Error messages

As of transformers v4.44, default chat template is no longer allowed, so you must provide a chat template if the tokenizer does not define one.

As of transformers v4.44, default chat template is no longer allowed, so you must provide a chat template if the tokenizer does not define one.

The server starts fine, then the first request to /v1/chat/completions fails and the log shows this:

vllm.entrypoints.chat_utils.ChatTemplateResolutionError: As of transformers v4.44, default chat template is no longer allowed, so you must provide a chat template if the tokenizer does not define one.

Older releases print the same text as a plain ValueError. A chat template is a small Jinja program, usually stored with the tokenizer, that turns your list of messages into the single string the model was trained to read. Transformers used to guess one when a model had none. It stopped doing that in v4.44, and vLLM keeps the wording, so without a template there's nothing to build the prompt from. We read the code at vLLM 0.30.0 (September 22, 2026), and the string still lives in vllm/renderers/hf.py.

Where vLLM looks before it gives up

It tries four places in order. First, a template you passed with --chat-template or in the request. Then the model's processor (skipped when the request carries tools), then the tokenizer. Last, vLLM has a short built-in list keyed on model type, mostly vision and OCR models like PaliGemma, SigLIP, BLIP-2 and DeepSeek-OCR. Only if all four come back empty do you get this error.

Why does the server start at all? vLLM tries a warm-up chat render at startup, but when it hits this error it only logs "This model does not support chat template." at debug level and carries on. So the problem shows up on the first real request, as an HTTP 400 with type BadRequestError.

The usual causes

A base model is the commonest one. Pretrained checkpoints often ship no template because they were never trained on chat turns. In issue #39279 (April 2026) someone served the base gemma-4-31B and the reply from another user was to use gemma-4-31B-it instead. If you actually want raw text continuation from a base model, use the completions route: the vLLM docs say llm.generate doesn't apply a chat template, while llm.chat does.

Some chat models arrive without a Jinja template at all. DeepSeek-V3.2 hit this on December 2, 2025 in #29849, and a maintainer replied with PR #29837, which added a V3.2 chat template to vLLM itself and was merged on December 3, 2025. For a brand-new model, check whether a newer vLLM already carries its template before writing one.

A broken download can do it too. The reporter of #18140, on Llama-3.2-3B-Instruct, closed it after re-downloading: the first copy was missing config files, and the tokenizer config is where the template lives.

How to pass a template with --chat-template

The vLLM docs give this form, which takes a file path or the template itself as a string:

vllm serve <model> --chat-template ./path-to-chat-template.jinja

The repo's examples folder has ready-made ones such as template_chatml.jinja and template_alpaca.jinja. A bare file name that isn't on disk is also looked up in vLLM's own bundled template folder, so --chat-template template_chatml.jinja works from any directory. Pick the format the model was actually fine-tuned on (the model card usually says). For Vicuna, a maintainer's advice in #17977 was to write your own Jinja from the examples or borrow one from a Hugging Face repo that already defines it.

A wrong template won't raise this error. It'll just produce worse answers, so check the model's format before reaching for ChatML by default.

Demand, measured September 26, 2026: 21 issues in vllm-project/vllm quote the full message, and 2,059 mention "chat template".