Vai_q_pytorch displays a warning message if an issue could lead to problems or incompleteness in the quantization result (r to the message text for details). The message format is [VAIQ_WARN][MESSAGE_ID]: message text. The quantization process can still proceed to completion despite the warning.
List important warning messages in the following table:
|BatchNorm OP attribute affine=False has been replaced by affine=True when parsing the model.
|Bit width setting in the configuration file. If it conflicts with that from torch_quantizer API, the setting in the configuration file is used.
|Convert to XMODEL failed. Check the message text to locate the reason.
|CUDA (HIP) is not available. Change the device to CPU.
|Data parallel is not supported. The wrapper 'torch.nn.DataParallel' has been removed in vai_q_pytorch.
|Only quantization aware training process has a deployable model.
|The input arguments device mismatches with the quantizer device type.
|Failed to generate XMODEL due to some reasons. Refer to the message text.
|Fast fine-tuning function is ignored in test mode.
|vai_q_pytorch recognizes the list OP as a float operator by default.
|The OP might not be fused by the compiler and is assigned to DPU.
|Force to change the negative_slope of LeakyReLU to 0.1015625 because DPU only supports this value. It is recommended to change all negative_slope of LeakyReLU to 0.1015625 and re-train the float model for better-deployed model accuracy.
|matplotlib is needed for visualization but not found. It needs to be installed.
|There is not enough memory for fast finetune, and this process is ignored. Try to use a smaller calibration dataset.
|Cannot find the XIR package in the environment. It needs to be installed.
|ReLU6 has been replaced by ReLU.
|Sigmoid has been replaced by Hardsigmoid.
|SiLU has been replaced by Hardswish.
|Quantization scale is too large or too small.
|Some tensors are not quantized. Check their particularity.
|The tensor type of the node cannot be quantized. Only support float32/double/float16 quantization.
|The tensor has an "inf" or "nan" value. Quantization for this tensor is ignored.
|Only support exporting TorchScript with PyTorch 1.10 and later versions.
|XIR version does not match the current vai_q_pytorch.
|Not support to dump XMODEL when the target device is not DPU.
|Reused module might lead to low accuracy of QAT. Ensure this is what you expect. Refer to the message text to locate the module with the issue.
|The argument device is no longer needed. Device information is obtained from the model directly.
|Exported scale values are not trained.