Installation

Contents

Installation#

To use MASE, you can easily create an environment with all required dependendies using Docker or Anaconda. Follow the instructions on the links below according to your preferred method.

Hint

Some parts of the flow may assume you have a version of Vivado/Vitis installed. Other parts, such as the emit verilog flow, require Verilator, which is included in the Docker container, but not on the Conda environment. If you prefer using Conda, you can just install Verilator locally.

Import a model#

To import a model into MASE and use all its features, the following options are available.

  • Generate a MaseGraph from a torch.nn.Module instance.

    • Since the MASE IR is based on Torch FX, symbolic tracing limitations apply to this method, namely models with control flow cannot be traced (see documentation).

    • Exising Pytorch models can be patched to remove control and run symbolic tracing (see here for examples).

import torch.nn as nn
from chop import MaseGraph

class MyModel(nn.Module):
    def __init__(self):
        ...

    def forward(self):
        ...

model = MyModel()
mg = MaseGraph(model)

  • Import a model using the ONNX backend.

    • HuggingFace models can be imported using the MaseOnnxGraph.from_pretrained method, which first exports an ONNX representation of the model, then imports it into MASE. See list of supported models here.

    • Any other ONNX models can be directly imported using the export_fx_graph analysis pass.

from chop import MaseOnnxGraph
from chop.passes export_fx_graph_analysis_pass

pretrained = "bert-base-uncased"
og = MaseOnnxGraph.from_pretrained(pretrained)
mg, _ = export_fx_graph_analysis_pass(og)
Contents