Add Qwen-3 chat template and Ollama template support

[kiankyars]

This pull request is in regards to GH-2521.

I have added the templates for both qwen3 and qwen3 ollama.

Please let me know if there's anything else to add since I'd like to use these templates ASAP.

[kiankyars]

@davedgd and @shimmyshimmer Just tagging you guys since I'm not sure if you get a notification from the issue itself.

[davedgd]

@davedgd and @shimmyshimmer Just tagging you guys since I'm not sure if you get a notification from the issue itself.

Thank you @kiankyars — much appreciated!

[Datta0]

Hey @kiankyars , thanks for this.
Can you please tell me what is the requirement for having chat template here? We can directly access it via the tokenizer right?

I understand if ollama_template is included as it might not be easily accessible.
Please correct me if I'm wrong

[kiankyars]

Hey @kiankyars , thanks for this. Can you please tell me what is the requirement for having chat template here? We can directly access it via the tokenizer right?

I understand if ollama_template is included as it might not be easily accessible. Please correct me if I'm wrong

You raise a good point. The qwen3_template is included mostly to ensure consistency with other templates. in chat_templates.py

[davedgd]

@kiankyars: There seems to be issue with the qwen-3 template that was added. Specifically, it reliably produces the following error:

TemplateSyntaxError: expected token 'end of print statement', got 'name'

I've created a minimal example here that works with qwen-2.5 but fails with qwen-3:

from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-0.6B-unsloth-bnb-4bit",
    max_seq_length = 8192,
    load_in_4bit = True,
)

tokenizer = get_chat_template(
    tokenizer,
    chat_template = "qwen-3", # works with qwen-2.5 but not qwen-3
)

messages = [{"role": "system", "content": "You are an expert assistant."}, 
            {"role": "user", "content": "Say Hello!"}, 
            {"role": "assistant", "content": "Hello!"}]

inputs = tokenizer.apply_chat_template(messages, 
                                       tokenize = False, 
                                       add_generation_prompt = False)

print(inputs)

With qwen-2.5 as the template, the result is as expected:

<|im_start|>system
You are an expert assistant.<|im_end|>
<|im_start|>user
Say Hello!<|im_end|>
<|im_start|>assistant
Hello!<|im_end|>

However, with qwen-3 as in the example, you will get the error noted earlier:

TemplateSyntaxError: expected token 'end of print statement', got 'name'

[shimmyshimmer]

@kiankyars: There seems to be issue with the qwen-3 template that was added. Specifically, it reliably produces the following error:

TemplateSyntaxError: expected token 'end of print statement', got 'name'

I've created a minimal example here that works with qwen-2.5 but fails with qwen-3:

from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-0.6B-unsloth-bnb-4bit",
    max_seq_length = 8192,
    load_in_4bit = True,
)

tokenizer = get_chat_template(
    tokenizer,
    chat_template = "qwen-3", # works with qwen-2.5 but not qwen-3
)

messages = [{"role": "system", "content": "You are an expert assistant."}, 
            {"role": "user", "content": "Say Hello!"}, 
            {"role": "assistant", "content": "Hello!"}]

inputs = tokenizer.apply_chat_template(messages, 
                                       tokenize = False, 
                                       add_generation_prompt = False)

print(inputs)

With qwen-2.5 as the template, the result is as expected:

<|im_start|>system
You are an expert assistant.<|im_end|>
<|im_start|>user
Say Hello!<|im_end|>
<|im_start|>assistant
Hello!<|im_end|>

However, with qwen-3 as in the example, you will get the error noted earlier:

TemplateSyntaxError: expected token 'end of print statement', got 'name'

@davedgd oh great point! would you like to add a PR so I can accept it! Thank you :)

[davedgd]

@davedgd oh great point! would you like to add a PR so I can accept it! Thank you :)

@shimmyshimmer: Done! :) I traced down the issue to a couple of missing escapes on the double quotes. Seems to work fine with the patch in place:

<|im_start|>system
You are an expert assistant.<|im_end|>
<|im_start|>user
Say Hello!<|im_end|>
<|im_start|>assistant
<think>

</think>

Hello!<|im_end|>

[kiankyars]

@davedgd oh great point! would you like to add a PR so I can accept it! Thank you :)

@shimmyshimmer: Done! :) I traced down the issue to a couple of missing escapes on the double quotes. Seems to work fine with the patch in place:

<|im_start|>system
You are an expert assistant.<|im_end|>
<|im_start|>user
Say Hello!<|im_end|>
<|im_start|>assistant
<think>

</think>

Hello!<|im_end|>

Thanks for pointing that out! I gotta figure out how my tests did not reveal that.