peterjc123
peterjc123
@spacycoder I'm glad it works and it looks cleaner.
@spacycoder Yes, both quantization for either `norm` or `RMSNorm` are unsupported at the moment. I wonder if you could actually do that using TFLite. But anyway, we should safely skip...
@spacycoder OP-wise speaking, yes, we may go through `MUL -> MEAN -> RSQRT -> MUL`. But the quantization errors can't be ignored I guess, especially for `pow` and `rsqrt`. Also,...
with #356, at least it won't throw an error for the models you provided. Quantization for those ops are still skipped.
Fixed via https://github.com/alibaba/TinyNeuralNetwork/pull/390
Just noticed that you are not using the Quantized graph rewrite of TinyNN as I can see the following option in your code. `"rewrite_graph": False` and `torch.quantization.prepare_qat`. Just FYI, the...
We will take a look.
是我们graph optimizer的问题，没有对fuse_quant_dequant=True做特殊处理，可以先暂时通过optimize=4绕过
可以试一下这个选项 https://github.com/alibaba/TinyNeuralNetwork/blob/main/tinynn/converter/base.py#L58