapter.adapters:
-        reshape = None
+            adapter = adapter_cls.load(x, lora, alpha, dora_scale, loaded_keys)
-        if reshape_name in lora.keys():
+            if adapter is not None:
-            try:
+                patch_dict[to_load[x]] = adapter
-                reshape = lora[reshape_name].tolist()
+                loaded_keys.update(adapter.loaded_keys)
-                loaded_keys.add(reshape_name)
+                continue
-            except:
-                pass
-
-        regular_lora = "{}.lora_up.weight".format(x)
-        diffusers_lora = "{}_lora.up.weight".format(x)
-        diffusers2_lora = "{}.lora_B.weight".format(x)
-        diffusers3_lora = "{}.lora.up.weight".format(x)
-        mochi_lora = "{}.lora_B".format(x)
-        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
-        A_name = None
-
-        if regular_lora in lora.keys():
-            A_name = regular_lora
-            B_name = "{}.lora_down.weight".format(x)
-            mid_name = "{}.lora_mid.weight".format(x)
-        elif diffusers_lora in lora.keys():
-            A_name = diffusers_lora
-            B_name = "{}_lora.down.weight".format(x)
-            mid_name = None
-        elif diffusers2_lora in lora.keys():
-            A_name = diffusers2_lora
-            B_name = "{}.lora_A.weight".format(x)
-            mid_name = None
-        elif diffusers3_lora in lora.keys():
-            A_name = diffusers3_lora
-            B_name = "{}.lora.down.weight".format(x)
-            mid_name = None
-        elif mochi_lora in lora.keys():
-            A_name = mochi_lora
-            B_name = "{}.lora_A".format(x)
-            mid_name = None
-        elif transformers_lora in lora.keys():
-            A_name = transformers_lora
-            B_name ="{}.lora_linear_layer.down.weight".format(x)
-            mid_name = None
-
-        if A_name is not None:
-            mid = None
-            if mid_name is not None and mid_name in lora.keys():
-                mid = lora[mid_name]
-                loaded_keys.add(mid_name)
-            patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale, reshape))
-            loaded_keys.add(A_name)
-            loaded_keys.add(B_name)
-
-
-        ######## loha
-        hada_w1_a_name = "{}.hada_w1_a".format(x)
-        hada_w1_b_name = "{}.hada_w1_b".format(x)
-        hada_w2_a_name = "{}.hada_w2_a".format(x)
-        hada_w2_b_name = "{}.hada_w2_b".format(x)
-        hada_t1_name = "{}.hada_t1".format(x)
-        hada_t2_name = "{}.hada_t2".format(x)
-        if hada_w1_a_name in lora.keys():
-            hada_t1 = None
-            hada_t2 = None
-            if hada_t1_name in lora.keys():
-                hada_t1 = lora[hada_t1_name]
-                hada_t2 = lora[hada_t2_name]
-                loaded_keys.add(hada_t1_name)
-                loaded_keys.add(hada_t2_name)
-
-            patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale))
-            loaded_keys.add(hada_w1_a_name)
-            loaded_keys.add(hada_w1_b_name)
-            loaded_keys.add(hada_w2_a_name)
-            loaded_keys.add(hada_w2_b_name)
-
-
-        ######## lokr
-        lokr_w1_name = "{}.lokr_w1".format(x)
-        lokr_w2_name = "{}.lokr_w2".format(x)
-        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
-        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
-        lokr_t2_name = "{}.lokr_t2".format(x)
-        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
-        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
-
-        lokr_w1 = None
-        if lokr_w1_name in lora.keys():
-            lokr_w1 = lora[lokr_w1_name]
-            loaded_keys.add(lokr_w1_name)
-
-        lokr_w2 = None
-        if lokr_w2_name in lora.keys():
-            lokr_w2 = lora[lokr_w2_name]
-            loaded_keys.add(lokr_w2_name)
-
-        lokr_w1_a = None
-        if lokr_w1_a_name in lora.keys():
-            lokr_w1_a = lora[lokr_w1_a_name]
-            loaded_keys.add(lokr_w1_a_name)
-
-        lokr_w1_b = None
-        if lokr_w1_b_name in lora.keys():
-            lokr_w1_b = lora[lokr_w1_b_name]
-            loaded_keys.add(lokr_w1_b_name)
-
-        lokr_w2_a = None
-        if lokr_w2_a_name in lora.keys():
-            lokr_w2_a = lora[lokr_w2_a_name]
-            loaded_keys.add(lokr_w2_a_name)
-
-        lokr_w2_b = None
-        if lokr_w2_b_name in lora.keys():
-            lokr_w2_b = lora[lokr_w2_b_name]
-            loaded_keys.add(lokr_w2_b_name)
-
-        lokr_t2 = None
-        if lokr_t2_name in lora.keys():
-            lokr_t2 = lora[lokr_t2_name]
-            loaded_keys.add(lokr_t2_name)
-
-        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
-            patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale))
-
-        #glora
-        a1_name = "{}.a1.weight".format(x)
-        a2_name = "{}.a2.weight".format(x)
-        b1_name = "{}.b1.weight".format(x)
-        b2_name = "{}.b2.weight".format(x)
-        if a1_name in lora:
-            patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale))
-            loaded_keys.add(a1_name)
-            loaded_keys.add(a2_name)
-            loaded_keys.add(b1_name)
-            loaded_keys.add(b2_name)
 
