tiny-d1-omni-600M-for-testing

This is a modified, damaged copy of LiquidAI/d1-omni-600M, made only to test software. Its output is gibberish. Do not use it for anything else.

It keeps the architecture, tokenizer, decision head and temperatures of d1-omni-600M, so it runs the same code paths as the real model in llama.cpp (lfm2-d1-omni decision type: bidirectional LFM2 trunk, decision head blocks, option markers, image and audio input), but it is small enough for CI.

GGUF version: ggml-org-dev/tiny-d1-omni-600M-for-testing-GGUF

Changes from the original model

Sliced from the safetensors with slice_d1omni600m.py, which follows the slice-model skill of llama.cpp:

  • Trunk: 4 of 16 layers (0, 2, 3, 5), hidden size 1024 -> 128, heads 16 -> 2, KV heads 8 -> 1, FFN 4608 -> 256 (intermediate_size 6656 -> 384)
  • Decision head: both layers kept, width 1024 -> 128 (heads 16 -> 2, MLP 4096 -> 512), scorer 1024 -> 128
  • Vision encoder (SigLIP2): 2 of 12 layers, hidden size 768 -> 128, heads 12 -> 2, FFN 3072 -> 256, projector 2048 -> 256
  • Audio encoder (FastConformer): 1 of 17 layers, FFN 2048 -> 512 (ff_expansion_factor 4 -> 1), subsampling channels 256 -> 32, residual width 512 -> 128. The width (d_model 512) is kept, llama.cpp hardcodes it
  • The weights are stored in F32 as in the original; the Python code of the original repo is not included

License

Derived from d1-omni-600M by Liquid AI, under the LFM Open License v1.0.

Downloads last month
-
Safetensors
Model size
13.5M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ggml-org-dev/tiny-d1-omni-600M-for-testing

Finetuned
(11)
this model
Quantizations
1 model