Fix persist_layer_norm override: set args.no_persist_layer_norm=True (Megatron uses the negated form)

Pre-install pybind11 via current interpreter + pre-compile Megatron dataset helpers at module load

Revert UI to plain Gradio; add args.persist_layer_norm=False (torch_norm asserts it false)

Override TE/Apex fusion flags after initialize_megatron — switch transformer_impl to local, disable all fusion paths

Add pybind11 (Megatron datasets builder needs python3 -m pybind11 --includes)

Initial best-effort ZeroGPU port of Khala song generation

initial commit