Context

For tool calling to work end-to-end during training, a chat template needs to satisfy several requirements:

1. Supports tools: the template can render a full tool-calling conversation (user → assistant with tool_calls → tool → ...) without error, and the tool content actually appears in the rendered output.
2. Prefix-preserving: appending messages doesn't change the rendering of earlier messages. This is required by _get_tool_suffix_ids, which extracts tool response formatting tokens by comparing tokenizations with and without tool messages appended.
3. Response schema: a schema that allows parse_response to extract tool calls from the raw assistant text. No tokenizer ships with one built-in yet; TRL provides them via add_response_schema for supported templates.

Current status

Notes

• DeepseekV3: actually supports tools, but its template concatenates arguments directly as a string (+ tool['function']['arguments']), so it requires arguments to be a JSON string rather than a dict. This is a template quirk that will need a patch.
• "Supports tools = No" can mean different things:
• Template rejects the role sequence entirely (Cohere, FalconMamba, Gemma family)
• Template indexes into content as a list for all roles including tool (Idefics2, Idefics3, LlavaNext, SmolVLM)
• Template silently ignores tool messages — renders without error but tool content is missing from output (Cohere2, Phi3)
• Qwen3ForCausalLM uses {% generation %} / {% endgeneration %} tags (newer transformers feature), while Qwen3MoeForCausalLM uses the older template style. They need separate patches.
• Response schemas are not built-in to any tokenizer yet. TRL provides them via add_response_schema() for Qwen3 (matches Qwen3Moe) and Qwen3.5.

EDIT: Qwen2-VL and 2.5-VL actually don't support tool calling