pymllm.orchestrator.detokenizer_process¶

DetokenizerProcess – subprocess that converts token IDs back to text.

Receives BatchTokenIDOut-style dicts from the SchedulerProcess, detokenizes them, and forwards the decoded strings to the RequestResponseProcess.

Attributes¶

logger

Classes¶

DetokenizerProcess

Runs inside a subprocess. Detokenizes finished outputs.

Functions¶

run_detokenizer_process(recv_from_scheduler_addr, ...)

Entry point for torch.multiprocessing.Process(target=...).

Module Contents¶

pymllm.orchestrator.detokenizer_process.logger¶

class pymllm.orchestrator.detokenizer_process.DetokenizerProcess(recv_from_scheduler_addr, send_to_rr_addr, tokenizer_cfg=None)¶

Runs inside a subprocess. Detokenizes finished outputs.

Parameters:

recv_from_scheduler_addr (str)
send_to_rr_addr (str)
tokenizer_cfg (Optional[Dict[str, Any]])

init_sockets()¶

Return type:: None

init_tokenizer()¶

Load the tokenizer from the configured path.

Return type:: None

event_loop()¶

Infinite loop: recv token IDs -> detokenize -> send text to RR.

Return type:: None

shutdown()¶

Return type:: None

pymllm.orchestrator.detokenizer_process.run_detokenizer_process(recv_from_scheduler_addr, send_to_rr_addr, pipe_writer, tokenizer_cfg=None)¶

Entry point for torch.multiprocessing.Process(target=...).

Parameters:

recv_from_scheduler_addr (str)
send_to_rr_addr (str)
pipe_writer (multiprocessing.connection.Connection)
tokenizer_cfg (Optional[Dict[str, Any]])

Return type:

None