260 lines
9.9 KiB
Python
260 lines
9.9 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 and persist complete P9.4 managed evidence."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import hashlib
|
|
import json
|
|
from pathlib import Path
|
|
from typing import cast
|
|
|
|
from tools.anima_attention_coupling_workflow import (
|
|
SAMPLER,
|
|
SCHEDULER,
|
|
SEED,
|
|
STEPS,
|
|
BuiltAnimaAttentionCouplingWorkflow,
|
|
)
|
|
from tools.comfy_api import ImageReference, JsonObject
|
|
|
|
from .fixture import TextEncoderLoraFixtureIdentity
|
|
from .history import extract_text_encoder_lora_history
|
|
from .image_validation import TEXT_ENCODER_LORA_IMAGE_VALIDATOR
|
|
from .matrix import TextEncoderLoraCase
|
|
|
|
|
|
class TextEncoderLoraResultRecorder:
|
|
"""Own ordered artifacts, acceptance, and terminal P9.4 publication."""
|
|
|
|
def __init__(self, root: Path) -> None:
|
|
"""Retain one already-created collision-resistant artifact root."""
|
|
|
|
self._root = root.resolve()
|
|
if not self._root.is_dir():
|
|
raise ValueError("P9.4 result root must already exist.")
|
|
self._observations: list[JsonObject] = []
|
|
self._image_paths: dict[str, Path] = {}
|
|
self._public_metadata: JsonObject = {}
|
|
self._fixture_identity: JsonObject | None = None
|
|
|
|
def record_fixture(self, identity: TextEncoderLoraFixtureIdentity) -> None:
|
|
"""Retain one already-validated external fixture identity."""
|
|
|
|
if not isinstance(identity, TextEncoderLoraFixtureIdentity):
|
|
raise TypeError("P9.4 result fixture requires a fixture identity.")
|
|
if self._fixture_identity is not None:
|
|
raise ValueError("P9.4 text-encoder LoRA fixture was already recorded.")
|
|
self._fixture_identity = identity.as_record()
|
|
|
|
def record_metadata(self, node_id: str, metadata: JsonObject) -> None:
|
|
"""Retain validated live public-node metadata by exact node ID."""
|
|
|
|
if not node_id or node_id in self._public_metadata:
|
|
raise ValueError(
|
|
"P9.4 public-node metadata identity is invalid or repeated."
|
|
)
|
|
self._public_metadata[node_id] = metadata
|
|
self._write_json(
|
|
self._root / "public-node-metadata.json", self._public_metadata
|
|
)
|
|
|
|
def record_case(
|
|
self,
|
|
case: TextEncoderLoraCase,
|
|
workflow: BuiltAnimaAttentionCouplingWorkflow,
|
|
*,
|
|
prompt_id: str,
|
|
history: JsonObject,
|
|
reference: ImageReference,
|
|
image_bytes: bytes,
|
|
wall_runtime_ms: float,
|
|
) -> Path:
|
|
"""Validate one history and persist its labeled original image."""
|
|
|
|
if any(item.get("case_id") == case.case_id for item in self._observations):
|
|
raise ValueError(f"P9.4 case was already recorded: {case.case_id}")
|
|
if wall_runtime_ms <= 0:
|
|
raise ValueError("P9.4 wall runtime must be positive.")
|
|
evidence = extract_text_encoder_lora_history(history, workflow)
|
|
self._validate_adapter_uses(case, evidence.diagnostics)
|
|
image_path = self._root / f"{case.case_id}.png"
|
|
image_path.write_bytes(image_bytes)
|
|
self._image_paths[case.case_id] = image_path
|
|
workflow_path = self._root / f"{case.case_id}.workflow.json"
|
|
history_path = self._root / f"{case.case_id}.history.json"
|
|
self._write_json(workflow_path, workflow.prompt)
|
|
self._write_json(history_path, history)
|
|
self._observations.append(
|
|
{
|
|
"case_id": case.case_id,
|
|
"label": case.label,
|
|
"spatial_mode": case.spatial_mode.value,
|
|
"global_text_lora": case.global_text_lora,
|
|
"regional_text_lora": case.regional_text_lora,
|
|
"regional_model_lora": case.regional_model_lora,
|
|
"comparison_case_id": case.comparison_case_id,
|
|
"prompt_id": prompt_id,
|
|
"metrics": evidence.metrics,
|
|
"conditioning_batch_snapshot": evidence.conditioning_batch_snapshot,
|
|
"diagnostic_record_count": evidence.diagnostics["record_count"],
|
|
"wall_runtime_ms": wall_runtime_ms,
|
|
"image_file": image_path.name,
|
|
"image_sha256": hashlib.sha256(image_bytes).hexdigest(),
|
|
"image_size_bytes": len(image_bytes),
|
|
"image_reference": {
|
|
"filename": reference.filename,
|
|
"subfolder": reference.subfolder,
|
|
"type": reference.output_type,
|
|
},
|
|
"workflow_file": workflow_path.name,
|
|
"history_file": history_path.name,
|
|
}
|
|
)
|
|
self._persist("in_progress", {})
|
|
return image_path
|
|
|
|
def finalize(
|
|
self,
|
|
definitions: tuple[TextEncoderLoraCase, ...],
|
|
*,
|
|
system_stats: JsonObject,
|
|
cleanup_verified: bool,
|
|
masks_removed: bool,
|
|
) -> Path:
|
|
"""Publish completion after calls, pixels, metadata, and cleanup pass."""
