Add LFM2.5-Embedding-350M dense bi-encoder support#67
Open
franckverrot wants to merge 1 commit into
Open
Conversation
Extend the existing lfm2.py (previously ColBERT-only) to support the LiquidAI/LFM2.5-Embedding-350M dense bi-encoder, which shares model_type "lfm2". The dense variant is detected via the presence of 1_Pooling/config.json and differs from the late-interaction ColBERT variant in three ways: - CLS pooling (via pool_by_config) instead of mean pooling - no 1_Dense projection; outputs the raw 1024-dim hidden state - bidirectional attention with a non-causal key-padding mask The bidirectional models use a centered short convolution over t-1..t+1, whereas mlx_lm's generative ShortConv left-pads causally, so had to swap a NonCausalShortConv into the conv layers for the dense variant. Verified we still produce embeddings numerically identical to the reference implementation. AI (Opus 4.8) was used to assist in the refactor/debugging.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Extend the existing lfm2.py (previously ColBERT-only) to support the LiquidAI/LFM2.5-Embedding-350M dense bi-encoder, which shares model_type "lfm2".
The dense variant is detected via the presence of 1_Pooling/config.json and differs from the late-interaction ColBERT variant in three ways:
The bidirectional models use a centered short convolution over t-1..t+1, whereas mlx_lm's generative ShortConv left-pads causally, so had to swap a NonCausalShortConv into the conv layers for the dense variant. Verified we still produce embeddings numerically identical to the reference implementation.
AI (Opus 4.8) was used to assist in the refactor/debugging.