smk2295/JSQ-Qwen3-8B-Base-uns70-int4

JSQ — Qwen3-8B-Base (70% unstructured sparsity + INT4, symmetric)

Baseline compressed checkpoint for compression research.

• Method: JSQ (joint sparsification + quantization, Wanda-style pruning + weight quant)
• Sparsity: 70% unstructured
• Quantization: INT4, per-group 128, symmetric (absmax)
• Base: Qwen/Qwen3-8B-Base

Files

• model.safetensors: sparse + fake-quantized weights in fp16, loadable via AutoModelForCausalLM.
• compression/scales.safetensors: per-group-128 symmetric scales [out, in/128] per layer (scale=absmax/7).
• compression_config.json: method / sparsity / bits / granularity / symmetric.

Notes

• Mask recoverable as weight == 0. Symmetric: w=round(x/scale)*scale, levels in [-7,7].