import copy
import logging
import sys
import unittest
from unittest.mock import MagicMock, patch
import torch
import torch.nn as nn
from mock_torch_npu import (
mock_npu,
mock_npu_convert_weight_to_int4pack,
mock_npu_dtype_cast,
mock_npu_dynamic_mx_quant,
mock_npu_format_cast,
mock_npu_quant_matmul,
mock_npu_quantize,
mock_npu_weight_quant_batchmatmul,
)
from utils import TestModel
from amct_pytorch import convert, quantize
logger = logging.getLogger(__name__)
LINEAR_AW_QUANT = 'LinearAWQuant'
LINEAR = 'Linear'
NPU_WEIGHT_QUANTIZED_LINEAR = 'NpuWeightQuantizedLinear'
torch.manual_seed(0)
class TestAwq(unittest.TestCase):
'''
ST FOR AWQ ALGORITHM
'''
@classmethod
def setUpClass(cls):
cls.test_model = TestModel().to(torch.bfloat16)
cls.inputs = torch.randn(64, 64).to(torch.bfloat16)
cls.ori_out = cls.test_model(cls.inputs)
logger.info('TestAwq START!')
@classmethod
def tearDownClass(cls):
logger.info('TestAwq END!')
def setUp(self):
mock_torch_npu = MagicMock()
sys.modules['torch_npu'] = mock_torch_npu
def tearDown(self):
del sys.modules['torch_npu']
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int8_tensor_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int8',
'symmetric': True,
'strategy': 'tensor',
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR_AW_QUANT)
self.assertEqual(model.linear1.scale_w.shape[0], 1)
self.assertIsNone(model.linear1.offset_w)
self.assertIsNone(model.linear1.scale_d)
self.assertIsNone(model.linear1.offset_d)
torch.Tensor.npu = mock_npu
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int8_tensor_asym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int8',
'symmetric': False,
'strategy': 'tensor',
},
},
'algorithm': {'awq': {'grids_num': 10}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR_AW_QUANT)
self.assertEqual(model.linear1.scale_w.shape[0], 1)
self.assertIsNotNone(model.linear1.offset_w)
self.assertIsNone(model.linear1.scale_d)
self.assertIsNone(model.linear1.offset_d)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int8_channel_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int8',
'symmetric': True,
'strategy': 'channel',
},
},
'algorithm': {'awq': {'grids_num': 300}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR_AW_QUANT)
self.assertEqual(model.linear1.scale_w.shape[0], 64)
self.assertIsNone(model.linear1.offset_w)
self.assertIsNone(model.linear1.scale_d)
self.assertIsNone(model.linear1.offset_d)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int8_channel_asym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int8',
'symmetric': False,
'strategy': 'channel',
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR_AW_QUANT)
self.assertEqual(model.linear1.scale_w.shape[0], 64)
self.assertIsNotNone(model.linear1.offset_w)
self.assertIsNone(model.linear1.scale_d)
self.assertIsNone(model.linear1.offset_d)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int8_group_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int8',
'symmetric': True,
'strategy': 'group',
'group_size': 32
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
self.assertEqual(model.linear1.scale_w.shape[0], 64)
self.assertEqual(model.linear1.scale_w.shape[1], 2)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int8_group_asym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int8',
'symmetric': False,
'strategy': 'group',
'group_size': 32
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
self.assertEqual(model.linear1.scale_w.shape[0], 64)
self.assertEqual(model.linear1.scale_w.shape[1], 2)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int4_tensor_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int4',
'symmetric': True,
'strategy': 'tensor',
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int4_tensor_asym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int4',
'symmetric': False,
'strategy': 'tensor',
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int4_channel_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int4',
'symmetric': True,
'strategy': 'channel',
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int4_channel_asym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int4',
'symmetric': False,
'strategy': 'channel',
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int4_group_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int4',
'symmetric': True,
'strategy': 'group',
'group_size': 32
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_int4_group_asym_awq_success(self, mock_1, mock_2, mock_3, mock_4):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'int4',
'symmetric': False,
'strategy': 'group',
'group_size': 32
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear3).__name__, LINEAR)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch('torch_npu.npu_format_cast', wraps=mock_npu_format_cast)
@patch('torch_npu.npu_dtype_cast', wraps=mock_npu_dtype_cast)
@patch('torch_npu.npu_dynamic_mx_quant', wraps=mock_npu_dynamic_mx_quant)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_fp4_group_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4, mock_5, mock_6, mock_7):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'float4_e2m1',
'symmetric': True,
'strategy': 'group',
'group_size': 32
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
self.assertIsNotNone(model.linear1.scale_w)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
self.assertIsNotNone(model.linear1.scale_w)
@patch('torch_npu.npu_quantize', wraps=mock_npu_quantize)
@patch('torch_npu.npu_quant_matmul', wraps=mock_npu_quant_matmul)
@patch('torch_npu.npu_weight_quant_batchmatmul', wraps=mock_npu_weight_quant_batchmatmul)
@patch('torch_npu.npu_convert_weight_to_int4pack', wraps=mock_npu_convert_weight_to_int4pack)
@patch('torch_npu.npu_format_cast', wraps=mock_npu_format_cast)
@patch('torch_npu.npu_dtype_cast', wraps=mock_npu_dtype_cast)
@patch('torch_npu.npu_dynamic_mx_quant', wraps=mock_npu_dynamic_mx_quant)
@patch(
'amct_pytorch.classic.deploy_op.weight_npu_quant_module.check_parameters_in_schema',
MagicMock(return_value=True),
)
def test_mxfp4_group_sym_awq_success(self, mock_1, mock_2, mock_3, mock_4, mock_5, mock_6, mock_7):
cfg = {
'batch_num': 1,
'quant_cfg': {
'weights': {
'type': 'mxfp4_e2m1',
'symmetric': True,
'strategy': 'group',
'group_size': 32
},
},
'algorithm': {'awq': {'grids_num': 20}},
}
model = copy.deepcopy(self.test_model).to(torch.bfloat16)
quantize(model, cfg)
model(self.inputs)
torch.Tensor.npu = mock_npu
self.assertEqual(type(model.linear1).__name__, LINEAR_AW_QUANT)
self.assertEqual(type(model.linear2).__name__, LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
self.assertIsNone(model.linear1.scale_w)
convert(model)
quant_out = model(self.inputs.npu())
self.assertEqual(type(model.linear1).__name__, NPU_WEIGHT_QUANTIZED_LINEAR)
self.assertEqual(type(model.linear2).__name__, LINEAR)
self.assertEqual(type(model.linear3).__name__, LINEAR)
self.assertIsNotNone(model.linear1.scale_w)