Assistant messages carrying `tool_calls` render to nothing under the current
template: the message loop ends in a single `{{- content + "<|im_end|>\n" -}}`
with no branch for tool calls. Two consequences, both reproduced on
llama.cpp `server-rocm-b10088`:
1. llama.cpp probes the template at startup, sees that a tool call renders to
nothing, and reports `supports_tool_calls: false` /
`supports_parallel_tool_calls: false` in `/props`. It then constrains
generation to at most one call, so the second call of a parallel request is
never sampled.
2. Replaying a conversation that already contains an assistant tool call
produces an empty assistant turn, while the following `tool` result is still
rendered -- the model sees an answer to a question it has no record of asking.
**This PR** adds the `format_arg_value` / `render_tool_calls` macros (taken
verbatim from the latest LFM2.5 template) and one branch that emits them, so a
tool call renders as
`<|tool_call_start|>[name(ar…
- Fix typo in README, there's no LLM with "2.6" parameters! (f70d22bde5a09942036d545417b1dd0aca6e23be)
Co-authored-by: Mihai Popa <MihaiPopa-1@users.noreply.huggingface.co>
Yuri Khrustalev +1: Xenova Verified signature- Update chat_template.jinja (16d3c9b08df3621b720407b9a85b4d4f2e1a2ebd) - Update chat template (7212ba35e776fbd3375be66b8f51fcfe80d7a569) Co-authored-by: Joshua <Xenova@users.noreply.huggingface.co>b9e9a59