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 fab3e2a0bf
commit 524f13ac8a
13 files changed
+278 -49

No files matched your search

+29
View File
@@ -4,6 +4,35 @@
---
## 2026-09-16 · 集成智谱 CogView-3-Flash(文生图)+ CogVideoX-Flash(文生视频)
**一句话**:给上一轮的智谱供应商补上媒体生成——图像页用 `cogview-3-flash`、视频页用 `cogvideox-flash`,前后端按「选到的供应商」自动切到智谱契约(不动 Agnes)。
### 契约(官网 + 真 Key 实测确认)
- **图像 cogview-3-flash**:`POST {base}/images/generations`,`{model, prompt, size}`(size 为像素档:1024x1024/768x1344/864x1152/1344x768/1152x864/1440x720/720x1440)→ 返回 `data[].url`;**无 ratio、不支持图生图**。
- **视频 cogvideox-flash**:`POST {base}/videos/generations`(**异步**)`{model, prompt, quality, with_audio, size, fps:30/60}` → 返回 `task_status:PROCESSING`;再 `GET /async-result/{id}` 轮询到 `SUCCESS` 取 `video_result[].url`。仅文生视频。
### 后端
- `data/gateway.yaml` bigmodel 加 `media`(图像端点+像素档、视频 `/videos/generations`+`/async-result/{id}`,新增 `video_style:"zhipu"`),并注册 `cogview-3-flash`(image)、`cogvideox-flash`(video)(顺带修复本轮早前编辑误吞的 `server:` 顶层键)。
- `config.py`:`MediaConfig` 新增 `video_style` / `video_default_size`。
- 适配器 `create_video` / `poll_video`:按 `video_style=="zhipu"` 走智谱异步契约(fps/quality/with_audio + 路径参数轮询);Agnes 分支不变(`__init__.py` docstring 改四家)。
- `main.py`:新增 `_video_status`(兼容 `task_status`,智谱才识得 `SUCCESS/FAIL` 终态)与 `_video_terminal`,替换 3 处视频状态判定;`_validate_video_request` 按供应商分流(智谱仅 text 文生视频,size/quality/fps 合规校验,非法回干净 400 不进熔断);`videos_create` 改为先 `_resolve_media_model` 再按供应商校验。
- `settings.py`:面板 `MEDIA_KEYS` 补 `video_style`/`video_default_size`(否则「网关配置」面板往返会丢这两个字段)。
- 顺带修正 `DELETE /feedback` 过时 docstring(admin 可真删 admin_delete)。
### 前端(按所选模型 `provider==='bigmodel'` 自动切)
- `Videos.vue`:选智谱视频模型显示专用表单(分辨率/质量 speed·quality/帧率 30·60/配音 switch,固定文生视频),body 用 `extra:{quality,fps,with_audio}`;选 Agnes 原表单不变;切智谱模型时强制 `mode='text'`。
- `Images.vue`:选 cogview 时 size 换成像素档、隐藏 ratio 与参考图(cogview 无图生图)。
- `media.ts`:`findVideoUrl` 优先键补 `video_result`。
- i18n:新增 `videos.zSize/zQuality/zSpeed/zQualityHigh/zFps/zAudio/zTip` 7 键,中英对称。
### 验证
- 单测:新增/修正 bigmodel 媒体契约断言、视频轮询必经 `?model=`(现两家视频供应商)、配置往返含新 media 键 → **155 passed / 0 failed**;前端 `npm run build` 过(typecheck + i18n 对称),新入口 `index-CxdHy_Sw.js`。
- 真机隔离(uvicorn:8899 临时 users/history,连真实智谱):`/v1/models` 列出 8 个智谱模型;图像 → 200 `data[].url`;视频创建 → `video_id`+`task_status:PROCESSING`;轮询 → `status:processing`;非法参数(`mode=keyframe`)→ 400 干净拒绝。
- 上线:`supervisorctl restart api-gateway:api-gateway_00`(代码改动需重启),前端 dist 已 build 覆盖。
---
## 2026-09-16 · 新增智谱 GLM(open.bigmodel.cn)供应商
**一句话**:接入智谱开放平台作为新的 OpenAI 兼容供应商,收录 6 个 Flash 模型(3 文本 + 3 视觉),密钥走 `data/.env` 的 `ZHIPU_API_KEY`,真图实测确认 3 个 V 型号支持图像输入。
+1 -1
View File
@@ -46,7 +46,7 @@ api-gateway/
| Agnes AI | `https://api.agnes-ai.cn/v1` | Bearer | `agnes-3.0-flash`(text)、`agnes-image-2.5-flash`(image)、`agnes-video-2.5-flash`(video) |
| b.ai | `https://api.b.ai/v1` | Bearer | `qwen3.8-flash`、`hy3`(text) |
| 火山方舟 Coding | `https://ark.cn-beijing.volces.com/api/coding/v3` | Bearer | `ark-code-latest`(text) |
