feat: 集成智谱 CogView-3-Flash 文生图 + CogVideoX-Flash 文生视频

This commit is contained in:
api-gateway committed 2026-09-16 10:11:20 +08:00
1 parent c58b29a06f
commit 403676619e
13 files changed
+278 -49

No files matched your search

+32
View File
@@ -12,6 +12,7 @@ from __future__ import annotations
import json
from typing import Any, AsyncIterator
from urllib.parse import quote
from app.adapters.base import BaseAdapter, json_or_raise
@@ -184,6 +185,26 @@ class OpenAICompatAdapter(BaseAdapter):
size: str = "720P",
extra: dict[str, Any] | None = None,
) -> dict[str, Any]:
# 智谱视频契约:POST /videos/generations(异步)——参数为 quality/with_audio/fps/像素 size,
# 不需要 agnes 的 seconds/mode/aspect_ratio/n。
if self.media.video_style == "zhipu":
endpoint = self.media.video_create_endpoint or "/videos/generations"
x = extra or {}
payload: dict[str, Any] = {
"model": model,
"prompt": prompt,
"size": size or self.media.video_default_size or "1920x1080",
"quality": x.get("quality") or "speed",
"fps": int(x.get("fps") or 30),
"with_audio": bool(x.get("with_audio") or False),
}
async with self._lease() as client:
resp = await client.post(endpoint, json=payload, headers=self._headers)
resp.raise_for_status()
data = json_or_raise(resp)
return {"video_id": data.get("id") or data.get("request_id") or data.get("task_id"),
"raw": data}
# ── Agnes 风格(默认):POST {base}/videos,seconds/mode/size/aspect_ratio/n ──
endpoint = self.media.video_create_endpoint or "/videos"
payload: dict[str, Any] = {
"model": model,
@@ -211,6 +232,17 @@ class OpenAICompatAdapter(BaseAdapter):
# ── 视频轮询 ────────────────────────────────────
async def poll_video(self, video_id: str, model: str = "", extra: dict[str, Any] | None = None) -> dict[str, Any]:
# 智谱视频异步查询:GET {poll_base}/async-result/{video_id}(路径参数,无 query)
if self.media.video_style == "zhipu":
pb = (self.media.video_poll_base or self.base_url).rstrip("/")
pp = (self.media.video_poll_path or "/async-result").rstrip("/")
url = f"{pb}{pp}/{quote(str(video_id), safe='')}"
async with self._lease() as client:
resp = await client.get(url, headers=self._headers)
resp.raise_for_status()
data = json_or_raise(resp)
return {"raw": data}
mname = self.media.video_model_name or model or ""
poll_base = self.media.video_poll_base or self.base_url
poll_path = self.media.video_poll_path or "/agnesapi"
+5 -1
View File
@@ -54,7 +54,9 @@ class MediaConfig:
video_create_endpoint: str | None = None # 相对 base_url,默认 /videos
video_poll_base: str | None = None # 绝对 base(视频查询,默认同 base_url 去掉 /v1)
video_poll_path: str | None = None # 默认 /agnesapi(Agnes)
video_model_name: str | None = None # 轮询用的 model_name 参数
video_model_name: str | None = None # 轮询用的 model_name 参数(Agnes)
video_style: str | None = None # 视频契约风格:"zhipu"=智谱异步(videos/generations + /async-result/{id});None=Agnes 风格
video_default_size: str | None = None # 视频默认尺寸(智谱为像素串如 1920x1080;Agnes 用 size 档位)
@property
def supports_image(self) -> bool:
@@ -304,6 +306,8 @@ def _parse_media(pval: dict[str, Any]) -> MediaConfig:
video_poll_base=m.get("video_poll_base"),
video_poll_path=m.get("video_poll_path"),
