Can fairseq save the model parameters as fp16?

[SefaZeng]

❓ Questions and Help

Before asking:

1. search the issues.
2. search the docs.

What is your question?

I noticed fairseq can train with fp16 but the saved model parameters are still fp32. But it seems fp32 model will have a larger accuracy loss when using dynamic quant from pytorch. Can fairseq save model parameters as fp16 directly?

Code

What have you tried?

What's your environment?

• fairseq Version (e.g., 1.0 or main):
• PyTorch Version (e.g., 1.0)
• OS (e.g., Linux):
• How you installed fairseq (pip, source):
• Build command you used (if compiling from source):
• Python version:
• CUDA/cuDNN version:
• GPU models and configuration:
• Any other relevant information: