Files
ComfyUI/comfy_execution/graph.py
Deep Mehta af7b4a921d feat: Add CacheProvider API for external distributed caching (#12056)
* feat: Add CacheProvider API for external distributed caching

Introduces a public API for external cache providers, enabling distributed
caching across multiple ComfyUI instances (e.g., Kubernetes pods).

New files:
- comfy_execution/cache_provider.py: CacheProvider ABC, CacheContext/CacheValue
  dataclasses, thread-safe provider registry, serialization utilities

Modified files:
- comfy_execution/caching.py: Add provider hooks to BasicCache (_notify_providers_store,
  _check_providers_lookup), subcache exclusion, prompt ID propagation
- execution.py: Add prompt lifecycle hooks (on_prompt_start/on_prompt_end) to
  PromptExecutor, set _current_prompt_id on caches

Key features:
- Local-first caching (check local before external for performance)
- NaN detection to prevent incorrect external cache hits
- Subcache exclusion (ephemeral subgraph results not cached externally)
- Thread-safe provider snapshot caching
- Graceful error handling (provider errors logged, never break execution)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix: use deterministic hash for cache keys instead of pickle

Pickle serialization is NOT deterministic across Python sessions due
to hash randomization affecting frozenset iteration order. This causes
distributed caching to fail because different pods compute different
hashes for identical cache keys.

Fix: Use _canonicalize() + JSON serialization which ensures deterministic
ordering regardless of Python's hash randomization.

This is critical for cross-pod cache key consistency in Kubernetes
deployments.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* test: add unit tests for CacheProvider API

- Add comprehensive tests for _canonicalize deterministic ordering
- Add tests for serialize_cache_key hash consistency
- Add tests for contains_nan utility
- Add tests for estimate_value_size
- Add tests for provider registry (register, unregister, clear)
- Move json import to top-level (fix inline import)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* style: remove unused imports in test_cache_provider.py

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix: move _torch_available before usage and use importlib.util.find_spec

Fixes ruff F821 (undefined name) and F401 (unused import) errors.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix: use hashable types in frozenset test and add dict test

Frozensets can only contain hashable types, so use nested frozensets
instead of dicts. Added separate test for dict handling via serialize_cache_key.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* refactor: expose CacheProvider API via comfy_api.latest.Caching

- Add Caching class to comfy_api/latest/__init__.py that re-exports
  from comfy_execution.cache_provider (source of truth)
- Fix docstring: "Skip large values" instead of "Skip small values"
  (small compute-heavy values are good cache targets)
- Maintain backward compatibility: comfy_execution.cache_provider
  imports still work

Usage:
    from comfy_api.latest import Caching

    class MyProvider(Caching.CacheProvider):
        def on_lookup(self, context): ...
        def on_store(self, context, value): ...

    Caching.register_provider(MyProvider())

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* docs: clarify should_cache filtering criteria

Change docstring from "Skip large values" to "Skip if download time > compute time"
which better captures the cost/benefit tradeoff for external caching.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* docs: make should_cache docstring implementation-agnostic

Remove prescriptive filtering suggestions - let implementations
decide their own caching logic based on their use case.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* feat: add optional ui field to CacheValue

- Add ui field to CacheValue dataclass (default None)
- Pass ui when creating CacheValue for external providers
- Use result.ui (or default {}) when returning from external cache lookup

This allows external cache implementations to store/retrieve UI data
if desired, while remaining optional for implementations that skip it.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* refactor: rename _is_cacheable_value to _is_external_cacheable_value

Clearer name since objects are also cached locally - this specifically
checks for external caching eligibility.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* refactor: async CacheProvider API + reduce public surface

- Make on_lookup/on_store async on CacheProvider ABC
- Simplify CacheContext: replace cache_key + cache_key_bytes with
  cache_key_hash (str hex digest)
- Make registry/utility functions internal (_prefix)
- Trim comfy_api.latest.Caching exports to core API only
- Make cache get/set async throughout caching.py hierarchy
- Use asyncio.create_task for fire-and-forget on_store
- Add NaN gating before provider calls in Core
- Add await to 5 cache call sites in execution.py

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: remove unused imports (ruff) and update tests for internal API

