Replies: 2 comments
|
这个"卡死"是可以理解的:图片一旦进了会话,就留在会话日志里了。所以第一次超过 5 张触发上限之后,后面每一轮请求都还带着那批图片,于是每次都撞同一堵墙——不是这一轮失败,是这条会话被钉住了。 先救急:这条会话救不回来,新开一个会话继续(旧的归档掉)。 再避免复发,三条按代价从低到高:
第 3 条我有个现成的实现(pi2dsh 的伴生路由 + 但要说清两点,别让我误导你:
利益相关:pi2dsh 是我维护的, |
|
Same failure class here — the GLM gateway caps at 8 images per request (yours: 5 via Claude), and the session gets permanently pinned because replayed context keeps every accumulated image; the gateway only returns a generic 400 hiding the real reason. I patched Settings example (fork patch — apply the branch, llm-pi-ai:
providers:
your-provider:
apiKeyEnv: YOUR_API_KEY_ENV
api: openai-completions
baseURL: https://your-gateway.example.com/v1
maxRequestImages: 5 # match your gateway's per-request image cap
models:
- id: your-vision-model
input: [text, image] # required for images to be sent at allBranch (reference / cherry-pick): https://github.com/ahrar-deriv/deepseek-harness/tree/request-image-count-cap — personal fork patch, not upstreamed yet. Until patched, the workarounds above (one-image-at-a-time tooling, image→text routing) are the stable options. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
一个目录下有多个图片,让他调用模型理解图片内容,他会把自己卡死。因为模型限制一次最多少能处理5张。
然后这个会话就等于死这了。
At most 5 image(s) may be provided in one prompt. (parameter=image)

All reactions