video_model_name=m.get("video_model_name"),
video_style=m.get("video_style"),
video_default_size=m.get("video_default_size"),
)
+64 -12
View File
@@ -1729,8 +1729,8 @@ async def _video_wait_and_callback(cfg, pkey, pconfig, pmodel_id, video_id, wait
await asyncio.sleep(2)
continue
raw = final.get("raw") or {}
status = raw.get("status") or raw.get("state")
if status in ("completed", "failed", "succeeded", "success"):
status = _video_status(raw)
if _video_terminal(status):
break
await asyncio.sleep(2)
# Webhook 回调
@@ -1745,8 +1745,8 @@ async def _video_wait_and_callback(cfg, pkey, pconfig, pmodel_id, video_id, wait
try:
# Bug 2 修复: 按真实状态上报,failed 不再误标成 completed
raw = (final.get("raw") if final else None) or {}
status = str(raw.get("status") or raw.get("state") or "")
if status not in ("completed", "failed", "succeeded", "success"):
status = _video_status(raw)
if not _video_terminal(status):
status = "timeout"
payload = {"video_id": video_id, "provider": pkey, "status": status,
"result": raw, "x_request_id": rid}
@@ -1764,12 +1764,63 @@ _VIDEO_SIZES = ("720P",)
_VIDEO_ASPECTS = ("21:9", "16:9", "4:3", "1:1", "3:4", "9:16")
def _validate_video_request(req) -> None:
def _video_status(raw: dict | None) -> str:
"""视频原始返回的状态取法(统一小写)。
智谱异步结果是 task_status(PROCESSING/SUCCESS/FAIL),Agnes 用 status / state。
"""
if not raw:
return ""
return str(raw.get("task_status") or raw.get("status") or raw.get("state") or "").lower()
def _video_terminal(status: str) -> bool:
"""是否已到终态(成功或失败),用于结束轮询。"""
s = status.lower()
return s in ("success", "succeeded", "completed", "done", "finished",
"failed", "fail", "error", "cancelled", "canceled")
# 智谱 CogVideoX-Flash 支持的分辨率(像素串),与 Agnes 的 720P 档位语义不同。
_ZHIPU_VIDEO_SIZES = {
"720x480", "1024x1024", "1280x960", "960x1280",
"1920x1080", "1080x1920", "2048x1080",
}
def _validate_video_request(req, media=None) -> None:
"""第十一轮 #3:按 Agnes Video 2.5 契约做**入站校验**,把明显不合法的 mode/媒体/时长
组合在网关挡下并回**干净 400**(不再白发上游浪费限流额度,也不进熔断)。
第十三轮补齐 size / aspect_ratio / n / videos 四项(文档「Flash 专属错误」与
「生成模式规则」里明确列出的限制),并保留原有 mode/媒体/时长校验。
目前只有 Agnes 配了视频,字段语义按它的文档来;接入其它视频供应商时需按各家契约放宽。"""
# ── 智谱 CogVideoX(CogVideoX-Flash):仅 text 文生视频;quality/fps/with_audio 走 req.extra ──
if media is not None and media.video_style == "zhipu":
if (req.mode or "text") != "text":
raise HTTPException(400, "智谱视频(CogVideoX)仅支持 text 文生视频模式")
if int(req.n or 1) != 1:
raise HTTPException(400, "视频接口一次只提交 1 个任务(n 必须为 1,收到 %s)" % req.n)
if req.videos or req.images or req.audios or req.first_frame or req.last_frame:
raise HTTPException(400, "智谱视频不支持首尾帧 / 参考图 / 参考音频(文生视频只填 prompt)")
if req.size and req.size not in _ZHIPU_VIDEO_SIZES:
raise HTTPException(400, "智谱视频 size 仅支持 %s,收到 %r" %
("/".join(sorted(_ZHIPU_VIDEO_SIZES)), req.size))
_x = req.extra or {}
_q = _x.get("quality")
if _q is not None and _q not in ("speed", "quality"):
raise HTTPException(400, "智谱视频 quality 仅支持 speed / quality,收到 %r" % _q)
_f = _x.get("fps")
if _f is not None:
try:
_fi = int(_f)
except (TypeError, ValueError):
_fi = -1
if _fi not in (30, 60):
