drawais's picture
Initial HQQ INT4 release
dcd22b9 verified
metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-Math-7B-Instruct
tags:
  - quantized
  - 4-bit
  - int4
  - qwen2.5
  - math
language:
  - en
pipeline_tag: text-generation

Qwen2.5-Math-7B-Instruct-HQQ-INT4

INT4 quantization of Qwen/Qwen2.5-Math-7B-Instruct. Calibration-free companion to drawais/Qwen2.5-Math-7B-Instruct-AWQ-INT4.

Footprint

Source params 7B (math-specialized)
Quantized weights ~5.3 GB on disk
Inference VRAM (incl. KV cache @ 32K context) ~10 GB

Quick start

from transformers import AutoTokenizer
from hqq.models.hf.base import AutoHQQHFModel

tok = AutoTokenizer.from_pretrained("drawais/Qwen2.5-Math-7B-Instruct-HQQ-INT4")
model = AutoHQQHFModel.from_quantized("drawais/Qwen2.5-Math-7B-Instruct-HQQ-INT4", device="cuda")

License

Apache 2.0 (inherits from base model).