• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright (c) Meta Platforms, Inc. and affiliates.
2# All rights reserved.
3#
4# This source code is licensed under the BSD-style license found in the
5# LICENSE file in the root directory of this source tree.
6
7import unittest
8
9import torch
10
11from executorch.backends.xnnpack.test.tester import Tester
12from executorch.examples.models.llama.model import Llama2Model
13
14
15class TestLlama2ETExample(unittest.TestCase):
16    def test_f32(self):
17        self._test()
18
19    def test_f16(self):
20        self._test(torch.float16)
21
22    # TODO - dynamic shape
23
24    def _test(self, dtype: torch.dtype = torch.float):
25        assert dtype in [
26            torch.float,
27            torch.float16,
28        ], f"Only fp32 and fp16 are supported, but got dtype: {dtype}"
29
30        llama2 = Llama2Model()
31        model = llama2.get_eager_model().to(dtype)
32
33        # Only convert fp32 inputs to dtype
34        example_inputs = tuple(
35            tensor.to(dtype) if tensor.dtype == torch.float32 else tensor
36            for tensor in llama2.get_example_inputs()
37        )
38
39        (
40            Tester(model, example_inputs)
41            .export()
42            .to_edge_transform_and_lower()
43            .to_executorch()
44            .serialize()
45            .run_method_and_compare_outputs(atol=5e-2, inputs=example_inputs)
46        )
47