This repository has been archived by the owner on Aug 7, 2024. It is now read-only.
Adding Float8 Linear variants supporting inference-only with lower overhead #283
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Outdated suggestions cannot be applied.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.
The changes include two new Float8 Linear implementations that removes some extra wiring in
Float8Linear
unnecessary for inference-only use cases to result in lower latency.Float8SWLinear
supports direct fp8 type direct downcast for activation, and Static per-tensor scale for Weight. Our analysis shows that using this results in no loss of accuracy in Llama models.Float8DASWLinear
supports Dynamic per-tensor scale for Activation, and Static per-tensor scale for Weight. This is used when activation tensor requires dynamic scaling. Compared toFloat8SWLinear
, this has higher overhead introduced by dynamic activation tensor scale calculation. The overhead can be mitigated when used withtorch.compile
.cc: @ani300