         w_norm_name = "{}.w_norm".format(x)
         b_norm_name = "{}.b_norm".format(x)
@@ -408,26 +282,6 @@ def model_lora_keys_unet(model, key_map={}):
     return key_map
 
 
-def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function):
-    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
-    lora_diff *= alpha
-    weight_calc = weight + function(lora_diff).type(weight.dtype)
-    weight_norm = (
-        weight_calc.transpose(0, 1)
-        .reshape(weight_calc.shape[1], -1)
-        .norm(dim=1, keepdim=True)
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
-        .transpose(0, 1)
-    )
-
-    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
-    if strength != 1.0:
-        weight_calc -= weight
-        weight += strength * (weight_calc)
-    else:
-        weight[:] = weight_calc
-    return weight
-
 def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
     """
     Pad a tensor to a new shape with zeros.
@@ -482,6 +336,16 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
         if isinstance(v, list):
             v = (calculate_weight(v[1:], v[0][1](comfy.model_management.cast_to_device(v[0][0], weight.device, intermediate_dtype, copy=True), inplace=True), key, intermediate_dtype=intermediate_dtype), )
 
+        if isinstance(v, weight_adapter.WeightAdapterBase):
+            output = v.calculate_weight(weight, key, strength, strength_model, offset, function, intermediate_dtype, original_weights)
+            if output is None:
+                logging.warning("Calculate Weight Failed: {} {}".format(v.name, key))
+            else:
+                weight = output
+                if old_weight is not None:
+                    weight = old_weight
+            continue
+
         if len(v) == 1:
             patch_type = "diff"
         elif len(v) == 2:
@@ -508,157 +372,6 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             diff_weight = comfy.model_management.cast_to_device(target_weight, weight.device, intermediate_dtype) - \
                           comfy.model_management.cast_to_device(original_weights[key][0][0], weight.device, intermediate_dtype)
             weight += function(strength * comfy.model_management.cast_to_device(diff_weight, weight.device, weight.dtype))
-        elif patch_type == "lora": #lora/locon
-            mat1 = comfy.model_management.cast_to_device(v[0], weight.device, intermediate_dtype)
-            mat2 = comfy.model_management.cast_to_device(v[1], weight.device, intermediate_dtype)
-            dora_scale = v[4]
-            reshape = v[5]
-
-            if reshape is not None:
-                weight = pad_tensor_to_shape(weight, reshape)
-
-            if v[2] is not None:
-                alpha = v[2] / mat2.shape[0]
-            else:
-                alpha = 1.0
-
-            if v[3] is not None:
-                #locon mid weights, hopefully the math is fine because I didn't properly test it
-                mat3 = comfy.model_management.cast_to_device(v[3], weight.device, intermediate_dtype)
-                final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-            try:
-                lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "lokr":
-            w1 = v[0]
-            w2 = v[1]
-            w1_a = v[3]
-            w1_b = v[4]
-            w2_a = v[5]
-            w2_b = v[6]
-            t2 = v[7]
-            dora_scale = v[8]
-            dim = None
-
-            if w1 is None:
-                dim = w1_b.shape[0]
-                w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
-            else:
-                w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
-
-            if w2 is None:
-                dim = w2_b.shape[0]
-                if t2 is None:
-                    w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
-                else:
-                    w2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                        comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
-                                        comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
-                                        comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
-            else:
-                w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
-
-            if len(w2.shape) == 4:
-                w1 = w1.unsqueeze(2).unsqueeze(2)
-            if v[2] is not None and dim is not None:
-                alpha = v[2] / dim
-            else:
-                alpha = 1.0
-
-            try:
-                lora_diff = torch.kron(w1, w2).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "loha":
-            w1a = v[0]
-            w1b = v[1]
-            if v[2] is not None:
-                alpha = v[2] / w1b.shape[0]
-            else:
-                alpha = 1.0
-
-            w2a = v[3]
-            w2b = v[4]
-            dora_scale = v[7]
-            if v[5] is not None: #cp decomposition
-                t1 = v[5]
-                t2 = v[6]
-                m1 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                    comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
-
-                m2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
-            else:
-                m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
-                m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
-
-            try:
-                lora_diff = (m1 * m2).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "glora":
-            dora_scale = v[5]
-
-            old_glora = False
-            if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]:
-                rank = v[0].shape[0]
-                old_glora = True
-
-            if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]:
-                if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]:
-                    pass
-                else:
-                    old_glora = False
-                    rank = v[1].shape[0]
-
-            a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
-            a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
-            b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
-            b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
-
-            if v[4] is not None:
-                alpha = v[4] / rank
-            else:
-                alpha = 1.0
-
-            try:
-                if old_glora:
-                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora
-                else:
-                    if weight.dim() > 2:
-                        lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
-                    else:
-                        lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
-                    lora_diff += torch.mm(b1, b2).reshape(weight.shape)
-
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
         else:
             logging.warning("patch type not recognized {} {}".format(patch_type, key))
 

