Files
InvokeAI/invokeai/app/cli_app.py
psychedelicious 593604bbba feat(nodes): add invocation cache
The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned.

## Results

This feature provides anywhere some significant to massive performance improvement.

The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal.

## Overview

A new `invocation_cache` service is added to handle the caching. There's not much to it.

All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching.

The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic.

To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key.

## In-Memory Implementation

An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts.

Max node cache size is added as `node_cache_size` under the `Generation` config category.

It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher.

Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them.

## Node Definition

The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`.

Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`.

The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something.

## One Gotcha

Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again.

If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit.

## Linear UI

The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs.

This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default.

This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`.

## Workflow Editor

All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user.

The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes.

Users should consider saving their workflows after loading them in and having them updated.

## Future Enhancements - Callback

A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not.

This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field.

## Future Enhancements - Persisted Cache

Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future.
2023-09-18 13:41:19 +10:00

486 lines
19 KiB
Python

# Copyright (c) 2022 Kyle Schouviller (https://github.com/kyle0654)
from invokeai.app.services.invocation_cache.invocation_cache_memory import MemoryInvocationCache
from .services.config import InvokeAIAppConfig
# parse_args() must be called before any other imports. if it is not called first, consumers of the config
# which are imported/used before parse_args() is called will get the default config values instead of the
# values from the command line or config file.
config = InvokeAIAppConfig.get_config()
config.parse_args()
if True: # hack to make flake8 happy with imports coming after setting up the config
import argparse
import re
import shlex
import sys
import time
from typing import Optional, Union, get_type_hints
import torch
from pydantic import BaseModel, ValidationError
from pydantic.fields import Field
import invokeai.backend.util.hotfixes # noqa: F401 (monkeypatching on import)
from invokeai.app.services.board_image_record_storage import SqliteBoardImageRecordStorage
from invokeai.app.services.board_images import BoardImagesService, BoardImagesServiceDependencies
from invokeai.app.services.board_record_storage import SqliteBoardRecordStorage
from invokeai.app.services.boards import BoardService, BoardServiceDependencies
from invokeai.app.services.image_record_storage import SqliteImageRecordStorage
from invokeai.app.services.images import ImageService, ImageServiceDependencies
from invokeai.app.services.invocation_stats import InvocationStatsService
from invokeai.app.services.resource_name import SimpleNameService
from invokeai.app.services.urls import LocalUrlService
from invokeai.backend.util.logging import InvokeAILogger
from invokeai.version.invokeai_version import __version__
from .cli.commands import BaseCommand, CliContext, ExitCli, SortedHelpFormatter, add_graph_parsers, add_parsers
from .cli.completer import set_autocompleter
from .invocations.baseinvocation import BaseInvocation
from .services.default_graphs import create_system_graphs, default_text_to_image_graph_id
from .services.events import EventServiceBase
from .services.graph import (
Edge,
EdgeConnection,
GraphExecutionState,
GraphInvocation,
LibraryGraph,
are_connection_types_compatible,
)
from .services.image_file_storage import DiskImageFileStorage
from .services.invocation_queue import MemoryInvocationQueue
