brevitas
brevitas copied to clipboard
Brevitas: neural network quantization in PyTorch
Until now, channel splitting was supported only before quantization. This PR adds support for applying channel splitting after basic quantization of a model.
https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/nn/equalized_layer.py#L19 Has to be done similarly to: https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/graph/equalize.py#L639
Speeding up GPFQ with separate forward passes for quantized and float input. I avoided offloading the float input to disc and instead saved them under an attribute for GPFQ and...
I'm trying to package `brevitas` on conda-forge and the requirement `dependencies==2.0.1` is an issue as it's old. Is it possible to update this requirement to `7.7.0`?
Hi, First of all congratulation for the amazing work that you have done so far re-implementing all the Tensor operations in the quantized domain. I saw that the documentation is...
Adds the following changes in reference to Issue #975: - Updated type hints for `get_max_available_float` - Expanded `test_max_value` for FNUZ minifloat format and for minifloats with custom exponent bit-widths -...
Adds the following fixes / enhancements: - [x] Set the device correct when checkpoints are loaded - [x] Fix issue setting the bitwidth for the output of linear layers when...
Expand our minifloat testing with the following: - Testing unscaled minifloat quantizers (https://github.com/Xilinx/brevitas/blob/df1a137aa089434e2bccb1cb87c9cb74c610d7a8/src/brevitas/quant/experimental/float_base.py#L30 and https://github.com/Xilinx/brevitas/blob/df1a137aa089434e2bccb1cb87c9cb74c610d7a8/src/brevitas/quant/experimental/float_base.py#L34) - Testing float proxies as done for int ones here https://github.com/Xilinx/brevitas/blob/dev/tests/brevitas/proxy/test_proxy.py - Add test for...
This implements: - New GroupwiseQuantTensor for Int and Float - Relevant Proxy classes - MX Float based quantizers - One notebook to test instantiation and execution Missing: - Export -...
I am using a custom CNN architecture and trained my model using qnn.QuantConv2d layers for QAT. When I inspect the stored weights of model, to load it to hardware it...