Artificial Sweetener
017e3fc7fe
Keep compatible parallel regional paths on one inherited model lineage, including NegPip interoperability, while retaining bounded fused and optional Triton execution paths. Expand graph-shape, lifecycle, memory-safety, and runtime regressions across the supported attention families.
361 lines
14 KiB
Python
361 lines
14 KiB
Python
# SimpleSyrup - workflow-focused ComfyUI extensions for image generation
|
|
# Copyright (C) 2026 Artificial Sweetener and contributors
|
|
# SPDX-License-Identifier: AGPL-3.0-or-later
|
|
|
|
"""Validate collaborator-owned MODEL modifier state before regional derivation."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import logging
|
|
from dataclasses import dataclass
|
|
from enum import StrEnum
|
|
|
|
from comfy.model_patcher import ModelPatcher
|
|
from comfy.patcher_extension import WrappersMP
|
|
|
|
from ..domain.processed_regional_attention import ProcessedRegionalAttentionPlan
|
|
from ..domain.regional_attention_execution import RegionalAttentionExecutionMode
|
|
from ..domain.regional_model_capabilities import (
|
|
RegionalModelCapabilities,
|
|
RegionalModelFamily,
|
|
RegionalPatchConflict,
|
|
)
|
|
from .ppm_negpip_interop import (
|
|
PPM_NEGPIP_INTEROP_VALIDATOR,
|
|
PpmNegpipInterop,
|
|
)
|
|
|
|
LOGGER = logging.getLogger(__name__)
|
|
|
|
_EASYCACHE_OPTION = "easycache"
|
|
_ATTN2_PATCH_CONFLICTS = {
|
|
RegionalPatchConflict.ATTN2_INPUT_PATCH: "attn2_patch",
|
|
RegionalPatchConflict.ATTN2_OUTPUT_PATCH: "attn2_output_patch",
|
|
}
|
|
|
|
|
|
class RegionalPreservedModelModifier(StrEnum):
|
|
"""Identify one admitted upstream modifier category without local paths."""
|
|
|
|
MODEL_FUNCTION_WRAPPER = "model_function_wrapper"
|
|
DIFFUSION_MODEL_WRAPPER = "diffusion_model_wrapper"
|
|
OPTIMIZED_ATTENTION_OVERRIDE = "optimized_attention_override"
|
|
OBJECT_PATCH = "object_patch"
|
|
MODEL_WEIGHT_PATCH = "model_weight_patch"
|
|
EASYCACHE = "easycache"
|
|
|
|
|
|
@dataclass(frozen=True, slots=True)
|
|
class RegionalModelPatchInteropReport:
|
|
"""Publish immutable modifier categories preserved by one admitted MODEL."""
|
|
|
|
model_family: RegionalModelFamily
|
|
preserved_modifiers: tuple[RegionalPreservedModelModifier, ...]
|
|
negpip: PpmNegpipInterop | None = None
|
|
|
|
def __post_init__(self) -> None:
|
|
"""Require a typed family and canonical unique modifier order."""
|
|
|
|
if not isinstance(self.model_family, RegionalModelFamily):
|
|
raise TypeError("Regional interop report family has an invalid type.")
|
|
if not isinstance(self.preserved_modifiers, tuple) or any(
|
|
not isinstance(modifier, RegionalPreservedModelModifier)
|
|
for modifier in self.preserved_modifiers
|
|
):
|
|
raise TypeError("Regional interop report modifiers have invalid types.")
|
|
if len(set(self.preserved_modifiers)) != len(self.preserved_modifiers):
|
|
raise ValueError("Regional interop report modifiers must be unique.")
|
|
if self.negpip is not None and not isinstance(self.negpip, PpmNegpipInterop):
|
|
raise TypeError("Regional interop report NegPiP state has an invalid type.")
|
|
|
|
@property
|
|
def cache_modifier(self) -> RegionalPreservedModelModifier | None:
|
|
"""Return admitted EasyCache state when installed."""
|
|
|
|
modifier = RegionalPreservedModelModifier.EASYCACHE
|
|
if modifier in self.preserved_modifiers:
|
|
return modifier
|
|
return None
|
|
|
|
|
|
class RegionalModelPatchInteropValidator:
|
|
"""Admit preserved modifier surfaces and reject known branch collisions."""