| 智谱 GLM | `https://open.bigmodel.cn/api/paas/v4` | Bearer | `GLM-4.7-Flash`、`GLM-4-Flash-250414`、`GLM-Z1-Flash`、`GLM-4.6V-Flash`、`GLM-4V-Flash`、`GLM-4.1V-Thinking-Flash`(text;其中 `GLM-4.6V-Flash`/`GLM-4V-Flash`/`GLM-4.1V-Thinking-Flash` 支持图像输入) |
| 智谱 GLM | `https://open.bigmodel.cn/api/paas/v4` | Bearer | 文本:`GLM-4.7-Flash`、`GLM-4-Flash-250414`、`GLM-Z1-Flash`、`GLM-4.6V-Flash`、`GLM-4V-Flash`、`GLM-4.1V-Thinking-Flash`(其中 3 个 V 型号支持图像输入);图像:`cogview-3-flash`;视频:`cogvideox-flash`(文生视频,异步 `/videos/generations` + `/async-result/{id}` 轮询) |
> Agnes 的图像/视频走自有端点,通过供应商配置里的 `media` 段指定。
+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")
+19 -2
View File
@@ -74,7 +74,19 @@ providers:
auth:
token_env: "ZHIPU_API_KEY"
extra_headers: {}
media: {}
# 智谱媒体生成(openai 兼容 + 异步查询):
# - 图像 cogview-3-flash:POST /images/generations,size 为像素档,返回 data[].url
# - 视频 cogvideox-flash:POST /videos/generations(异步),GET /async-result/{id} 轮询;
# 完成时 task_status=SUCCESS、video_result[].url。
# video_style: zhipu 让适配器按智谱契约构造视频 payload 与轮询路径(不走 agnes 那套 seconds/mode/aspect)。
media:
image_endpoint: "/images/generations"
image_default_size: "1024x1024"
video_create_endpoint: "/videos/generations"
video_poll_base: "https://open.bigmodel.cn/api/paas/v4"
video_poll_path: "/async-result"
video_style: "zhipu"
video_default_size: "1920x1080"
models:
- id: "GLM-4.7-Flash"
display_name: "GLM-4.7-Flash"
@@ -97,7 +109,12 @@ providers:
display_name: "GLM-4.1V-Thinking-Flash"
type: "text"
vision: true
- id: "cogview-3-flash"
display_name: "CogView-3-Flash"
type: "image"
- id: "cogvideox-flash"
display_name: "CogVideoX-Flash"
type: "video"
server:
host: "127.0.0.1"
port: 8808
+1 -1
View File
@@ -26,7 +26,7 @@ export function findVideoUrl(obj: any): string {
return ''
}
if (typeof obj === 'object') {
for (const k of ['video_url', 'play_url', 'url', 'html_url', 'video', 'media_url', 'result', 'data', 'mp4']) {
for (const k of ['video_result', 'video_url', 'play_url', 'url', 'html_url', 'video', 'media_url', 'result', 'data', 'mp4']) {
if (k in obj) { const u = findVideoUrl(obj[k]); if (u) return u }
}
for (const v of Object.values(obj)) { const u = findVideoUrl(v); if (u) return u }
+7
View File
@@ -303,6 +303,13 @@ export default {
videos: {
taskTitle: 'Tasks',
fixed: 'Fixed',
zSize: 'Resolution',
zQuality: 'Quality',
zSpeed: 'Speed first',
zQualityHigh: 'Quality first',
zFps: 'FPS',
zAudio: 'Audio',
zTip: 'Zhipu CogVideoX supports text-to-video only (text mode); generation takes a while, please wait.',
title: 'Video generation',
model: 'Video model',
mode: 'Scenario',
+7
View File
@@ -303,6 +303,13 @@ export default {
videos: {
taskTitle: '任务列表',
fixed: '固定',
zSize: '分辨率',
zQuality: '质量',
zSpeed: '速度优先',
zQualityHigh: '质量优先',
zFps: '帧率',
zAudio: '配音',
zTip: '智谱 CogVideoX 仅支持文生视频(text 模式);生成耗时较长,请耐心等待。',
title: '视频生成',
model: '视频模型',
mode: '场景',
+19 -5
View File
@@ -14,10 +14,10 @@
<div class="field">
<label>{{ t('images.size') }}</label>
<el-select v-model="size" class="fill">
<el-option v-for="s in sizes" :key="s" :value="s" :label="s" />