- Remove unused CacheContext and _serialize_cache_key imports from
  caching.py (now handled by _build_context helper)
- Update test_cache_provider.py to use _-prefixed internal names
- Update tests for new CacheContext.cache_key_hash field (str)
- Make MockCacheProvider methods async to match ABC

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: address coderabbit review feedback

- Add try/except to _build_context, return None when hash fails
- Return None from _serialize_cache_key on total failure (no id()-based fallback)
- Replace hex-like test literal with non-secret placeholder

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: use _-prefixed imports in _notify_prompt_lifecycle

The lifecycle notification method was importing the old non-prefixed
names (has_cache_providers, get_cache_providers, logger) which no
longer exist after the API cleanup.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: add sync get_local/set_local for graph traversal

ExecutionList in graph.py calls output_cache.get() and .set() from
sync methods (is_cached, cache_link, get_cache). These cannot await
the now-async get/set. Add get_local/set_local that bypass external
providers and only access the local dict — which is all graph
traversal needs.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: remove cloud-specific language from cache provider API

Make all docstrings and comments generic for the OSS codebase.
Remove references to Kubernetes, Redis, GCS, pods, and other
infrastructure-specific terminology.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* style: align documentation with codebase conventions

Strip verbose docstrings and section banners to match existing minimal
documentation style used throughout the codebase.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: add usage example to Caching class, remove pickle fallback

- Add docstring with usage example to Caching class matching the
  convention used by sibling APIs (Execution.set_progress, ComfyExtension)
- Remove non-deterministic pickle fallback from _serialize_cache_key;
  return None on JSON failure instead of producing unretrievable hashes
- Move cache_provider imports to top of execution.py (no circular dep)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor: move public types to comfy_api, eager provider snapshot

Address review feedback:
- Move CacheProvider/CacheContext/CacheValue definitions to
  comfy_api/latest/_caching.py (source of truth for public API)
- comfy_execution/cache_provider.py re-exports types from there
- Build _providers_snapshot eagerly on register/unregister instead
  of lazy memoization in _get_cache_providers

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: generalize self-inequality check, fail-closed canonicalization

Address review feedback from guill:
- Rename _contains_nan to _contains_self_unequal, use not (x == x)
  instead of math.isnan to catch any self-unequal value
- Remove Unhashable and repr() fallbacks from _canonicalize; raise
  ValueError for unknown types so _serialize_cache_key returns None
  and external caching is skipped (fail-closed)
- Update tests for renamed function and new fail-closed behavior

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: suppress ruff F401 for re-exported CacheContext

CacheContext is imported from _caching and re-exported for use by
caching.py. Add noqa comment to satisfy the linter.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: enable external caching for subcache (expanded) nodes

Subcache nodes (from node expansion) now participate in external
provider store/lookup. Previously skipped to avoid duplicates, but
the cost of missing partial-expansion cache hits outweighs redundant
stores — especially with looping behavior on the horizon.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: wrap register/unregister as explicit static methods

Define register_provider and unregister_provider as wrapper functions
in the Caching class instead of re-importing. This locks the public
API signature in comfy_api/ so internal changes can't accidentally
break it.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: use debug-level logging for provider registration

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: follow ProxiedSingleton pattern for Caching class

Add Caching as a nested class inside ComfyAPI_latest inheriting from
ProxiedSingleton with async instance methods, matching the Execution
and NodeReplacement patterns. Retains standalone Caching class for
direct import convenience.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: inline registration logic in Caching class

Follow the Execution/NodeReplacement pattern — the public API methods
contain the actual logic operating on cache_provider module state,
not wrapper functions delegating to free functions.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: single Caching definition inside ComfyAPI_latest

Remove duplicate standalone Caching class. Define it once as a nested
class in ComfyAPI_latest (matching Execution/NodeReplacement pattern),
with a module-level alias for import convenience.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: remove prompt_id from CacheContext, type-safe canonicalization

Remove prompt_id from CacheContext — it's not relevant for cache
matching and added unnecessary plumbing (_current_prompt_id on every
cache). Lifecycle hooks still receive prompt_id directly.