|
|
|
|
def validate(
|
|
self,
|
|
model: object,
|
|
capabilities: RegionalModelCapabilities,
|
|
) -> RegionalModelPatchInteropReport:
|
|
"""Return preserved categories without changing supplied MODEL state."""
|
|
|
|
if not isinstance(model, ModelPatcher):
|
|
raise TypeError("Regional patch interop requires a Comfy ModelPatcher.")
|
|
if not isinstance(capabilities, RegionalModelCapabilities):
|
|
raise TypeError("Regional patch interop requires model capabilities.")
|
|
|
|
model_options = _require_dictionary_attribute(model, "model_options")
|
|
transformer_options = _require_nested_dictionary(
|
|
model_options,
|
|
"transformer_options",
|
|
owner_label="MODEL",
|
|
)
|
|
wrappers = _require_wrapper_state(model)
|
|
object_patches = _require_dictionary_attribute(model, "object_patches")
|
|
model_weight_patches = _require_dictionary_attribute(model, "patches")
|
|
patches = _require_optional_patch_state(transformer_options)
|
|
|
|
negpip = PPM_NEGPIP_INTEROP_VALIDATOR.validate(
|
|
capabilities.model_family,
|
|
model_options=model_options,
|
|
wrappers=wrappers,
|
|
object_patches=object_patches,
|
|
transformer_patches=patches,
|
|
)
|
|
cache_modifier = self._validate_cache_state(
|
|
transformer_options,
|
|
wrappers,
|
|
)
|
|
self._reject_attention_collisions(
|
|
patches,
|
|
capabilities,
|
|
admitted_negpip=negpip,
|
|
)
|
|
|
|
modifiers: list[RegionalPreservedModelModifier] = []
|
|
model_wrapper = model_options.get("model_function_wrapper")
|
|
if model_wrapper is not None:
|
|
if not callable(model_wrapper):
|
|
raise TypeError("MODEL model_function_wrapper must be callable.")
|
|
modifiers.append(RegionalPreservedModelModifier.MODEL_FUNCTION_WRAPPER)
|
|
if any(wrappers[wrapper_type] for wrapper_type in wrappers):
|
|
if (
|
|
WrappersMP.DIFFUSION_MODEL in wrappers
|
|
and wrappers[WrappersMP.DIFFUSION_MODEL]
|
|
):
|
|
modifiers.append(RegionalPreservedModelModifier.DIFFUSION_MODEL_WRAPPER)
|
|
optimized_override = transformer_options.get("optimized_attention_override")
|
|
if optimized_override is not None:
|
|
if not callable(optimized_override):
|
|
raise TypeError("MODEL optimized_attention_override must be callable.")
|
|
modifiers.append(
|
|
RegionalPreservedModelModifier.OPTIMIZED_ATTENTION_OVERRIDE
|
|
)
|
|
if object_patches:
|
|
modifiers.append(RegionalPreservedModelModifier.OBJECT_PATCH)
|
|
if model_weight_patches:
|
|
modifiers.append(RegionalPreservedModelModifier.MODEL_WEIGHT_PATCH)
|
|
if cache_modifier is not None:
|
|
modifiers.append(cache_modifier)
|
|
|
|
report = RegionalModelPatchInteropReport(
|
|
capabilities.model_family,
|
|
tuple(modifiers),
|
|
negpip,
|
|
)
|
|
LOGGER.info(
|
|
"Regional MODEL patch interoperability admitted",
|
|
extra={
|
|
"model_family": report.model_family.value,
|
|
"preserved_modifiers": tuple(
|
|
modifier.value for modifier in report.preserved_modifiers
|
|
),
|
|
},
|
|
)
|
|
return report
|
|
|
|
def validate_execution(
|
|
self,
|
|
report: RegionalModelPatchInteropReport,
|
|
processed_plan: ProcessedRegionalAttentionPlan,
|
|
execution_mode: RegionalAttentionExecutionMode,
|
|
) -> None:
|
|
"""Reject cache state that cannot identify spatial or scheduled changes."""
|
|
|
|
if not isinstance(report, RegionalModelPatchInteropReport):
|
|
raise TypeError("Regional execution interop requires an admission report.")
