[VAIQ/XIR] Serialization Error: Graph exceeded maximum protobuf size of 2GB in Vitis AI 3.0

[xenbabala]

Environment:

• Vitis AI Version: 3.0 (GPU Docker)

• Target Architecture: DPUCVDX8G (VCK190)

• Model Type: Video Diffusion Model (U-ViT)

Description:
I am quantizing a large video diffusion model using vai_q_pytorch. After the calibration (calib) phase, I encounter a fatal error during the export_xmodel step (test mode).

Error Log:

[libprotobuf ERROR google/protobuf/message_lite.cc:410] serial_v2.Graph exceeded maximum protobuf size of 2GB: 2623119139
F20260203 06:39:53.074004 49063 serialize_v2.cpp:727] [UNILOG][FATAL][XIR_WRITE_PB_FAILURE][failed to write pb file] file = quantize_result/UViT3DPose_int.xmodel

Questions:

1. Is the 2GB limit for .xmodel serialization fixed in Vitis AI 3.0?

2. If it is fixed, what is the recommended workflow for massive models?

• Should I partition the model into multiple sub-graphs manually?

• Is there a way to externalize constant buffers (like RoPE frequencies) so they aren't serialized within the protobuf graph?

3. I've already used @register_custom_op to reduce node count, but the size remains >2GB. Are there other strategies to compress the XIR graph before serialization?