tool-call-parser for DeepSeek-V3

[RedBeamGH]

I would like to use DeepSeek-V3 with a tool calls, what tool call parser and chat template should I use for this?

[Shyboy0499]

Use the dedicated DeepSeek parser plus the matching chat template (docs/features/tool_calling.md). For DeepSeek-V3:

vllm serve deepseek-ai/DeepSeek-V3-0324 \
  --enable-auto-tool-choice \
  --tool-call-parser deepseek_v3 \
  --chat-template examples/tool_chat_template_deepseekv3.jinja

Details per model family:

• DeepSeek-V3 (deepseek-ai/DeepSeek-V3-0324) → --tool-call-parser deepseek_v3 with tool_chat_template_deepseekv3.jinja.
• DeepSeek-R1 (deepseek-ai/DeepSeek-R1-0528) → same deepseek_v3 parser but tool_chat_template_deepseekr1.jinja.
• DeepSeek-V3.1 → --tool-call-parser deepseek_v31 with tool_chat_template_deepseekv31.jinja.
• DeepSeek-V3.2 / V4 → separate parsers exist (deepseek_v3.2 / deepseek_v4 in vllm/parser/); check the tool_calling docs for the exact flag for your version.

Two things to get right:

1. --enable-auto-tool-choice is required — without it, tool calling isn't activated even with a parser set.
2. The parser and chat template must be paired. DeepSeek's tool calls are encoded in the chat template (via <|startofthink|>/<|tool_use|> style tokens for R1, or the function-call format for V3), and the parser expects exactly that template. Mixing e.g. deepseek_v3 parser with the default template will misparse tool calls.