86c6268d5cWhen an empty assistant turn is dropped (a 0-token model reply, or a Stop-button
sentinel handled by _drop_empty_assistant_sentinels) the GGUF chat history can be
left with two user turns back to back. Strict Jinja chat templates (Gemma 3, some
Mistral variants) call raise_exception("Conversation roles must alternate ...") on
the first role-parity break, so llama-server returns a 400 and the thread becomes
unsendable.
Add _coalesce_consecutive_user_turns (merging only adjacent user turns, never
assistant/tool turns, preserving multimodal parts) and apply it in
_openai_messages_for_gguf_chat. The tool path inherits the fix for free because it
rebuilds from this same normalized history via _set_or_prepend_system_message. The
passthrough path is left untouched (it forwards messages verbatim). No-op for
already-alternating histories. Adds unit and end-to-end coverage.
|
||
|---|---|---|
| .. | ||
| assets | ||
| auth | ||
| core | ||
| hub | ||
| loggers | ||
| models | ||
| plugins | ||
| requirements | ||
| routes | ||
| state | ||
| storage | ||
| tests | ||
| utils | ||
| __init__.py | ||
| _platform_compat.py | ||
| cloudflare_tunnel.py | ||
| colab.py | ||
| main.py | ||
| run.py | ||
| startup_banner.py | ||