PTQ with IPEX backend and XPU device is not working

[paguilomanas]

Hi all,

I have been trying to apply post-training-quantization to a custom vision model (pretrained vgg16 model) which I have already finetuned using "xpu" (Intel GPU Max Series). I have saved the resulting model weights from this finetuning in "pt_training_xpu_none.pt" (I cannot attach the file :( it is too big).

The issue takes place when loading the weights for quantizing the model and afterwards running inference.
The neural_compressor.quantization is not working and returns several unexpected errors when using this configuration conf=PostTrainingQuantConfig(backend="ipex", device="xpu").

The main errors are the following (the full output trace is below):

• [ERROR] Unexpected exception AttributeError("'_OpNamespace' 'torch_ipex' object has no attribute 'merged_embeddingbag_cat_forward'") happened during tuning.
  Traceback (most recent call last):
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/utils/utility.py", line 103, in getattr
  mod = getattr(self.module, name)

• AttributeError: module 'intel_extension_for_pytorch' has no attribute 'quantization'

• RuntimeError: No such operator torch_ipex::merged_embeddingbag_cat_forward

• AttributeError(AttributeError: '_OpNamespace' 'torch_ipex' object has no attribute 'merged_embeddingbag_cat_forward'
  2024-06-28 09:59:49 [ERROR] Specified timeout or max trials is reached! Not found any quantized model which meet accuracy goal. Exit.

These are the library versions I am using (I guess there might be some incompatibilities between IPEX and INC versions):

neural_compressor: 2.6
intel_extension_for_pytorch: 2.1.30.post0+xpu
torch: 2.1.0.post2
torchaudio: 2.1.0.post2
torchvision: 0.16.0.post2

This is a simplified version of my code:

import torch

from neural_compressor.config import PostTrainingQuantConfig
from neural_compressor import quantization
import intel_extension_for_pytorch as ipex

from modules import img_generator_pytorch
from modules import VGG16Custom

quantize = True

model_path = "./pt_training_xpu_none.pt"
val_dataset_path = "./dataset/val"
q_model_path = "./pt_inference_xpu_inc"

model = VGG16Custom()
model.load_state_dict(torch.load(model_path))
model.to("xpu")
model.eval()

# if model is not quantized and saved already
if quantize: 
        data_loader_calib = img_generator_pytorch(val_dataset_path, batch_size=batch_size, shuffle=True, drop_remainder=True)
        conf = PostTrainingQuantConfig(backend="ipex", device="xpu")
        q_model= quantization.fit(model=model,  conf=conf, calib_dataloader=data_loader_calib)
        q_model.save(q_model_path)

# once model is saved and quantize we just load the saved model (no need to quantize it everytime)
q_model = load(q_model_path, model)

This is the output trace with the errors:

2024-06-28 09:59:49 [INFO] Start auto tuning.
2024-06-28 09:59:49 [INFO] Quantize model without tuning!
2024-06-28 09:59:49 [INFO] Quantize the model with default configuration without evaluating the model.                To perform the tuning process, please either provide an eval_func or provide an                    eval_dataloader an eval_metric.
2024-06-28 09:59:49 [INFO] Adaptor has 5 recipes.
2024-06-28 09:59:49 [INFO] 0 recipes specified by user.
2024-06-28 09:59:49 [INFO] 3 recipes require future tuning.
2024-06-28 09:59:49 [WARNING] Fail to remove /home/cic/intel_sustainable_AI_phase2/nc_workspace/2024-06-28_09-59-45/ipex_config_tmp.json.
2024-06-28 09:59:49 [INFO] *** Initialize auto tuning
2024-06-28 09:59:49 [INFO] {
2024-06-28 09:59:49 [INFO]     'PostTrainingQuantConfig': {
2024-06-28 09:59:49 [INFO]         'AccuracyCriterion': {
2024-06-28 09:59:49 [INFO]             'criterion': 'relative',
2024-06-28 09:59:49 [INFO]             'higher_is_better': True,
2024-06-28 09:59:49 [INFO]             'tolerable_loss': 0.01,
2024-06-28 09:59:49 [INFO]             'absolute': None,
2024-06-28 09:59:49 [INFO]             'keys': <bound method AccuracyCriterion.keys of <neural_compressor.config.AccuracyCriterion object at 0x7ff8ee2040a0>>,
2024-06-28 09:59:49 [INFO]             'relative': 0.01
2024-06-28 09:59:49 [INFO]         },
2024-06-28 09:59:49 [INFO]         'approach': 'post_training_static_quant',
2024-06-28 09:59:49 [INFO]         'backend': 'ipex',
2024-06-28 09:59:49 [INFO]         'calibration_sampling_size': [
2024-06-28 09:59:49 [INFO]             100
2024-06-28 09:59:49 [INFO]         ],
2024-06-28 09:59:49 [INFO]         'device': 'xpu',
2024-06-28 09:59:49 [INFO]         'diagnosis': False,
2024-06-28 09:59:49 [INFO]         'domain': 'auto',
2024-06-28 09:59:49 [INFO]         'example_inputs': 'Not printed here due to large size tensors...',
2024-06-28 09:59:49 [INFO]         'excluded_precisions': [
2024-06-28 09:59:49 [INFO]         ],
2024-06-28 09:59:49 [INFO]         'framework': 'pytorch_ipex',
2024-06-28 09:59:49 [INFO]         'inputs': [
2024-06-28 09:59:49 [INFO]         ],
2024-06-28 09:59:49 [INFO]         'model_name': '',
2024-06-28 09:59:49 [INFO]         'ni_workload_name': 'quantization',
2024-06-28 09:59:49 [INFO]         'op_name_dict': None,
2024-06-28 09:59:49 [INFO]         'op_type_dict': None,
2024-06-28 09:59:49 [INFO]         'outputs': [
2024-06-28 09:59:49 [INFO]         ],
2024-06-28 09:59:49 [INFO]         'quant_format': 'default',
2024-06-28 09:59:49 [INFO]         'quant_level': 'auto',
2024-06-28 09:59:49 [INFO]         'recipes': {
2024-06-28 09:59:49 [INFO]             'smooth_quant': False,
2024-06-28 09:59:49 [INFO]             'smooth_quant_args': {
2024-06-28 09:59:49 [INFO]             },
2024-06-28 09:59:49 [INFO]             'layer_wise_quant': False,
2024-06-28 09:59:49 [INFO]             'layer_wise_quant_args': {
2024-06-28 09:59:49 [INFO]             },
2024-06-28 09:59:49 [INFO]             'fast_bias_correction': False,
2024-06-28 09:59:49 [INFO]             'weight_correction': False,
2024-06-28 09:59:49 [INFO]             'gemm_to_matmul': True,
2024-06-28 09:59:49 [INFO]             'graph_optimization_level': None,
2024-06-28 09:59:49 [INFO]             'first_conv_or_matmul_quantization': True,
2024-06-28 09:59:49 [INFO]             'last_conv_or_matmul_quantization': True,
2024-06-28 09:59:49 [INFO]             'pre_post_process_quantization': True,
2024-06-28 09:59:49 [INFO]             'add_qdq_pair_to_weight': False,
2024-06-28 09:59:49 [INFO]             'optypes_to_exclude_output_quant': [
2024-06-28 09:59:49 [INFO]             ],
2024-06-28 09:59:49 [INFO]             'dedicated_qdq_pair': False,
2024-06-28 09:59:49 [INFO]             'rtn_args': {
2024-06-28 09:59:49 [INFO]             },
2024-06-28 09:59:49 [INFO]             'awq_args': {
2024-06-28 09:59:49 [INFO]             },
2024-06-28 09:59:49 [INFO]             'gptq_args': {
2024-06-28 09:59:49 [INFO]             },
2024-06-28 09:59:49 [INFO]             'teq_args': {
2024-06-28 09:59:49 [INFO]             },
2024-06-28 09:59:49 [INFO]             'autoround_args': {
2024-06-28 09:59:49 [INFO]             }
2024-06-28 09:59:49 [INFO]         },
2024-06-28 09:59:49 [INFO]         'reduce_range': None,
2024-06-28 09:59:49 [INFO]         'TuningCriterion': {
2024-06-28 09:59:49 [INFO]             'max_trials': 100,
2024-06-28 09:59:49 [INFO]             'objective': [
2024-06-28 09:59:49 [INFO]                 'performance'
2024-06-28 09:59:49 [INFO]             ],
2024-06-28 09:59:49 [INFO]             'strategy': 'basic',
2024-06-28 09:59:49 [INFO]             'strategy_kwargs': None,
2024-06-28 09:59:49 [INFO]             'timeout': 0
2024-06-28 09:59:49 [INFO]         },
2024-06-28 09:59:49 [INFO]         'use_bf16': True
2024-06-28 09:59:49 [INFO]     }
2024-06-28 09:59:49 [INFO] }
2024-06-28 09:59:49 [WARNING] [Strategy] Please install `mpi4py` correctly if using distributed tuning; otherwise, ignore this warning.
2024-06-28 09:59:49 [INFO] Attention Blocks: 0
2024-06-28 09:59:49 [INFO] FFN Blocks: 0
2024-06-28 09:59:49 [ERROR] Unexpected exception AttributeError("'_OpNamespace' 'torch_ipex' object has no attribute 'merged_embeddingbag_cat_forward'") happened during tuning.
Traceback (most recent call last):
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/utils/utility.py", line 103, in __getattr__
    mod = getattr(self.module, name)
AttributeError: module 'intel_extension_for_pytorch' has no attribute 'quantization'
 
During handling of the above exception, another exception occurred:
 
Traceback (most recent call last):
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/torch/_ops.py", line 757, in __getattr__
    op, overload_names = torch._C._jit_get_operation(qualified_op_name)
RuntimeError: No such operator torch_ipex::merged_embeddingbag_cat_forward
 
