Deduplicated dequantization code #1453
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Outdated suggestions cannot be applied.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.
For my GPU acceleration PR #1412 I used a template to decouple the code for matrix vector multiplication and dequantization. This PR applies the same principle to the dequantization during prompt processing: the same dequantization kernels can be reused with a different template. This allows for the deduplication of CUDA code to ensure consistency. As a side effect the new kernels are also slightly faster on my hardware.
Performance numbers for perplexity calculation on the first 100 lines of wikitext:
I will also add GTX 1070 numbers once I have them.Done.The goal of this PR is not to optimize performance. The goal is to simplify the code base to allow for easier development. If the new kernels don't cause a performance regression I consider that good enough.