<el-option v-for="s in displaySizes" :key="s" :value="s" :label="s" />
</el-select>
</div>
<div class="field">
<div v-if="!isZhipu" class="field">
<label>{{ t('images.ratio') }}</label>
<el-select v-model="ratio" class="fill">
<el-option v-for="r in ratios" :key="r" :value="r" :label="r" />
@@ -44,7 +44,7 @@
/>
</div>
<div class="field mb-field">
<div v-if="!isZhipu" class="field mb-field">
<label>{{ t('images.refsLabel') }}</label>
<!-- 第十三轮:图生图 / 多图合成 —— 支持直接上传(浏览器内压成 data URI),也支持粘 URL -->
<div class="ref-btns">
@@ -166,8 +166,22 @@ const models = ref<any[]>([])
const model = ref('')
const sizes = ['1K', '2K', '3K', '4K']
const ratios = ['1:1', '3:4', '4:3', '16:9', '9:16', '2:3', '3:2', '21:9']
// 智谱 CogView-3-Flash:文生图,size 用像素档、天然无 ratio、不支持参考图(图生图)
const zhipuSizes = ['1024x1024', '768x1344', '864x1152', '1344x768', '1152x864', '1440x720', '720x1440']
const isZhipu = computed(() => models.value.find((m: any) => m.id === model.value)?.provider === 'bigmodel')
const displaySizes = computed(() => (isZhipu.value ? zhipuSizes : sizes))
const size = ref('1K')
const ratio = ref('1:1')
watch(model, (m) => {
const mo = models.value.find((x: any) => x.id === m)
const z = mo?.provider === 'bigmodel'
if (z) {
if (!zhipuSizes.includes(size.value)) size.value = '1024x1024'
ratio.value = '' // cogview 的宽高由 size 像素档决定,无 ratio
} else if (ratio.value === '') {
ratio.value = '1:1'
}
})
const count = ref(1) // 逐张生成:每张一次请求(上游 n 必须为 1)
const prompt = ref('')
// 参考图(图生图 / 多图合成):上传的走 data URI,URL 的直接透传,两类可混用
@@ -274,8 +288,8 @@ async function generate() {
progress.value = i
const t0 = performance.now()
// 上游图像接口的 n 必须为 1:多张就在这里逐张调用,每张出来就上屏(进度可见)
const body: any = { model: model.value, prompt: prompt.value, size: size.value, ratio: ratio.value, n: 1 }
if (list.length) body.image = list // 图生图 / 多图合成(网关转成上游的 extra_body.image)
const body: any = { model: model.value, prompt: prompt.value, size: size.value, ratio: isZhipu.value ? '' : ratio.value, n: 1 }
if (list.length && !isZhipu.value) body.image = list // 图生图 / 多图合成(网关转成上游的 extra_body.image)
const j = await api('/v1/images/generations', { method: 'POST', body: JSON.stringify(body) })
const ms = (performance.now() - t0) | 0
providerName.value = j.provider_name || j.provider || ''
+69 -17
View File
@@ -10,7 +10,9 @@
</el-select>
</div>
<!-- 第十一轮 #3:场景模式(对齐 Agnes Video 契约 text/keyframe/reference) -->
<!-- 智谱 CogVideoX-Flash / Agnes 两种视频契约:Agnes 用模式+时长+720P;智谱用 quality/fps/audio/像素 size -->
<template v-if="!isZhipu">
<!-- 第十一轮 #3:场景模式(对齐 Agnes Video 契约 text/keyframe/reference) -->
<div class="field mb-field">
<label>{{ t('videos.mode') }}</label>
<div class="modes">
@@ -54,6 +56,39 @@
:placeholder="t('videos.seedPH')" class="fill" />
</div>
</div>
</template>
<!-- 智谱 CogVideoX-Flash:文生视频专用(quality/fps/with_audio/像素 size,固定 text 模式) -->
<template v-else>
<div class="form-row mb-field">
<div class="field">
<label>{{ t('videos.zSize') }}</label>
<el-select v-model="zsize" class="fill">
<el-option v-for="s in zhipuSizes" :key="s" :value="s" :label="s" />
