Skip to content

fix fp8 accuracy drop issue#1611

Merged
wenhuach21 merged 13 commits intomainfrom
mengni/fp8_fix
Mar 26, 2026
Merged

fix fp8 accuracy drop issue#1611
wenhuach21 merged 13 commits intomainfrom
mengni/fp8_fix

Conversation

@mengniwang95
Copy link
Copy Markdown
Contributor

Description

fp8 quant_func only use weight.max() for scale calculation when tuning

Type of Change

  • Bug fix
  • New feature
  • Documentation update
  • Performance improvement
  • Code refactoring
  • Other (please specify):

acc before:

Tasks Version Filter n-shot Metric Value Stderr
gsm8k 3 flexible-extract 5 exact_match 0.7286 ± 0.0122
strict-match 5 exact_match 0.7316 ± 0.0122

acc after:

Tasks Version Filter n-shot Metric Value Stderr
gsm8k 3 flexible-extract 5 exact_match 0.7589 ± 0.0118
strict-match 5 exact_match 0.7619 ± 0.0117

@chensuyue chensuyue added this to the 0.12.0 milestone Mar 25, 2026
@wenhuach21
Copy link
Copy Markdown
Contributor

Really nice catch! Thanks!

mengniwang95 and others added 4 commits March 25, 2026 12:00
Signed-off-by: Mengni Wang <mengni.wang@intel.com>
Signed-off-by: Mengni Wang <mengni.wang@intel.com>
@wenhuach21
Copy link
Copy Markdown
Contributor

based on the data,I still recommend using rtn

@mengniwang95
Copy link
Copy Markdown
Contributor Author

based on the data,I still recommend using rtn

okay

@wenhuach21
Copy link
Copy Markdown
Contributor

/azp run Unit-Test-CUDA-AutoRound

@azure-pipelines
Copy link
Copy Markdown

Azure Pipelines successfully started running 1 pipeline(s).

@wenhuach21 wenhuach21 changed the title fix fp8 quant_func fix fp8 accuracy drop issue Mar 26, 2026
@wenhuach21 wenhuach21 merged commit b78453c into main Mar 26, 2026
40 checks passed
@wenhuach21 wenhuach21 deleted the mengni/fp8_fix branch March 26, 2026 06:04
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.

4 participants