The above exception was the direct cause of the following exception:
 
Traceback (most recent call last):
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/quantization.py", line 234, in fit
    strategy.traverse()
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/strategy/auto.py", line 140, in traverse
    super().traverse()
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/strategy/strategy.py", line 487, in traverse
    self._prepare_tuning()
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/strategy/strategy.py", line 383, in _prepare_tuning
    self.capability = self.capability or self.adaptor.query_fw_capability(self.model)
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/utils/utility.py", line 347, in fi
    res = func(*args, **kwargs)
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/adaptor/pytorch.py", line 3069, in query_fw_capability
    return self._get_quantizable_ops(model.model)
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/adaptor/pytorch.py", line 1052, in _get_quantizable_ops
    self._get_quantizable_ops_recursively(model, "", quantizable_ops)
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/adaptor/pytorch.py", line 3186, in _get_quantizable_ops_recursively
    model = ipex.quantization.prepare(
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/neural_compressor/utils/utility.py", line 107, in __getattr__
    spec.loader.exec_module(mod)
  File "<frozen importlib._bootstrap_external>", line 850, in exec_module
  File "<frozen importlib._bootstrap>", line 228, in _call_with_frames_removed
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/intel_extension_for_pytorch/quantization/__init__.py", line 1, in <module>
    from ._quantize import prepare, convert
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/intel_extension_for_pytorch/quantization/__init__.py", line 1, in <module>
    from ._quantize import prepare, convert
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/intel_extension_for_pytorch/quantization/_quantize.py", line 20, in <module>
    from ._quantize_utils import auto_prepare, auto_convert, copy_prepared_model
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/intel_extension_for_pytorch/quantization/_quantize_utils.py", line 10, in <module>
    from ._utils import (
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/intel_extension_for_pytorch/quantization/_utils.py", line 15, in <module>
    from ._quantization_state_utils import QTensorInfo
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/intel_extension_for_pytorch/quantization/_quantization_state_utils.py", line 57, in <module>
    torch.ops.torch_ipex.merged_embeddingbag_cat_forward,
  File "/opt/anaconda3/envs/gpu-pt/lib/python3.9/site-packages/torch/_ops.py", line 761, in __getattr__
    raise AttributeError(
AttributeError: '_OpNamespace' 'torch_ipex' object has no attribute 'merged_embeddingbag_cat_forward'
2024-06-28 09:59:49 [ERROR] Specified timeout or max trials is reached! Not found any quantized model which meet accuracy goal. Exit.

These are my custom modules:

def img_generator_pytorch(dataset_path: str, batch_size: int, shuffle: bool, drop_remainder: bool = False):

    data_transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])

    dataset = ImageFolder(dataset_path, transform=data_transform)
    data_loader = DataLoader(dataset, 
                             batch_size, 
                             shuffle=shuffle, 
                             drop_last=drop_remainder,
                             pin_memory=True)
    
    return data_loader

class VGG16Custom(nn.Module):

    def __init__(self) -> None:
        super(VGG16Custom, self).__init__()
        self.model = models.vgg16(weights="VGG16_Weights.DEFAULT")
        
        for param in self.model.parameters():
            param.requires_grad = False

        in_features = self.model.classifier[0].in_features
        self.model.classifier = self.VGG16Classifier(in_features)

    def __call__(self, *args, **kwargs):
        """Call model for inference."""
        return self.model(*args, **kwargs)
        
    def forward(self, x):
        # Pass the input through the model
        x = self.model(x)
        return x

    class VGG16Classifier(torch.nn.Module):
        """Definition of classifier."""

        def __init__(self, in_features):

            super().__init__()
            self.fc1 = nn.Linear(in_features, 1024)
            self.act1 = nn.ReLU()
            self.dropout1 = nn.Dropout(0.4)
            self.fc2 = nn.Linear(1024, 512)
            self.act2 = nn.ReLU()
            self.dropout2 = nn.Dropout(0.4)
            self.predictions = nn.Linear(512, NUM_CLASSES)

        def forward(self, x):
            """Forward pass of classifier."""
            x = self.fc1(x)
            x = self.act1(x)
            x = self.dropout1(x)
            x = self.fc2(x)
            x = self.act2(x)
            x = self.dropout2(x)
            x = self.predictions(x)
            return x

[paguilomanas]

I also wanted to mention that I have been following this main.py example code stored in "examples/pytorch/image_recognition/torchvision_models/quantization/ptq/cpu/ipex/main.py" which sets "xpu" as optional device and runs a benchmark afterwards. It was updated 7 months ago #1348 commited as "support ipex xpu".

[violetch24]

hi @paguilomanas , since IPEX-xpu has recently changed quantization interface while INC has not, there are some issues when using the latest version of IPEX. Please try with an older version v2.0.110+xpu, while we are working on adapting IPEX's new API to INC.