espnet2.speechlm.model.speechlm.parallel_utils.qwen3.apply_torch_compile_qwen3

source

espnet2.speechlm.model.speechlm.parallel_utils.qwen3.apply_torch_compile_qwen3(model: Module, titan_config: Dict[str, Any]) → Module

Apply torch.compile to transformer layers.

Compiles each transformer layer individually. Must be applied after activation checkpointing and before FSDP.

• Parameters:
  • model – HuggingFace Qwen3 model
  • titan_config – Configuration dict
• Returns: Model with compiled transformer layers