From ce91173a6c41fe71e764eb968e484c2d6e9241f6 Mon Sep 17 00:00:00 2001 From: "mula.liu" Date: Sat, 7 Nov 2026 00:48:17 +0800 Subject: [PATCH] v 0.1.2 --- backend/app/api/pipeline.py | 32 +- backend/app/api/platform.py | 209 +-------- backend/app/constants.py | 6 +- backend/app/db.py | 115 ++++- backend/app/jobs/handlers.py | 45 +- backend/app/jobs/tasks.py | 7 + backend/app/models/pipeline.py | 23 +- backend/app/providers/comfy_common.py | 20 + backend/app/services/capacity.py | 2 +- backend/app/services/dispatch.py | 126 ++++-- backend/app/services/execution.py | 4 +- backend/app/services/platform_generation.py | 405 ++++++++++++++++++ backend/app/services/progress.py | 71 +++ backend/app/services/status.py | 60 ++- backend/app/worker.py | 3 +- backend/tests/conftest.py | 19 + backend/tests/test_direct_connection.py | 5 +- backend/tests/test_platform_assets.py | 26 +- backend/tests/test_platform_version_reuse.py | 11 +- backend/tests/test_startup_lock.py | 8 +- backend/tests/test_task_activity.py | 261 +++++++++++ backend/tests/test_voice_timbre_gender.py | 17 +- docs/UI规范.md | 83 ++++ frontend/src/app/App.tsx | 10 + frontend/src/app/layout/ActivityCenter.tsx | 203 +++++++++ frontend/src/app/layout/AppHeader.tsx | 5 + frontend/src/hooks/index.ts | 1 + frontend/src/hooks/use-auto-refresh.ts | 10 +- frontend/src/hooks/use-polling.ts | 7 +- frontend/src/hooks/use-refresh-on-activity.ts | 39 ++ frontend/src/pages/AssetLibrary.tsx | 103 ++++- frontend/src/stores/activity.ts | 230 ++++++++++ 32 files changed, 1871 insertions(+), 295 deletions(-) create mode 100644 backend/app/services/platform_generation.py create mode 100644 backend/app/services/progress.py create mode 100644 backend/tests/test_task_activity.py create mode 100644 frontend/src/app/layout/ActivityCenter.tsx create mode 100644 frontend/src/hooks/use-refresh-on-activity.ts create mode 100644 frontend/src/stores/activity.ts diff --git a/backend/app/api/pipeline.py b/backend/app/api/pipeline.py index 9f6a4d5..3550196 100644 --- a/backend/app/api/pipeline.py +++ b/backend/app/api/pipeline.py @@ -10,7 +10,7 @@ from app import schemas from app.api import deps from app.constants import ShotStatus, TaskKind, TaskStatus from app.db import get_db -from app.models import ReviewRecord, Shot, Task, now_utc +from app.models import Project, ReviewRecord, Shot, Task, now_utc from app.services import daemon, dispatch, events, quality, status, storage router = APIRouter(tags=["pipeline"]) @@ -88,6 +88,20 @@ def list_tasks(project_id: int | None = Query(None), status_filter: str = Query( return {"tasks": [status.task_dict(t) for t in rows], "count": len(rows), "counts": counts} +# 注意顺序:这条必须排在 /api/tasks/{task_id} 前面,否则 activity 会被当成 task_id 解析 +@router.get("/api/tasks/activity") +def get_task_activity(active_only: bool = Query(False), owner_type: str = Query(""), + limit: int = Query(status.ACTIVITY_LIMIT), db: Session = Depends(get_db)): + """平台动态:此刻在跑什么、刚刚跑完了什么(项目侧 + 平台侧同源一条接口)。 + + 顶栏每 3 秒问一次,所以它只扫任务表里进行中的那几行 + 最近 6 小时的落定记录, + 不碰项目统计。前端据此决定要不要重新拉列表——用户不必再手动刷新页面去猜 + 「图到底出没了」。 + """ + return {"ok": True, "result": status.activity(db, limit=limit, owner_type=owner_type, + active_only=active_only)} + + @router.get("/api/tasks/{task_id}") def get_task(task_id: int, db: Session = Depends(get_db)): return {"task": status.task_dict(deps.get_task(db, task_id))} @@ -96,7 +110,8 @@ def get_task(task_id: int, db: Session = Depends(get_db)): @router.post("/api/tasks/{task_id}/retry") def retry_task(task_id: int, data: schemas.RetryOptions, db: Session = Depends(get_db)): task = deps.get_task(db, task_id) - project = deps.get_project(db, task.project_id) + # 平台资产生成不属于任何项目:按项目取父记录会凭空 404,界面就永远没法重试 + project = db.get(Project, task.project_id) if task.project_id else None if task.status not in (TaskStatus.FAILED, TaskStatus.CANCELLED): raise HTTPException(409, f"当前状态 {task.status} 不能重试") task.status = TaskStatus.QUEUED @@ -113,18 +128,27 @@ def retry_task(task_id: int, data: schemas.RetryOptions, db: Session = Depends(g shot.status = ShotStatus.QUEUED shot.stage = "queued" db.flush() - events.log_event(db, "task.requeued", project_id=project.id, task_id=task.id, shot_id=task.shot_id, + events.log_event(db, "task.requeued", project_id=project.id if project else None, + task_id=task.id, shot_id=task.shot_id, message=f"任务 #{task.id} 人工重试入队", entity_type="task", entity_id=task.id) + if project is None: + db.commit() + # 用户是点了「重试」才来的,不该再等一轮调度扫描 + dispatch.dispatch_task(db, db.get(Task, task.id)) + return {"ok": True, "result": status.task_dict(db.get(Task, task.id))} return deps.done(db, project, {"task_id": task.id, "status": task.status}) @router.post("/api/tasks/{task_id}/cancel") def cancel_task(task_id: int, data: schemas.CancelOptions, db: Session = Depends(get_db)): task = deps.get_task(db, task_id) - project = deps.get_project(db, task.project_id) + project = db.get(Project, task.project_id) if task.project_id else None if task.status in TaskStatus.TERMINAL: raise HTTPException(409, f"任务已结束({task.status}),无需取消") dispatch.cancel_task(db, task, reason=data.reason) + if project is None: + db.commit() + return {"ok": True, "result": status.task_dict(db.get(Task, task.id))} return deps.done(db, project, {"task_id": task.id, "status": task.status}) diff --git a/backend/app/api/platform.py b/backend/app/api/platform.py index f6225dc..0eaade2 100644 --- a/backend/app/api/platform.py +++ b/backend/app/api/platform.py @@ -1,10 +1,8 @@ """平台级数字资产 API:跨项目人物、场景、声线的管理、生成与引入。""" from __future__ import annotations -import hashlib import re from collections import Counter -from pathlib import Path from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile from sqlalchemy import select @@ -17,8 +15,14 @@ from app.db import get_db from app.models import (Asset, Character, PlatformCharacter, PlatformMedia, PlatformMediaVersion, PlatformScene, PlatformVoice, Project, Scene, VoiceAsset, now_utc) from app.providers import registry as model_registry -from app.providers.base import AssetGenRequest, ProviderError, VoiceGenRequest -from app.services import audio, events, storage, timbre, versions +from app.providers.base import AssetGenRequest +from app.services import (audio, events, platform_generation, status as status_service, storage, + timbre, versions) +from app.services.platform_generation import (build_owner_prompt as _owner_prompt, + get_owner as _owner, + normalize_type as _normalize_type, + resolve_reference_media as _reference_media, + type_label as _type_label) router = APIRouter(tags=["platform-assets"]) @@ -30,28 +34,6 @@ PLATFORM_KINDS: dict[str, set[str]] = { } -def _normalize_type(owner_type: str) -> str: - return { - "characters": "human", "character": "human", "human": "human", "humans": "human", - "scene": "scene", "scenes": "scene", "voice": "voice", "voices": "voice", - }.get(owner_type, owner_type) - - -def _owner(db: Session, owner_type: str, owner_id: int): - owner_type = _normalize_type(owner_type) - if owner_type == "human": - row = db.get(PlatformCharacter, owner_id) - elif owner_type == "scene": - row = db.get(PlatformScene, owner_id) - elif owner_type == "voice": - row = db.get(PlatformVoice, owner_id) - else: - raise HTTPException(400, f"不支持的平台资产类型:{owner_type}") - if row is None: - raise HTTPException(404, "平台资产不存在") - return row - - def _media(db: Session, owner_type: str, owner_id: int) -> list[dict]: rows = db.query(PlatformMedia).filter( PlatformMedia.owner_type == owner_type, PlatformMedia.owner_id == owner_id, @@ -59,31 +41,6 @@ def _media(db: Session, owner_type: str, owner_id: int) -> list[dict]: return [storage.platform_media_dict(row) for row in rows] -def _reference_media(db: Session, owner_type: str, owner_id: int, ids: list[int], - *, kind: str) -> list[PlatformMedia]: - """校验平台生成参考素材只能来自当前资产,避免跨资产误读本地路径。""" - rows: list[PlatformMedia] = [] - seen: set[int] = set() - image_suffixes = {".png", ".jpg", ".jpeg", ".webp", ".bmp", ".tif", ".tiff"} - audio_suffixes = {".wav", ".mp3", ".m4a", ".aac", ".flac", ".ogg", ".opus"} - for media_id in ids: - if media_id in seen: - continue - seen.add(media_id) - row = db.get(PlatformMedia, media_id) - if (row is None or row.owner_type != owner_type or row.owner_id != owner_id): - raise HTTPException(400, "参考素材不属于当前平台资产") - if not storage.exists(row.path): - raise HTTPException(400, "参考素材文件不存在,请重新上传") - suffix = Path(row.path).suffix.lower() - if kind == "image" and suffix not in image_suffixes: - raise HTTPException(400, "图像参考素材必须是图片文件") - if kind == "audio" and suffix not in audio_suffixes: - raise HTTPException(400, "音色参考素材必须是音频文件") - rows.append(row) - return rows - - def _human_dict(db: Session, row: PlatformCharacter, media: list[dict] | None = None) -> dict: project = db.get(Project, row.project_id) if row.project_id else None return { @@ -920,39 +877,6 @@ def publish_project_asset(data: schemas.PlatformPublish, db: Session = Depends(g "media_version": max(versions) if versions else 0}} -def produced_media_metadata(media, reference: dict, check: dict, text: str) -> None: - """把「这次到底用了谁的嗓子、念了什么、音色对不对」写进素材,界面才说得清楚。""" - metadata = dict(media.metadata_json or {}) - metadata.update({ - "text": text, - "reference_audio": reference.get("path") or reference.get("attached") or "", - "reference_source": reference.get("source") or "", - "reference_note": reference.get("note") or "使用本声线已绑定的参考音频", - "timbre_check": check, - }) - media.metadata_json = metadata - note = str(media.note or "") - tail = check.get("note") or "" - if tail and "音色自检" not in note: - media.note = f"{note}|音色自检:{tail}".strip("|") - - -def _owner_prompt(project: Project | None, owner_type: str, owner, kind: str, custom: str) -> str: - if custom.strip(): - return custom.strip() - bits = [] - if owner_type == "human": - bits = [f"{owner.name},电影级角色设定图", owner.appearance, owner.persona_prompt, - "五官、发型、体型、服装与配饰保持一致,正面清晰,材质细节可信"] - elif owner_type == "scene": - bits = [f"{owner.name},电影级场景概念图", owner.prompt or owner.description, - owner.lighting, owner.time_of_day, owner.palette, - "空间结构清晰,透视准确,建筑与陈设保持一致"] - if project: - bits.append(project.style_guide) - return "。".join(str(part).strip() for part in bits if part) - - @router.get("/api/platform/{owner_type}/{asset_id}/reference-status") def platform_reference_status(owner_type: str, asset_id: int, db: Session = Depends(get_db)): """开「生成试听」抽屉时先问一句:现在这点生成会拿到谁的嗓子。 @@ -969,109 +893,24 @@ def platform_reference_status(owner_type: str, asset_id: int, db: Session = Depe db, owner, clone=timbre.clone_model(spec), instruct=timbre.instruct_model(spec))} -@router.post("/api/platform/{owner_type}/{asset_id}/generate") +@router.post("/api/platform/{owner_type}/{asset_id}/generate", status_code=202) def generate_platform_asset(owner_type: str, asset_id: int, data: schemas.PlatformGenerate, db: Session = Depends(get_db)): + """发起一次生成:当场回 202 与任务号,出图/出试听在后台跑。 + + 以前这里是同步等出图的:实测 Qwen-Image 出图 46~49s,反代默认 60s 读超时, + 余量只有十几秒,同卡一排队就直接 504(前端表现为「生成失败」,图却照样落盘)。 + 现在它登记成统一任务表里的一条 platform_asset:便宜的校验同步做完才返 202 + (表单该红的框立刻红),进度与产物由 ``GET /api/tasks/activity`` 供 + 顶栏「平台动态」轮询,刷新页面、换台设备都还看得见同一条进度。 + """ owner_type = _normalize_type(owner_type) owner = _owner(db, owner_type, asset_id) - if owner_type == "voice": - spec, _adapter = model_registry.resolve_voice(db, data.provider or None, data.model or None) - if not spec.enabled: - raise HTTPException(409, f"模型 {spec.provider}/{spec.model} 已停用") - reference_ids = list(data.reference_media_ids) - if data.reference_media_id: - reference_ids.append(data.reference_media_id) - references = _reference_media(db, "voice", owner.id, reference_ids, kind="audio") - reference_media = references[0] if references else None - # 试听文本是「模型要念出来的台词」,不是声线设定。