comfy/weight_adapter/__init__.py

@@ -0,0 +1,13 @@
+from .base import WeightAdapterBase
+from .lora import LoRAAdapter
+from .loha import LoHaAdapter
+from .lokr import LoKrAdapter
+from .glora import GLoRAAdapter
+
+
+adapters: list[type[WeightAdapterBase]] = [
+    LoRAAdapter,
+    LoHaAdapter,
+    LoKrAdapter,
+    GLoRAAdapter,
+]

comfy/weight_adapter/base.py

@@ -0,0 +1,94 @@
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+import comfy.model_management
+
+
+class WeightAdapterBase:
+    name: str
+    loaded_keys: set[str]
+    weights: list[torch.Tensor]
+
+    @classmethod
+    def load(cls, x: str, lora: dict[str, torch.Tensor]) -> Optional["WeightAdapterBase"]:
+        raise NotImplementedError
+
+    def to_train(self) -> "WeightAdapterTrainBase":
+        raise NotImplementedError
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        raise NotImplementedError
+
+
+class WeightAdapterTrainBase(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    # [TODO] Collaborate with LoRA training PR #7032
+
+
+def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function):
+    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
+    lora_diff *= alpha
+    weight_calc = weight + function(lora_diff).type(weight.dtype)
+    weight_norm = (
+        weight_calc.transpose(0, 1)
+        .reshape(weight_calc.shape[1], -1)
+        .norm(dim=1, keepdim=True)
+        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+        .transpose(0, 1)
+    )
+
+    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
+    if strength != 1.0:
+        weight_calc -= weight
+        weight += strength * (weight_calc)
+    else:
+        weight[:] = weight_calc
+    return weight
+
+
+def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
+    """
+    Pad a tensor to a new shape with zeros.
+
+    Args:
+        tensor (torch.Tensor): The original tensor to be padded.
+        new_shape (List[int]): The desired shape of the padded tensor.
+
+    Returns:
+        torch.Tensor: A new tensor padded with zeros to the specified shape.
+
+    Note:
+        If the new shape is smaller than the original tensor in any dimension,
+        the original tensor will be truncated in that dimension.
+    """
+    if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]):
+        raise ValueError("The new shape must be larger than the original tensor in all dimensions")
+
+    if len(new_shape) != len(tensor.shape):
+        raise ValueError("The new shape must have the same number of dimensions as the original tensor")
+
+    # Create a new tensor filled with zeros
+    padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device)
+
+    # Create slicing tuples for both tensors
+    orig_slices = tuple(slice(0, dim) for dim in tensor.shape)
+    new_slices = tuple(slice(0, dim) for dim in tensor.shape)
+
+    # Copy the original tensor into the new tensor
+    padded_tensor[new_slices] = tensor[orig_slices]
+
+    return padded_tensor

