我已经按照GitHub上的示例，使用torch-mlir将一个torch模型转换为MLIR：

def save_module():
    resnet18 = models.resnet18(pretrained=True)
    resnet18.eval()

    module = torch_mlir.compile(resnet18, torch.ones(1, 3, 224, 224), output_type="torch")
    open("resnet18torch.mlir", "w").write(str(module))

之后，我修改了MLIR，现在我想做相反的操作，在我的Python代码中将这个MLIR加载为一个模块，以便继续编译它，像这样：

src = open("resnet18torch.mlir", "r").read()
#transform src to module
backend = refbackend.RefBackendLinalgOnTensorsBackend()
compiled = backend.compile(module)
jit_module = backend.load(compiled)
predictions(resnet18.forward, jit_module.forward, img, labels)

我只需要在第二段代码中添加一行来加载MLIR作为模块，但我在网上找不到任何相关信息。有人知道如何做吗？

回答：

我最终在LLVM discord的帮助下，弄清楚了如何加载torch mlir并如何重用它。以下是使用torch_mlir示例（torchscript_resnet18.py）的用法：

import torch_mlir
def load_module():
    #load the torch mlir
    src = open("resnet18torch.mlir", "r").read()
    with torch_mlir.ir.Context() as ctx:
        torch_mlir.dialects.torch.register_dialect(ctx)
        with torch_mlir.ir.Location.unknown() as loc:
            module = torch_mlir.ir.Module.parse(src)
    #translate the torch-mlir to the linalg-on-tensors dialect
    torch_mlir.run_pipeline_with_repro_report(module, "builtin.module(torch-backend-to-linalg-on-tensors-backend-pipeline)", "Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR")
    #print(module)
    #compile the module
    backend = refbackend.RefBackendLinalgOnTensorsBackend()
    compiled = backend.compile(module)
    jit_module = backend.load(compiled)
    return jit_module
#save_module()
jit_module = load_module()
predictions(jit_module.forward, img, labels)