</el-select>
</div>
<div class="field">
<label>{{ t('videos.zQuality') }}</label>
<el-select v-model="zquality" class="fill">
<el-option value="speed" :label="t('videos.zSpeed')" />
<el-option value="quality" :label="t('videos.zQualityHigh')" />
</el-select>
</div>
</div>
<div class="form-row mb-field">
<div class="field">
<label>{{ t('videos.zFps') }}</label>
<el-select v-model="zfps" class="fill">
<el-option :value="30" label="30fps" />
<el-option :value="60" label="60fps" />
</el-select>
</div>
<div class="field">
<label>{{ t('videos.zAudio') }}</label>
<el-switch v-model="zaudio" />
</div>
</div>
<div class="hint tip">{{ t('videos.zTip') }}</div>
</template>
<div class="field mb-field">
<label>{{ t('videos.promptRow') }}</label>
@@ -220,7 +255,7 @@
</template>
<script setup lang="ts">
import { computed, onMounted, onUnmounted, ref } from 'vue'
import { computed, onMounted, onUnmounted, ref, watch } from 'vue'
import { useRouter } from 'vue-router'
import { useI18n } from 'vue-i18n'
import { ElMessage, ElMessageBox } from 'element-plus'
@@ -256,6 +291,13 @@ const router = useRouter()
const { t } = useI18n()
const models = ref<any[]>([])
const model = ref('')
// 所选视频模型是否智谱 CogVideoX(文生视频专用表单)
const isZhipu = computed(() => models.value.find((m: any) => m.id === model.value)?.provider === 'bigmodel')
// 切到智谱视频模型时强制 text 文生视频模式(智谱无 keyframe/reference)
watch(model, (m) => {
const mo = models.value.find((x: any) => x.id === m)
if (mo?.provider === 'bigmodel') mode.value = 'text'
})
const prompt = ref('')
const seconds = ref('5')
const aspect = ref('16:9')
@@ -263,6 +305,12 @@ const quality = ref('720P')
const seed = ref<number | null>(null)
// 第十一轮 #3:场景模式与对应素材
const mode = ref<'text' | 'keyframe' | 'reference'>('text')
// 智谱 CogVideoX-Flash:text 文生视频专用参数(quality/fps/with_audio/像素 size)
const zhipuSizes = ['720x480', '1024x1024', '1280x960', '960x1280', '1920x1080', '1080x1920', '2048x1080']
const zsize = ref('1920x1080')
const zquality = ref<'speed' | 'quality'>('speed')
const zfps = ref(30)
const zaudio = ref(true)
const frameEl = ref<HTMLInputElement | null>(null)
const refImgEl = ref<HTMLInputElement | null>(null)
const refAudEl = ref<HTMLInputElement | null>(null)
@@ -466,21 +514,25 @@ async function create() {
const t0 = Date.now()
const summary = mediaSummary()
try {
const body: any = {
model: model.value,
prompt: prompt.value,
seconds: seconds.value,
size: quality.value,
aspect_ratio: aspect.value,
mode: mode.value,
}
if (seed.value != null) body.seed = seed.value
if (mode.value === 'keyframe') {
if (firstFrame.value) body.first_frame = firstFrame.value
if (lastFrame.value) body.last_frame = lastFrame.value
} else if (mode.value === 'reference') {
if (imageList.value.length) body.images = imageList.value
if (audioList.value.length) body.audios = audioList.value
const body: any = { model: model.value, prompt: prompt.value }
if (isZhipu.value) {
// 智谱 CogVideoX-Flash:text 文生视频;quality/fps/with_audio 走 extra(后端透传给上游)
body.size = zsize.value