from .services.invocation_services import InvocationServices
from .services.invoker import Invoker
from .services.latent_storage import DiskLatentsStorage, ForwardCacheLatentsStorage
from .services.model_manager_service import ModelManagerService
from .services.processor import DefaultInvocationProcessor
from .services.sqlite import SqliteItemStorage
if torch.backends.mps.is_available():
import invokeai.backend.util.mps_fixes # noqa: F401 (monkeypatching on import)
logger = InvokeAILogger().getLogger(config=config)
class CliCommand(BaseModel):
command: Union[BaseCommand.get_commands() + BaseInvocation.get_invocations()] = Field(discriminator="type") # type: ignore
class InvalidArgs(Exception):
pass
def add_invocation_args(command_parser):
# Add linking capability
command_parser.add_argument(
"--link",
"-l",
action="append",
nargs=3,
help="A link in the format 'source_node source_field dest_field'. source_node can be relative to history (e.g. -1)",
)
command_parser.add_argument(
"--link_node",
"-ln",
action="append",
help="A link from all fields in the specified node. Node can be relative to history (e.g. -1)",
)
def get_command_parser(services: InvocationServices) -> argparse.ArgumentParser:
# Create invocation parser
parser = argparse.ArgumentParser(formatter_class=SortedHelpFormatter)
def exit(*args, **kwargs):
raise InvalidArgs
parser.exit = exit
subparsers = parser.add_subparsers(dest="type")
# Create subparsers for each invocation
invocations = BaseInvocation.get_all_subclasses()
add_parsers(subparsers, invocations, add_arguments=add_invocation_args)
# Create subparsers for each command
commands = BaseCommand.get_all_subclasses()
add_parsers(subparsers, commands, exclude_fields=["type"])
# Create subparsers for exposed CLI graphs
# TODO: add a way to identify these graphs
text_to_image = services.graph_library.get(default_text_to_image_graph_id)
add_graph_parsers(subparsers, [text_to_image], add_arguments=add_invocation_args)
return parser
class NodeField:
alias: str
node_path: str
field: str
field_type: type
def __init__(self, alias: str, node_path: str, field: str, field_type: type):
self.alias = alias
self.node_path = node_path
self.field = field
self.field_type = field_type
def fields_from_type_hints(hints: dict[str, type], node_path: str) -> dict[str, NodeField]:
return {k: NodeField(alias=k, node_path=node_path, field=k, field_type=v) for k, v in hints.items()}
def get_node_input_field(graph: LibraryGraph, field_alias: str, node_id: str) -> NodeField:
"""Gets the node field for the specified field alias"""
exposed_input = next(e for e in graph.exposed_inputs if e.alias == field_alias)
node_type = type(graph.graph.get_node(exposed_input.node_path))
return NodeField(
alias=exposed_input.alias,
node_path=f"{node_id}.{exposed_input.node_path}",
field=exposed_input.field,
field_type=get_type_hints(node_type)[exposed_input.field],
)
def get_node_output_field(graph: LibraryGraph, field_alias: str, node_id: str) -> NodeField:
"""Gets the node field for the specified field alias"""
exposed_output = next(e for e in graph.exposed_outputs if e.alias == field_alias)
node_type = type(graph.graph.get_node(exposed_output.node_path))
node_output_type = node_type.get_output_type()
return NodeField(
alias=exposed_output.alias,
node_path=f"{node_id}.{exposed_output.node_path}",
field=exposed_output.field,
field_type=get_type_hints(node_output_type)[exposed_output.field],
)
def get_node_inputs(invocation: BaseInvocation, context: CliContext) -> dict[str, NodeField]:
"""Gets the inputs for the specified invocation from the context"""
node_type = type(invocation)
if node_type is not GraphInvocation:
return fields_from_type_hints(get_type_hints(node_type), invocation.id)
else:
graph: LibraryGraph = context.invoker.services.graph_library.get(context.graph_nodes[invocation.id])
return {e.alias: get_node_input_field(graph, e.alias, invocation.id) for e in graph.exposed_inputs}
def get_node_outputs(invocation: BaseInvocation, context: CliContext) -> dict[str, NodeField]:
"""Gets the outputs for the specified invocation from the context"""
node_type = type(invocation)
if node_type is not GraphInvocation:
return fields_from_type_hints(get_type_hints(node_type.get_output_type()), invocation.id)
else:
graph: LibraryGraph = context.invoker.services.graph_library.get(context.graph_nodes[invocation.id])