raise HTTPException(400, "智谱视频 fps 仅支持 30 / 60,收到 %r" % _f)
return
# ── Agnes 契约(默认)───────────────────────────────
mode = req.mode or "text"
if mode not in _VIDEO_MODES:
raise HTTPException(400, f"视频 mode 仅支持 {'/'.join(_VIDEO_MODES)},收到 {mode!r}")
@@ -1811,12 +1862,13 @@ async def videos_create(request: Request, req: VideoCreateRequestExtended, _auth
cfg = _cfg()
rid = getattr(request.state, "rid", "-")
caller = getattr(request.state, "caller", "anonymous")
_validate_video_request(req) # 第十一轮 #3:入站按上游契约校验 mode/媒体/时长
# Bug 4: 支持路由别名(如 wjd)→ 解析到第一个 type==video 的模型
# Bug 4: 支持路由别名(如 wjd)→ 解析到第一个 type==video 的模型。
# 解析放在入站校验之前,好让 _validate_video_request 按供应商媒体契约(Agnes / 智谱)分流。
resolved = _resolve_media_model(cfg, req.model, "video")
if resolved is None:
raise HTTPException(404, f"未知视频模型/路由: {req.model}(没有可用的 video 类型模型)")
pkey, pconfig, minfo, token, real_model = resolved
_validate_video_request(req, pconfig.media) # 入站按供应商媒体契约校验
# 顺序修复(本轮):先能力检查(501),再校验回调地址(400),最后限流 —— 避免无效请求白扣 token。
if not pconfig.media.supports_video:
raise HTTPException(501, f"供应商 {pconfig.name} 不支持视频创建")
@@ -1872,7 +1924,7 @@ async def videos_create(request: Request, req: VideoCreateRequestExtended, _auth
# 本轮修复: status 同时看 status/state(此前只看 status,而上游若用 state
# 会被显示成 "pending" —— 即使其实已完成,与等待循环的判定不一致)
raw_final = (final.get("raw") or {}) if final else {}
resp["status"] = (raw_final.get("status") or raw_final.get("state") or "pending") if final else "submitted"
resp["status"] = (_video_status(raw_final) or "pending") if final else "submitted"
elif req.callback_url:
# 第六轮修复: 此前只传 callback_url(不传 wait)时,_video_wait_and_callback 第一行
# 就 `if not wait: return None` → **webhook 被静默丢弃、永不回调**,而 README 却把
@@ -2026,8 +2078,8 @@ async def videos_poll(request: Request, video_id: str, model: str = "", wait: bo
log.warning("[rid=%s] 轮询视频 %s 瞬时失败: %s", rid, _logsafe(video_id), e)
await asyncio.sleep(2)
continue
status = (result.get("raw") or {}).get("status") or (result.get("raw") or {}).get("state")
if status in ("completed", "failed", "succeeded", "success"):
status = _video_status(result.get("raw"))
if _video_terminal(status):
break
await asyncio.sleep(2)
if result is None:
@@ -2043,12 +2095,12 @@ async def videos_poll(request: Request, video_id: str, model: str = "", wait: bo
# 第十轮: 轮询不单独占一行日志,改为按 video_id 快照累积,由 /logs 合成到创建那条里
video_poll_push(video_id, {
"ts": round(time.time(), 3), "ok": True,
"status": (raw.get("status") or raw.get("state") or ""),
"status": _video_status(raw),
"progress": raw.get("progress"),
"raw": json.dumps(raw, ensure_ascii=False)[:1200],
})
return {"video_id": video_id, "provider": pkey, "served_model": pmodel_id, "raw": raw,
"status": raw.get("status") or raw.get("state"), "x_request_id": rid}
"status": _video_status(raw), "x_request_id": rid}
except AuthConfigError as e:
cb.release_probe()
raise HTTPException(503, f"供应商 {pconfig.name} 未配置可用凭证: {e}")
+2
View File
@@ -61,6 +61,8 @@ MEDIA_KEYS = (
"video_poll_base",
"video_poll_path",
"video_model_name",
"video_style",
"video_default_size",
)
MODEL_TYPES = ("text", "image", "video")