embeddinggemma-2-UD-Q4_K_XL.gguf is an unmodified copy of the file from
unsloth/embeddinggemma-2-GGUF
at revision ba3888272494be64ed88c9eb536ddc61a1be73d5.
ea905fd08e8061db77a0031cbb5096d459d7cfd2bdd0ad2ba66e483f16719493The base model is Google's EmbeddingGemma 2, distributed under the Apache 2.0 license.
embeddinggemma-2-512d-UD-Q4_K_XL.ggufThe file above with one tensor changed, giving 512-dimensional vectors. Its output projection output.weight
(512 to 768, Q6_K) was dequantized and factored in FP64 as W = U diag(S) Vh, and replaced by the 512 x 512
F32 matrix diag(S) Vh (the method of tomaarsen/embeddinggemma-2-512d).
Because U has orthonormal columns, the vectors keep the norms and cosine similarities of the 768-dimensional
originals, up to floating-point rounding. Every other tensor is byte-identical to the file above, and
gemma-embedding2.embedding_length_out is 512. These vectors cannot be truncated below 512 dimensions:
the factorization does not keep the Matryoshka structure.
6e535e12366cf8ebf6635b1c1c0079610c09a36c1b0ec4220009aa40f2b54f58embeddinggemma-2-512d-Q8_0.ggufMade the same way from the Q8_0 file of unsloth/embeddinggemma-2-GGUF
at revision 031f0d4b35536f69ab3509d4893c923264fcf253 (embeddinggemma-2-Q8_0.gguf, 309855520 bytes, SHA-256
6f1bd4ac6c5df7444f9cca7ca36cafe6cfa34cd6f49fefb1e0b4be8143aed8bc): its output projection output.weight
(512 to 768, Q8_0) was dequantized and factored in FP64 as W = U diag(S) Vh, and replaced by the 512 x 512 F32
matrix diag(S) Vh. Every other tensor is byte-identical to that file, and gemma-embedding2.embedding_length_out
is 512. Each factorization fixes its own basis, so these vectors cannot be mixed with those of the UD-Q4_K_XL
512-dimensional file above.
20bd7c54d03634f599e8a9a35a399d2d2c3573097c3404efd2f76150c554bc95google/embeddinggemma-2