torch-mlir/projects/pt1/examples/torchscript_stablehlo_backend_tinybert.py

import torch
from torch_mlir import torchscript

from transformers import BertForMaskedLM


# Wrap the bert model to avoid multiple returns problem
class BertTinyWrapper(torch.nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.bert = BertForMaskedLM.from_pretrained(
            "prajjwal1/bert-tiny", return_dict=False
        )

    def forward(self, data):
        return self.bert(data)[0]


model = BertTinyWrapper()
model.eval()
data = torch.randint(30522, (2, 128))
out_stablehlo_mlir_path = "./bert_tiny_stablehlo.mlir"

module = torchscript.compile(
    model, data, output_type=torchscript.OutputType.STABLEHLO, use_tracing=True
)
with open(out_stablehlo_mlir_path, "w", encoding="utf-8") as outf:
    outf.write(module.operation.get_asm())

print(f"StableHLO IR of tiny bert successfully written into {out_stablehlo_mlir_path}")