remove unused imports

* 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/ldm/chroma/layers.py

@@ -1,119 +1,29 @@
-import math
-from dataclasses import dataclass
-
 import torch
 from torch import Tensor, nn
 
-from .math import attention, rope
+from .math import attention
-import comfy.ops
+from comfy.ldm.flux.layers import (
-import comfy.ldm.common_dit
+    MLPEmbedder,
-
+    RMSNorm,
-
+    QKNorm,
-class EmbedND(nn.Module):
+    SelfAttention,
-    def __init__(self, dim: int, theta: int, axes_dim: list):
+    ModulationOut,
-        super().__init__()
-        self.dim = dim
-        self.theta = theta
-        self.axes_dim = axes_dim
-
-    def forward(self, ids: Tensor) -> Tensor:
-        n_axes = ids.shape[-1]
-        emb = torch.cat(
-            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
-            dim=-3,
 )
 
-        return emb.unsqueeze(1)
 
 
-def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
+class ChromaModulationOut(ModulationOut):
-    """
+    @classmethod
-    Create sinusoidal timestep embeddings.
+    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
-    :param t: a 1-D Tensor of N indices, one per batch element.
+        return cls(
-                      These may be fractional.
+            shift=tensor[:, offset : offset + 1, :],
-    :param dim: the dimension of the output.
+            scale=tensor[:, offset + 1 : offset + 2, :],
-    :param max_period: controls the minimum frequency of the embeddings.
+            gate=tensor[:, offset + 2 : offset + 3, :],
-    :return: an (N, D) Tensor of positional embeddings.
-    """
-    t = time_factor * t
-    half = dim // 2
-    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half)
-
-    args = t[:, None].float() * freqs[None]
-    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
-    if dim % 2:
-        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
-    if torch.is_floating_point(t):
-        embedding = embedding.to(t)
-    return embedding
-
-class MLPEmbedder(nn.Module):
-    def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
-        self.silu = nn.SiLU()
-        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
-
-    def forward(self, x: Tensor) -> Tensor:
-        return self.out_layer(self.silu(self.in_layer(x)))
-
-
-class RMSNorm(torch.nn.Module):
-    def __init__(self, dim: int, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
-
-    def forward(self, x: Tensor):
-        return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6)
-
-
-class QKNorm(torch.nn.Module):
-    def __init__(self, dim: int, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
-        self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
-
-    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
-        q = self.query_norm(q)
-        k = self.key_norm(k)
-        return q.to(v), k.to(v)
-
-
-class SelfAttention(nn.Module):
-    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.num_heads = num_heads
-        head_dim = dim // num_heads
-
-        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
-        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
-        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
-
-
-@dataclass
-class ModulationOut:
-    shift: Tensor
-    scale: Tensor
-    gate: Tensor
-
-
-class Modulation(nn.Module):
-    def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.is_double = double
-        self.multiplier = 6 if double else 3
-        self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
-
-    def forward(self, vec: Tensor) -> tuple:
-        out = self.lin(nn.functional.silu(vec)).chunk(self.multiplier, dim=-1)
-
-        return (
-            ModulationOut(*out[:3]),
-            ModulationOut(*out[3:]) if self.is_double else None,
         )
 
 
 
+
 class Approximator(nn.Module):
     def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None):
         super().__init__()

comfy/ldm/chroma/model.py

@@ -7,15 +7,17 @@ from torch import Tensor, nn
 from einops import rearrange, repeat
 import comfy.ldm.common_dit
 
+from comfy.ldm.flux.layers import (
+    EmbedND,
+    timestep_embedding,
+)
+
 from .layers import (
     DoubleStreamBlock,
-    EmbedND,
     LastLayer,
-    MLPEmbedder,
     SingleStreamBlock,
-    timestep_embedding,
     Approximator,
-    ModulationOut
+    ChromaModulationOut,
 )
 
 
@@ -39,14 +41,6 @@ class ChromaParams:
     n_layers: int
 
 
-class ChromaModulationOut(ModulationOut):
-    @classmethod
-    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
-        return cls(
-            shift=tensor[:, offset : offset + 1, :],
-            scale=tensor[:, offset + 1 : offset + 2, :],
-            gate=tensor[:, offset + 2 : offset + 3, :],
-        )
 
 
 class Chroma(nn.Module):
@@ -77,7 +71,6 @@ class Chroma(nn.Module):
         self.n_layers = params.n_layers
         self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
         self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
-        self.time_in = MLPEmbedder(in_dim=64, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
         self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
         # set as nn identity for now, will overwrite it later.
         self.distilled_guidance_layer = Approximator(
@@ -88,9 +81,6 @@ class Chroma(nn.Module):
                     dtype=dtype, device=device, operations=operations
                 )
 
-        self.guidance_in = (
-            MLPEmbedder(in_dim=64, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if self.distilled_guidance_layer else nn.Identity()
-        )
 
         self.double_blocks = nn.ModuleList(
             [

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):
@@ -60,26 +60,32 @@ class IsChangedCache:
         return self.is_changed[node_id]
 
 class CacheSet:
-    def __init__(self, lru_size=None):
+    def __init__(self, lru_size=None, cache_none=False):
-        if lru_size is None or lru_size == 0:
+        if cache_none:
+            self.init_dependency_aware_cache()
+        elif 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
-    # blowing away the cache every time
-    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)
-
     # Performs like the old cache -- dump data ASAP
     def init_classic_cache(self):
         self.outputs = HierarchicalCache(CacheKeySetInputSignature)
         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 +420,14 @@ 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, lru_size=None, cache_none=False):
         self.lru_size = lru_size
+        self.cache_none = cache_none
         self.server = server
         self.reset()
 
     def reset(self):
-        self.caches = CacheSet(self.lru_size)
+        self.caches = CacheSet(self.lru_size, self.cache_none)
         self.status_messages = []
         self.success = True
 

main.py

@@ -156,7 +156,7 @@ 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)
+    e = execution.PromptExecutor(server_instance, lru_size=args.cache_lru, cache_none=args.cache_none)
     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