Include type name in canonicalized primitives so that int 7 and
str "7" produce distinct hashes. Also canonicalize dict keys properly
instead of str() coercion.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: address review feedback on cache provider API

- Hold references to pending store tasks to prevent "Task was destroyed
  but it is still pending" warnings (bigcat88)
- Parallel cache lookups with asyncio.gather instead of sequential
  awaits for better performance (bigcat88)
- Delegate Caching.register/unregister_provider to existing functions
  in cache_provider.py instead of reimplementing (bigcat88)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>
2026-03-12 16:09:07 -07:00

338 lines
14 KiB
Python

from __future__ import annotations
from typing import Type, Literal
import nodes
import asyncio
import inspect
from comfy_execution.graph_utils import is_link, ExecutionBlocker
from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, InputTypeOptions
# NOTE: ExecutionBlocker code got moved to graph_utils.py to prevent torch being imported too soon during unit tests
ExecutionBlocker = ExecutionBlocker
class DependencyCycleError(Exception):
pass
class NodeInputError(Exception):
pass
class NodeNotFoundError(Exception):
pass
class DynamicPrompt:
def __init__(self, original_prompt):
# The original prompt provided by the user
self.original_prompt = original_prompt
# Any extra pieces of the graph created during execution
self.ephemeral_prompt = {}
self.ephemeral_parents = {}
self.ephemeral_display = {}
def get_node(self, node_id):
if node_id in self.ephemeral_prompt:
return self.ephemeral_prompt[node_id]
if node_id in self.original_prompt:
return self.original_prompt[node_id]
raise NodeNotFoundError(f"Node {node_id} not found")
def has_node(self, node_id):
return node_id in self.original_prompt or node_id in self.ephemeral_prompt
def add_ephemeral_node(self, node_id, node_info, parent_id, display_id):
self.ephemeral_prompt[node_id] = node_info
self.ephemeral_parents[node_id] = parent_id
self.ephemeral_display[node_id] = display_id
def get_real_node_id(self, node_id):
while node_id in self.ephemeral_parents:
node_id = self.ephemeral_parents[node_id]
return node_id
def get_parent_node_id(self, node_id):
return self.ephemeral_parents.get(node_id, None)
def get_display_node_id(self, node_id):
while node_id in self.ephemeral_display:
node_id = self.ephemeral_display[node_id]
return node_id
def all_node_ids(self):
return set(self.original_prompt.keys()).union(set(self.ephemeral_prompt.keys()))
def get_original_prompt(self):
return self.original_prompt
def get_input_info(
class_def: Type[ComfyNodeABC],
input_name: str,
valid_inputs: InputTypeDict | None = None
) -> tuple[str, Literal["required", "optional", "hidden"], InputTypeOptions] | tuple[None, None, None]:
"""Get the input type, category, and extra info for a given input name.
Arguments:
class_def: The class definition of the node.
input_name: The name of the input to get info for.
valid_inputs: The valid inputs for the node, or None to use the class_def.INPUT_TYPES().
Returns:
tuple[str, str, dict] | tuple[None, None, None]: The input type, category, and extra info for the input name.
"""
valid_inputs = valid_inputs or class_def.INPUT_TYPES()
input_info = None
input_category = None
if "required" in valid_inputs and input_name in valid_inputs["required"]:
input_category = "required"
input_info = valid_inputs["required"][input_name]
elif "optional" in valid_inputs and input_name in valid_inputs["optional"]:
input_category = "optional"
input_info = valid_inputs["optional"][input_name]
elif "hidden" in valid_inputs and input_name in valid_inputs["hidden"]:
input_category = "hidden"
input_info = valid_inputs["hidden"][input_name]
if input_info is None:
return None, None, None
input_type = input_info[0]
if len(input_info) > 1:
extra_info = input_info[1]
else:
extra_info = {}
# if input_type is a list, it is a Combo defined in outdated format; convert it.
# NOTE: uncomment this when we are confident old format going away won't cause too much trouble.
# if isinstance(input_type, list):
# extra_info["options"] = input_type
# input_type = IO.Combo.io_type
return input_type, input_category, extra_info