|
|
if not isinstance(processed_plan, ProcessedRegionalAttentionPlan):
|
|
raise TypeError("Regional execution interop requires a processed plan.")
|
|
if not isinstance(execution_mode, RegionalAttentionExecutionMode):
|
|
raise TypeError("Regional execution interop requires an execution mode.")
|
|
cache_modifier = report.cache_modifier
|
|
if cache_modifier is None:
|
|
return
|
|
cache_name = cache_modifier.value
|
|
if execution_mode is not RegionalAttentionExecutionMode.FULL:
|
|
raise ValueError(
|
|
f"Attention Coupling cannot compose {cache_name} with "
|
|
f"{execution_mode.value} spatial views because the installed "
|
|
"cache identity does not encode view coordinates. Remove the "
|
|
"cache modifier before using this sampler."
|
|
)
|
|
if not processed_plan.is_time_invariant:
|
|
raise ValueError(
|
|
f"Attention Coupling cannot compose {cache_name} with scheduled "
|
|
"regional execution because the installed cache identity does "
|
|
"not encode regional context or LoRA schedule state. Remove the "
|
|
"cache modifier or use time-invariant regional conditioning."
|
|
)
|
|
LOGGER.info(
|
|
"Regional MODEL cache execution admitted",
|
|
extra={
|
|
"model_family": report.model_family.value,
|
|
"cache_modifier": cache_name,
|
|
"execution_mode": execution_mode.value,
|
|
},
|
|
)
|
|
|
|
@staticmethod
|
|
def _validate_cache_state(
|
|
transformer_options: dict[object, object],
|
|
wrappers: dict[str, dict[object, list[object]]],
|
|
) -> RegionalPreservedModelModifier | None:
|
|
"""Require one complete core cache owner or no cache state."""
|
|
|
|
easy_surfaces = (
|
|
_has_wrapper(wrappers, WrappersMP.OUTER_SAMPLE, "easycache"),
|
|
_has_wrapper(wrappers, WrappersMP.CALC_COND_BATCH, "easycache"),
|
|
_has_wrapper(wrappers, WrappersMP.DIFFUSION_MODEL, "easycache"),
|
|
)
|
|
lazy_surfaces = (
|
|
_has_wrapper(wrappers, WrappersMP.OUTER_SAMPLE, "lazycache"),
|
|
_has_wrapper(wrappers, WrappersMP.PREDICT_NOISE, "lazycache"),
|
|
)
|
|
has_easy = any(easy_surfaces)
|
|
has_lazy = any(lazy_surfaces)
|
|
if has_easy and has_lazy:
|
|
raise ValueError(
|
|
"EasyCache and LazyCache cannot both own one MODEL; use exactly "
|
|
"one cache modifier."
|
|
)
|
|
if has_easy and not all(easy_surfaces):
|
|
raise ValueError("MODEL contains an incomplete EasyCache wrapper set.")
|
|
if has_lazy and not all(lazy_surfaces):
|
|
raise ValueError("MODEL contains an incomplete LazyCache wrapper set.")
|
|
cache_state_exists = _EASYCACHE_OPTION in transformer_options
|
|
if (has_easy or has_lazy) != cache_state_exists:
|
|
raise ValueError(
|
|
"MODEL cache wrappers and transformer option state are inconsistent."
|
|
)
|
|
if has_lazy:
|
|
raise ValueError(
|
|
"Attention Coupling does not support LazyCache because it skips "
|
|
"complete denoiser evaluations and reuses previous-step prediction "
|
|
"deltas, so skipped steps cannot execute exact regional "
|
|
"conditioning or LoRA math. Remove LazyCache before the Attention "
|
|
"Coupling sampler."
|
|
)
|
|
if has_easy:
|
|
return RegionalPreservedModelModifier.EASYCACHE
|
|
return None
|
|
|
|
@staticmethod
|
|
def _reject_attention_collisions(
|
|
patches: dict[str, list[object]],
|
|
capabilities: RegionalModelCapabilities,
|
|
*,
|
|
admitted_negpip: PpmNegpipInterop | None,
|
|
) -> None:
|
|
"""Reject every populated attention surface owned by the backend."""