comfy/weight_adapter/glora.py

@@ -0,0 +1,93 @@
+import logging
+from typing import Optional
+
+import torch
+import comfy.model_management
+from .base import WeightAdapterBase, weight_decompose
+
+
+class GLoRAAdapter(WeightAdapterBase):
+    name = "glora"
+
+    def __init__(self, loaded_keys, weights):
+        self.loaded_keys = loaded_keys
+        self.weights = weights
+
+    @classmethod
+    def load(
+        cls,
+        x: str,
+        lora: dict[str, torch.Tensor],
+        alpha: float,
+        dora_scale: torch.Tensor,
+        loaded_keys: set[str] = None,
+    ) -> Optional["GLoRAAdapter"]:
+        if loaded_keys is None:
+            loaded_keys = set()
+        a1_name = "{}.a1.weight".format(x)
+        a2_name = "{}.a2.weight".format(x)
+        b1_name = "{}.b1.weight".format(x)
+        b2_name = "{}.b2.weight".format(x)
+        if a1_name in lora:
+            weights = (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale)
+            loaded_keys.add(a1_name)
+            loaded_keys.add(a2_name)
+            loaded_keys.add(b1_name)
+            loaded_keys.add(b2_name)
+            return cls(loaded_keys, weights)
+        else:
+            return None
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        v = self.weights
+        dora_scale = v[5]
+
+        old_glora = False
+        if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]:
+            rank = v[0].shape[0]
+            old_glora = True
+
+        if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]:
+            if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]:
+                pass
+            else:
+                old_glora = False
+                rank = v[1].shape[0]
+
+        a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
+        a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
+        b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
+        b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
+
+        if v[4] is not None:
+            alpha = v[4] / rank
+        else:
+            alpha = 1.0
+
+        try:
+            if old_glora:
+                lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora
+            else:
+                if weight.dim() > 2:
+                    lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
+                else:
+                    lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
+                lora_diff += torch.mm(b1, b2).reshape(weight.shape)
+
+            if dora_scale is not None:
+                weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
+            else:
+                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+        except Exception as e:
+            logging.error("ERROR {} {} {}".format(self.name, key, e))
+        return weight

comfy/weight_adapter/loha.py

@@ -0,0 +1,100 @@
+import logging
+from typing import Optional
+
+import torch
+import comfy.model_management
+from .base import WeightAdapterBase, weight_decompose
+
+
+class LoHaAdapter(WeightAdapterBase):
+    name = "loha"
+
+    def __init__(self, loaded_keys, weights):
+        self.loaded_keys = loaded_keys
+        self.weights = weights
+
+    @classmethod
+    def load(
+        cls,
+        x: str,
+        lora: dict[str, torch.Tensor],
+        alpha: float,
+        dora_scale: torch.Tensor,
+        loaded_keys: set[str] = None,
+    ) -> Optional["LoHaAdapter"]:
+        if loaded_keys is None:
+            loaded_keys = set()
+
+        hada_w1_a_name = "{}.hada_w1_a".format(x)
+        hada_w1_b_name = "{}.hada_w1_b".format(x)
+        hada_w2_a_name = "{}.hada_w2_a".format(x)
+        hada_w2_b_name = "{}.hada_w2_b".format(x)
+        hada_t1_name = "{}.hada_t1".format(x)
+        hada_t2_name = "{}.hada_t2".format(x)
+        if hada_w1_a_name in lora.keys():
+            hada_t1 = None
+            hada_t2 = None
+            if hada_t1_name in lora.keys():
+                hada_t1 = lora[hada_t1_name]
+                hada_t2 = lora[hada_t2_name]
+                loaded_keys.add(hada_t1_name)
+                loaded_keys.add(hada_t2_name)
+
+            weights = (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale)
+            loaded_keys.add(hada_w1_a_name)
+            loaded_keys.add(hada_w1_b_name)
+            loaded_keys.add(hada_w2_a_name)
+            loaded_keys.add(hada_w2_b_name)
+            return cls(loaded_keys, weights)
+        else:
+            return None
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        v = self.weights
+        w1a = v[0]
+        w1b = v[1]
+        if v[2] is not None:
+            alpha = v[2] / w1b.shape[0]
+        else:
+            alpha = 1.0
+
+        w2a = v[3]
+        w2b = v[4]
+        dora_scale = v[7]
+        if v[5] is not None: #cp decomposition
+            t1 = v[5]
+            t2 = v[6]
+            m1 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
+
+            m2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
+        else:
+            m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
+                            comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
+            m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
+                            comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
+
+        try:
+            lora_diff = (m1 * m2).reshape(weight.shape)
+            if dora_scale is not None:
+                weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
+            else:
+                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+        except Exception as e:
+            logging.error("ERROR {} {} {}".format(self.name, key, e))
+        return weight

