nan problem of Qwen2-72B quantization #519
Merged
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Outdated suggestions cannot be applied.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.
change weight scaling formulation, fix nan problem when quantize Qwen2-72B model
For #498 , casper-hansen #516 and Qwen team yangyo@32bf03c?diff=split&w=1 fix this problem by set nan or inf to 1 to walkaround it, But I think this is unreasonble.
I found the occurence of nan was caused by the process of weight scaling, where part of some weights like mlp.gate_proj, mlp.up_proj exceed the range of float16. So those weights become to 0 when loading model. The nans occur when cacluating 0/0. Add some small value to denominator can solve this problem.