旧实现拿 owner.description 兜底, - # 于是用户填的「温和、语速中等,沉稳」被当成台词念了出来,试听样本一听就是设定说明文。 - text = (data.sample_text or "").strip() - if not text: - raise HTTPException(422, "请先填写试听文本:它是模型要朗读的台词," - "系统不会用「音色描述」代替") - out = storage.platform_dir("voice", owner.id, "generated") / f"sample-{now_utc().timestamp():.0f}.wav" - # 参考音频决定音色,而表单里的「性别」对 IndexTTS2 只是元数据。 - # 旧实现缺参考时直接回退到模型默认参考(女声),男声声线就这样被生成成了女声。 - reference = timbre.resolve_reference( - db, owner, - explicit=str(storage.to_absolute(reference_media.path)) if reference_media else "", - clone=timbre.clone_model(spec), instruct=timbre.instruct_model(spec)) - if not reference["path"] and reference.get("source") == "none": - raise HTTPException(409, reference["note"]) - # 交给模型的是绝对路径(供应商要直接读文件);声线列里存 media_root 相对路径, - # 换机器 / 进容器时不会因为写死绝对路径而失效。与任务链路 handlers 保持一致。 - reference_path = reference["path"] or owner.provider_voice_id - request = VoiceGenRequest( - text=text, voice_id=owner.name, provider_voice_id=reference_path, - language=owner.language, speed=float((owner.tuning or {}).get("speed", 1.0)), - seed=data.seed, params=(owner.tuning or {}).get("params") or {}, - # 没有参考音频时靠它出音,有参考音频时用它微调音色(见 comfy_assets.graph) - instruct_text=timbre.instruct_text_of(owner) if timbre.instruct_model(spec) else "", - idempotency_key=f"platform-voice:{owner.id}:{asset_id}", - label=f"PLATFORM-VOICE-{owner.id}", output_path=out, - ) - try: - result = model_registry.generate_voice(db, request, spec.provider, spec.model) - except ProviderError as exc: - raise HTTPException(502, f"{exc.message}({exc.code})") from exc - kind = "voice_sample" - owner.provider = spec.provider - if reference["path"] and timbre.clone_model(spec): - # 把真正参与克隆的那段参考音频绑回声线:否则试听样本是 A 的嗓子, - # 之后项目配音又落回模型默认参考,变成另一种嗓子。 - owner.provider_voice_id = str(storage.to_relative(reference["path"])) - else: - spec, _adapter = model_registry.resolve_asset(db, "image", data.provider or None, data.model or None) - if not spec.enabled: - raise HTTPException(409, f"模型 {spec.provider}/{spec.model} 已停用") - references = _reference_media(db, owner_type, owner.id, data.reference_media_ids, kind="image") - prompt = _owner_prompt(None, owner_type, owner, data.kind, data.prompt) - out = storage.platform_dir(owner_type, owner.id, "generated") / f"{data.kind}.png" - request = AssetGenRequest( - kind=data.kind, prompt=prompt, negative_prompt=data.negative_prompt, - width=data.width, height=data.height, seed=data.seed, - reference_images=[str(storage.to_absolute(row.path)) for row in references], - params={"model": spec.model}, - idempotency_key=( - f"platform-image:{owner.id}:{data.kind}:" - f"{now_utc().isoformat()}:{data.prompt}:{','.join(map(str, data.reference_media_ids))}" - ), - label=f"PLATFORM-{owner_type.upper()}-{owner.id}", output_path=out, - ) - try: - result = model_registry.generate_asset(db, request, "image", spec.provider, spec.model) - except ProviderError as exc: - raise HTTPException(502, f"{exc.message}({exc.code})") from exc - kind = data.kind - produced = Path(result.file_path) - if not produced.exists() or produced.stat().st_size < 128: - raise HTTPException(502, "模型未返回有效文件") - info = storage.save_platform_local_file( - produced, owner_type=owner_type, owner_id=owner.id, kind=kind, - stem=f"{owner_type}-{owner.id}-{kind}-ai", - ) - if getattr(result, "duration_ms", 0): - info["duration_ms"] = int(result.duration_ms) - media = storage.register_platform_media( - db, owner_type=owner_type, owner_id=owner.id, kind=kind, path=info["path"], - name=f"{owner.name} · AI {AssetKind.LABELS.get(kind, kind)}", - note=f"平台资产生成:{spec.provider}/{spec.model}", source="provider", info=info, - metadata={"generated": True, "provider": spec.provider, "model": spec.model, - "prompt": getattr(request, "prompt", None) or getattr(request, "text", ""), - "reference_media_ids": ([reference_media.id] if owner_type == "voice" and reference_media - else [row.id for row in references]), - "task_id": None}, - ) - if owner_type == "voice": - # 音色自检:量一遍实际基频,与设定性别比对,结果进素材元数据供界面标注。 - check = timbre.check_voice(produced, owner.gender) - produced_media_metadata(media, reference, check, text) - db.commit() - events.log_event(db, "platform.asset.generated", - message=f"平台{_type_label(owner_type)}「{owner.name}」已生成" - f"{AssetKind.LABELS.get(kind, kind)}({spec.provider}/{spec.model})", - entity_type=f"platform_{owner_type}", entity_id=owner.id) - return {"ok": True, "result": storage.platform_media_dict(media)} + spec = platform_generation.preflight(db, owner_type, owner, data) + kind = "voice_sample" if owner_type == "voice" else data.kind + task = platform_generation.enqueue(db, owner_type=owner_type, owner=owner, spec=spec, + data=data, asset_kind=kind) + return {"ok": True, "result": status_service.task_dict(task)} @router.post("/api/platform/{owner_type}/{asset_id}/text-assist") @@ -1107,7 +946,3 @@ def platform_text_assist(owner_type: str, asset_id: int, data: schemas.PlatformT raise HTTPException(502, "大模型未返回有效内容") return {"ok": True, "result": {"text": result.text.strip(), "provider": spec.provider, "model": spec.model}} - - -def _type_label(owner_type: str) -> str: - return {"human": "人物", "scene": "场景", "voice": "声线"}.get(owner_type, owner_type) diff --git a/backend/app/constants.py b/backend/app/constants.py index 748e397..3dada26 100644 --- a/backend/app/constants.py +++ b/backend/app/constants.py @@ -24,8 +24,12 @@ class TaskKind: COMPOSE = "compose" # 按时间线合成 + 混音 + 导出 IMPORT = "import" # 分镜/脚本导入 ASSET = "asset" # 资产衍生(占位,供扩展) + # 平台资产库的 AI 生成:出定妆图 / 场景图 / 试听样本。 + # 它与 ASSET 的区别只在「不属于任何项目」——正因如此才必须单独一种 kind, + # 否则按 kind 过滤的队列视图会把平台侧的生成混进某个项目的资产任务里。 + PLATFORM_ASSET = "platform_asset" - ALL = [GENERATE_SHOT, VOICEOVER, QC_SHOT, COMPOSE, IMPORT, ASSET] + ALL = [GENERATE_SHOT, VOICEOVER, QC_SHOT, COMPOSE, IMPORT, ASSET, PLATFORM_ASSET] class TaskStatus: diff --git a/backend/app/db.py b/backend/app/db.py index 46c0db2..fa7d70b 100644 --- a/backend/app/db.py +++ b/backend/app/db.py @@ -1,6 +1,7 @@ """数据库会话与基类。支持 PostgreSQL(生产)与 SQLite(本地自检)。""" from __future__ import annotations +import logging import os import threading import time @@ -15,6 +16,8 @@ from sqlalchemy.pool import NullPool, StaticPool from app.config import settings +log = logging.getLogger("nex.db") + class Base(DeclarativeBase): pass @@ -241,13 +244,18 @@ def init_db() -> None: 创建新表,再对比 ORM 元数据与真实列;只补缺失列,不删改已有数据。 新补列一律按 nullable 创建,便于旧数据行平滑升级。 - 建表/补列/回填三步都走 ``wait_for_lock``:抢不到文件锁时等上一个进程退出, + 建表之后的补列/放开非空/回填三步都走 ``wait_for_lock``:抢不到文件锁时等上一个进程退出, 而不是让整个后端起不来(起不来时端口还占着,表现是「全站无响应」)。 + + ``_relax_nullability`` 是这里唯一会**改已有列**的一步,因为只有它避不开: + 平台资产接入统一任务表之后 ``tasks.project_id`` 必须可空, + 而 SQLite 的 ALTER TABLE 改不了列约束——只能整表重建(数据逐列搬过去)。 """ from app import models # noqa: F401 确保模型全部注册 wait_for_lock("建表 create_all", lambda: Base.metadata.create_all(bind=engine)) _maintenance(_add_missing_columns, "补齐新增列") + _maintenance(_relax_nullability, "放开新增可空列") _maintenance(_backfill_defaults, "回填新增列默认值") @@ -269,6 +277,108 @@ def _add_missing_columns() -> None: conn.execute(text(f'ALTER TABLE "{table.name}" ADD COLUMN {ddl}')) +def _relax_nullability() -> None: + """把「ORM 已放开、老库里却还是 NOT NULL」的列改成可空。 + + 为什么只此一列要命:``tasks.project_id`` 建库时是必填,因为当时所有异步工作 + 都挂在某个项目下;平台资产库的 AI 生成接入同一张任务表之后,它没有项目可挂 + (平台母版本就不属于任何项目)。硬塞 0/-1 会让每一条按项目统计的 SQL + (容量、成本、流程进度)多背一个特例,所以老老实实放开约束。 + + SQLite 没有 ``ALTER COLUMN ... DROP NOT NULL``,只能整表重建: + 按 ORM 元数据建新表 → 逐列搬数据(只搬两边都有的列)→ 删旧表 → 改名 → 重建索引。 + 搬完再核对行数,对不上就抛错:宁可后端起不来,也不能悄悄少几十条任务履历。 + """ + inspector = inspect(engine) + existing_tables = set(inspector.get_table_names()) + dialect = engine.dialect.name + for table in Base.metadata.sorted_tables: + if table.name not in existing_tables: + continue + db_columns = {column["name"]: column for column in inspector.get_columns(table.name)} + stale = [column.name for column in table.columns + if column.nullable and column.name in db_columns + and not db_columns[column.name]["nullable"] and not column.primary_key] + if not stale: + continue + if dialect == "sqlite": + with engine.begin() as conn: + before = conn.execute(text(f'SELECT COUNT(*) FROM "{table.name}"')).scalar() or 0 + _rebuild_sqlite_table(conn, inspector, table, set(db_columns)) + after = conn.execute(text(f'SELECT COUNT(*) FROM "{table.name}"')).scalar() or 0 + if before != after: + raise RuntimeError(f"重建 {table.name} 后行数不一致({before} → {after}),已中止启动") + log.info("已放开 %s.%s 的 NOT NULL(整表重建,%d 行数据已搬运)", + table.name, "、".join(stale), after) + elif dialect in ("postgresql", "mysql", "mariadb"): + with engine.begin() as conn: + for name in stale: + conn.execute(text(f'ALTER TABLE "{table.name}" ALTER COLUMN "{name}" DROP NOT NULL')) + log.info("已放开 %s.%s 的 NOT NULL", table.name, "、".join(stale)) + else: + log.warning("方言 %s 未实现放开非空,%s.%s 仍是 NOT NULL", + dialect, table.name, "、".join(stale)) + + +def _literal(column) -> str: + """按列类型给一个「空值当没有」的兜底字面量(与 Python 侧 default 同语义)。""" + from sqlalchemy import Boolean, Date, DateTime, Float, Integer, Interval, Numeric, Time + + if isinstance(column.type, Interval): + return "'00:00:00'" + if isinstance(column.type, (DateTime, Date, Time)): + return "'1970-01-01 00:00:00'" + if isinstance(column.type, Boolean): + return "0" + if isinstance(column.type, (Integer, Float, Numeric)): + return "0" + if column.type.__class__.__name__ in ("JSON",): + return "'{}'" + return "''" + + +def _rebuild_sqlite_table(conn, inspector, table, db_columns: set[str]) -> None: + """按 ORM 元数据重建一张 SQLite 表(列定义以模型为准,数据逐列搬运)。""" + from sqlalchemy.schema import CreateIndex, CreateTable + + temp = f"_nex_migrate_{table.name}" + # 镜像表借 Base.metadata 建:新建的 MetaData 里没有 projects/shots, + # 编译外键时会报 NoReferencedTableError。用完即摘,别把临时表留在元数据里。 + mirror = table.to_metadata(Base.metadata, name=temp) + # 只搬两边都有的列:库里多出来的历史列(早已从模型里删掉)不值得再留一份空壳 + db_null = {column["name"] for column in inspector.get_columns(table.name) if column["nullable"]} + shared = [column.name for column in table.columns if column.name in db_columns] + quote = lambda name: '"' + name.replace('"', '""') + '"' + + def source(column) -> str: + """搬运表达式:刚补进来的列在旧行里全是 NULL,而新表按模型是 NOT NULL。 + + ``_add_missing_columns`` 只能把新列加成可空(SQLite 的限制), + 所以这里必须给个兜底值,否则「补列 → 重建」这两步自己打架: + 一条 NOT NULL 的老库记录搬进新表就会当场抛 IntegrityError。 + """ + name = quote(column.name) + if column.nullable or column.name not in db_null: + return name + return f"COALESCE({name}, {_literal(column)})" + + try: + conn.execute(text(str(CreateTable(mirror).compile(dialect=engine.dialect)))) + finally: + Base.metadata.remove(mirror) + columns = ", ".join(quote(column.name) for column in table.columns + if column.name in db_columns) + values = ", ".join(source(column) for column in table.columns if column.name in db_columns) + conn.execute(text(f'INSERT INTO {quote(temp)} ({columns}) ' + f'SELECT {values} FROM {quote(table.name)}')) + conn.execute(text(f'DROP TABLE {quote(table.name)}')) + # 改名放在建索引之前:索引挂在表对象上,DDL 里写的是正式表名, + # 先建就会指向还不存在的表。 + conn.execute(text(f'ALTER TABLE {quote(temp)} RENAME TO {quote(table.name)}')) + for index in table.indexes: + conn.execute(text(str(CreateIndex(index).compile(dialect=engine.dialect)))) + + def _backfill_defaults() -> None: """只回填代码依赖的新增标量列;JSON 列由读取侧继续兼容 null。""" statements = [ @@ -278,6 +388,9 @@ def _backfill_defaults() -> None: "UPDATE characters SET platform_source_type='' WHERE platform_source_type IS NULL", "UPDATE scenes SET platform_source_type='' WHERE platform_source_type IS NULL", "UPDATE voice_assets SET platform_source_type='' WHERE platform_source_type IS NULL", + # 统一任务表新列:旧行没有归属与状态说明,补齐空串,读取侧不必再判 null + "UPDATE tasks SET owner_type='' WHERE owner_type IS NULL", + "UPDATE tasks SET message='' WHERE message IS NULL", ] with engine.begin() as conn: for sql in statements: diff --git a/backend/app/jobs/handlers.py b/backend/app/jobs/handlers.py index 5645887..4981b38 100644 --- a/backend/app/jobs/handlers.py +++ b/backend/app/jobs/handlers.py @@ -25,6 +25,7 @@ from app.models import Asset, Character, Project, Scene, Shot, Task, Timeline, V from app.providers import registry as model_registry from app.providers.base import ProviderError from app.services import audio, compose, costing, dispatch, events, ffmpeg as ff +from app.services import progress from app.services import prompt as prompt_service from app.services import quality, status as status_service, storage, timbre @@ -169,8 +170,10 @@ def handle_generate_shot(task_id: int) -> dict: 进度条,也不该把它倒回去——倒回去看起来像任务被打回起点,用户会以为卡住了。 """ value = float(ratio) - progress = None if value < 0 else 0.1 + 0.6 * max(0.0, min(1.0, value)) - dispatch.heartbeat(db, task, progress=progress, note=note) + # 别名叫 _pct:模块里另有一个 services.progress(run 里绑任务上下文), + # 局部变量压住同名导入,将来在闭包里调 progress.report() 会撞成 AttributeError + _pct = None if value < 0 else 0.1 + 0.6 * max(0.0, min(1.0, value)) + dispatch.heartbeat(db, task, progress=_pct, note=note) db.commit() started = time.monotonic() @@ -692,6 +695,38 @@ def handle_asset(task_id: int) -> dict: db.close() +def handle_platform_asset(task_id: int) -> dict: + """平台资产库的 AI 生成(定妆图 / 场景图 / 试听样本)。 + + 执行体与项目侧任务完全同构:认领 → 跑模型 → 落素材 → 回写任务行。 + 真正的业务在 services/platform_generation(那里同时被接口层复用做参数校验), + 这一层只负责把异常收口成「重试或失败」,绝不向上抛给执行后端。 + """ + from app.services import platform_generation + + db, task = _load(task_id) + if db is None or task is None: + return {"skipped": "任务不存在"} + if not _begin(db, task): + return {"skipped": "任务已处于终态"} + try: + result = platform_generation.generate(db, task) + if _was_cancelled(db, task.id): + # 出图已经落盘,但用户明确取消过:任务行保持「已取消」,别再写成成功 + return {"cancelled": True, "media": result.get("media")} + media = result.get("media") or {} + dispatch.mark_succeeded(db, task, result=result, output_path=media.get("path") or None) + # 落定之后把「此刻在做什么」换成结果一句话:执行体最后写的那条进度是 + # 「算力已就绪:模型出图中」,留在已完成那一行上就成了自相矛盾的回显。 + task.message = platform_generation.landed_message(media) + db.commit() + return result + except Exception as exc: + return _finish_failure(db, task, exc, shot=None) + finally: + db.close() + + HANDLERS = { TaskKind.GENERATE_SHOT: handle_generate_shot, TaskKind.VOICEOVER: handle_voiceover, @@ -699,6 +734,7 @@ HANDLERS = { TaskKind.COMPOSE: handle_compose, TaskKind.IMPORT: handle_import, TaskKind.ASSET: handle_asset, + TaskKind.PLATFORM_ASSET: handle_platform_asset, } @@ -739,6 +775,10 @@ def run(kind: str, task_id: int) -> dict: _reconcile_project(task_id) return {"failed": True} try: + # progress.bind 让埋在适配器深处的代码(providers/comfy_common.gpu_slot) + # 能把「排在别人后面」和「模型正在出图」写回任务行——没有这一句, + # 动辄几分钟的任务在界面上就只有一个转圈的图标。 + progress.bind(task_id) with _Heartbeat(task_id): return handler(task_id) except Exception as exc: # handler 内部已收口,这里只做最后兜底 @@ -749,4 +789,5 @@ def run(kind: str, task_id: int) -> dict: db.close() return {"failed": True, "error": str(exc)} finally: + progress.unbind() _reconcile_project(task_id) # 队列动过,项目阶段就得跟上 diff --git a/backend/app/jobs/tasks.py b/backend/app/jobs/tasks.py index cc6d097..9e46b62 100644 --- a/backend/app/jobs/tasks.py +++ b/backend/app/jobs/tasks.py @@ -51,6 +51,13 @@ def asset(self, task_id: int) -> dict: return handlers.run("asset", task_id) +@celery_app.task(name="app.jobs.platform_asset", bind=True, acks_late=True) +def platform_asset(self, task_id: int) -> dict: + from app.jobs import handlers + + return handlers.run("platform_asset", task_id) + + @celery_app.task(name="app.jobs.dispatch_due") def dispatch_due() -> dict: """周期调度器:扫描到期任务并按模型限流额度派发。""" diff --git a/backend/app/models/pipeline.py b/backend/app/models/pipeline.py index 7cc7079..29adf5b 100644 --- a/backend/app/models/pipeline.py +++ b/backend/app/models/pipeline.py @@ -12,21 +12,38 @@ from app.db import Base class Task(TimestampMixin, Base): - """统一任务表:所有异步工作(生成/配音/质检/合成/导入)都走同一编排。""" + """统一任务表:**凡是「要跑模型、要等一会儿」的活都在这一张表里**。 + + 项目侧的生成/配音/质检/合成/导入走它,平台资产库里「点一下 AI 生成定妆图」 + 也走它(``kind=platform_asset``、``project_id`` 为空、``owner_type/owner_id`` + 指向平台母版)。这样「平台此刻在忙什么」只需要问一处, + 顶栏的活动面板、限流派发、失败重试、断点续跑、成本台账全都自动对平台任务生效, + 不必为平台侧另立一套并行体系(那套体系当年就是为了让用户看得见进度才建的)。 + """ __tablename__ = "tasks" __table_args__ = (Index("ix_tasks_status_priority", "status", "priority", "id"), - Index("ix_tasks_provider_status", "provider", "status")) + Index("ix_tasks_provider_status", "provider", "status"), + Index("ix_tasks_owner_status", "owner_type", "owner_id", "status")) id: Mapped[int] = mapped_column(primary_key=True) - project_id: Mapped[int] = mapped_column(ForeignKey("projects.id", ondelete="CASCADE"), index=True) + # 可空:平台资产母版不属于任何项目(老库里这一列是 NOT NULL, + # 由 db._relax_nullability 在启动时改成可空,不需要删库重建) + project_id: Mapped[int | None] = mapped_column( + ForeignKey("projects.id", ondelete="CASCADE"), nullable=True, index=True) shot_id: Mapped[int | None] = mapped_column(ForeignKey("shots.id", ondelete="CASCADE"), nullable=True, index=True) timeline_id: Mapped[int | None] = mapped_column(Integer, nullable=True, index=True) + # 非项目归属(平台母版):卡片要按「这条资产正在生成」亮状态,重复点击要按它去重 + owner_type: Mapped[str] = mapped_column(String(20), default="", index=True) + owner_id: Mapped[int | None] = mapped_column(Integer, nullable=True) kind: Mapped[str] = mapped_column(String(30), default=TaskKind.GENERATE_SHOT, index=True) name: Mapped[str] = mapped_column(String(200), default="") stage: Mapped[str] = mapped_column(String(40), default="dispatch") status: Mapped[str] = mapped_column(String(30), default=TaskStatus.QUEUED, index=True) + # 此刻在做什么的一句话说明(「排队中:等待算力空闲」「模型出图中」)。 + # 只有进度条时用户分不清「模型在算」和「排在别人后面」,只能一遍遍刷新页面。 + message: Mapped[str] = mapped_column(String(400), default="") priority: Mapped[int] = mapped_column(Integer, default=5) progress: Mapped[float] = mapped_column(Float, default=0.0) diff --git a/backend/app/providers/comfy_common.py b/backend/app/providers/comfy_common.py index 1df18c3..fc481d0 100644 --- a/backend/app/providers/comfy_common.py +++ b/backend/app/providers/comfy_common.py @@ -68,13 +68,33 @@ def lock_keys(config: dict | None, base_url: str) -> list[str]: return keys +def slot_note(message: str) -> None: + """把算力锁上的等待写成任务行上的一行状态;不在任务上下文里时静默返回。 + + 延迟导入:本模块被各适配器**顶层**导入,而 providers 不该在导入期依赖 services + (services 反过来要读注册表)。这一句拿不到也不影响出图,吞掉即可。 + """ + try: + from app.services import progress + + progress.report(message) + except Exception: # noqa: BLE001 - 状态回写不该弄挂生成 + pass + + @contextmanager def gpu_slot(config: dict, base_url: str) -> Iterator[None]: # 锁键先排序再逐个获取,保证全局一致的加锁顺序,避免死锁。 locks = [resource_lock(key) for key in sorted(lock_keys(config, base_url))] with ExitStack() as stack: for lock in locks: + # 排队这件事必须说给用户听:一次出图动辄十几分钟,而 video / image / tts + # 打的是同一台 ComfyUI,在这把锁上无声串行(见 lock_keys)。以前界面上 + # 只有「进行中」三个字在转,用户分不清「模型正在算」还是「排在别人后面」, + # 只能一遍遍刷新页面。 + slot_note("排队中:等待算力空闲") stack.enter_context(lock) + slot_note("算力已就绪:模型出图中") yield diff --git a/backend/app/services/capacity.py b/backend/app/services/capacity.py index 8573f55..8a1e45f 100644 --- a/backend/app/services/capacity.py +++ b/backend/app/services/capacity.py @@ -19,7 +19,7 @@ from app.models import ModelSpec, Task # 找不到模型规格时,用任务类型推断媒体类型。资产任务默认按图像处理; # 若使用 TTS 生成声线样本,任务行仍会带上已解析的具体模型,因此会优先命中规格表。 KIND_MEDIA_TYPE = {TaskKind.GENERATE_SHOT: "video", TaskKind.VOICEOVER: "tts", - TaskKind.ASSET: "image"} + TaskKind.ASSET: "image", TaskKind.PLATFORM_ASSET: "image"} @dataclass diff --git a/backend/app/services/dispatch.py b/backend/app/services/dispatch.py index 9aa8cb3..4425ca2 100644 --- a/backend/app/services/dispatch.py +++ b/backend/app/services/dispatch.py @@ -343,52 +343,104 @@ def dispatch_due(db: Session, *, limit: int | None = None, project_id: int | Non blocked[key] = reason if slots[key] <= 0: continue - # 先「原子认领」再提交执行:条件更新只有一个会话能成功, - # 因此守护线程、drain、人工 dispatch 并发时同一任务绝不会被跑两次。 - claimed = db.execute(update(Task) - .where(Task.id == task.id, - Task.status.in_([TaskStatus.QUEUED, TaskStatus.RETRYING])) - # worker_id 先清空:认领到执行体真正起跑之间有个短窗口, - # 期间它还属于「没有 runner」而不是「上一代进程的僵尸」, - # 否则断点续跑可能把刚认领的任务再捞回去跑两遍。 - .values(status=TaskStatus.RUNNING, started_at=now, heartbeat_at=now, - attempts=Task.attempts + 1, next_retry_at=None, error_message="", - worker_id="", execution_backend=execution.backend_name()) - .execution_options(synchronize_session=False)).rowcount - db.commit() - if not claimed: + if not claim_task(db, task, now=now): continue # 已被其它派发者认领,跳过 - try: - handle = execution.submit(task) - except Exception as exc: # 后端不可用 → 归还队列并稍后重试 - log.exception("任务派发失败 task=%s", task.id) - task = db.get(Task, task.id) - task.status = TaskStatus.QUEUED - task.attempts = max(0, (task.attempts or 1) - 1) - task.started_at = None - task.heartbeat_at = None - task.error_message = f"任务派发失败:{exc}"[:4000] - task.next_retry_at = now + timedelta(seconds=30) - events.log_event(db, "task.dispatch_failed", project_id=task.project_id, task_id=task.id, - level="error", message=task.error_message, entity_type="task", entity_id=task.id) - db.commit() - continue + handle = submit_claimed(db, task, now=now) + if not handle: + continue # 派发失败,已归还队列 slots[key] -= 1 - task = db.get(Task, task.id) - task.celery_task_id = handle - task.execution_backend = execution.backend_name() - task.queued_at = task.queued_at or now + mark_dispatched(db, task, handle, now=now) dispatched.append({"task_id": task.id, "kind": task.kind, "shot_id": task.shot_id, "provider": task.provider, "handle": handle}) - events.log_event(db, "task.dispatched", project_id=task.project_id, task_id=task.id, - shot_id=task.shot_id, message=f"任务已派发到 {task.execution_backend} 执行后端", - entity_type="task", entity_id=task.id, - data={"provider": task.provider, "model": task.model, "handle": handle}) db.commit() return {"dispatched": dispatched, "count": len(dispatched), "blocked": [ {"provider_model": k, "reason": v} for k, v in blocked.items()]} +def claim_task(db: Session, task: Task, *, now: datetime | None = None) -> bool: + """原子认领一条排队中的任务:条件更新只有一个会话能成功。 + + 守护线程、drain、人工 dispatch、平台侧「点一下生成」四条路都会走到这里, + 因此同一任务绝不会被跑两次——这是「不重复烧 GPU」的唯一保证,别换成先读后写。 + """ + now = now or now_utc() + claimed = db.execute(update(Task) + .where(Task.id == task.id, + Task.status.in_([TaskStatus.QUEUED, TaskStatus.RETRYING])) + # worker_id 先清空:认领到执行体真正起跑之间有个短窗口, + # 期间它还属于「没有 runner」而不是「上一代进程的僵尸」, + # 否则断点续跑可能把刚认领的任务再捞回去跑两遍。 + .values(status=TaskStatus.RUNNING, started_at=now, heartbeat_at=now, + attempts=Task.attempts + 1, next_retry_at=None, error_message="", + worker_id="", execution_backend=execution.backend_name()) + .execution_options(synchronize_session=False)).rowcount + db.commit() + return bool(claimed) + + +def submit_claimed(db: Session, task: Task, *, now: datetime | None = None) -> str: + """把已认领的任务交给执行后端,返回句柄;派发失败时归还队列并稍后重试,返回空串。""" + now = now or now_utc() + try: + return execution.submit(task) + except Exception as exc: # 后端不可用 → 归还队列并稍后重试 + log.exception("任务派发失败 task=%s", task.id) + task = db.get(Task, task.id) + task.status = TaskStatus.QUEUED + task.attempts = max(0, (task.attempts or 1) - 1) + task.started_at = None + task.heartbeat_at = None + task.error_message = f"任务派发失败:{exc}"[:4000] + task.next_retry_at = now + timedelta(seconds=30) + events.log_event(db, "task.dispatch_failed", project_id=task.project_id, task_id=task.id, + level="error", message=task.error_message, entity_type="task", entity_id=task.id) + db.commit() + return "" + + +def mark_dispatched(db: Session, task: Task, handle: str, *, now: datetime | None = None) -> Task: + """记下执行句柄并留一条事件(不提交:批量派发时由调用方统一收口)。""" + now = now or now_utc() + task = db.get(Task, task.id) + task.celery_task_id = handle + task.execution_backend = execution.backend_name() + task.queued_at = task.queued_at or now + events.log_event(db, "task.dispatched", project_id=task.project_id, task_id=task.id, + shot_id=task.shot_id, message=f"任务已派发到 {task.execution_backend} 执行后端", + entity_type="task", entity_id=task.id, + data={"provider": task.provider, "model": task.model, "handle": handle}) + return task + + +def dispatch_task(db: Session, task: Task) -> dict: + """立刻派发**这一条**任务:有额度当场起跑,没额度就留在队列里等守护线程。 + + 平台资产生成走这里。用户点完按钮不该干等 dispatch_interval_seconds(默认 5s) + 才知道排上了没有;但也不能绕过限流插队——出图、出片、配音打的是同一台 + ComfyUI,插队的代价由正在跑的项目买单。所以这里查同样的 headroom、 + 走同样的原子认领,只是不排队等下一轮扫描。 + 返回值只用于日志与接口提示,真正的进度一律以任务行为准。 + """ + now = now_utc() + available, reason = capacity.headroom(db, task.provider or "", task.model or None) + if available <= 0: + return {"dispatched": False, "queued": True, "task_id": task.id, + "status": task.status, "reason": reason or "算力已满,已进入队列"} + if not claim_task(db, task, now=now): + fresh = db.get(Task, task.id) + return {"dispatched": False, "queued": True, "task_id": task.id, + "status": fresh.status, "reason": "任务已被其它派发者认领"} + handle = submit_claimed(db, task, now=now) + if not handle: + fresh = db.get(Task, task.id) + return {"dispatched": False, "queued": True, "task_id": task.id, + "status": fresh.status, "reason": fresh.error_message or "任务派发失败"} + fresh = mark_dispatched(db, task, handle, now=now) + db.commit() + return {"dispatched": True, "queued": True, "task_id": fresh.id, + "status": fresh.status, "handle": handle, "reason": ""} + + # ---------------------------------------------------------------- 状态回写 def _write_through(db: Session, *, commit: bool = False, attempts: int = 4) -> None: """把状态回写落库,遇到 SQLite 抢锁先自己退避重试。 diff --git a/backend/app/services/execution.py b/backend/app/services/execution.py index 75e0e2d..cfe6040 100644 --- a/backend/app/services/execution.py +++ b/backend/app/services/execution.py @@ -23,10 +23,12 @@ CELERY_TASK_NAMES = { "compose": "app.jobs.compose", "import": "app.jobs.import_package", "asset": "app.jobs.asset", + "platform_asset": "app.jobs.platform_asset", } +# 平台资产生成(出图/出试听)与导入同属 io:都不占大 GPU 时长,但都会被算力锁挡住 QUEUES = {"generate_shot": "generate", "voiceover": "audio", "qc_shot": "qc", "compose": "compose", - "import": "io", "asset": "io"} + "import": "io", "asset": "io", "platform_asset": "io"} def celery_app(): diff --git a/backend/app/services/platform_generation.py b/backend/app/services/platform_generation.py new file mode 100644 index 0000000..8675268 --- /dev/null +++ b/backend/app/services/platform_generation.py @@ -0,0 +1,405 @@ +r"""平台资产库的 AI 生成:登记成任务 → 统一调度执行 → 进度写回任务行。 + +为什么要走任务表(2026-11 · 504 那一串): + 出图实测 46~49s(docs/MiniMaxH3_Spark_性能诊断与优化方案.md:681),反代 + ``proxy_read_timeout`` 出厂 60s,余量只有十几秒;同卡一排队就直接 504。 + 而 504 只是**看起来**失败了——出图线程并没有被取消,图照样落盘, + 用户看不见结果,只能手动刷新页面去猜它到底成没成。 + 现在接口只做「校验 + 登记 + 派发」,当场 202 返回任务号, + 进度、产物、失败原因全都在 ``tasks`` 那一行里,换设备、刷新页面都接着看得到。 + +为什么并入 ``tasks`` 而不是另立一张操作表: + 「所有要跑模型、要等一会儿的活」本该只有一个地方能问。并表之后 + 限流派发(capacity.headroom)、失败重试、断点续跑(resume_interrupted)、 + 心跳与 stale、成本与事件台账,平台侧生成全部自动继承, + 不必再维护第二套进行中转圈。代价只是 ``project_id`` 变成可空, + 由 ``db._relax_nullability`` 在启动时原地放开。 + +执行体在 ``jobs/handlers.handle_platform_asset`` 里,与项目侧任务同一条通道; +深处的算力排队说明由 ``services/progress`` 写回(见 providers/comfy_common.gpu_slot)。 +""" +from __future__ import annotations + +import logging +from pathlib import Path + +from fastapi import HTTPException +from sqlalchemy import select +from sqlalchemy.orm import Session + +from app import schemas +from app.constants import AssetKind, TaskKind, TaskStatus +from app.models import (PlatformCharacter, PlatformMedia, PlatformScene, PlatformVoice, + Project, Task, now_utc) +from app.providers import registry as model_registry +from app.providers.base import AssetGenRequest, ProviderError, VoiceGenRequest +from app.services import costing, dispatch, events, progress, storage, timbre + +log = logging.getLogger("nex.platform_generation") + +#: 一条平台生成任务最多能被写成多少次状态说明(message 列宽 400) +MAX_LABEL_CHARS = 200 + + +def landed_message(media: dict) -> str: + """任务成功落定时要写进 ``tasks.message`` 的那句话。 + + ``progress.report`` 写的是「此刻在做什么」,成功之后没人覆盖它, + 于是顶栏「平台动态 · 最近落定」会出现「已完成 + 算力已就绪:模型出图中」 + 这种自相矛盾的回显(状态胶囊说完了,正文还在说进行中)。 + """ + bits = [f"已生成{media.get('kind_label') or '新素材'}"] + if media.get("version"): + bits.append(f"第 {media['version']} 版") + if media.get("duration_ms"): + bits.append(f"{round(media['duration_ms'] / 1000, 1)}s") + elif media.get("size_bytes"): + bits.append(f"{round(media['size_bytes'] / 1024 / 1024, 2)} MB") + return " · ".join(bits)[:MAX_LABEL_CHARS] + + +# ---------------------------------------------------------------- 母版与素材的公共读法 +def normalize_type(owner_type: str) -> str: + return { + "characters": "human", "character": "human", "human": "human", "humans": "human", + "scene": "scene", "scenes": "scene", "voice": "voice", "voices": "voice", + }.get(owner_type, owner_type) + + +def get_owner(db: Session, owner_type: str, owner_id: int): + owner_type = normalize_type(owner_type) + if owner_type == "human": + row = db.get(PlatformCharacter, owner_id) + elif owner_type == "scene": + row = db.get(PlatformScene, owner_id) + elif owner_type == "voice": + row = db.get(PlatformVoice, owner_id) + else: + raise HTTPException(400, f"不支持的平台资产类型:{owner_type}") + if row is None: + raise HTTPException(404, "平台资产不存在") + return row + + +def type_label(owner_type: str) -> str: + return {"human": "人物", "scene": "场景", "voice": "声线"}.get(owner_type, owner_type) + + +def resolve_reference_media(db: Session, owner_type: str, owner_id: int, ids: list[int], + *, kind: str) -> list[PlatformMedia]: + """校验平台生成参考素材只能来自当前资产,避免跨资产误读本地路径。""" + rows: list[PlatformMedia] = [] + seen: set[int] = set() + image_suffixes = {".png", ".jpg", ".jpeg", ".webp", ".bmp", ".tif", ".tiff"} + audio_suffixes = {".wav", ".mp3", ".m4a", ".aac", ".flac", ".ogg", ".opus"} + for media_id in ids: + if media_id in seen: + continue + seen.add(media_id) + row = db.get(PlatformMedia, media_id) + if (row is None or row.owner_type != owner_type or row.owner_id != owner_id): + raise HTTPException(400, "参考素材不属于当前平台资产") + if not storage.exists(row.path): + raise HTTPException(400, "参考素材文件不存在,请重新上传") + suffix = Path(row.path).suffix.lower() + if kind == "image" and suffix not in image_suffixes: + raise HTTPException(400, "图像参考素材必须是图片文件") + if kind == "audio" and suffix not in audio_suffixes: + raise HTTPException(400, "音色参考素材必须是音频文件") + rows.append(row) + return rows + + +def build_owner_prompt(project: Project | None, owner_type: str, owner, kind: str, custom: str) -> str: + if custom.strip(): + return custom.strip() + bits = [] + if owner_type == "human": + bits = [f"{owner.name},电影级角色设定图", owner.appearance, owner.persona_prompt, + "五官、发型、体型、服装与配饰保持一致,正面清晰,材质细节可信"] + elif owner_type == "scene": + bits = [f"{owner.name},电影级场景概念图", owner.prompt or owner.description, + owner.lighting, owner.time_of_day, owner.palette, + "空间结构清晰,透视准确,建筑与陈设保持一致"] + if project: + bits.append(project.style_guide) + return "。".join(str(part).strip() for part in bits if part) + + +def resolve_spec(db: Session, owner_type: str, data: schemas.PlatformGenerate): + """按资产类型解析这次生成实际会派给哪个模型(留空即跟随平台默认)。""" + if owner_type == "voice": + spec, _adapter = model_registry.resolve_voice(db, data.provider or None, data.model or None) + else: + spec, _adapter = model_registry.resolve_asset(db, "image", + data.provider or None, data.model or None) + return spec + + +def resolve_voice_reference(db: Session, owner, spec, + data: schemas.PlatformGenerate) -> tuple[PlatformMedia | None, dict]: + """把「这次生成会拿到谁的嗓子」算出来:参考音频决定音色,表单里的性别只是元数据。""" + reference_ids = list(data.reference_media_ids) + if data.reference_media_id: + reference_ids.append(data.reference_media_id) + references = resolve_reference_media(db, "voice", owner.id, reference_ids, kind="audio") + reference_media = references[0] if references else None + reference = timbre.resolve_reference( + db, owner, + explicit=str(storage.to_absolute(reference_media.path)) if reference_media else "", + clone=timbre.clone_model(spec), instruct=timbre.instruct_model(spec)) + return reference_media, reference + + +# ---------------------------------------------------------------- 同步校验 +def preflight(db: Session, owner_type: str, owner, data: schemas.PlatformGenerate): + """开工前能当场说清的错误,绝不留到五分钟以后。 + + 异步化之后接口立刻返回,如果连「没填试听文本」都要等轮询才发现, + 用户面对的是一个跑完了才知道白跑的进度条。所以这些**便宜且确定**的校验 + 留在请求里同步做(422/409 原样吐给表单),耗时部分才交给后台。 + 后台执行时会把同样的校验再跑一遍:排到 GPU 时模型可能已被停用、 + 参考音频可能已被删掉,以及时报错的是供应商而不是表单。 + """ + spec = resolve_spec(db, owner_type, data) + if not spec.enabled: + raise HTTPException(409, f"模型 {spec.provider}/{spec.model} 已停用") + if owner_type != "voice": + return spec + if not (data.sample_text or "").strip(): + raise HTTPException(422, "请先填写试听文本:它是模型要朗读的台词," + "系统不会用「音色描述」代替") + # 嗓子来源必须当场说清:克隆模型缺参考音频时会静默回退到模型默认参考(女声), + # 男声声线就这样被生成成了女声——等四十秒后试听才发现,等于白跑一轮。 + _media, reference = resolve_voice_reference(db, owner, spec, data) + if not reference["path"] and reference.get("source") == "none": + raise HTTPException(409, reference["note"]) + return spec + + +# ---------------------------------------------------------------- 登记与执行 +def active_task(db: Session, owner_type: str, owner_id: int, asset_kind: str) -> Task | None: + """这条母版的这一类素材是否已经有生成任务在跑(挡连点:三次点击只该出一张图)。""" + rows = db.scalars(select(Task).where( + Task.owner_type == owner_type, Task.owner_id == owner_id, + Task.kind == TaskKind.PLATFORM_ASSET, Task.status.in_(TaskStatus.ACTIVE) + ).order_by(Task.id)).all() + for row in rows: + if str((row.payload_json or {}).get("asset_kind") or "") == asset_kind: + return row + return None + + +def enqueue(db: Session, *, owner_type: str, owner, spec, data: schemas.PlatformGenerate, + asset_kind: str) -> Task: + """登记一条平台生成任务并立刻尝试派发;返回任务行(202 的响应体就是它)。""" + running = active_task(db, owner_type, owner.id, asset_kind) + if running is not None: + raise HTTPException(409, f"「{owner.name}」的{AssetKind.LABELS.get(asset_kind, asset_kind)}" + f"正在生成中(任务 #{running.id}),完成前不能重复发起") + label = f"生成「{owner.name}」的{AssetKind.LABELS.get(asset_kind, asset_kind)}"[:MAX_LABEL_CHARS] + cost = (costing.estimate_voice_cost(spec, len((data.sample_text or "").strip())) + if owner_type == "voice" else 0.0) + task = Task( + project_id=None, shot_id=None, kind=TaskKind.PLATFORM_ASSET, + name=label, stage="platform_generate", status=TaskStatus.QUEUED, + priority=5, provider=spec.provider, model=spec.model, + mode="voice" if owner_type == "voice" else data.kind, + owner_type=owner_type, owner_id=owner.id, + # 同一母版同一素材类型的重复点击靠这条键挡,不给它配项目侧那套镜头幂等语义 + idempotency_key=f"platform:{owner_type}:{owner.id}:{asset_kind}:{int(now_utc().timestamp())}", + # 只留一次自动重试:连不上算力这类瞬时故障值得再试一把(生成失败没有产物, + # 不存在重复扣费);参数与配置错误在 generate 里标了 retryable=False,当场判死。 + max_attempts=2, payload_json=_payload(owner_type, owner, asset_kind, data), + cost=cost, queued_at=now_utc(), message="排队中:等待调度") + db.add(task) + db.flush() + events.log_event(db, "task.queued", project_id=None, task_id=task.id, + message=f"平台{type_label(owner_type)}「{owner.name}」{label}", + entity_type="task", entity_id=task.id, + data={"owner_type": owner_type, "owner_id": owner.id, "asset_kind": asset_kind, + "provider": spec.provider, "model": spec.model}) + db.commit() + result = dispatch.dispatch_task(db, task) + log.info("平台生成任务 #%s 已登记:%s(派发 %s)", task.id, label, + "立即开始" if result["dispatched"] else f"排队等待,{result['reason']}") + return db.get(Task, task.id) + + +def _payload(owner_type: str, owner, asset_kind: str, + data: schemas.PlatformGenerate) -> dict: + """任务参数:请求体原样收在 ``request`` 里,外层放调度与展示要用的字段。 + + ``owner_name`` 只是给顶栏「平台动态」的「查看」用的:那条深链按名字搜资产库, + 而 ``task.name`` 是「生成「俊杰」的定妆图」这样的一句话,拿它当搜索词永远 0 条。 + 存名字而不是回表查,是为了不给 ``task_dict`` 加一次查询——它全站每条任务都要过。 + """ + return {"operation": "generate", "owner_type": owner_type, "owner_id": owner.id, + "owner_name": getattr(owner, "name", ""), + "asset_kind": asset_kind, "request": data.model_dump()} + + +def request_from(task: Task) -> tuple[str, int, str, schemas.PlatformGenerate]: + """从任务行还原「当时点的是什么」(执行体只拿得到 task_id)。""" + payload = task.payload_json or {} + raw = dict(payload.get("request") or {}) + # 老任务行可能带着外层多写的字段,Strict 模型会当场拒绝——只留请求体认识的键 + fields = set(schemas.PlatformGenerate.model_fields) + data = schemas.PlatformGenerate(**{k: v for k, v in raw.items() if k in fields}) + return (str(payload.get("owner_type") or ""), int(payload.get("owner_id") or 0), + str(payload.get("asset_kind") or data.kind), data) + + +def produced_media_metadata(media, reference: dict, check: dict, text: str) -> None: + """把「这次到底用了谁的嗓子、念了什么、音色对不对」写进素材,界面才说得清楚。""" + metadata = dict(media.metadata_json or {}) + metadata.update({ + "text": text, + "reference_audio": reference.get("path") or reference.get("attached") or "", + "reference_source": reference.get("source") or "", + "reference_note": reference.get("note") or "使用本声线已绑定的参考音频", + "timbre_check": check, + }) + media.metadata_json = metadata + note = str(media.note or "") + tail = check.get("note") or "" + if tail and "音色自检" not in note: + media.note = f"{note}|音色自检:{tail}".strip("|") + + +def generate(db: Session, task: Task) -> dict: + """真正跑一次平台资产生成并落库;失败一律抛 ``ProviderError`` 交给任务层归类。""" + owner_type, owner_id, asset_kind, data = request_from(task) + owner = get_owner(db, owner_type, owner_id) + progress.report("准备生成参数") + if owner_type == "voice": + result = _generate_voice(db, owner, data, task) + else: + result = _generate_image(db, owner_type, owner, data, task) + db.commit() + events.log_event(db, "platform.asset.generated", project_id=None, task_id=task.id, + message=f"平台{type_label(owner_type)}「{owner.name}」已生成" + f"{AssetKind.LABELS.get(asset_kind, asset_kind)}" + f"({result['provider']}/{result['model']})", + entity_type=f"platform_{owner_type}", entity_id=owner.id, + data={"media_id": result["media"].get("id"), "task_id": task.id}) + db.commit() + return result + + +def _config_error(message: str, code: str) -> ProviderError: + """参数/配置类错误:重试一百次也是同一个结果,直接判死并把原因说清。""" + return ProviderError(message, code=code, retryable=False) + + +def _generate_image(db: Session, owner_type: str, owner, + data: schemas.PlatformGenerate, task: Task) -> dict: + spec = resolve_spec(db, owner_type, data) + if not spec.enabled: + raise _config_error(f"模型 {spec.provider}/{spec.model} 已停用", "model_disabled") + references = resolve_reference_media(db, owner_type, owner.id, data.reference_media_ids, + kind="image") + prompt = build_owner_prompt(None, owner_type, owner, data.kind, data.prompt) + out = storage.platform_dir(owner_type, owner.id, "generated") / f"{data.kind}.png" + request = AssetGenRequest( + kind=data.kind, prompt=prompt, negative_prompt=data.negative_prompt, + width=data.width, height=data.height, seed=data.seed, + reference_images=[str(storage.to_absolute(row.path)) for row in references], + params={"model": spec.model}, + idempotency_key=( + f"platform-image:{owner.id}:{data.kind}:" + f"{now_utc().isoformat()}:{data.prompt}:{','.join(map(str, data.reference_media_ids))}" + ), + label=f"PLATFORM-{owner_type.upper()}-{owner.id}", output_path=out, + ) + progress.report(f"调用 {spec.provider}/{spec.model} 出图") + result = model_registry.generate_asset(db, request, "image", spec.provider, spec.model) + produced = _require_file(result, "图像") + kind = data.kind + info = _store(produced, owner_type=owner_type, owner_id=owner.id, kind=kind, + stem=f"{owner_type}-{owner.id}-{kind}-ai", result=result) + media = storage.register_platform_media( + db, owner_type=owner_type, owner_id=owner.id, kind=kind, path=info["path"], + name=f"{owner.name} · AI {AssetKind.LABELS.get(kind, kind)}", + note=f"平台资产生成:{spec.provider}/{spec.model}", source="provider", info=info, + metadata={"generated": True, "provider": spec.provider, "model": spec.model, + "prompt": request.prompt, + "reference_media_ids": [row.id for row in references], + "task_id": task.id}, + ) + return {"media": storage.platform_media_dict(media), "kind": kind, + "provider": spec.provider, "model": spec.model} + + +def _generate_voice(db: Session, owner, data: schemas.PlatformGenerate, task: Task) -> dict: + spec = resolve_spec(db, "voice", data) + if not spec.enabled: + raise _config_error(f"模型 {spec.provider}/{spec.model} 已停用", "model_disabled") + # 试听文本是「模型要念出来的台词」,不是声线设定。旧实现拿 owner.description 兜底, + # 于是用户填的「温和、语速中等,沉稳」被当成台词念了出来,试听样本一听就是设定说明文。 + text = (data.sample_text or "").strip() + if not text: + raise _config_error("请先填写试听文本:它是模型要朗读的台词,系统不会用「音色描述」代替", + "missing_sample_text") + reference_media, reference = resolve_voice_reference(db, owner, spec, data) + # 旧实现缺参考时直接回退到模型默认参考(女声),男声声线就这样被生成成了女声。 + if not reference["path"] and reference.get("source") == "none": + raise _config_error(reference["note"], "missing_reference_audio") + out = storage.platform_dir("voice", owner.id, "generated") / f"sample-{now_utc().timestamp():.0f}.wav" + # 交给模型的是绝对路径(供应商要直接读文件);声线列里存 media_root 相对路径, + # 换机器 / 进容器时不会因为写死绝对路径而失效。与任务链路 handlers 保持一致。 + reference_path = reference["path"] or owner.provider_voice_id + request = VoiceGenRequest( + text=text, voice_id=owner.name, provider_voice_id=reference_path, + language=owner.language, speed=float((owner.tuning or {}).get("speed", 1.0)), + seed=data.seed, params=(owner.tuning or {}).get("params") or {}, + # 没有参考音频时靠它出音,有参考音频时用它微调音色(见 comfy_assets.graph) + instruct_text=timbre.instruct_text_of(owner) if timbre.instruct_model(spec) else "", + idempotency_key=f"platform-voice:{owner.id}:{task.id}", + label=f"PLATFORM-VOICE-{owner.id}", output_path=out, + ) + progress.report(f"调用 {spec.provider}/{spec.model} 出试听") + result = model_registry.generate_voice(db, request, spec.provider, spec.model) + produced = _require_file(result, "音频") + info = _store(produced, owner_type="voice", owner_id=owner.id, kind="voice_sample", + stem=f"voice-{owner.id}-ai", result=result) + media = storage.register_platform_media( + db, owner_type="voice", owner_id=owner.id, kind="voice_sample", path=info["path"], + name=f"{owner.name} · AI {AssetKind.LABELS.get('voice_sample', 'voice_sample')}", + note=f"平台资产生成:{spec.provider}/{spec.model}", source="provider", info=info, + metadata={"generated": True, "provider": spec.provider, "model": spec.model, + "prompt": request.text, + "reference_media_ids": [reference_media.id] if reference_media else [], + "task_id": task.id}, + ) + owner.provider = spec.provider + if reference["path"] and timbre.clone_model(spec): + # 把真正参与克隆的那段参考音频绑回声线:否则试听样本是 A 的嗓子, + # 之后项目配音又落回模型默认参考,变成另一种嗓子。 + owner.provider_voice_id = str(storage.to_relative(reference["path"])) + # 音色自检:量一遍实际基频,与设定性别比对,结果进素材元数据供界面标注。 + check = timbre.check_voice(produced, owner.gender) + produced_media_metadata(media, reference, check, text) + return {"media": storage.platform_media_dict(media), "kind": "voice_sample", + "provider": spec.provider, "model": spec.model} + + +def _require_file(result, what: str) -> Path: + produced = Path(result.file_path) + if not produced.exists() or produced.stat().st_size < 128: + raise ProviderError(f"模型未返回有效{what}文件", code="missing_output") + return produced + + +def _store(produced: Path, *, owner_type: str, owner_id: int, kind: str, stem: str, + result) -> dict: + info = storage.save_platform_local_file( + produced, owner_type=owner_type, owner_id=owner_id, kind=kind, stem=stem) + if getattr(result, "duration_ms", 0): + info["duration_ms"] = int(result.duration_ms) + return info + + +__all__ = ["active_task", "build_owner_prompt", "enqueue", "generate", "get_owner", + "normalize_type", "preflight", "produced_media_metadata", "request_from", + "resolve_reference_media", "resolve_spec", "resolve_voice_reference", "type_label"] diff --git a/backend/app/services/progress.py b/backend/app/services/progress.py new file mode 100644 index 0000000..c297c65 --- /dev/null +++ b/backend/app/services/progress.py @@ -0,0 +1,71 @@ +"""进行中的任务如何把自己「此刻在做什么」说出来。 + +为什么需要这一层:出图动辄四十几秒、出片动辄十几分钟,而调用模型的那段代码埋在 +适配器深处(``providers/comfy_common.py``),手里既没有会话也不知道 task_id。 +以前它什么都说不出来,界面上就只剩一个转圈的图标——用户分不清「模型正在算」 +和「排在别人的任务后面」,只能一遍遍刷新页面去猜。 + +做法:执行线程起跑时把 task_id 绑到 thread-local(见 ``jobs/handlers.run``), +深处任何一层调 ``report("排队中:等待算力空闲")`` 就能写回任务行; +不在任务上下文里(普通请求线程)静默返回,调用点因此不需要判空。 + +写入走**另一条会话**并立刻提交:执行线程自己那条会话正握着业务事务, +在深处顺手 commit 会把事务提前收掉,半途失败的素材状态就会对不上。 +""" +from __future__ import annotations + +import logging +import threading +from typing import Any + +from sqlalchemy import update + +from app.db import SessionLocal +from app.models import Task, now_utc + +log = logging.getLogger("nex.progress") + +_local = threading.local() + +#: message 列宽 400,超了按字符砍掉——宁可截断也不要当场报错 +MAX_MESSAGE_CHARS = 400 + + +def bind(task_id: int) -> None: + """声明「当前线程正在执行这条任务」;由 handlers.run 统一调用。""" + _local.task_id = int(task_id) + + +def unbind() -> None: + _local.task_id = None + + +def current_task_id() -> int | None: + return getattr(_local, "task_id", None) + + +def report(message: str, *, progress: float | None = None) -> None: + """给进行中的任务补一句解释(可顺带推进度);不在任务上下文里时什么都不做。 + + 这里**只写不读**、且永远不改 status:状态归执行体自己收口, + 深处这一层只负责把过程说清楚,失败了也不该影响正在跑的模型调用。 + """ + task_id = current_task_id() + if not task_id: + return + patch: dict[str, Any] = {"message": message[:MAX_MESSAGE_CHARS], "heartbeat_at": now_utc()} + if progress is not None: + patch["progress"] = max(0.0, min(1.0, float(progress))) + session = SessionLocal() + try: + session.execute(update(Task).where(Task.id == task_id).values(**patch) + .execution_options(synchronize_session=False)) + session.commit() + except Exception as exc: # noqa: BLE001 - 说不出来不等于做错了 + log.debug("任务 #%s 状态说明回写失败:%s", task_id, exc) + session.rollback() + finally: + session.close() + + +__all__ = ["MAX_MESSAGE_CHARS", "bind", "current_task_id", "report", "unbind"] diff --git a/backend/app/services/status.py b/backend/app/services/status.py index c3c7b0b..e4b5e9a 100644 --- a/backend/app/services/status.py +++ b/backend/app/services/status.py @@ -5,7 +5,7 @@ """ from __future__ import annotations -from datetime import datetime, timezone +from datetime import datetime, timedelta, timezone from sqlalchemy import func, select from sqlalchemy.orm import Session, selectinload @@ -80,6 +80,10 @@ def task_dict(task: Task) -> dict: return { "id": task.id, "project_id": task.project_id, "shot_id": task.shot_id, "timeline_id": task.timeline_id, "kind": task.kind, "name": task.name, "stage": task.stage, + # 非项目归属(平台资产母版):卡片按它亮「正在生成」,活动面板按它跳回资产 + "owner_type": task.owner_type or "", "owner_id": task.owner_id, + # 「此刻在做什么」优先取执行体自己说的,退回深处 progress.report 写的那句 + "message": _task_message(task), "status": task.status, "priority": task.priority, "progress": round(task.progress or 0, 4), "provider": task.provider, "model": task.model, "mode": task.mode, "execution_backend": task.execution_backend, "celery_task_id": task.celery_task_id, @@ -100,6 +104,60 @@ def task_dict(task: Task) -> dict: } +def _task_message(task: Task) -> str: + """任务此刻在做什么的一句话:优先 ``task.message``(执行体与算力锁写的), + + 退回 ``checkpoint_json.last_note``(项目侧 handler 一直在往这里写阶段说明), + 两边都有时取更新的那一条——两侧写入者不同,谁最后落地谁就是真相。 + """ + note = str((task.checkpoint_json or {}).get("last_note") or "") + mine = str(task.message or "") + if not note: + return mine + if not mine: + return note + at = (task.checkpoint_json or {}).get("at") or "" + heartbeat = task.heartbeat_at.isoformat() if task.heartbeat_at else "" + return note if str(at) >= str(heartbeat or "") else mine + + +#: 活动面板默认回吐多少条历史:够翻完一个下午的生成,又不至于让 3 秒一次的轮询变重 +ACTIVITY_LIMIT = 40 +#: 已落定的任务在活动面板里保留多久(与 ACTIVITY_LIMIT 一起构成「刚刚完成」那一段) +ACTIVITY_WINDOW = timedelta(hours=6) + + +def activity(db: Session, *, limit: int = ACTIVITY_LIMIT, owner_type: str = "", + active_only: bool = False) -> dict: + """顶栏「平台动态」要的那一份:进行中的全部 + 最近落定的若干条。 + + 数据源就是统一任务表,所以项目侧的出片与平台侧的生成天然同屏—— + 用户问的是「平台现在在忙什么」,不关心这条活算不算某个项目的。 + ``limit`` 只截历史:进行中的任务一条都不能少,否则用户在面板里 + 看不到自己刚点的那次,比没有面板更糟(他会以为排队排丢了)。 + """ + stmt = select(Task).order_by(Task.id.desc()) + if owner_type: + stmt = stmt.where(Task.owner_type == owner_type) + if active_only: + rows = db.scalars(stmt.where(Task.status.in_(TaskStatus.ACTIVE))).all() + else: + active = db.scalars(stmt.where(Task.status.in_(TaskStatus.ACTIVE))).all() + settled = db.scalars( + stmt.where(Task.status.notin_(TaskStatus.ACTIVE), + Task.updated_at >= now_utc() - ACTIVITY_WINDOW) + .limit(min(200, max(1, limit))) + ).all() + rows = list(active) + list(settled) + items = [task_dict(t) for t in rows] + counts: dict[str, int] = {} + for row in rows: + counts[row.status] = counts.get(row.status, 0) + 1 + return {"tasks": items, + "active": sum(counts.get(s, 0) for s in TaskStatus.ACTIVE), + "counts": counts, "window_hours": ACTIVITY_WINDOW.total_seconds() / 3600} + + def shot_dict(db: Session, shot: Shot, *, assets: bool = True) -> dict: from app.services import preview as preview_service from app.services import prompt as prompt_service diff --git a/backend/app/worker.py b/backend/app/worker.py index 90e0d20..c673a3e 100644 --- a/backend/app/worker.py +++ b/backend/app/worker.py @@ -13,7 +13,7 @@ from app.config import settings # 队列划分:不同工种的资源画像不同(GPU/网络受限的生成 vs CPU 密集的合成) QUEUES = {"generate_shot": "generate", "voiceover": "audio", "qc_shot": "qc", - "compose": "compose", "import": "io", "asset": "io"} + "compose": "compose", "import": "io", "asset": "io", "platform_asset": "io"} TASK_ROUTES = { "app.jobs.generate_shot": {"queue": "generate"}, @@ -22,6 +22,7 @@ TASK_ROUTES = { "app.jobs.compose": {"queue": "compose"}, "app.jobs.import_package": {"queue": "io"}, "app.jobs.asset": {"queue": "io"}, + "app.jobs.platform_asset": {"queue": "io"}, } celery_app = Celery("nex_animation", broker=settings.broker_url, backend=settings.result_backend) diff --git a/backend/tests/conftest.py b/backend/tests/conftest.py index 53c56e6..86b716d 100644 --- a/backend/tests/conftest.py +++ b/backend/tests/conftest.py @@ -153,6 +153,25 @@ def api(client): body = self.post("/api/projects", json=payload) return body["result"]["id"] + def generate(self, url, json, *, timeout=120, expect=202, task_only=False): + """发起一次平台资产生成,并**等它真的跑完**,返回生成出来的素材字典。 + + 接口只登记任务并立刻返 202(同步等出图那一版会被反代 60s 读超时打断), + 所以自检不能再直接读 POST 响应体:必须等 inline 线程池跑空、 + 再回读任务行,否则断言读到的是「排队中」的空结果。 + ``expect`` 非 202 时(表单校验、去重拦截)原样返回响应体,不做等待。 + """ + body = self.call("POST", url, expect=expect, json=json) + if expect != 202: + return body + task_id = body["result"]["id"] + assert execution.wait_idle(timeout), f"平台生成任务 #{task_id} 在 {timeout}s 内没有跑完" + task = self.get(f"/api/tasks/{task_id}")["task"] + assert task["status"] == "succeeded", ( + f"平台生成任务 #{task_id} 未成功:{task['status']} / " + f"{task['error_code']} {task['error_message']}") + return task if task_only else task["result"]["media"] + def drain(self, project_id, timeout=600): return self.post(f"/api/projects/{project_id}/dispatch", json={"drain": True, "timeout": timeout}) diff --git a/backend/tests/test_direct_connection.py b/backend/tests/test_direct_connection.py index 34f097b..2db2213 100644 --- a/backend/tests/test_direct_connection.py +++ b/backend/tests/test_direct_connection.py @@ -144,7 +144,10 @@ def test_dead_endpoint_is_reported_as_refused_not_as_outage(shell_proxy): comfy_common.require_reachable(dead, stage="开工前") message = str(caught.value) assert caught.value.code == "provider_unreachable" - assert "502" not in message + # 断言前先把地址从文本里摘掉:``_free_base()`` 拿的是随机空闲端口, + # 50227 这种端口号自带「502」三个字,会让这条用例每跑几千次红一次。 + haystack = message.replace(dead, "").replace(dead.rsplit(":", 1)[-1], "") + assert "502" not in haystack assert ("Connection refused" in message or "ConnectError" in message), message assert "模型管理中心" in message diff --git a/backend/tests/test_platform_assets.py b/backend/tests/test_platform_assets.py index 176e281..07ec91b 100644 --- a/backend/tests/test_platform_assets.py +++ b/backend/tests/test_platform_assets.py @@ -31,11 +31,11 @@ def test_platform_assets_upload_import_publish_and_generate(api): )["result"] assert uploaded["url"].startswith("/media/platform/") - generated = api.post( + generated = api.generate( f"/api/platform/{'scene'}/{scene['id']}/generate", - json={"kind": "scene_image", "provider": "mock", "model": "mock-image-1", - "prompt": "冷蓝全息舰桥", "width": 320, "height": 180}, - )["result"] + {"kind": "scene_image", "provider": "mock", "model": "mock-image-1", + "prompt": "冷蓝全息舰桥", "width": 320, "height": 180}, + ) assert generated["metadata"]["generated"] is True imported = api.post("/api/platform/import", json={ @@ -126,12 +126,12 @@ def test_platform_image_generation_accepts_owned_reference_media(api): files={"file": ("reference.png", _make_image_png(320, 320, "0x335577"), "image/png")}, data={"kind": "reference", "name": "参考定妆"}, )["result"] - generated = api.post( + generated = api.generate( f"/api/platform/human/{character['id']}/generate", - json={"kind": "portrait", "provider": "mock", "model": "mock-image-1", - "prompt": "保持参考图中的五官与服装", "width": 320, "height": 320, - "reference_media_ids": [reference["id"]]}, - )["result"] + {"kind": "portrait", "provider": "mock", "model": "mock-image-1", + "prompt": "保持参考图中的五官与服装", "width": 320, "height": 320, + "reference_media_ids": [reference["id"]]}, + ) assert generated["metadata"]["reference_media_ids"] == [reference["id"]] @@ -166,11 +166,11 @@ def test_platform_voice_generation_uses_uploaded_reference(api, db, monkeypatch) monkeypatch.setattr(platform_api.model_registry, "is_available", lambda *_args, **_kwargs: True) monkeypatch.setattr(platform_api.model_registry, "generate_voice", fake_generate) - generated = api.post( + generated = api.generate( f"/api/platform/voice/{voice['id']}/generate", - json={"kind": "voice_sample", "provider": "comfyui", "model": "indextts-test", - "sample_text": "参考音色已经接通。", "reference_media_id": reference["id"]}, - )["result"] + {"kind": "voice_sample", "provider": "comfyui", "model": "indextts-test", + "sample_text": "参考音色已经接通。", "reference_media_id": reference["id"]}, + ) assert generated["kind"] == "voice_sample" assert captured["request"].provider_voice_id.endswith(reference["path"]) assert Path(captured["request"].provider_voice_id).is_absolute() diff --git a/backend/tests/test_platform_version_reuse.py b/backend/tests/test_platform_version_reuse.py index 935f684..b2ce4e2 100644 --- a/backend/tests/test_platform_version_reuse.py +++ b/backend/tests/test_platform_version_reuse.py @@ -146,13 +146,16 @@ def test_upload_after_master_deletion_reuses_id_without_integrity_error(api): def test_generate_after_reused_id_does_not_500(api): - """端到端复现报错接口:主键复用 + 遗留孤儿行时 generate 必须 200 且版本号正确递增。""" + """端到端复现报错接口:主键复用 + 遗留孤儿行时 generate 必须成功且版本号正确递增。 + + 三次生成串行等待(api.generate),因为同一条母版同一类素材在生成中会被去重挡住。 + """ human = api.post("/api/platform/characters", json={ "name": "生成复用", "appearance": "银色短发,深蓝机甲", "persona_prompt": "冷静果决的领航员"})["result"] payload = {"kind": "portrait", "provider": "mock", "model": "mock-image-1", "prompt": "冷蓝定妆", "width": 240, "height": 240} - generated = api.post(f"/api/platform/human/{human['id']}/generate", json=payload)["result"] + generated = api.generate(f"/api/platform/human/{human['id']}/generate", payload) media_id = generated["id"] assert generated["version"] == 1 and _version_nos("platform_media_versions", "media_id", media_id) == [1] @@ -164,13 +167,13 @@ def test_generate_after_reused_id_does_not_500(api): "name": "生成复用", "appearance": "银色短发,深蓝机甲", "persona_prompt": "冷静果决的领航员"})["result"] assert reuse["id"] != human["id"] - second = api.post(f"/api/platform/human/{reuse['id']}/generate", json=payload)["result"] + second = api.generate(f"/api/platform/human/{reuse['id']}/generate", payload) assert second["id"] == media_id, "新母版复用同一 media_id,正是当年 500 的触发条件" assert second["version"] == 2, "首版号按子表现状推导,跳过孤儿行占用的 version_no=1" assert _version_nos("platform_media_versions", "media_id", media_id) == [1, 2] assert Path(storage.to_absolute(second["path"])).exists() - third = api.post(f"/api/platform/human/{reuse['id']}/generate", json=payload)["result"] + third = api.generate(f"/api/platform/human/{reuse['id']}/generate", payload) assert third["version"] == 3 and _version_nos("platform_media_versions", "media_id", media_id) == [1, 2, 3] diff --git a/backend/tests/test_startup_lock.py b/backend/tests/test_startup_lock.py index 441c569..2ef12ba 100644 --- a/backend/tests/test_startup_lock.py +++ b/backend/tests/test_startup_lock.py @@ -106,8 +106,8 @@ def test_real_sqlite_write_lock_is_waited_for(tmp_path: Path): def test_init_db_survives_a_busy_database(monkeypatch): - """把 init_db 的三步全换成「先抛两次锁」,确认启动路径整体是等锁而不是崩掉。""" - counters = {"create_all": 0, "columns": 0, "backfill": 0} + """把 init_db 的四步全换成「先抛两次锁」,确认启动路径整体是等锁而不是崩掉。""" + counters = {"create_all": 0, "columns": 0, "relax": 0, "backfill": 0} def flaky(key: str): def run(*_args, **_kwargs) -> None: @@ -119,8 +119,10 @@ def test_init_db_survives_a_busy_database(monkeypatch): metadata = type("M", (), {"create_all": staticmethod(flaky("create_all"))})() monkeypatch.setattr(dbmod, "Base", type("B", (), {"metadata": metadata})()) monkeypatch.setattr(dbmod, "_add_missing_columns", flaky("columns")) + # tasks.project_id 放开非空要整表重建 tasks,这一步同样在启动路径上,必须一起等锁 + monkeypatch.setattr(dbmod, "_relax_nullability", flaky("relax")) monkeypatch.setattr(dbmod, "_backfill_defaults", flaky("backfill")) monkeypatch.setattr(dbmod, "STARTUP_LOCK_BUDGET_SECONDS", 10.0) dbmod.init_db() - assert counters == {"create_all": 3, "columns": 3, "backfill": 3} + assert counters == {"create_all": 3, "columns": 3, "relax": 3, "backfill": 3} diff --git a/backend/tests/test_task_activity.py b/backend/tests/test_task_activity.py new file mode 100644 index 0000000..1a22998 --- /dev/null +++ b/backend/tests/test_task_activity.py @@ -0,0 +1,261 @@ +"""统一任务表与「平台动态」:所有跑模型的活都必须能被问出进度。 + +用户的原话(2026-11):「平台应该有个进度,所有跑模型的任务都应该写到一个任务表中。 +异步,可以请求进度。然后界面上可以整体展示。」 + +在那之前平台资产库的 AI 生成是一个同步 HTTP 请求:出图实测四十几秒,反代 60s 读 +超时就把连接掐了(504),而**出图线程并没有被取消**——图照样落盘,用户看不见, +只能手动刷新页面去猜它到底成没成。现在接口只做「校验 + 登记 + 派发」, +当场返回任务号,进度与产物全在 ``tasks`` 那一行里,换设备、刷新页面都接着看得到。 + +这一组用例锁的是三件容易被改回去的事: + 1. 平台生成必须落到同一张 ``tasks`` 表,并带上母版归属(``owner_type/owner_id``) + —— 顶栏要按它点亮卡片、活动面板要按它跳回资产; + 2. ``GET /api/tasks/activity`` 必须看得见它,且进行中的任务一条都不能被条数截掉; + 3. 同一条母版同一类素材在跑的时候再点一次必须挡下来(409), + 否则三次点击就是三次 GPU 排队、三份产物、三倍扣费。 +""" +from __future__ import annotations + +import threading + +from app.constants import TaskKind, TaskStatus +from app.providers import registry as model_registry +from app.providers.base import AssetGenResult, ProviderError +from tests.conftest import _make_image_png + + +def _scene(api, name: str = "星舰舰桥") -> dict: + return api.post("/api/platform/scenes", + json={"name": name, "prompt": "环形舰桥,全息星图", "lighting": "冷蓝"})["result"] + + +def _png(width: int = 320, height: int = 180) -> bytes: + """一张能过「产物校验」的图:模型跑完后平台会拒绝小于 128 字节的文件。 + + 本机没有 ffmpeg 时 ``_make_image_png`` 会退回 70 字节的 1x1 图, + 那种图在这里会被判成「模型未返回有效图像文件」,用例就变成测 ffmpeg 装没装。 + """ + payload = _make_image_png(width, height, "0x2f6d4f") + return payload + b"\x00" * max(0, 256 - len(payload)) + + +def _body(prompt: str = "冷蓝全息舰桥") -> dict: + return {"kind": "scene_image", "provider": "mock", "model": "mock-image-1", + "prompt": prompt, "width": 320, "height": 180} + + +def _poll(api, task_id: int, done, what: str, timeout: float = 20.0) -> dict: + """轮询任务行直到 ``done(task)`` 成立——用于观察「正在跑」这一中间态。""" + import time + + deadline = time.monotonic() + timeout + task: dict = {} + while time.monotonic() < deadline: + task = api.get(f"/api/tasks/{task_id}")["task"] + if done(task): + return task + time.sleep(0.05) + raise AssertionError(f"任务 #{task_id} 未在 {timeout}s 内{what},当前 " + f"{task.get('status')} / {task.get('message')}") + + +# ---------------------------------------------------------------- 落进统一任务表 +def test_platform_generation_lands_in_the_unified_task_table(api): + """生成不再是一个「请求-等待-响应」,而是一条可查询、可重试、可取消的任务。""" + scene = _scene(api) + task = api.generate(f"/api/platform/scene/{scene['id']}/generate", _body(), task_only=True) + + assert task["kind"] == TaskKind.PLATFORM_ASSET + # 母版本就不属于任何项目:挂 0/-1 会让每一条按项目统计的 SQL 多背一个特例 + assert task["project_id"] is None and task["shot_id"] is None + assert task["owner_type"] == "scene" and task["owner_id"] == scene["id"] + assert task["status"] == TaskStatus.SUCCEEDED + assert scene["name"] in task["name"] + assert task["output_url"].startswith("/media/platform/") + # 产物反向指回任务号:素材出问题时能从文件一路追回是哪一次生成 + assert task["result"]["media"]["metadata"]["task_id"] == task["id"] + assert task["elapsed_ms"] >= 0 and task["payload"]["asset_kind"] == "scene_image" + + +def test_activity_shows_finished_platform_work(api): + """顶栏那一份「平台动态」就是统一任务表的切片,不需要第二个数据源。""" + scene = _scene(api) + task = api.generate(f"/api/platform/scene/{scene['id']}/generate", _body(), task_only=True) + + panel = api.get("/api/tasks/activity")["result"] + assert panel["active"] == 0 + assert panel["counts"].get(TaskStatus.SUCCEEDED) == 1 + assert panel["window_hours"] > 0 + + row = next(item for item in panel["tasks"] if item["id"] == task["id"]) + assert row["owner_type"] == "scene" and row["owner_id"] == scene["id"] + assert row["project_id"] is None + # 面板上必须能说出「这是哪条母版的什么」,只剩一个转圈图标等于没说 + assert scene["name"] in row["name"] + assert row["message"] and row["result"]["media"]["id"] + + +def test_activity_ignores_settled_work_outside_the_window(api): + """历史只留最近一段时间:三天前跑成功的那条不该一直挂在顶栏角标旁边。""" + from datetime import timedelta + + from app.db import SessionLocal + from app.models import Task, now_utc + from app.services import status as status_service + + scene = _scene(api) + task = api.generate(f"/api/platform/scene/{scene['id']}/generate", _body(), task_only=True) + with SessionLocal() as session: + row = session.get(Task, task["id"]) + # 落定时间推到窗口之外:它已经从「刚刚完成」那一格里掉出去了 + row.updated_at = now_utc() - timedelta(hours=status_service.ACTIVITY_WINDOW.total_seconds() / 3600 + 1) + session.commit() + + panel = api.get("/api/tasks/activity")["result"] + assert [item["id"] for item in panel["tasks"]] == [] + assert panel["counts"] == {} + + +# ---------------------------------------------------------------- 进行中的可见性 +def _blocking_adapter(release: threading.Event, notes: list[str]): + """替身模型:卡在一个事件上,并在深处报一句「算力排队」——真实链路正是这个样子。 + + ``services/progress`` 靠 thread-local 里的 task_id 写回任务行, + 跑在 inline 线程池里的执行体如果没绑定 task_id,这里写进去的就是空气, + 界面上也永远看不到「排在别人后面」和「模型正在算」的区别。 + """ + from app.services import progress + + class Blocking: + name = "blocking" + + def __call__(self, db, request, media_type, provider, model): + notes.append(progress.current_task_id() and "bound" or "unbound") + progress.report("排队中:等待算力空闲") + assert release.wait(20), "替身模型没有被放行,用例会挂在这里" + with open(request.output_path, "wb") as handle: + handle.write(_png()) + return AssetGenResult(file_path=str(request.output_path), metadata={"model": model}) + + return Blocking() + + +def test_a_running_generation_is_visible_and_blocks_the_second_click(api, monkeypatch): + """跑模型期间:界面看得见它、卡片点不动它,一次点击只该有一份产物。""" + release = threading.Event() + notes: list[str] = [] + monkeypatch.setattr(model_registry, "generate_asset", _blocking_adapter(release, notes)) + scene = _scene(api) + + # 202 而不是 200:接口只做「校验 + 登记 + 派发」,出图这件事交给后台 + first = api.post(f"/api/platform/scene/{scene['id']}/generate", json=_body(), expect=202) + assert first["result"]["status"] in (TaskStatus.QUEUED, TaskStatus.RUNNING) + task_id = first["result"]["id"] + # 深处那句排队说明必须能到达任务行:出图四十多秒里,用户看的就只有这一行字 + running = _poll(api, task_id, lambda t: t["message"] == "排队中:等待算力空闲", + "报出算力排队说明") + assert running["status"] == TaskStatus.RUNNING + assert notes == ["bound"], "执行体必须把 task_id 绑进线程,否则 progress.report 写不进来" + + panel = api.get("/api/tasks/activity")["result"] + assert panel["active"] == 1 + assert panel["tasks"][0]["id"] == task_id + + # 用户没等完又点了一次:挡下来,而不是悄悄排第二份 + again = api.post(f"/api/platform/scene/{scene['id']}/generate", + json=_body("换个提示词再试"), expect=409) + assert "正在生成中" in again["detail"] and scene["name"] in again["detail"] + + release.set() + from app.services import execution + + assert execution.wait_idle(30) + done = api.get(f"/api/tasks/{task_id}")["task"] + assert done["status"] == TaskStatus.SUCCEEDED + assert api.get("/api/tasks/activity")["result"]["active"] == 0 + + +def test_a_running_platform_task_can_be_cancelled_without_a_project(api, monkeypatch): + """取消按钮过去靠「按项目取父记录」拿权限,平台任务没有项目,会凭空 404。""" + release = threading.Event() + monkeypatch.setattr(model_registry, "generate_asset", _blocking_adapter(release, [])) + scene = _scene(api) + task_id = api.post(f"/api/platform/scene/{scene['id']}/generate", json=_body(), + expect=202)["result"]["id"] + _poll(api, task_id, lambda t: t["status"] == TaskStatus.RUNNING, "开始执行") + + cancelled = api.post(f"/api/tasks/{task_id}/cancel", json={"reason": "提示词写错了"})["result"] + assert cancelled["status"] == TaskStatus.CANCELLED + + release.set() + from app.services import execution + + assert execution.wait_idle(30) + # 出图线程跑完了也不许把「已取消」改写成成功:用户明确说过不要 + final = api.get(f"/api/tasks/{task_id}")["task"] + assert final["status"] == TaskStatus.CANCELLED + + +def test_a_transient_failure_stays_on_the_panel_while_it_waits_to_retry(api, monkeypatch): + """连不上算力是瞬时故障:平台自己会再试一次,这期间它在面板上仍然是「进行中」。 + + 挡在这里的是「重试中的任务从角标上消失了」这一类回归——用户看到的是 + 进度条归零、什么都没发生,于是以为失败了,其实排队等下一次派发。 + """ + def busy(db, request, media_type, provider, model): + raise ProviderError("算力服务无响应:http://comfyui:9200", code="gpu_unreachable") + + monkeypatch.setattr(model_registry, "generate_asset", busy) + scene = _scene(api) + created = api.post(f"/api/platform/scene/{scene['id']}/generate", json=_body(), + expect=202)["result"] + + from app.services import execution + + assert execution.wait_idle(30) + task = api.get(f"/api/tasks/{created['id']}")["task"] + assert task["status"] == TaskStatus.RETRYING + assert task["error_code"] == "gpu_unreachable" + assert task["max_attempts"] == 2 and task["next_retry_at"] is not None + + panel = api.get("/api/tasks/activity")["result"] + assert panel["active"] == 1 + row = next(item for item in panel["tasks"] if item["id"] == created["id"]) + # 面板上直接把供应商那句原话摊开,用户才知道去模型管理中心查地址,而不是重传素材 + assert "算力服务无响应" in row["error_message"] + + +def test_a_failed_platform_task_can_be_retried_from_the_panel(api, monkeypatch): + """面板上的「重试」要能立刻再排一次,而不是等下一轮调度扫描(本机 5s)。 + + 这里用 retryable=False 的错误(提示词超限这一类):平台自己不会反复撞墙, + 改好之后由人点重试——那正是活动面板上那颗按钮的用途。 + """ + fixed = {"ok": False} + + def adapter(db, request, media_type, provider, model): + if not fixed["ok"]: + raise ProviderError("提示词超出模型长度上限", code="prompt_too_long", retryable=False) + with open(request.output_path, "wb") as handle: + handle.write(_png()) + return AssetGenResult(file_path=str(request.output_path), metadata={"model": model}) + + monkeypatch.setattr(model_registry, "generate_asset", adapter) + scene = _scene(api) + task_id = api.post(f"/api/platform/scene/{scene['id']}/generate", json=_body(), + expect=202)["result"]["id"] + + from app.services import execution + + assert execution.wait_idle(30) + failed = api.get(f"/api/tasks/{task_id}")["task"] + assert failed["status"] == TaskStatus.FAILED and failed["attempts"] == 1 + + fixed["ok"] = True + retried = api.post(f"/api/tasks/{task_id}/retry", json={})["result"] + assert retried["status"] in (TaskStatus.QUEUED, TaskStatus.RUNNING, TaskStatus.SUCCEEDED) + assert execution.wait_idle(30) + done = api.get(f"/api/tasks/{task_id}")["task"] + assert done["status"] == TaskStatus.SUCCEEDED + assert done["result"]["media"]["metadata"]["task_id"] == task_id diff --git a/backend/tests/test_voice_timbre_gender.py b/backend/tests/test_voice_timbre_gender.py index ce253b7..f7e2bf6 100644 --- a/backend/tests/test_voice_timbre_gender.py +++ b/backend/tests/test_voice_timbre_gender.py @@ -190,9 +190,8 @@ def test_male_voice_without_reference_borrows_male_reference(api, db): assert resolved["path"].endswith(".wav") assert male_master - created = api.post(f"/api/platform/voice/{newcomer}/generate", - json={"kind": "voice_sample", "sample_text": "我是俊杰,公司的产品专家。"}) - media = created["result"] + media = api.generate(f"/api/platform/voice/{newcomer}/generate", + {"kind": "voice_sample", "sample_text": "我是俊杰,公司的产品专家。"}) assert media["metadata"]["reference_source"] == "gender_sibling" assert media["metadata"]["text"] == "我是俊杰,公司的产品专家。" assert "timbre_check" in media["metadata"] @@ -265,8 +264,8 @@ def test_sample_text_is_required_and_never_the_description(api, db): json={"kind": "voice_sample", "sample_text": " \n "}, expect=422) assert "试听文本" in blank["detail"] - produced = api.post(f"/api/platform/voice/{voice_id}/generate", - json={"kind": "voice_sample", "sample_text": "你好,我是紫小慧。"})["result"] + produced = api.generate(f"/api/platform/voice/{voice_id}/generate", + {"kind": "voice_sample", "sample_text": "你好,我是紫小慧。"}) assert produced["metadata"]["text"] == "你好,我是紫小慧。" assert "音色温暖柔和" not in produced["metadata"]["text"] @@ -308,8 +307,8 @@ def test_ai_sample_never_becomes_this_voice_s_own_reference(api, db): _clone_tts(db) _platform_voice(api, name="洪七公 · 声线", gender="male", reference_hz=105) newcomer = _platform_voice(api, name="李俊杰", gender="male", description="低沉浑厚的男声") - produced = api.post(f"/api/platform/voice/{newcomer}/generate", - json={"kind": "voice_sample", "sample_text": "我是李俊杰。"})["result"] + produced = api.generate(f"/api/platform/voice/{newcomer}/generate", + {"kind": "voice_sample", "sample_text": "我是李俊杰。"}) assert produced["metadata"]["generated"] is True # 模拟旧链路的泄漏现场:声线绑定的其实就是那个 AI 样本 @@ -364,8 +363,8 @@ def test_importing_a_voice_keeps_the_ai_provenance(api, db, studio): _clone_tts(db) master = _platform_voice(api, name="周伯通 · 声线", gender="male", reference_hz=118) - api.post(f"/api/platform/voice/{master}/generate", - json={"kind": "voice_sample", "sample_text": "老顽童来也。"}) + api.generate(f"/api/platform/voice/{master}/generate", + {"kind": "voice_sample", "sample_text": "老顽童来也。"