Revert #79 #192
Merged
Revert #79 #192
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
While testing potential improvements of
IQ1_S_R4quantization, I ran into NaNs while running a DeepSeek-Lite perplexity calculation. I did agrep -ron a folder with many big files while running the calculation and suddenly I got a NaN PPL. I repeated the calculation without doing anything else at the same time and the NaN did not happen. I then ran with 32 threads on a 16-core system and was able to reliably get a NaN at some random chunk.This means there is a race.
The race was most likely introduced in #79 (avoid repeating already done quantizations of activations). I honestly do not understand why there could be a race, or even less do I understand why it would only happen for DeepSeek-Lite quantized with
IQ1_S_R4. I have done countless runs since #79 and never observed anything suspicious.Either way, this PR reverts #79. After doing so, there aren't any NaNs no matter how busy I make the system while running DeepSeek-Lite inference. Hopefully this will also fix the NaNs @saood06 gets with
IQ1_S_R4quantized DeepSeek-R1 (see discussion in #185).