class TopologicalSort:
def __init__(self, dynprompt):
self.dynprompt = dynprompt
self.pendingNodes = {}
self.blockCount = {} # Number of nodes this node is directly blocked by
self.blocking = {} # Which nodes are blocked by this node
self.externalBlocks = 0
self.unblockedEvent = asyncio.Event()
def get_input_info(self, unique_id, input_name):
class_type = self.dynprompt.get_node(unique_id)["class_type"]
class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
return get_input_info(class_def, input_name)
def make_input_strong_link(self, to_node_id, to_input):
inputs = self.dynprompt.get_node(to_node_id)["inputs"]
if to_input not in inputs:
raise NodeInputError(f"Node {to_node_id} says it needs input {to_input}, but there is no input to that node at all")
value = inputs[to_input]
if not is_link(value):
raise NodeInputError(f"Node {to_node_id} says it needs input {to_input}, but that value is a constant")
from_node_id, from_socket = value
self.add_strong_link(from_node_id, from_socket, to_node_id)
def add_strong_link(self, from_node_id, from_socket, to_node_id):
if not self.is_cached(from_node_id):
self.add_node(from_node_id)
if to_node_id not in self.blocking[from_node_id]:
self.blocking[from_node_id][to_node_id] = {}
self.blockCount[to_node_id] += 1
self.blocking[from_node_id][to_node_id][from_socket] = True
def add_node(self, node_unique_id, include_lazy=False, subgraph_nodes=None):
node_ids = [node_unique_id]
links = []
while len(node_ids) > 0:
unique_id = node_ids.pop()
if unique_id in self.pendingNodes:
continue
self.pendingNodes[unique_id] = True
self.blockCount[unique_id] = 0
self.blocking[unique_id] = {}
inputs = self.dynprompt.get_node(unique_id)["inputs"]
for input_name in inputs:
value = inputs[input_name]
if is_link(value):
from_node_id, from_socket = value
if subgraph_nodes is not None and from_node_id not in subgraph_nodes:
continue
_, _, input_info = self.get_input_info(unique_id, input_name)
is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"]
if (include_lazy or not is_lazy):
if not self.is_cached(from_node_id):
node_ids.append(from_node_id)
links.append((from_node_id, from_socket, unique_id))
for link in links:
self.add_strong_link(*link)
def add_external_block(self, node_id):
assert node_id in self.blockCount, "Can't add external block to a node that isn't pending"
self.externalBlocks += 1
self.blockCount[node_id] += 1
def unblock():
self.externalBlocks -= 1
self.blockCount[node_id] -= 1
self.unblockedEvent.set()
return unblock
def is_cached(self, node_id):
return False
def get_ready_nodes(self):
return [node_id for node_id in self.pendingNodes if self.blockCount[node_id] == 0]
def pop_node(self, unique_id):
del self.pendingNodes[unique_id]
for blocked_node_id in self.blocking[unique_id]:
self.blockCount[blocked_node_id] -= 1
del self.blocking[unique_id]
def is_empty(self):
return len(self.pendingNodes) == 0
class ExecutionList(TopologicalSort):
"""
ExecutionList implements a topological dissolve of the graph. After a node is staged for execution,
it can still be returned to the graph after having further dependencies added.
"""
def __init__(self, dynprompt, output_cache):
super().__init__(dynprompt)
self.output_cache = output_cache
self.staged_node_id = None
self.execution_cache = {}
self.execution_cache_listeners = {}
def is_cached(self, node_id):
return self.output_cache.get_local(node_id) is not None
def cache_link(self, from_node_id, to_node_id):
if to_node_id not in self.execution_cache:
self.execution_cache[to_node_id] = {}
self.execution_cache[to_node_id][from_node_id] = self.output_cache.get_local(from_node_id)
if from_node_id not in self.execution_cache_listeners:
self.execution_cache_listeners[from_node_id] = set()
self.execution_cache_listeners[from_node_id].add(to_node_id)
def get_cache(self, from_node_id, to_node_id):
if to_node_id not in self.execution_cache:
return None
value = self.execution_cache[to_node_id].get(from_node_id)
if value is None:
return None
#Write back to the main cache on touch.
self.output_cache.set_local(from_node_id, value)
return value
def cache_update(self, node_id, value):
