Much faster prompt processing for IQ1_S and IQ1_M on ARM_NEON#553
Merged
Conversation
added 2 commits
June 24, 2025 13:27
This file contains hidden or 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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 PR corresponds to PRs #531, #533, #534, #546, #549, #550, #552, and applies the on-the-fly repacking technique to
the 1-bit quants
IQ1_SandIQ1_MonARM_NEON.Here is a PP-512 performance comparison between the main branch and this PR for LlaMA-3.1-8B-Instruct on M2-Max
IQ1_Mdid not have a fasterIQKimplementation, so the 19 t/s is what one has within the standardggmlGEMM framework.