README.md

metadata

license: mit
base_model:
  - deepseek-ai/DeepSeek-V3
pipeline_tag: text-generation
library_name: transformers

DeepSeek V3 FP16 Atten NaN

This is a minimal reproduceable sample to let the final layer of DeepSeek V3's attention output NaNs when using data type float16.

Run the run.py to see the NaNs.

Weights are converted to bfloat16 using the original float8 e4m3fn, then converted to float16, then extracted from the final layer's attention.