| {# Direct-tokenization template for vLLM 0.27 embedding requests. | |
| Qwen's processor post-processor removes the role newline before images, | |
| keeps it before text/video, and appends <embedding> without a preceding | |
| newline. vLLM 0.27 adds each frame's vision boundaries itself, so video | |
| uses a bare <video_pad> placeholder. #} | |
| {%- for message in messages %} | |
| {{- '<|im_start|>' + message.role }} | |
| {%- if message.content is string %} | |
| {{- '\n' + message.content }} | |
| {%- else %} | |
| {%- for item in message.content %} | |
| {%- if item.type in ['image', 'image_url'] or 'image' in item or 'image_url' in item %} | |
| {{- '<|vision_start|><|image_pad|><|vision_end|>' }} | |
| {%- elif item.type in ['video', 'video_url'] or 'video' in item or 'video_url' in item %} | |
| {{- '\n<|video_pad|>' }} | |
| {%- elif item.type == 'text' or 'text' in item %} | |
| {{- ('\n' if loop.first else '') + item.text }} | |
| {%- else %} | |
| {{- raise_exception('Unsupported embedding content type') }} | |
| {%- endif %} | |
| {%- endfor %} | |
| {%- endif %} | |
| {{- '<|im_end|>' }} | |
| {%- if not loop.last %} | |
| {{- '\n' }} | |
| {%- endif %} | |
| {%- endfor %} | |
| {{- '<embedding>' }} | |