body.mode = 'text'
body.extra = { quality: zquality.value, fps: zfps.value, with_audio: zaudio.value }
} else {
body.seconds = seconds.value
body.size = quality.value
body.aspect_ratio = aspect.value
body.mode = mode.value
if (seed.value != null) body.seed = seed.value
if (mode.value === 'keyframe') {
if (firstFrame.value) body.first_frame = firstFrame.value
if (lastFrame.value) body.last_frame = lastFrame.value
} else if (mode.value === 'reference') {
if (imageList.value.length) body.images = imageList.value
if (audioList.value.length) body.audios = audioList.value
}
}
if (!stream.value) body.wait = true // 同步等结果
const j = await api('/v1/videos', { method: 'POST', body: JSON.stringify(body) })
+23 -10
View File
@@ -94,14 +94,26 @@ def test_load_config_providers():
# 智谱:base_url + token_env + 6 个模型 + V 型号视觉标记
assert cfg.providers["bigmodel"].base_url == "https://open.bigmodel.cn/api/paas/v4"
assert cfg.providers["bigmodel"].auth.token_env == "ZHIPU_API_KEY"
bm_ids = [m.id for m in cfg.providers["bigmodel"].models]
assert bm_ids == ["GLM-4.7-Flash", "GLM-4-Flash-250414", "GLM-Z1-Flash",
"GLM-4.6V-Flash", "GLM-4V-Flash", "GLM-4.1V-Thinking-Flash"]
for m in cfg.providers["bigmodel"].models:
if m.id in ("GLM-4.6V-Flash", "GLM-4V-Flash", "GLM-4.1V-Thinking-Flash"):
assert m.vision is True
else:
assert m.vision is False
bm = cfg.providers["bigmodel"]
text_ids = [m.id for m in bm.models if m.type == "text"]
assert text_ids == ["GLM-4.7-Flash", "GLM-4-Flash-250414", "GLM-Z1-Flash",
"GLM-4.6V-Flash", "GLM-4V-Flash", "GLM-4.1V-Thinking-Flash"]
# 智谱媒体模型:cogview-3-flash(图像)、cogvideox-flash(视频)
assert [m.id for m in bm.models if m.type == "image"] == ["cogview-3-flash"]
assert [m.id for m in bm.models if m.type == "video"] == ["cogvideox-flash"]
for m in bm.models:
if m.type != "text":
continue
want = m.id in ("GLM-4.6V-Flash", "GLM-4V-Flash", "GLM-4.1V-Thinking-Flash")
assert m.vision is want, f"{m.id} vision 应为 {want}"
# 智谱媒体契约(图像像素档 + 视频 zhipu 异步风格)
md = bm.media
assert md.supports_image and md.supports_video
assert md.image_default_size == "1024x1024"
assert md.video_style == "zhipu"
assert md.video_default_size == "1920x1080"
assert md.video_create_endpoint == "/videos/generations"
assert md.video_poll_path == "/async-result"
def test_load_config_routes():
@@ -2244,8 +2256,9 @@ def test_video_polls_merged_into_one_log_row_with_response():
"prompt": "x", "aspect_ratio": "16:9"}, headers=ah)
assert r.status_code == 200, r.text
vid = r.json()["video_id"]
client.get(f"/v1/videos/{vid}", headers=ah) # 第 1 次轮询 processing
client.get(f"/v1/videos/{vid}", headers=ah) # 第 2 次轮询 completed
poll_q = f"/v1/videos/{vid}?model=agnes-video-2.5-flash" # 多视频供应商(agnes+智谱)时必须显式指定 model
client.get(poll_q, headers=ah) # 第 1 次轮询 processing
client.get(poll_q, headers=ah) # 第 2 次轮询 completed
logs = client.get("/logs?n=200", headers=ah).json()["entries"]
rows = [e for e in logs if e.get("video_id") == vid]
assert len(rows) == 1, f"轮询应合成到创建那条里(实际 {len(rows)} 行): {[r['path'] for r in logs]}"