if node_id in self.execution_cache_listeners:
for to_node_id in self.execution_cache_listeners[node_id]:
if to_node_id in self.execution_cache:
self.execution_cache[to_node_id][node_id] = value
def add_strong_link(self, from_node_id, from_socket, to_node_id):
super().add_strong_link(from_node_id, from_socket, to_node_id)
self.cache_link(from_node_id, to_node_id)
async def stage_node_execution(self):
assert self.staged_node_id is None
if self.is_empty():
return None, None, None
available = self.get_ready_nodes()
while len(available) == 0 and self.externalBlocks > 0:
# Wait for an external block to be released
await self.unblockedEvent.wait()
self.unblockedEvent.clear()
available = self.get_ready_nodes()
if len(available) == 0:
cycled_nodes = self.get_nodes_in_cycle()
# Because cycles composed entirely of static nodes are caught during initial validation,
# we will 'blame' the first node in the cycle that is not a static node.
blamed_node = cycled_nodes[0]
for node_id in cycled_nodes:
display_node_id = self.dynprompt.get_display_node_id(node_id)
if display_node_id != node_id:
blamed_node = display_node_id
break
ex = DependencyCycleError("Dependency cycle detected")
error_details = {
"node_id": blamed_node,
"exception_message": str(ex),
"exception_type": "graph.DependencyCycleError",
"traceback": [],
"current_inputs": []
}
return None, error_details, ex
self.staged_node_id = self.ux_friendly_pick_node(available)
return self.staged_node_id, None, None
def ux_friendly_pick_node(self, node_list):
# If an output node is available, do that first.
# Technically this has no effect on the overall length of execution, but it feels better as a user
# for a PreviewImage to display a result as soon as it can
# Some other heuristics could probably be used here to improve the UX further.
def is_output(node_id):
class_type = self.dynprompt.get_node(node_id)["class_type"]
class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
if hasattr(class_def, 'OUTPUT_NODE') and class_def.OUTPUT_NODE == True:
return True
return False
# If an available node is async, do that first.
# This will execute the asynchronous function earlier, reducing the overall time.
def is_async(node_id):
class_type = self.dynprompt.get_node(node_id)["class_type"]
class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
return inspect.iscoroutinefunction(getattr(class_def, class_def.FUNCTION))
for node_id in node_list:
if is_output(node_id) or is_async(node_id):
return node_id
#This should handle the VAEDecode -> preview case
for node_id in node_list:
for blocked_node_id in self.blocking[node_id]:
if is_output(blocked_node_id):
return node_id
#This should handle the VAELoader -> VAEDecode -> preview case
for node_id in node_list:
for blocked_node_id in self.blocking[node_id]:
for blocked_node_id1 in self.blocking[blocked_node_id]:
if is_output(blocked_node_id1):
return node_id
#TODO: this function should be improved
return node_list[0]
def unstage_node_execution(self):
assert self.staged_node_id is not None
self.staged_node_id = None
def complete_node_execution(self):
node_id = self.staged_node_id
self.pop_node(node_id)
self.execution_cache.pop(node_id, None)
self.execution_cache_listeners.pop(node_id, None)
self.staged_node_id = None
def get_nodes_in_cycle(self):
# We'll dissolve the graph in reverse topological order to leave only the nodes in the cycle.
# We're skipping some of the performance optimizations from the original TopologicalSort to keep
# the code simple (and because having a cycle in the first place is a catastrophic error)
blocked_by = { node_id: {} for node_id in self.pendingNodes }
for from_node_id in self.blocking:
for to_node_id in self.blocking[from_node_id]:
if True in self.blocking[from_node_id][to_node_id].values():
blocked_by[to_node_id][from_node_id] = True
to_remove = [node_id for node_id in blocked_by if len(blocked_by[node_id]) == 0]
while len(to_remove) > 0:
for node_id in to_remove:
for to_node_id in blocked_by:
if node_id in blocked_by[to_node_id]:
del blocked_by[to_node_id][node_id]
del blocked_by[node_id]
to_remove = [node_id for node_id in blocked_by if len(blocked_by[node_id]) == 0]
return list(blocked_by.keys())