Skip to content
This repository was archived by the owner on Jun 10, 2021. It is now read-only.

[WIP] read quantizedWeight file#32

Open
jsenellart wants to merge 3 commits intoOpenNMT:masterfrom
jsenellart:quantizeWeights
Open

[WIP] read quantizedWeight file#32
jsenellart wants to merge 3 commits intoOpenNMT:masterfrom
jsenellart:quantizeWeights

Conversation

@jsenellart
Copy link
Copy Markdown
Contributor

@jsenellart jsenellart commented May 14, 2018

quick and dirty implementation - for reading quantized weights.

  • for lookup tables, rows are converted on the fly back to float
  • for linear weights, the only adjustment is a possible memory alignment
  • for linear biases, convert on the fly to float
@jsenellart jsenellart requested a review from guillaumekln May 14, 2018 22:17
Jean A. Senellart added 2 commits May 15, 2018 07:49
more strict test to check if row has been unquantized
6fc8654
Merge branch 'master' into quantizeWeights
597c46f 
* master: unroll more main matrix mult loop with AVX512 for 10% additional efficiency
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

1 participant

@jsenellart