Transformers getting started

Basics: Transformer Neural Networks: A Step-by-Step Breakdown

Install all python dependencies (tested on python 3.11)

1. clone the repo, go the folder

2. Create a python venv python3.11 -m venv .transformersEnv

3. source the venv source .transformersEnv/bin/activate

4. install dependencies pip install -r requirements.txt

5. run convert script python bert-convert.py or python bert_pretrained.py
   bert-convert.py script creates a onnx folder with Bert onnx model bert-base-cased.onnx with precision (float32)
   bert_pretrained.py script creates a onnx_models folder with Bert onnx model i.e., bert-base-cased-squad.onnx
   Use Netron to verify and visualise onnx model

onnx to MLIR (LLVM dialect)

Use onnx-mlir to convert onnx models to LLVM dialect.

1. Build on local or for quick conversions use Docker img Instructions
2. Follow Environment Variables Setup
3. To generate LLVM Dialectonnx-mlir -O3 --EmitLLVMIR <onnx model file>
4. (optional) creating executable binary example

LLVM Dialect to LLVM IR

1. Build MLIR on local (preffered debug version) MLIR getting started
2. (Optional) set path variable to mlir project bin folder
3. use mlir-translate <input-file-LLVM-dialect> --mlir-to-llvmir -o <output.ll> to get LLVM IR.

Other useful docs

1. Transformer Model Optimization (onnx runtime)

2. Pytorch to onnx export