This repo contains serialized blobs of an up projection layer of llama3-8B (oc=14336, ic=4096). The linear layer has been quantized (GPTQ W4 Sym with group size 32) and sparsified by 50%.
├── sparse_w4
│ ├── linear_bitmap_int32.bin
│ ├── linear_compressed_qweight_int32.bin
│ ├── linear_nnz_int16.bin
│ ├── linear_scales_float16.bin
│ └── linear_zeros_int32.bin
Usage
The following script shows how to process the blobs in python. It shows unpacking, zero location recovery, as well as weight dequantization process.
python unpack_blobs.py
you can ignore
internal/