Module executor 

Model executor implementations.

Decoder-only LLMs go through LlmExecutor (wrapping any Box<dyn DecoderOnlyLLM>). Per-modality executors (Bert / Clip / Whisper / Tts) remain separate — they have different forward contracts that don’t fit the prefill/decode interface.

Re-exports

pub use bert_executor::BertModelExecutor;

pub use clip_executor::ClipModelExecutor;

pub use llm_executor::LlmExecutor;

pub use stub_executor::StubModelExecutor;

pub use tts_executor::TtsModelExecutor;

pub use whisper_executor::WhisperModelExecutor;

Modules

bert_executor

BERT Model Executor for embeddings

clip_executor

CLIP Model Executor for multimodal embeddings.

common

Common executor utilities — extracted from duplicated code across Qwen3, Qwen2, and Llama executors.

llm_executor

LlmExecutor<M> — adapts a DecoderOnlyLLM to the ModelExecutor trait the engine scheduler calls.

stub_executor

Stub model executor for MVP testing and development

tts_executor

Qwen3-TTS Executor — text-to-speech pipeline wiring Talker LM + Vocoder.

whisper_executor

Whisper ASR Executor — full decode pipeline matching Python whisper.