deepseek-ai

/

DeepSeek-V3.2

like 676

Follow

DeepSeek 109k

but tool call changed a lot

I tried the v3.2-exp template, and it looks normal in scenarios not involving tool calls.

Is there anyone have a solution?

Quoting from the README:

import transformers
# encoding/encoding_dsv32.py
from encoding_dsv32 import encode_messages, parse_message_from_completion_text

tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.2")

messages = [
    {"role": "user", "content": "hello"},
    {"role": "assistant", "content": "Hello! I am DeepSeek.", "reasoning_content": "thinking..."},
    {"role": "user", "content": "1+1=?"}
]
encode_config = dict(thinking_mode="thinking", drop_thinking=True, add_default_bos_token=True)

# messages -> string
prompt = encode_messages(messages, **encode_config)
# Output: "<｜begin▁of▁sentence｜><｜User｜>hello<｜Assistant｜></think>Hello! I am DeepSeek.<｜end▁of▁sentence｜><｜User｜>1+1=?<｜Assistant｜><think>"

# string -> tokens
tokens = tokenizer.encode(prompt)
# Output: [0, 128803, 33310, 128804, 128799, 19923, 3, 342, 1030, 22651, 4374, 1465, 16, 1, 128803, 19, 13, 19, 127252, 128804, 128798]

This release does not include a Jinja-format chat template. Please refer to the Python code mentioned above.

The encode_messagesfunction acts roughly the same way as a typical Jinja2 chat template, where given messages and other args, in return, you'll get the "rendered" prompt string that you can send for Completions. A slight thing to look out for are for tool definitions, response format, etc, where this new approach defines them as messages attributes.

I encourage you to read through the render_messages() function. It is very intuitive and easy to read. For me personally, this is a better approach and is better than reading through Jinja2 template. For the most part, the encoding_dsv32.py file still contains quite a lot of template strings. But the logic and branching is easier to read through and debug, because all is done in standard Python.

Hope this helps.