locodellm.session.session_state#
Holds the state of a generation session.
- class locodellm.session.session_state.SessionState(model: Any, tokenizer: Any, verbose: int = 0, model_path: str | None = None, chat_template: str | None = None)#
Holds the state of a generation session.
Use
create_session()to build an instance from a model path or an already-loadedonnxruntime_genai.Model.- model#
The loaded
onnxruntime_genai.Model.
- tokenizer#
The tokenizer bound to model.
- tokens#
All token ids accumulated so far (prompt + generated).
- text#
The text generated during the last
generate()call.
- verbose#
Verbosity level (0 = silent, 1+ = print progress).
- generate(prompt: str, max_length: int = 200, **search_options: Any) SessionState#
Generate text from prompt, appending to the conversation history.
The full token history (previous turns + prompt) is replayed so the model sees the complete context.
- Parameters:
prompt – The text prompt to send to the model.
max_length – Maximum number of tokens (including all tokens accumulated across turns).
**search_options – Extra search options forwarded to
GeneratorParams.set_search_options(e.g.temperature,top_k,top_p).
- Returns:
- locodellm.session.session_state.create_session(model: str | Any, providers: list[str] | None = None, verbose: int = 0, chat_template: str | None = None) SessionState#
Create a
SessionStatefrom a model path or loaded model.- Parameters:
model – A path (
str) to the model directory or an already-loadedonnxruntime_genai.Modelinstance.providers – Ordered list of execution providers, e.g.
["CUDAExecutionProvider", "CPUExecutionProvider"]. When None, onnxruntime-genai picks its default provider. Ignored when model is not a path.verbose – Verbosity level (0 = silent, 1+ = print progress messages during model loading and generation).
chat_template – Chat template to apply around prompts. Use
"chatml"for Qwen and other ChatML-based instruct models. When None, prompts are sent as-is.
- Returns:
A new
SessionStateready forSessionState.generate().