comfy/weight_adapter/lokr.py

@@ -0,0 +1,133 @@
+import logging
+from typing import Optional
+
+import torch
+import comfy.model_management
+from .base import WeightAdapterBase, weight_decompose
+
+
+class LoKrAdapter(WeightAdapterBase):
+    name = "lokr"
+
+    def __init__(self, loaded_keys, weights):
+        self.loaded_keys = loaded_keys
+        self.weights = weights
+
+    @classmethod
+    def load(
+        cls,
+        x: str,
+        lora: dict[str, torch.Tensor],
+        alpha: float,
+        dora_scale: torch.Tensor,
+        loaded_keys: set[str] = None,
+    ) -> Optional["LoKrAdapter"]:
+        if loaded_keys is None:
+            loaded_keys = set()
+        lokr_w1_name = "{}.lokr_w1".format(x)
+        lokr_w2_name = "{}.lokr_w2".format(x)
+        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
+        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
+        lokr_t2_name = "{}.lokr_t2".format(x)
+        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
+        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
+
+        lokr_w1 = None
+        if lokr_w1_name in lora.keys():
+            lokr_w1 = lora[lokr_w1_name]
+            loaded_keys.add(lokr_w1_name)
+
+        lokr_w2 = None
+        if lokr_w2_name in lora.keys():
+            lokr_w2 = lora[lokr_w2_name]
+            loaded_keys.add(lokr_w2_name)
+
+        lokr_w1_a = None
+        if lokr_w1_a_name in lora.keys():
+            lokr_w1_a = lora[lokr_w1_a_name]
+            loaded_keys.add(lokr_w1_a_name)
+
+        lokr_w1_b = None
+        if lokr_w1_b_name in lora.keys():
+            lokr_w1_b = lora[lokr_w1_b_name]
+            loaded_keys.add(lokr_w1_b_name)
+
+        lokr_w2_a = None
+        if lokr_w2_a_name in lora.keys():
+            lokr_w2_a = lora[lokr_w2_a_name]
+            loaded_keys.add(lokr_w2_a_name)
+
+        lokr_w2_b = None
+        if lokr_w2_b_name in lora.keys():
+            lokr_w2_b = lora[lokr_w2_b_name]
+            loaded_keys.add(lokr_w2_b_name)
+
+        lokr_t2 = None
+        if lokr_t2_name in lora.keys():
+            lokr_t2 = lora[lokr_t2_name]
+            loaded_keys.add(lokr_t2_name)
+
+        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
+            weights = (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale)
+            return cls(loaded_keys, weights)
+        else:
+            return None
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        v = self.weights
+        w1 = v[0]
+        w2 = v[1]
+        w1_a = v[3]
+        w1_b = v[4]
+        w2_a = v[5]
+        w2_b = v[6]
+        t2 = v[7]
+        dora_scale = v[8]
+        dim = None
+
+        if w1 is None:
+            dim = w1_b.shape[0]
+            w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
+                            comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
+        else:
+            w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
+
+        if w2 is None:
+            dim = w2_b.shape[0]
+            if t2 is None:
+                w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
+            else:
+                w2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
+        else:
+            w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
+
+        if len(w2.shape) == 4:
+            w1 = w1.unsqueeze(2).unsqueeze(2)
+        if v[2] is not None and dim is not None:
+            alpha = v[2] / dim
+        else:
+            alpha = 1.0
+
+        try:
+            lora_diff = torch.kron(w1, w2).reshape(weight.shape)
+            if dora_scale is not None:
+                weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
+            else:
+                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+        except Exception as e:
+            logging.error("ERROR {} {} {}".format(self.name, key, e))
+        return weight