}) api.post("/api/platform/import", json={"project_id": studio["project_id"], "asset_type": "voice", "asset_id": master}) diff --git a/docs/UI规范.md b/docs/UI规范.md index 27351ad..f08116f 100644 --- a/docs/UI规范.md +++ b/docs/UI规范.md @@ -1337,6 +1337,89 @@ chromeless 页面(不套顶栏 / 侧栏,也不拉需要鉴权的 `/api/meta` 只剩 `extra` 的卡头少一个样本。它量的是渲染结果,不要求每页样本数相同—— 看的是 `ALIGN_PROBED` 总有没有掉到地板(本轮实测 181 / 底线 40),不是数标题。 +### 13.7 模型任务统一进任务表:顶栏「平台动态」是唯一进度源(2026-11-06 · 用户第 2、3 条 + 架构指令) + +**症状**:资产库点「AI 生成」是一个同步 HTTP 请求,出图 40 秒、出片 6 分钟, +请求挂着的时候界面上什么都没有;生成完了页面也不知道,用户只能自己按 F5 猜 +(原话「很多资产的生成需要手动刷新页面」)。更根本的是**平台在跑什么看不见**—— +三条生成同时在进行,界面上没有任何一处说得出「现在有几条活、跑到哪了、谁在跑」。 + +用户给的口径是一句架构指令,不是一处样式: + +> 平台应该有个进度,所有跑模型的任务都应该写到一个任务表中。异步,可以请求进度。然后界面上可以整体展示。 + +落到实现上是三条规矩: + +**① 任务表是唯一进度源,不新建第二张表。** +既有 `tasks` 表(调度台在用的那张)加三列就够用:`owner_type` / `owner_id` 把任务挂到 +「哪条母版」,`message` 放人类可读的进度文案(「正在出图 3/4 · 分镜 2」)。 +生成接口从「同步跑完再返回」改成 **202 + 排队**: +`POST /api/platform/{owner_type}/{asset_id}/generate` 立刻返回整条 task dict, +真正的执行走 `jobs/handlers.py`,与调度台共用 `dispatch_task` 和 `io` 队列。 +**同步接口留在 `services/platform_generation.py::generate_*` 里**给测试与内部调用用, +接口层只是不再有人等着它返回。 + +**② 全站只有一个地方展示进度:顶栏的活动中心。** +`GET /api/tasks/activity` 返回 `tasks / active / counts`(只回 6 小时内的活,上限 40 条), +`app/layout/ActivityCenter.tsx` 订它:**有活在跑**时按钮变成 `` + 「生成中」+ 数量 `Badge`, +**空闲时**是 `IconClock` + 「平台动态」,永远占同一个位置(按钮不随状态出现或消失, +顶栏宽度就不抖)。抽屉分「进行中 / 最近落定」两段,每行给:名字、``、 +`message`、`ProgressBar`、`provider/model · 已用多久`,动作是取消 / 重试 / 查看。 +`stale`(心跳超时)单独用 warn 档,**卡在 90% 不动**和**失败**是两件事,别混成一个红。 + +「查看」跳 `/assets?type=…&q=…`,资产库是唯一认这份深链的人:读完立刻 +`clearDeepLink({},{replace:true})` 把 query 擦掉——否则同一颗胶囊连点两次, +URL 没变,effect 不再触发。 + +**③ 「自动刷新」开关只管页面重拉数据,不管进度轮询。** +开关(`Check`,存 `nex.autoRefresh`,默认开)落在活动中心的吸底条上, +一句话说明白了它管什么:**关掉只停「页面自己重拉列表」,顶栏进度照旧在跑**。 +- 受它管的:`useAutoRefresh` 的兜底周期、`usePolling`(`intervalMs` 那个定时器)。 +- 不受它管的:活动中心的 3 秒轮询。关掉自动刷新不等于让平台停止干活—— + 任务还在跑,进度当然还得看得见,否则关掉之后连「跑到哪了」都查不到。 + +那颗「立即刷新」按钮**绕开这个开关**(`stores/activity.ts::refreshNow` → `manualTick` +1): +开关管的是「页面自己偷偷重拉」,人已经按下按钮了还要先看开关脸色,等于这颗按钮白点—— +开关关掉期间落定的那几笔,正是靠它补上的。`useRefreshOnActivity` 因此订两个水位: +`revision`(落定,受开关管)与 `manualTick`(手动,不受开关管)。 +双向判据(`/api/platform/assets` 的请求计数 + 卡上定妆图的 `version`)实测: +开关 ON 时任务落定后计数 1→2、卡图 v10→v11(没人按 F5); +开关 OFF 时落定后 14 秒(覆盖 3s 快轮询 + 20s 兜底周期)计数恒为 3、数据停在旧快照, +点「立即刷新」立刻 3→4 并补回 v12。**两头都要量**:只验 A 侧, +「开关是个摆设」这种实现照样全绿。 + +连线是 `stores/activity.ts` 的 `revision`:某条任务从进行中变成落定,`revision += 1`, +挂了的页面各自重拉一次(`hooks/use-refresh-on-activity.ts`)。**落定那一刻刷**比任何 +固定周期都准,也只在真要刷的时候刷。两条实现细节: +- 首帧不刷:`seen` 初值 `-1`,挂载时只把当前 `revision` 记成基线。 + 页面自己刚拉过一次,`revision` 里还带着历史落定数,不记基线就会一进页面就刷两下。 +- 开关关掉时**照样推进水位**:否则重新打开的瞬间会补放积压的那几笔, + 页面会在人没反应的时候自己连刷好几下。 + +**卡底动作的配套改动**(用户第 1 条:「类型和上传、引入是配对的,但 AI 生成是单独的逻辑」): +母版卡底部从两轨改**三轨** `grid-cols-[auto_minmax(0,1fr)_auto]`—— + +| 轨 | 内容 | 为什么 | +|---|---|---| +| 左 | 只有「AI 生成」 | 它是「让模型跑一张」,走排队任务,结果什么时候回来不由这张卡决定 | +| 中 | `ActionGroup[类型 Select + 上传 (+ 录制)]` + `ActionGroup[引入项目 + 删除]` | 前两颗是同一个动作的两半(传成哪一类 → 传),后两颗是母版级的 | +| 右 | 版本 / 更新时间 | 元数据不是动作,永远不换到动作下面(§7.1) | + +「有文件」那一侧左缘那条 `border-s border-fill-secondary` 是**分界线不是装饰**:画在中轨 +而不是左轨右缘,是为了让它的长度天然等于被分开的那组动作的高度,不必再凑 `items-stretch` ++ `self-center`(那两个工具类还会在 `classaudit` 里新引入 `self` 这个工具类根, +把 `UserAdmin.tsx` 里那个同名变量 `self` 误判成死类名)。 +`max-lg` 单列时连同 `ps-3` 一起撤掉(`max-lg:border-s-0 max-lg:ps-0`)——竖线在只有一列的时候没有分界对象。 +生成中这张卡:按钮 `loading`、卡头多一枚 `Tag tone="accent" icon={}` +写「生成中 · 已用 Xs」。卡片不必自己记「我刚点了哪一张」——`byOwner` 是顶栏那份任务表 +按 owner 建的索引,切换筛选、翻页、甚至整页刷新之后,正在生成那张卡照样亮着。 +列表视图的动作列同一套分轨(竖排三段,中间一条 `border-t` 就是那条分界线的竖排写法)。 + +**探针侧新增的盲区**:三轨栅格让 `btnclustercheck` 的「父元素是 grid → 不是它的对象」 +这条豁免正好接管卡底(三条轨各是 grid item),于是它量不到轨内。轨内的按钮各自 +关在 `.nex-action-group` 里,归 `btnrowcheck` 管。**别为了过闸把 ActionGroup 拆成裸按钮**—— +拆了以后它们成了 grid item 的兄弟,两道闸都扫不到,反而把盲区放大。 + --- ## 14. 说明文字与表单增删(G1/G2/G3/G4/G5) diff --git a/frontend/src/app/App.tsx b/frontend/src/app/App.tsx index 0713eba..1cd01d9 100644 --- a/frontend/src/app/App.tsx +++ b/frontend/src/app/App.tsx @@ -21,6 +21,7 @@ import { Navigate, Outlet, useLocation, useMatches } from 'react-router-dom' import { Button, Flex, Layout, Spin } from 'antd' import { Notice } from '../components/ui' import { IconReload } from '../components/icons' +import { startActivityFeed, stopActivityFeed } from '../stores/activity' import { authState, refreshState, useAuthStore, useIsAuthenticated } from '../stores/auth' import { loadSystem, useSystemStore } from '../stores/system' import { toggleTheme } from '../stores/theme' @@ -108,6 +109,15 @@ export default function App() { } }, [chromeless, isAuthenticated]) + /* 顶栏「平台动态」的订阅:整站只有这一条任务表轮询(有活 3s / 空闲 15s), + 各页面的自动刷新都挂在它落定的那一刻,不再各拉各的。 + 登录页不订阅:没鉴权时 /api/tasks/activity 会 401,那条 401 还会把人 + 从登录页踢去 next= 一圈(AuthGate 的 401 回调),像登录页自己抽风。 */ + useEffect(() => { + if (chromeless || !isAuthenticated) stopActivityFeed() + else startActivityFeed() + }, [chromeless, isAuthenticated]) + /* 全局快捷键:⌘/Ctrl+Shift+L 切明暗;输入框里打字时不抢键 */ useEffect(() => { function onKey(event: KeyboardEvent) { diff --git a/frontend/src/app/layout/ActivityCenter.tsx b/frontend/src/app/layout/ActivityCenter.tsx new file mode 100644 index 0000000..644237b --- /dev/null +++ b/frontend/src/app/layout/ActivityCenter.tsx @@ -0,0 +1,203 @@ +/** + * 顶栏「平台动态」:全站所有跑模型的任务在同一处看得见、点得动。 + * + * 症状(用户原话三条,2026-11): + * · 「当有 AI 正在生成资产时,没有地方可以看到平台正在做的操作」—— + * 出图 40 秒、出片 6 分钟,界面却只有一个转圈的按钮,甚至常常连它都没有; + * · 「很多资产的生成需要手动刷新页面」——生成早落定了,列表还停在点按钮之前; + * · 「是否应该有个统一的自动刷新开关」——刷新这件事此前散在 12 个页面里, + * 谁也关不掉。 + * 做法:后端把每一次跑模型都登记进统一任务表(``tasks``,平台生成与项目出片同源), + * 这一颗按钮是它的订阅端:数字 = 此刻进行中的条数,点开是全部进行中 + + * 最近 6 小时落定的流水,每条能取消 / 重试 / 跳回它所属的资产或项目; + * 全站自动刷新开关也落在这块面板上——它是「刷新策略」的唯一入口。 + * + * 为什么不放进侧栏或者做成一张页:这不是一个功能模块,是常驻的运行状态, + * 和顶栏已有的「在线/离线」是同一族信息(AppHeader 右侧那段)。 + * 数据由 stores/activity 单例轮询供给(有活 3 秒、空闲 15 秒), + * 这里只读 store,不再发第二份请求。 + */ +import { useMemo, useState } from 'react' +import { useNavigate } from 'react-router-dom' +import { Flex, Tooltip, Typography } from 'antd' +import { Badge, Button, Check, Drawer, Empty, ProgressBar, Spinner, Status, Tag } from '../../components/ui' +import { IconCaution, IconClock, IconClose, IconExternal, IconRetry } from '../../components/icons' +import { fmtAgo, fmtMs } from '../../lib/format' +import { cancel, refreshNow, retry, setAutoRefresh, useActivityStore, type ActivityTask } from '../../stores/activity' +import { notifyBad } from '../../stores/notice' + +const ACTIVE = ['queued', 'running', 'retrying'] + +/** 任务类型 → 一句话:面板上一眼分得清这是出图、出片还是配音 */ +const KIND_LABEL: Record = { + platform_asset: '平台资产生成', generate_shot: '镜头视频', voiceover: '配音', + qc_shot: '质量检查', compose: '时间线合成', asset: '素材处理', import: '导入' +} + +/** 母版生成跳回资产库并带上它的分类与名字;项目任务跳回调度中心 */ +function targetOf(task: ActivityTask) { + if (task.owner_type && task.owner_id) { + // 搜索词用母版名(登记时顺手存在 payload 里),不是 task.name: + // 「生成「俊杰」的定妆图」拿去搜资产库永远 0 条。 + // 改名之后点「查看」会搜不到——那时落在正确的分类里也比停在原页面有用。 + const q = new URLSearchParams({ type: task.owner_type }) + const name = String(task.payload?.owner_name || '').trim() + if (name) q.set('q', name) + return `/assets?${q.toString()}` + } + if (task.project_id) return `/projects/${task.project_id}/dispatch` + return '' +} + +function Row({ task }: { task: ActivityTask }) { + const navigate = useNavigate() + const [busy, setBusy] = useState('') + const active = ACTIVE.includes(task.status) + const href = targetOf(task) + const run = async (what: string, fn: () => Promise) => { + setBusy(what) + try { await fn() } catch (exc: any) { notifyBad({ title: what + '失败', desc: exc?.message || String(exc) }) } finally { setBusy('') } + } + + return ( +
+ {/* 第一行只回答「是哪件事、到哪一步了」:标题 + 状态,图标负责跑动 */} + + {active && } + + {task.name} + + + + + {/* 第二行是执行体自己写的那句「此刻在做什么」——排队、传参考图、 + 等算力锁、下载回传都在这一行,用户不必猜那 40 秒里发生了什么 */} + + {task.message || KIND_LABEL[task.kind] || task.kind} + + + {active && task.progress > 0 && } + + {/* 第三行:元数据 + 动作。元数据是「谁在跑、跑了多久」, + 动作只有取消 / 重试 / 跳过去看,全部是 text 档,不跟标题抢注意力 */} + + + {(task.provider || task.model) ? `${task.provider}/${task.model}` : KIND_LABEL[task.kind] || task.kind} + {/* 进行中说「已用多久」,落定说「什么时候的事」——同一句「已用 刚刚」两头不靠 */} + {active ? ` · 已用 ${fmtMs(task.elapsed_ms)}` : ` · ${fmtAgo(task.finished_at || task.updated_at)}`} + {task.attempts > 1 ? ` · 第 ${task.attempts} 次` : ''} + + {task.stale && active && ( + } + title={`执行体心跳已停止 ${Math.round((task.heartbeat_age_seconds ?? 0) / 60)} 分钟,平台会自动接管续跑`}> + 心跳中断 + + )} + + {active && ( + + )} + {!active && (task.status === 'failed' || task.status === 'cancelled') && ( + + )} + {href && ( + + )} + + + + {task.status === 'failed' && task.error_message && ( + + {task.error_code ? `${task.error_code}:` : ''}{task.error_message} + + )} +
+ ) +} + +export function ActivityCenter() { + const [open, setOpen] = useState(false) + const tasks = useActivityStore((s) => s.tasks) + const active = useActivityStore((s) => s.active) + const autoRefresh = useActivityStore((s) => s.autoRefresh) + + const { running, settled } = useMemo(() => ({ + running: tasks.filter((row) => ACTIVE.includes(row.status)), + settled: tasks.filter((row) => !ACTIVE.includes(row.status)).slice(0, 12) + }), [tasks]) + + const title = active > 0 ? `平台正在执行 ${active} 项任务` : '平台当前空闲' + + return ( + <> + 0 ? `${title} · 点击查看进度` : '平台动态 · 所有跑模型的任务都在这里'}> + {/* 空的时候它只是一颗安静的入口;有活时换转圈 + 计数, + 计数用 kit 的 Badge(顶栏流程控制台把 Tag 放按钮里是同一族写法) */} + + + + setOpen(false)} width={520} label="平台动态" + title="平台动态" + sub="所有跑模型的任务都在同一张任务表里排队:这里看得到它正在做什么、跑到哪、能不能停。" + actions={ + <> + + 自动刷新页面 + + + + }> +
+
+ + 进行中 + {running.length} + +
+ {running.map((task) => )} + {!running.length && ( +
+ 此刻没有任务在跑。发起一次生成,它会立刻出现在这里。 +
+ )} +
+
+ + {!!settled.length && ( +
+ + 最近落定 + {settled.length} + +
+ {settled.map((task) => )} +
+
+ )} + + {!tasks.length && 还没有生成记录} +
+
+ + ) +} + +export default ActivityCenter diff --git a/frontend/src/app/layout/AppHeader.tsx b/frontend/src/app/layout/AppHeader.tsx index 2cc58a0..daca77a 100644 --- a/frontend/src/app/layout/AppHeader.tsx +++ b/frontend/src/app/layout/AppHeader.tsx @@ -19,6 +19,7 @@ import { IconBook, IconMoon, IconSiderCollapse, IconSiderExpand, IconSun } from import { cn } from '../../lib/cn' import { toggleTheme, useAppliedTheme } from '../../stores/theme' import { useSystemStore } from '../../stores/system' +import ActivityCenter from './ActivityCenter' import PipelineConsole from './PipelineConsole' import UserMenu from './UserMenu' @@ -91,6 +92,10 @@ export function AppHeader({ projectId, collapsed, onCollapse, onOpenBrief, brief 后端离线} className="max-lg:hidden" /> )} + {/* 平台动态(跑模型的任务在做什么)不跟着 md 让位: + 窄屏它自己收成「转圈 + 计数」一颗图标,而它报的是「机器正占着活」—— + 这恰恰是手机上看板最需要知道的事,隐掉等于让人白等一次生成。 */} + {md && ( -
- label="上传" title="选完即作为新版本登记到这个母版下" - onFiles={(files) => files.forEach((file) => void uploadFile(row, file, uploadKindFor(row)))} /> + label="上传" + title="选完即作为新版本登记到这个母版下" + onFiles={(files) => files.forEach((file) => void uploadFile(row, file, uploadKindFor(row)))} /> +
-
-
- +
@@ -1293,6 +1335,7 @@ export default function AssetLibrary() { {rows.map((row: any, idx: number) => { const audios = (row.assets || []).filter(isAudio) + const generating = generatingOf(row) return ( )} {!row.media_count && 待补素材} + {!!generating && (} + title={generating.message || '正在生成,进度见顶栏「平台动态」'}> + 生成中 · {fmtMs(generating.elapsed_ms)} + )}
{row.code || `${type}-${row.id}`}
@@ -1388,20 +1435,30 @@ export default function AssetLibrary() { ) })} {!row.assets?.length && ( -
+
暂无素材:用下方「AI 生成」出图 / 出试听,或直接上传现成文件。
)} - {/* 动作在左、版本元数据在右:钉成两列栅格。 - 以前靠 BarRight 的 ml-auto 自然排列,一超宽那段元数据就独自 - 折到第二行的最右端(实测左边缘比同排按钮低 310px),像掉出去的孤儿。 */} -
- - + {/* 「传成哪一类」和「上传」是同一个动作的两半,拆到两行就读不出主谓 */}