Expected closing statement token. OpenSquareBracket !== CloseStatement

[cwoolum]

System Info

"@huggingface/transformers": "^3.3.2",

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

When trying to parse messages using AutoTokenizer.from_pretrained with the model deepseek-ai/DeepSeek-R1-Distill-Llama-8B, I receive the error

Failed to get test result: Error: Parser Error: Expected closing statement token. OpenSquareBracket !== CloseStatement.

Others were seeing similar errors in other DeepSeek models but I'm not sure if that affects this package

https://huggingface.co/mlx-community/deepseek-r1-distill-qwen-1.5b/discussions/1

Reproduction

1. Load the tokenizer and try to generate a prompt

async function loadTokenizer() {
  const tokenizer = await AutoTokenizer.from_pretrained('deepseek-ai/DeepSeek-R1-Distill-Llama-8B');
  return tokenizer;
}

const tokenizer = await loadTokenizer();
const prompt = tokenizer.apply_chat_template([{ role: 'user', content: 'How are you today?' }], { tokenize: false, add_generation_prompt: true })

An error will be thrown when tokenizer.apply_chat_template is called.

[xenova]

Hi! This will be fixed by huggingface/huggingface.js#1142.

New release coming soon.

[Sofianel5]

@xenova in the meantime - can I install @huggingface/transformers via git? bun add git+https://github.com/huggingface/huggingface.js seems to be resulting in a different structure in node_modules from the normal npm import.

EDIT 1:

nvm - can just use latest jinja (0.3.3) on npm as override for transformers 0.3.2. Nonetheless, I'm still getting this error even with the latest jinja override with the same repro steps as above.

EDIT 2:

Nvm again, the override doesn't work bc @huggingface/transformers uses the built version which already includes the 0.3.2 jinja version inline.

[xenova]

Good news: This has now been fixed in Transformers.js v3.3.3 👍