|
|
|
|
expected = tuple(case.case_id for case in definitions)
|
|
observed = tuple(str(item["case_id"]) for item in self._observations)
|
|
if observed != expected:
|
|
raise ValueError("P9.4 observations are incomplete or out of order.")
|
|
if len(self._public_metadata) != 3:
|
|
raise ValueError("P9.4 requires metadata for all three public samplers.")
|
|
if self._fixture_identity is None:
|
|
raise ValueError(
|
|
"P9.4 requires the exact text-encoder LoRA fixture identity."
|
|
)
|
|
if not cleanup_verified or not masks_removed:
|
|
raise ValueError("P9.4 managed process, port, or mask cleanup failed.")
|
|
self._validate_call_counts(definitions)
|
|
comparisons = TEXT_ENCODER_LORA_IMAGE_VALIDATOR.validate(
|
|
definitions,
|
|
self._image_paths,
|
|
)
|
|
return self._persist(
|
|
"completed",
|
|
{
|
|
"image_comparisons": comparisons,
|
|
"system_stats": system_stats,
|
|
"cleanup": {
|
|
"process_stopped": True,
|
|
"port_available": True,
|
|
"masks_removed": True,
|
|
},
|
|
},
|
|
)
|
|
|
|
@staticmethod
|
|
def _validate_adapter_uses(
|
|
case: TextEncoderLoraCase,
|
|
diagnostics: JsonObject,
|
|
) -> None:
|
|
"""Require text-only hooks to stay out of diffusion diagnostics."""
|
|
|
|
snapshots = cast(list[JsonObject], diagnostics["snapshots"])
|
|
uses = [
|
|
use
|
|
for snapshot in snapshots
|
|
for use in cast(list[object], snapshot.get("adapter_uses", []))
|
|
if isinstance(use, dict)
|
|
]
|
|
conditioning_entries = [
|
|
snapshot.get("regional_conditioning_entries") for snapshot in snapshots
|
|
]
|
|
if any(
|
|
not isinstance(entries, list) or len(entries) != 2
|
|
for entries in conditioning_entries
|
|
):
|
|
raise ValueError("P9.4 diagnostics must retain both regional contexts.")
|
|
if not case.regional_model_lora:
|
|
if uses:
|
|
raise ValueError(
|
|
f"P9.4 text-only case {case.case_id!r} entered model execution."
|
|
)
|
|
return
|
|
if not uses:
|
|
raise ValueError(
|
|
f"P9.4 model-LoRA case {case.case_id!r} has no adapter uses."
|
|
)
|
|
if any(use.get("target_count") != 448 for use in uses):
|
|
raise ValueError("P9.4 ADAPTER_A model execution must retain all 448 targets.")
|
|
|
|
def _validate_call_counts(
|
|
self,
|
|
definitions: tuple[TextEncoderLoraCase, ...],
|
|
) -> None:
|
|
"""Require every LoRA comparison to preserve complete denoiser calls."""
|
|
|
|
observations = {str(item["case_id"]): item for item in self._observations}
|
|
for case in definitions:
|
|
if case.comparison_case_id is None:
|
|
continue
|
|
if self._model_calls(observations[case.case_id]) != self._model_calls(
|
|
observations[case.comparison_case_id]
|
|
):
|
|
raise ValueError(
|
|
f"P9.4 case {case.case_id!r} changed complete model-call count."
|
|
)
|
|
|
|
@staticmethod
|
|
def _model_calls(observation: JsonObject) -> int:
|
|
"""Return one narrowed measured complete model-call count."""
|
|
|
|
metrics = observation.get("metrics")
|
|
if not isinstance(metrics, dict):
|
|
raise TypeError("P9.4 observation metrics must be an object.")
|
|
value = metrics.get("model_call_count")
|
|
if not isinstance(value, int):
|
|
raise TypeError("P9.4 observation model-call count must be an integer.")
|
|
return value
|
|
|
|
def _persist(self, status: str, extra: JsonObject) -> Path:
|
|
"""Atomically replace the current machine-readable result."""
|
|
|
|
path = self._root / "p9.4-result.json"
|
|
payload: JsonObject = {
|
|
"schema_version": 1,
|
|
"phase": "p9.4",
|
|
"status": status,
|
|
"sampling": {
|
|
"seed": SEED,
|
|
"steps": STEPS,
|
|
"sampler": SAMPLER,
|
|
"scheduler": SCHEDULER,
|
|
"full_dimensions": [1024, 1024],
|
|
"upscale_dimensions": [1024, 1024, 1536, 1536],
|
|
},
|
|
"observations": self._observations,
|
|
"text_encoder_lora_fixture": self._fixture_identity,
|
|
"public_node_metadata_file": "public-node-metadata.json",
|
|
**extra,
|
|
}
|
|
temporary = path.with_suffix(".tmp")
|
|
temporary.write_text(
|
|
json.dumps(payload, indent=2, sort_keys=True) + "\n",
|
|
encoding="utf-8",
|
|
)
|
|
temporary.replace(path)
|
|
return path
|
|
|
|
@staticmethod
|
|
def _write_json(path: Path, value: object) -> None:
|
|
"""Write one deterministic human-reviewable sidecar."""
|
|
|
|
path.write_text(
|
|
json.dumps(value, indent=2, sort_keys=True) + "\n",
|
|
encoding="utf-8",
|
|
)
|