Your current environment

n/a - not an env specifc issue

🐛 Describe the bug

[Bug] /v2/embed with input_type returns misleading 400 on nemotron-embed-vl

Summary

The Cohere /v2/embed path synthesizes [system, user] messages internally when input_type maps to a task_instructions prefix (#38362). Chat templates that guard against messages | length > 1 then reject this 2-message input with an error that
reads like a caller-side problem. Today this breaks nvidia/llama-nemotron-embed-vl-1b-v2.

Reproducer

vllm serve nvidia/llama-nemotron-embed-vl-1b-v2 \
  --trust-remote-code --max-model-len 10240 \
  --chat-template examples/pooling/embed/template/nemotron_embed_vl.jinja

curl -sS -X POST http://localhost:8000/v2/embed \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/llama-nemotron-embed-vl-1b-v2",
    "texts": ["machine learning", "the cat sat on the mat"],
    "input_type": "query",
    "embedding_types": ["float"]
  }'

{"error":{"message":"Embedding models should only embed one message at a time","type":"BadRequestError","code":400}}

Root cause

EmbedIOProcessor._mixed_input_to_messages (vllm/entrypoints/pooling/embed/io_processor.py:298-321) builds a 2-message list [{role:"system", content:task_prefix}, {role:"user", content:text}] whenever a task_prefix is present. _pre_process_coh\ ere_online then routes that through _batch_render_chat, where the template's {% if messages | length > 1 %}{{ raise_exception(...) }}{% endif %} guard fires. The caller never sent messages and has no way to know the server is responsible for the 2-message shape.

Scope

The failure requires both: model declares task_instructions and is served with a length-guarded template. Bundled templates today:

Nemotron-embed-vl is the only model that currently hits both conditions. Future multimodal embedding models that declare task_instructions and ship with a guarded template will break on arrival.

Proposed direction

Move the multi-message policy out of jinja and into EmbedIOProcessor. Declare per-model/template capability (e.g. supports_multi_message_chat) and check it in _pre_process_cohere_online before rendering. For incompatible combinations, either fal
l back to inlining the task prefix into user content (pre-#38362 behavior) or return an error that names the real cause.

Related observations (not this issue)

• The error string "Embedding models should only embed one message at a time" is repeated across the three guarded templates and reads as universal — worth rewording independently of the fix.
• template_basic.jinja silently concatenates multi-message input for CLIP/Siglip/ColPali/PaliGemma/Chameleon, which is a separate footgun that may deserve its own issue.

Out of scope

Changing the /v1/embeddings messages contract (single conversation → single embedding). See #40042 discussion.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.