Merge the GGUF files (text encoder, Q8_0 and BF16 vision encoders) into the LMK repository

3ea3e57 verified 11 days ago

2.85 kB

license: apache-2.0
base_model: google/embeddinggemma-2
pipeline_tag: feature-extraction
library_name: lm-kit
tags:
  - embedding
  - sentence-similarity
  - multimodal-embedding
  - lmk
  - gguf
  - lm-kit
language:
  - multilingual

EmbeddingGemma 2 (LMK) for LM-Kit.NET

An LM-Kit model archive of google/embeddinggemma-2, the open multimodal embedding model from Google DeepMind, packaged for LM-Kit.NET. Text, code and images are embedded into one 768-dimensional vector space (Matryoshka truncation to 512, 256 or 128 dimensions, re-normalized), with an 8,192-token context.

File	Content	Size
`embeddinggemma-2-Q5_K_M.lmk`	LM-Kit archive: the text encoder (Q5_K_M) and the vision encoder (Q8_0) in one file. This is what `LM.LoadFromModelID("embeddinggemma-2")` loads	439 MB
`embeddinggemma-2-Q5_K_M.gguf`	Text encoder (270M parameters), Q5_K_M	213 MB
`embeddinggemma-2-mmproj-vision-Q8_0.gguf`	Vision encoder (170M parameters), Q8_0, audio tower removed	226 MB
`embeddinggemma-2-mmproj-vision-BF16.gguf`	Vision encoder, BF16 reference (full precision, for re-quantization)	368 MB

The archive's vision encoder reads only its header at load: its weights reach memory with the first image, so text-only use costs no vision memory. The GGUF files are the archive's contents, for runtimes that load GGUF directly.

Usage

using LMKit.Model;
using LMKit.Embeddings;
using LMKit.Media.Image;

// Text, code and images, in the same vector space
using LM model = LM.LoadFromModelID("embeddinggemma-2");
var embedder = new Embedder(model);

float[] query    = embedder.GetQueryEmbeddings("waterproof running shoes");
float[] document = embedder.GetEmbeddings("Our trail shoe keeps your feet dry on wet rock.");
float[] image    = embedder.GetEmbeddings(ImageBuffer.LoadAsRGB("shoe.jpg"));

LM-Kit.NET applies the model card's prompts by retrieval role (task: search result | query: for queries, title: none | text: for documents), frames every input with <bos> and <eos> as the publisher's tokenizer does, and sizes images with the model's own processor rule (280 soft tokens by default, aspect ratio preserved, bicubic resampling).

Conversion

Converted with LM-Kit's llama.cpp fork (upstream convert_hf_to_gguf.py, architecture gemma-embedding2, projector gemma4v). Every text tokenizes to exactly the publisher's token ids, and the vectors match the publisher's own pipeline (sentence-transformers) for text and images.

License

Apache 2.0, as the original model. See the Gemma 4 license and the original model card.