sherpa icon indicating copy to clipboard operation
sherpa copied to clipboard

Published 20 hours ago • verified publisher icon k2-fsa

Support automatic-mixed-precision (AMP)

Open csukuangfj opened this issue 2 years ago • 4 comments

csukuangfj avatar Dec 07 '22 04:12 csukuangfj

Hi, I was wondering if this PR is ready to test or not. I'd like checkout to csukuangfj:mixed-precision-2022-12-07 do a test.

yuekaizhang avatar Jan 12 '23 07:01 yuekaizhang

Hi, I was wondering if this PR is ready to test or not.

I have tested it locally but it does not seem to improve the speed, so it is still left to open.

csukuangfj avatar Jan 12 '23 07:01 csukuangfj

Hi, I was wondering if this PR is ready to test or not.

I have tested it locally but it does not seem to improve the speed, so it is still left to open.

I see. Did you try to use model.half() then export using libtorch? I think it should be pure fp16 mode.

yuekaizhang avatar Jan 12 '23 07:01 yuekaizhang

Hi, I was wondering if this PR is ready to test or not.

I have tested it locally but it does not seem to improve the speed, so it is still left to open.

I see. Did you try to use model.half() then export using libtorch? I think it should be pure fp16 mode.

No, not yet. I have not tried this approach.

csukuangfj avatar Jan 12 '23 07:01 csukuangfj