return {e.alias: get_node_output_field(graph, e.alias, invocation.id) for e in graph.exposed_outputs}
def generate_matching_edges(a: BaseInvocation, b: BaseInvocation, context: CliContext) -> list[Edge]:
"""Generates all possible edges between two invocations"""
afields = get_node_outputs(a, context)
bfields = get_node_inputs(b, context)
matching_fields = set(afields.keys()).intersection(bfields.keys())
# Remove invalid fields
invalid_fields = set(["type", "id"])
matching_fields = matching_fields.difference(invalid_fields)
# Validate types
matching_fields = [
f for f in matching_fields if are_connection_types_compatible(afields[f].field_type, bfields[f].field_type)
]
edges = [
Edge(
source=EdgeConnection(node_id=afields[alias].node_path, field=afields[alias].field),
destination=EdgeConnection(node_id=bfields[alias].node_path, field=bfields[alias].field),
)
for alias in matching_fields
]
return edges
class SessionError(Exception):
"""Raised when a session error has occurred"""
pass
def invoke_all(context: CliContext):
"""Runs all invocations in the specified session"""
context.invoker.invoke(context.session, invoke_all=True)
while not context.get_session().is_complete():
# Wait some time
time.sleep(0.1)
# Print any errors
if context.session.has_error():
for n in context.session.errors:
context.invoker.services.logger.error(
f"Error in node {n} (source node {context.session.prepared_source_mapping[n]}): {context.session.errors[n]}"
)
raise SessionError()
def invoke_cli():
logger.info(f"InvokeAI version {__version__}")
# get the optional list of invocations to execute on the command line
parser = config.get_parser()
parser.add_argument("commands", nargs="*")
invocation_commands = parser.parse_args().commands
# get the optional file to read commands from.
# Simplest is to use it for STDIN
if infile := config.from_file:
sys.stdin = open(infile, "r")
model_manager = ModelManagerService(config, logger)
events = EventServiceBase()
output_folder = config.output_path
# TODO: build a file/path manager?
if config.use_memory_db:
db_location = ":memory:"
else:
db_location = config.db_path
db_location.parent.mkdir(parents=True, exist_ok=True)
logger.info(f'InvokeAI database location is "{db_location}"')
graph_execution_manager = SqliteItemStorage[GraphExecutionState](
filename=db_location, table_name="graph_executions"
)
urls = LocalUrlService()
image_record_storage = SqliteImageRecordStorage(db_location)
image_file_storage = DiskImageFileStorage(f"{output_folder}/images")
names = SimpleNameService()
board_record_storage = SqliteBoardRecordStorage(db_location)
board_image_record_storage = SqliteBoardImageRecordStorage(db_location)
boards = BoardService(
services=BoardServiceDependencies(
board_image_record_storage=board_image_record_storage,
board_record_storage=board_record_storage,
image_record_storage=image_record_storage,
url=urls,
logger=logger,
)
)
board_images = BoardImagesService(
services=BoardImagesServiceDependencies(
board_image_record_storage=board_image_record_storage,
board_record_storage=board_record_storage,
image_record_storage=image_record_storage,
url=urls,
logger=logger,
)
)
images = ImageService(
services=ImageServiceDependencies(
board_image_record_storage=board_image_record_storage,
image_record_storage=image_record_storage,
image_file_storage=image_file_storage,
url=urls,
logger=logger,
names=names,
graph_execution_manager=graph_execution_manager,
)
)
services = InvocationServices(
model_manager=model_manager,
events=events,
latents=ForwardCacheLatentsStorage(DiskLatentsStorage(f"{output_folder}/latents")),
images=images,
boards=boards,
board_images=board_images,
queue=MemoryInvocationQueue(),
graph_library=SqliteItemStorage[LibraryGraph](filename=db_location, table_name="graphs"),
graph_execution_manager=graph_execution_manager,
processor=DefaultInvocationProcessor(),
performance_statistics=InvocationStatsService(graph_execution_manager),
logger=logger,
configuration=config,
invocation_cache=MemoryInvocationCache(max_cache_size=config.node_cache_size),
)
system_graphs = create_system_graphs(services.graph_library)
system_graph_names = set([g.name for g in system_graphs])
set_autocompleter(services)
invoker = Invoker(services)
session: GraphExecutionState = invoker.create_execution_state()
parser = get_command_parser(services)
re_negid = re.compile("^-[0-9]+$")
# Uncomment to print out previous sessions at startup
# print(services.session_manager.list())
context = CliContext(invoker, session, parser)
set_autocompleter(services)
command_line_args_exist = len(invocation_commands) > 0
done = False
while not done:
try:
if command_line_args_exist:
cmd_input = invocation_commands.pop(0)
done = len(invocation_commands) == 0
else:
cmd_input = input("invoke> ")
except (KeyboardInterrupt, EOFError):
# Ctrl-c exits
break
try:
# Refresh the state of the session
# history = list(get_graph_execution_history(context.session))
history = list(reversed(context.nodes_added))
# Split the command for piping
cmds = cmd_input.split("|")
start_id = len(context.nodes_added)
current_id = start_id
new_invocations = list()
for cmd in cmds:
if cmd is None or cmd.strip() == "":
raise InvalidArgs("Empty command")
# Parse args to create invocation
args = vars(context.parser.parse_args(shlex.split(cmd.strip())))
# Override defaults
for field_name, field_default in context.defaults.items():
if field_name in args:
args[field_name] = field_default
# Parse invocation
command: CliCommand = None # type:ignore
system_graph: Optional[LibraryGraph] = None
if args["type"] in system_graph_names:
system_graph = next(filter(lambda g: g.name == args["type"], system_graphs))
invocation = GraphInvocation(graph=system_graph.graph, id=str(current_id))
for exposed_input in system_graph.exposed_inputs:
if exposed_input.alias in args:
node = invocation.graph.get_node(exposed_input.node_path)
field = exposed_input.field
setattr(node, field, args[exposed_input.alias])
command = CliCommand(command=invocation)
context.graph_nodes[invocation.id] = system_graph.id
else:
args["id"] = current_id
command = CliCommand(command=args)
if command is None:
continue
# Run any CLI commands immediately
if isinstance(command.command, BaseCommand):
# Invoke all current nodes to preserve operation order
invoke_all(context)
# Run the command
command.command.run(context)
continue
# TODO: handle linking with library graphs
# Pipe previous command output (if there was a previous command)
edges: list[Edge] = list()
if len(history) > 0 or current_id != start_id:
from_id = history[0] if current_id == start_id else str(current_id - 1)
from_node = (
next(filter(lambda n: n[0].id == from_id, new_invocations))[0]
if current_id != start_id
else context.session.graph.get_node(from_id)
)
matching_edges = generate_matching_edges(from_node, command.command, context)
edges.extend(matching_edges)
# Parse provided links
if "link_node" in args and args["link_node"]:
for link in args["link_node"]:
node_id = link
if re_negid.match(node_id):
node_id = str(current_id + int(node_id))
link_node = context.session.graph.get_node(node_id)
matching_edges = generate_matching_edges(link_node, command.command, context)
matching_destinations = [e.destination for e in matching_edges]
edges = [e for e in edges if e.destination not in matching_destinations]
edges.extend(matching_edges)
if "link" in args and args["link"]:
for link in args["link"]:
edges = [
e
for e in edges
if e.destination.node_id != command.command.id or e.destination.field != link[2]
]
node_id = link[0]
if re_negid.match(node_id):
node_id = str(current_id + int(node_id))
# TODO: handle missing input/output
node_output = get_node_outputs(context.session.graph.get_node(node_id), context)[link[1]]
node_input = get_node_inputs(command.command, context)[link[2]]
edges.append(
Edge(
source=EdgeConnection(node_id=node_output.node_path, field=node_output.field),
destination=EdgeConnection(node_id=node_input.node_path, field=node_input.field),
)
)
new_invocations.append((command.command, edges))
current_id = current_id + 1
# Add the node to the session
context.add_node(command.command)
for edge in edges:
print(edge)
context.add_edge(edge)
# Execute all remaining nodes
invoke_all(context)
except InvalidArgs:
invoker.services.logger.warning('Invalid command, use "help" to list commands')
continue
except ValidationError:
invoker.services.logger.warning('Invalid command arguments, run "<command> --help" for summary')
except SessionError:
# Start a new session
invoker.services.logger.warning("Session error: creating a new session")
context.reset()
except ExitCli:
break
except SystemExit:
continue
invoker.stop()
if __name__ == "__main__":
if config.version:
print(f"InvokeAI version {__version__}")
else:
invoke_cli()