search

neuralmagic / sparsezoo

Neural network model repository for highly sparse and sparse-quantized models with matching sparsification recipes
Apache License 2.0
371 stars 25 forks source link

Omit scalar weight #424

Closed horheynm closed 10 months ago

horheynm commented 10 months ago

Description

Fixed logic to get the weights in the upstream code that calls to get the weight-name. Ideally want to fix the weight name to not to return any if the weight itself is a scalar, but at this level we cannot get the weight matrix, just the name.

Example of bug fix: Screenshot 2024-02-01 at 5 05 23 PM

matrix A and B are not accessible, so returns the zero points which are not size > 1 matrix, causes error when computing the quantization level (need >1 size, bc of the log -> one entry blows up log)

rahul-tuli commented 10 months ago

Should we change the base branch to #421 so that we can cherry pick in one go?