optimizer de-oscillation

[hann-wang]

Method

• At the start of each interval, store the FP32 master weights as a reference.
• During the interval, accumulate the movement of both:
  • FP32 master weights
  • FP4 quantized weights
• At the end of the interval, mark a weight as oscillatory if the FP4 weight moves significantly more than the FP32 master weight.
• For oscillatory weights, reset the FP32 master weight to the dequantized FP4 value to suppress quantization-induced instability.

Usage

Set env var to enable de-oscillation:

• DEOSC_STEP: de-oscillation is not enabled until this step.
• DEOSC_RATIO: threshold of dist(FP4)/dist(FP32).

The observation interval is currently hard-coded to 200.