3 ms·

GLM-130B: An Open Bilingual Pre-Trained Model

• homarp 4y ago
  "GLM-130B reaches INT4 quantization w/ no perf degradation, allowing effective inference on 43090 or 82080 Ti GPUs, the most ever affordable GPUs required for using 100B-scale models" (https://twitter.com/thukeg/status/1579449491316674560 https://twitter.com/thukeg/status/1579449491316674560) and by extrapolation "Presumably this means you can fit a solid 30B or so model on a single gaming GPU." (https://twitter.com/TheSeaMouse/status/1579494085211328512 https://twitter.com/TheSeaMouse/status/1579494085211328512)