|
|
|
|
conflicts = tuple(
|
|
patch_name
|
|
for conflict, patch_name in _ATTN2_PATCH_CONFLICTS.items()
|
|
if conflict in capabilities.known_patch_conflicts
|
|
and patches.get(patch_name)
|
|
and not (
|
|
patch_name == "attn2_patch"
|
|
and admitted_negpip is not None
|
|
and patches[patch_name] == [admitted_negpip.attention_patch]
|
|
)
|
|
)
|
|
if conflicts:
|
|
raise ValueError(
|
|
"Attention Coupling cannot compose with existing MODEL attention "
|
|
f"patches: {', '.join(conflicts)}. Remove the upstream attention "
|
|
"modifier; this backend requires exclusive branch-aware ownership."
|
|
)
|
|
|
|
|
|
def _require_dictionary_attribute(
|
|
owner: object,
|
|
attribute_name: str,
|
|
) -> dict[object, object]:
|
|
"""Return one required mutable host dictionary without coercion."""
|
|
|
|
value = getattr(owner, attribute_name, None)
|
|
if not isinstance(value, dict):
|
|
raise TypeError(f"MODEL {attribute_name} must be a dictionary.")
|
|
return value
|
|
|
|
|
|
def _require_nested_dictionary(
|
|
mapping: dict[object, object],
|
|
key: str,
|
|
*,
|
|
owner_label: str,
|
|
) -> dict[object, object]:
|
|
"""Return one required nested host dictionary without changing it."""
|
|
|
|
value = mapping.get(key)
|
|
if not isinstance(value, dict):
|
|
raise TypeError(f"{owner_label} {key} must be a dictionary.")
|
|
return value
|
|
|
|
|
|
def _require_wrapper_state(
|
|
model: ModelPatcher,
|
|
) -> dict[str, dict[object, list[object]]]:
|
|
"""Validate and return the installed keyed wrapper dictionary."""
|
|
|
|
raw = getattr(model, "wrappers", None)
|
|
if not isinstance(raw, dict):
|
|
raise TypeError("MODEL wrappers must be a dictionary.")
|
|
narrowed: dict[str, dict[object, list[object]]] = {}
|
|
for wrapper_type, keyed in raw.items():
|
|
if not isinstance(wrapper_type, str) or not isinstance(keyed, dict):
|
|
raise TypeError("MODEL wrapper types must map to dictionaries.")
|
|
narrowed_keyed: dict[object, list[object]] = {}
|
|
for key, callbacks in keyed.items():
|
|
if not isinstance(callbacks, list):
|
|
raise TypeError("MODEL wrapper callbacks must be a list.")
|
|
if any(not callable(callback) for callback in callbacks):
|
|
raise TypeError("MODEL wrapper callbacks must be callable.")
|
|
narrowed_keyed[key] = callbacks
|
|
narrowed[wrapper_type] = narrowed_keyed
|
|
return narrowed
|
|
|
|
|
|
def _require_optional_patch_state(
|
|
transformer_options: dict[object, object],
|
|
) -> dict[str, list[object]]:
|
|
"""Validate and return the optional installed transformer patch lists."""
|
|
|
|
raw = transformer_options.get("patches")
|
|
if raw is None:
|
|
return {}
|
|
if not isinstance(raw, dict):
|
|
raise TypeError("MODEL transformer patches must be a dictionary.")
|
|
narrowed: dict[str, list[object]] = {}
|
|
for patch_name, callbacks in raw.items():
|
|
if not isinstance(patch_name, str):
|
|
raise TypeError("MODEL transformer patch names must be strings.")
|
|
if not isinstance(callbacks, list):
|
|
raise TypeError("MODEL transformer patch callbacks must be a list.")
|
|
if any(not callable(callback) for callback in callbacks):
|
|
raise TypeError("MODEL transformer patch callbacks must be callable.")
|
|
narrowed[patch_name] = callbacks
|
|
return narrowed
|
|
|
|
|
|
def _has_wrapper(
|
|
wrappers: dict[str, dict[object, list[object]]],
|
|
wrapper_type: str,
|
|
key: str,
|
|
) -> bool:
|
|
"""Report whether one exact installed cache wrapper list is populated."""
|
|
|
|
return bool(wrappers.get(wrapper_type, {}).get(key, ()))
|
|
|
|
|
|
REGIONAL_MODEL_PATCH_INTEROP_VALIDATOR = RegionalModelPatchInteropValidator()
|