[ONNX Converter] Simple LLM inference support (ONNX format)

[EunjuYang]

ONNX Converter Support

We need to support ONNX conversion. Let's start from simple LLM model conversion and its test.

• Task1: Single token inference test @sumon-98
• Task2: Multi-token inference test sachin-ingh
• Task3: weight bin file loading @niket-agarwal

Important dates

Final demonstration 10/13
Intermediate PoC 9/30

Intermediate PoC requirements

• Running onnx converted 4 FC layers with weight loading.
• requirement: input output comparison with python ONNX format
• Intermediate PoC : 9/30

[sumon-98]

Task 1: Single token inference test (Issues with NNTrainer)

• Multiply layer: the second operand can only be broadcasted in NNTrainer
• Concat layer: input dimension is set incorrect (eg. 64 x 8 should be 512 but instead comes to be 1024)

[sachin-nntrainer]

Task 2: Multi-token inference test

• Python modelling for multi token generations: Asked for review.

• Inferencing on nntrainer: Working with Sumon to fix single token model issues.