Multihead arch cutlass int8 qkv average #2035
Draft
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
- skip connection add before layer norm now has a scaling factor (alpha) - replace conv layer of value and mlh heads with an embedding layer when attention body is used.
- also remove hardcoding.
- will be removed once it's fixed.
- also fix scratch space calculation.
- factor of sizeof(DataType) was missing.
get almaudoh's code
- to handle bigger/wider networks
1.3% improvement in BT2 on RTX 4090 15.6% improvement in test BT3 network with 64 heads.
- only tries doing the KQV dense layers in int8. - Accuracy seems reasonable. - Right now quantization isn't fused, and de-quantization is done with bias add. - Both the above can be possibly be fused with more work. - Also need to attempt INT8 for other dense layers (MHA dense, FFN1 and FFN2)
…rnels for clipping of inputs for non-int8 inference.
…mal fp16 weights.
…udoh/lc0 into multihead-arch-cutlass-int8
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Outdated suggestions cannot be applied.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.
This is a temporary PR to allow testing of a branch of cuda int8 that may not eventually be merged.