Inconsistent chat templates across the 3 models

[yeahdongcn]

I'd like to add support for Seed-Coder in llama.cpp: ggml-org/llama.cpp#13472, but I noticed that the chat templates differ across the three related models. Is this inconsistency expected, or should they be unified?

Base: {% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set role = message['role'] %}{{ bos_token + role + '\n' + message['content'] | trim + eos_token }}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\n'}}{% endif %}"

Instruct: {% set has_system = messages[0]['role'] == 'system' %}{% if not has_system %}{{ bos_token + 'system\nYou are an AI programming assistant, utilizing the Seed-Coder model, developed by ByteDance Seed, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer.\n\n' + eos_token }}{% endif %}{% for message in messages %}{{ bos_token + message['role'] + '\n' + message['content'] | trim + eos_token }}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\n'}}{% endif %}

Reasoning: {% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set role = message['role'] %}{{ bos_token + role + '\n' + message['content'] | trim + eos_token }}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\n'}}{% endif %}

[CISC]

They are also completely wrong. :)

The Base model shouldn't have a chat template , and the other two are not setting up an actual chat.

[shenz2-2000]

Thanks for raising this! Yes, this inconsistency is expected.

To clarify: the base model was not trained for chat and only supports completion-style prompting, so the chat_template you see there is just our default placeholder and isn't actively used. In contrast, the instruct and reasoning models were trained with different system prompts to reflect their intended use cases—so their templates differ accordingly.

Appreciate your interest in supporting Seed-Coder in llama.cpp—let me know if I can help with anything else!

[CISC]

To clarify: the base model was not trained for chat and only supports completion-style prompting, so the chat_template you see there is just our default placeholder and isn't actively used.

It is used (I know, for transformers you have to apply it, but this is not necessarily the case elsewhere), you should remove it completely so it's not auto-detected in other frameworks!

[shenz2-2000]

To clarify: the base model was not trained for chat and only supports completion-style prompting, so the chat_template you see there is just our default placeholder and isn't actively used.

It is used (I know, for transformers you have to apply it, but this is not necessarily the case elsewhere), you should remove it completely so it's not auto-detected in other frameworks!

Thanks for pointing that out.
I understand that including it can lead to confusion or unintended behavior in other frameworks where templates are auto-detected. We'll go ahead and remove the chat_template field from the base model to avoid any ambiguity.
Appreciate the careful review!