Is your feature request related to a problem? Please describe.
At the time of checkpoint saving, integer weights can be recomputed using SAWB quantized (instead of MAX) when they show narrow distributions that could cause problems in system with limited MAC accumulation length.
This feature is already covered by PR #95, but only when smoothquant was not used during the quantization process.
Describe the solution you'd like
Support ckpt saving with weight recomputation for models quantized with smoothquant enabled.
Is your feature request related to a problem? Please describe.
At the time of checkpoint saving, integer weights can be recomputed using SAWB quantized (instead of MAX) when they show narrow distributions that could cause problems in system with limited MAC accumulation length.
This feature is already covered by PR #95, but only when smoothquant was not used during the quantization process.
Describe the solution you'd like
Support ckpt saving with weight recomputation for models quantized with smoothquant enabled.