comfy/weight_adapter/lora.py

@@ -0,0 +1,142 @@
+import logging
+from typing import Optional
+
+import torch
+import comfy.model_management
+from .base import WeightAdapterBase, weight_decompose, pad_tensor_to_shape
+
+
+class LoRAAdapter(WeightAdapterBase):
+    name = "lora"
+
+    def __init__(self, loaded_keys, weights):
+        self.loaded_keys = loaded_keys
+        self.weights = weights
+
+    @classmethod
+    def load(
+        cls,
+        x: str,
+        lora: dict[str, torch.Tensor],
+        alpha: float,
+        dora_scale: torch.Tensor,
+        loaded_keys: set[str] = None,
+    ) -> Optional["LoRAAdapter"]:
+        if loaded_keys is None:
+            loaded_keys = set()
+
+        reshape_name = "{}.reshape_weight".format(x)
+        regular_lora = "{}.lora_up.weight".format(x)
+        diffusers_lora = "{}_lora.up.weight".format(x)
+        diffusers2_lora = "{}.lora_B.weight".format(x)
+        diffusers3_lora = "{}.lora.up.weight".format(x)
+        mochi_lora = "{}.lora_B".format(x)
+        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
+        A_name = None
+
+        if regular_lora in lora.keys():
+            A_name = regular_lora
+            B_name = "{}.lora_down.weight".format(x)
+            mid_name = "{}.lora_mid.weight".format(x)
+        elif diffusers_lora in lora.keys():
+            A_name = diffusers_lora
+            B_name = "{}_lora.down.weight".format(x)
+            mid_name = None
+        elif diffusers2_lora in lora.keys():
+            A_name = diffusers2_lora
+            B_name = "{}.lora_A.weight".format(x)
+            mid_name = None
+        elif diffusers3_lora in lora.keys():
+            A_name = diffusers3_lora
+            B_name = "{}.lora.down.weight".format(x)
+            mid_name = None
+        elif mochi_lora in lora.keys():
+            A_name = mochi_lora
+            B_name = "{}.lora_A".format(x)
+            mid_name = None
+        elif transformers_lora in lora.keys():
+            A_name = transformers_lora
+            B_name = "{}.lora_linear_layer.down.weight".format(x)
+            mid_name = None
+
+        if A_name is not None:
+            mid = None
+            if mid_name is not None and mid_name in lora.keys():
+                mid = lora[mid_name]
+                loaded_keys.add(mid_name)
+            reshape = None
+            if reshape_name in lora.keys():
+                try:
+                    reshape = lora[reshape_name].tolist()
+                    loaded_keys.add(reshape_name)
+                except:
+                    pass
+            weights = (lora[A_name], lora[B_name], alpha, mid, dora_scale, reshape)
+            loaded_keys.add(A_name)
+            loaded_keys.add(B_name)
+            return cls(loaded_keys, weights)
+        else:
+            return None
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        v = self.weights
+        mat1 = comfy.model_management.cast_to_device(
+            v[0], weight.device, intermediate_dtype
+        )
+        mat2 = comfy.model_management.cast_to_device(
+            v[1], weight.device, intermediate_dtype
+        )
+        dora_scale = v[4]
+        reshape = v[5]
+
+        if reshape is not None:
+            weight = pad_tensor_to_shape(weight, reshape)
+
+        if v[2] is not None:
+            alpha = v[2] / mat2.shape[0]
+        else:
+            alpha = 1.0
+
+        if v[3] is not None:
+            # locon mid weights, hopefully the math is fine because I didn't properly test it
+            mat3 = comfy.model_management.cast_to_device(
+                v[3], weight.device, intermediate_dtype
+            )
+            final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
+            mat2 = (
+                torch.mm(
+                    mat2.transpose(0, 1).flatten(start_dim=1),
+                    mat3.transpose(0, 1).flatten(start_dim=1),
+                )
+                .reshape(final_shape)
+                .transpose(0, 1)
+            )
+        try:
+            lora_diff = torch.mm(
+                mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)
+            ).reshape(weight.shape)
+            if dora_scale is not None:
+                weight = weight_decompose(
+                    dora_scale,
+                    weight,
+                    lora_diff,
+                    alpha,
+                    strength,
+                    intermediate_dtype,
+                    function,
+                )
+            else:
+                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+        except Exception as e:
+            logging.error("ERROR {} {} {}".format(self.name, key, e))
+        return weight

comfy_execution/caching.py

@@ -316,3 +316,156 @@ class LRUCache(BasicCache):
             self.children[cache_key].append(self.cache_key_set.get_data_key(child_id))
         return self
 
