Troubleshooting Guide for BigDL-Nano#
Refer to this section for common issues faced while using BigDL-Nano.
Installation#
Why I fail to install openvino-dev==2022.2 when pip install bigdl-nano[inference]
?#
Please check your system first as openvino-dev 2022.2 does not support centos. Refer this for more details. You can install bigdl-nano[inference] >= 2.2 instead, as bigdl-nano[inference] >= 2.2 use openvino-dev >= 2022.3 which supports centos again.
How to solve SYNK issue caused by onnx when pip install bigdl-nano[inference]
?#
We are trying to solve this issue by upgrading onnx to 1.13. But there exists conflict between onnx 1.13 and other dependencies (such as intel-tensorflow and pytorch-lightning), mainly because of protobuf version. If you are concerned about security of onnx, we recommend to pip install onnx==1.13
after pip install bigdl-nano[inference]
.
Inference#
assert precision in list(self.cur_config['ops'].keys())
when using ipex quantization with inc on machine with BF16 instruction set#
It’s known issue for Intel® Neural Compressor that they don’t deal with BF16 op well at version 1.13.1 . This has been fixed in version 2.0. You can install bigdl-nano[inference] >= 2.2 to fix this problem.
error message like set strict=False
when InferenceOptimizer.trace(accelerator='jit')
or InferenceOptimizer.quantize(accelerator='jit')
#
You can set strict=False
for torch.jit.trace
by setting jit_strict=False
in InferenceOptimizer.trace(accelerator='jit', xxx, jit_strict=False)
or InferenceOptimizer.quantize(accelerator='jit', xxx, jit_strict=False)
.
Refer API usage of torch.jit.trace for more details.
Why channels_last
option fails for my computer vision model?#
Please check the shape of your input data first, we don’t support channels_last
for 3D input now. If your model is a 3D model or your input data is not 4D Tensor, normally channels_last
option will fail.
Why InferenceOptimizer.load(dir)
fails to load my model saved by InferenceOptimizer.save(model, dir)
#
If you accelerate the model with accelerator=None
by InferenceOptimizer.trace
/InferenceOptimizer.quantize
or it’s just a normal torch.nn.Module, you have to pass original FP32 model to load pytorch model by InferenceOptimizer.load(dir, model=model)
.
Why my bf16 model is slower than fp32 model?#
You can first check whether your machine supports the bf16 instruction set first by lscpu | grep "bf16"
. If there is no avx512_bf16
or amx_bf16
in the output, then, without instruction set support, the performance of bf16 cannot be guaranteed, and generally, its performance will deteriorate.
INVALID_ARGUMENT : Got invalid dimensions for input
or [ PARAMETER_MISMATCH ] Can not clone with new dims.
when do inference with OpenVINO / ONNXRuntime accelerated model#
This error usually occurs when your dataset produces data with dynamic shape, and such case needs you to manually set dynamic_axes
parameter and pass dynamic_axes
to trace/quantize
.
For examples, if your forward function looks like def forward(x: torch.Tensor):
, and it recieves 4d Tensor as input. However, your input data’s shape will vary during inference, it will be (1, 3, 224, 224) or (1, 3, 256, 256), then in such case, you should:
dynamic_axes['x'] = {0: 'batch_size', 2: 'width', 3: 'height'} # this means the 0/2/3 dim of your input data may vary during inference
input_sample = torch.randn(1, 3, 224, 224)
acce_model = trace(model=model, input_sample=x, dynamic_axes=dynamic_axes)
You can refer to API usage of torch.onnx.export for more details.
Why jit didn’t work on my model?#
Please check first if you use patch_cuda(disable_jit=True)
command of Nano, if you have used it to disable cuda operation, it will disable jit at the same time by torch.jit._state.disable()
, so jit has no effect now.
How to cope with out-of-memory during workload with Intel® Extension for PyTorch*#
If you found the workload runs with Intel® Extension for PyTorch* occupies a remarkably large amount of memory, you can try to reduce the occupied memory size by setting weights_prepack=False
when calling InferenceOptimizer.trace
\ InferenceOptimizer.quantize
.
RuntimeError: Check ‘false’ failed at src/frontends/common/src/frontend.cpp#
You may see this error when you do inference with accelerator=OpenVINO
in keras. It only occurs when you use intel-tensorflow
>= 2.8 and you forget source bigdl-nano-init
. The way to solve this problem is just source bigdl-nano-init
or source bigdl-nano-init -j
.
TypeError: deprecated() got an unexpected keyword argument ‘name’#
If a version problem caused by too low cryptography version. You can fix it by just pip install cryptography==38.0.0
.