Skip to content

[quantization] [draft] Introduce smse_for_gptq#597

Draft
stamalakhov wants to merge 1 commit intoSamsung:mainfrom
stamalakhov:smse_for_gptq
Draft

[quantization] [draft] Introduce smse_for_gptq#597
stamalakhov wants to merge 1 commit intoSamsung:mainfrom
stamalakhov:smse_for_gptq

Conversation

@stamalakhov
Copy link
Copy Markdown
Contributor

@stamalakhov stamalakhov commented Mar 31, 2026

This PR introduces smse_for_gptq to improve accuracy.

Config ID train_PPL test_PPL
FP32 11.09 11.05
GPTQ_MSE_w4A16_smse_128_samples 12.25 12.12
GPTQ_MSE_w4A16_smse_for_gptq_128_samples 12.19 12.12
GPTQ_MSE_w4A16_smse_256_samples 12.25 12.12
GPTQ_MSE_w4A16_smse_512_samples 11.22 12.28
GPTQ_MSE_w4A16_smse_for_gptq_512_samples 12.16 12.13

where
train_ppl - perplexity evaluated on calibration dataset,
test_ppl is perplexity evaluated on test dataset.

TICO-DCO-1.0-Signed-off-by: s.malakhov s.malakhov@partner.samsung.com

@stamalakhov stamalakhov self-assigned this Mar 31, 2026
This PR introduces smse_for_gptq to improve accuracy.

TICO-DCO-1.0-Signed-off-by: s.malakhov <s.malakhov@partner.samsung.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant