-
Hello, In other quantization frameworks, e.g. TFLite(https://arxiv.org/pdf/1712.05877.pdf), Glow (https://github.com/pytorch/glow/blob/a07fbed5da37e6f9bf8b459d15b445067efa3173/docs/Quantization.md?plain=1#L184C28-L184C28), and also in PyTorch (https://discuss.pytorch.org/t/is-bias-quantized-while-doing-pytorch-static-quantization/146416/5), biases are typically either left as FLOAT32 or quantized to higher-precision int datatypes e.g. int32. Does NNCF support a similar thing? Thank you, |
Beta Was this translation helpful? Give feedback.
Answered by
alexsu52
Sep 4, 2023
Replies: 1 comment 5 replies
-
Hello @i3abghany , NNCF leaves biases in FP32. |
Beta Was this translation helpful? Give feedback.
5 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hello @i3abghany,
I think it is possible, but I would like to understand the user scenario and the benefits of this option in NNCF. Could you share your use case? Perhaps I can make some suggestion.