+
+class DependencyAwareCache(BasicCache):
+    """
+    A cache implementation that tracks dependencies between nodes and manages
+    their execution and caching accordingly. It extends the BasicCache class.
+    Nodes are removed from this cache once all of their descendants have been
+    executed.
+    """
+
+    def __init__(self, key_class):
+        """
+        Initialize the DependencyAwareCache.
+
+        Args:
+            key_class: The class used for generating cache keys.
+        """
+        super().__init__(key_class)
+        self.descendants = {}  # Maps node_id -> set of descendant node_ids
+        self.ancestors = {}    # Maps node_id -> set of ancestor node_ids
+        self.executed_nodes = set()  # Tracks nodes that have been executed
+
+    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        """
+        Clear the entire cache and rebuild the dependency graph.
+
+        Args:
+            dynprompt: The dynamic prompt object containing node information.
+            node_ids: List of node IDs to initialize the cache for.
+            is_changed_cache: Flag indicating if the cache has changed.
+        """
+        # Clear all existing cache data
+        self.cache.clear()
+        self.subcaches.clear()
+        self.descendants.clear()
+        self.ancestors.clear()
+        self.executed_nodes.clear()
+
+        # Call the parent method to initialize the cache with the new prompt
+        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+
+        # Rebuild the dependency graph
+        self._build_dependency_graph(dynprompt, node_ids)
+
+    def _build_dependency_graph(self, dynprompt, node_ids):
+        """
+        Build the dependency graph for all nodes.
+
+        Args:
+            dynprompt: The dynamic prompt object containing node information.
+            node_ids: List of node IDs to build the graph for.
+        """
+        self.descendants.clear()
+        self.ancestors.clear()
+        for node_id in node_ids:
+            self.descendants[node_id] = set()
+            self.ancestors[node_id] = set()
+
+        for node_id in node_ids:
+            inputs = dynprompt.get_node(node_id)["inputs"]
+            for input_data in inputs.values():
+                if is_link(input_data):  # Check if the input is a link to another node
+                    ancestor_id = input_data[0]
+                    self.descendants[ancestor_id].add(node_id)
+                    self.ancestors[node_id].add(ancestor_id)
+
+    def set(self, node_id, value):
+        """
+        Mark a node as executed and store its value in the cache.
+
+        Args:
+            node_id: The ID of the node to store.
+            value: The value to store for the node.
+        """
+        self._set_immediate(node_id, value)
+        self.executed_nodes.add(node_id)
+        self._cleanup_ancestors(node_id)
+
+    def get(self, node_id):
+        """
+        Retrieve the cached value for a node.
+
+        Args:
+            node_id: The ID of the node to retrieve.
+
+        Returns:
+            The cached value for the node.
+        """
+        return self._get_immediate(node_id)
+
+    def ensure_subcache_for(self, node_id, children_ids):
+        """
+        Ensure a subcache exists for a node and update dependencies.
+
+        Args:
+            node_id: The ID of the parent node.
+            children_ids: List of child node IDs to associate with the parent node.
+
+        Returns:
+            The subcache object for the node.
+        """
+        subcache = super()._ensure_subcache(node_id, children_ids)
+        for child_id in children_ids:
+            self.descendants[node_id].add(child_id)
+            self.ancestors[child_id].add(node_id)
+        return subcache
+
+    def _cleanup_ancestors(self, node_id):
+        """
+        Check if ancestors of a node can be removed from the cache.
+
+        Args:
+            node_id: The ID of the node whose ancestors are to be checked.
+        """
+        for ancestor_id in self.ancestors.get(node_id, []):
+            if ancestor_id in self.executed_nodes:
+                # Remove ancestor if all its descendants have been executed
+                if all(descendant in self.executed_nodes for descendant in self.descendants[ancestor_id]):
+                    self._remove_node(ancestor_id)
+
+    def _remove_node(self, node_id):
+        """
+        Remove a node from the cache.
+
+        Args:
+            node_id: The ID of the node to remove.
+        """
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        if cache_key in self.cache:
+            del self.cache[cache_key]
+        subcache_key = self.cache_key_set.get_subcache_key(node_id)
+        if subcache_key in self.subcaches:
+            del self.subcaches[subcache_key]
+
+    def clean_unused(self):
+        """
+        Clean up unused nodes. This is a no-op for this cache implementation.
+        """
+        pass
+
+    def recursive_debug_dump(self):
+        """
+        Dump the cache and dependency graph for debugging.
+
+        Returns:
+            A list containing the cache state and dependency graph.
+        """
+        result = super().recursive_debug_dump()
+        result.append({
+            "descendants": self.descendants,
+            "ancestors": self.ancestors,
+            "executed_nodes": list(self.executed_nodes),
+        })
+        return result

execution.py

@@ -15,7 +15,7 @@ import nodes
 import comfy.model_management
 from comfy_execution.graph import get_input_info, ExecutionList, DynamicPrompt, ExecutionBlocker
 from comfy_execution.graph_utils import is_link, GraphBuilder
-from comfy_execution.caching import HierarchicalCache, LRUCache, CacheKeySetInputSignature, CacheKeySetID
+from comfy_execution.caching import HierarchicalCache, LRUCache, DependencyAwareCache, CacheKeySetInputSignature, CacheKeySetID
 from comfy_execution.validation import validate_node_input
 
 class ExecutionResult(Enum):
@@ -59,20 +59,27 @@ class IsChangedCache:
             self.is_changed[node_id] = node["is_changed"]
         return self.is_changed[node_id]
 
