yumoqing/pipeline-platform
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(platform): _capability_guard输出侧媒体词对称排除——t2i描述含'图像/图片'(输出)被误判图像输入→反复纠正成i2i且overrides.model_capability压不住(2026-09-07 qwen-image-plus实测:文档明写输入文本输出图像,助手override t2i被守卫每次改回i2i死循环);此前只排video未排image,补_OUTPUT_KIND对称表

Browse Source
This commit is contained in:
yumoqing 2026-09-07 18:18:42 +08:00
parent 31834c812e
commit 48735b98ba
1 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
pipeline_platform/platform_ability.py
View File

@ -497,7 +497,15 @@ def _capability_guard(capability, description):
return None, None
has_ref = any(w in desc for w in _DESC_REF_WORDS)
kinds = [k for k, words in _DESC_MEDIA_WORDS.items() if any(w in desc for w in words)]
input_kinds = [k for k in kinds if k != 'video']
# 输出侧媒体词对称排除（2026-09-07 qwen-image-plus 实测根因）：纯文本生成
# 能力的描述几乎必含其输出媒体词——t2v 必含「视频」、t2i 必含「图像/图片」，
# 那是输出声明不是输入声明。此前只排 video（video 在 t2v/i2v 描述必现），
# 没排 image（image 在 t2i 描述必现）→「千问系列图像生成模型」被误判有
# 图像输入 → t2i 被反复纠正成 i2i，且 overrides.model_capability 压不住
# （守卫在 apply 前置阶段每次又改回去，用户纠正陷入死循环）。
_OUTPUT_KIND = {'t2v': 'video', 't2i': 'image'}
out_kind = _OUTPUT_KIND.get(capability)
input_kinds = [k for k in kinds if k != out_kind]
if not (has_ref or input_kinds):
return None, None
if capability == 't2t':