import torch
import numpy as np
import torch.nn.functional as F
import torch_npu
from torch_npu.testing.testcase import TestCase, run_tests
from torch_npu.testing.common_utils import create_common_tensor
from torch_npu.testing.common_distributed import skipIfUnsupportMultiNPU
class TestEmbedding(TestCase):
def cpu_op_exec(self, weight, indices):
weight.requires_grad_(True)
out = F.embedding(indices, weight, scale_grad_by_freq=True, padding_idx=37)
return out.detach().numpy()
def npu_op_exec(self, weight, indices):
weight.requires_grad_(True)
out = F.embedding(indices, weight, scale_grad_by_freq=True, padding_idx=37)
out_npu = out.to("cpu")
return out_npu.detach().numpy()
def test_shape_nz_format(self):
shape_format = [
[[np.float32, 29, [40, 32]], [np.int64, 0, [40]]],
[[np.float32, 29, [40, 1024]], [np.int64, 0, [40]]],
[[np.float32, 29, [40000, 1024]], [np.int64, 0, [3125]]],
[[np.float32, 29, [40000, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, 29, [40, 32]], [np.int64, 0, [40]]],
[[np.float16, 29, [40, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, 29, [33712, 1024]], [np.int64, 0, [64, 7]]],
[[np.float32, 29, [40, 32]], [np.int64, 0, [40]]],
[[np.float32, 29, [40, 1024]], [np.int64, 0, [40]]],
[[np.float32, 29, [40000, 1024]], [np.int64, 0, [3125]]],
[[np.float32, 29, [40000, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, 29, [40, 32]], [np.int64, 0, [40]]],
[[np.float16, 29, [40, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, 29, [33712, 1024]], [np.int64, 0, [64, 7]]]
]
for item in shape_format:
weight_cpu, weight_npu = create_common_tensor(item[0], 1, 1)
indices_cpu, indices_npu = create_common_tensor(item[1], 0, 1)
if weight_cpu.dtype == torch.float16:
weight_cpu = weight_cpu.to(torch.float32)
cpu_out = self.cpu_op_exec(weight_cpu, indices_cpu)
npu_out = self.npu_op_exec(weight_npu, indices_npu)
cpu_out = cpu_out.astype(npu_out.dtype)
self.assertEqual(cpu_out, npu_out)
def test_shape_format(self):
shape_format = [
[[np.float32, 0, [40, 32]], [np.int64, 0, [40]]],
[[np.float32, 0, [40, 1024]], [np.int64, 0, [40]]],
[[np.float32, 0, [40000, 1024]], [np.int64, 0, [3125]]],
[[np.float32, 0, [40000, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, 0, [40, 32]], [np.int64, 0, [40]]],
[[np.float16, 0, [40, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, 0, [33712, 1024]], [np.int64, 0, [64, 7]]],
[[np.float32, -1, [40, 32]], [np.int64, 0, [40]]],
[[np.float32, -1, [40, 1024]], [np.int64, 0, [40]]],
[[np.float32, -1, [40000, 1024]], [np.int64, 0, [3125]]],
[[np.float32, -1, [40000, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, -1, [40, 32]], [np.int64, 0, [40]]],
[[np.float16, -1, [40, 1024]], [np.int64, 0, [128, 8]]],
[[np.float16, -1, [33712, 1024]], [np.int64, 0, [64, 7]]]
]
for item in shape_format:
weight_cpu, weight_npu = create_common_tensor(item[0], 1, 1)
indices_cpu, indices_npu = create_common_tensor(item[1], 0, 1)
if weight_cpu.dtype == torch.float16:
weight_cpu = weight_cpu.to(torch.float32)
cpu_out = self.cpu_op_exec(weight_cpu, indices_cpu)
npu_out = self.npu_op_exec(weight_npu, indices_npu)
cpu_out = cpu_out.astype(npu_out.dtype)
self.assertEqual(cpu_out, npu_out)
@skipIfUnsupportMultiNPU(2)
def test_check_diff_device(self):
embedding_matrix = torch.randn(5, 3).to('npu:1')
word_indices = torch.tensor([1, 2, 3]).to('npu:0')
msg = "Expected all tensors to be on the same device, but found at least two devices, npu:"
with self.assertRaisesRegex(RuntimeError, msg):
torch.nn.functional.embedding(word_indices, embedding_matrix)
if __name__ == "__main__":
run_tests()