-class CacheSet:
-    def __init__(self, lru_size=None):
-        if lru_size is None or lru_size == 0:
-            self.init_classic_cache()
-        else:
-            self.init_lru_cache(lru_size)
-        self.all = [self.outputs, self.ui, self.objects]
 
-    # Useful for those with ample RAM/VRAM -- allows experimenting without
+class CacheType(Enum):
-    # blowing away the cache every time
+    CLASSIC = 0
-    def init_lru_cache(self, cache_size):
+    LRU = 1
-        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+    DEPENDENCY_AWARE = 2
-        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+
-        self.objects = HierarchicalCache(CacheKeySetID)
+
+class CacheSet:
+    def __init__(self, cache_type=None, cache_size=None):
+        if cache_type == CacheType.DEPENDENCY_AWARE:
+            self.init_dependency_aware_cache()
+            logging.info("Disabling intermediate node cache.")
+        elif cache_type == CacheType.LRU:
+            if cache_size is None:
+                cache_size = 0
+            self.init_lru_cache(cache_size)
+            logging.info("Using LRU cache")
+        else:
+            self.init_classic_cache()
+
+        self.all = [self.outputs, self.ui, self.objects]
 
     # Performs like the old cache -- dump data ASAP
     def init_classic_cache(self):
@@ -80,6 +87,17 @@ class CacheSet:
         self.ui = HierarchicalCache(CacheKeySetInputSignature)
         self.objects = HierarchicalCache(CacheKeySetID)
 
+    def init_lru_cache(self, cache_size):
+        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.objects = HierarchicalCache(CacheKeySetID)
+
+    # only hold cached items while the decendents have not executed
+    def init_dependency_aware_cache(self):
+        self.outputs = DependencyAwareCache(CacheKeySetInputSignature)
+        self.ui = DependencyAwareCache(CacheKeySetInputSignature)
+        self.objects = DependencyAwareCache(CacheKeySetID)
+
     def recursive_debug_dump(self):
         result = {
             "outputs": self.outputs.recursive_debug_dump(),
@@ -414,13 +432,14 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
     return (ExecutionResult.SUCCESS, None, None)
 
 class PromptExecutor:
-    def __init__(self, server, lru_size=None):
+    def __init__(self, server, cache_type=False, cache_size=None):
-        self.lru_size = lru_size
+        self.cache_size = cache_size
+        self.cache_type = cache_type
         self.server = server
         self.reset()
 
     def reset(self):
-        self.caches = CacheSet(self.lru_size)
+        self.caches = CacheSet(cache_type=self.cache_type, cache_size=self.cache_size)
         self.status_messages = []
         self.success = True
 

main.py

@@ -156,7 +156,13 @@ def cuda_malloc_warning():
 
 def prompt_worker(q, server_instance):
     current_time: float = 0.0
-    e = execution.PromptExecutor(server_instance, lru_size=args.cache_lru)
+    cache_type = execution.CacheType.CLASSIC
+    if args.cache_lru > 0:
+        cache_type = execution.CacheType.LRU
+    elif args.cache_none:
+        cache_type = execution.CacheType.DEPENDENCY_AWARE
+
+    e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru)
     last_gc_collect = 0
     need_gc = False
     gc_collect_interval = 10.0

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.14.6
+comfyui-frontend-package==1.15.13
 torch
 torchsde
 torchvision

server.py

@@ -48,7 +48,7 @@ async def send_socket_catch_exception(function, message):
 @web.middleware
 async def cache_control(request: web.Request, handler):
     response: web.Response = await handler(request)
-    if request.path.endswith('.js') or request.path.endswith('.css'):
+    if request.path.endswith('.js') or request.path.endswith('.css') or request.path.endswith('index.json'):
         response.headers.setdefault('Cache-Control', 'no-cache')
     return response