Ovis-Omni-Embedding-3B — experimental mixed W2/W4/W8
This is an experimental, model-free RTN compressed-tensors conversion of ATH-MaaS/Ovis-Omni-Embedding-3B. It is not a GPTQ checkpoint and has not been accuracy-calibrated.
The language-model transformer blocks use mixed weight-only compression:
- layers 0–13: W2A16
- layers 14–26: W4A16
- layer 27: W8A16
The multimodal audio/vision towers and auxiliary talker/token2wav weights are
preserved in BF16. The artifact is intended for vLLM's compressed-tensors
backend and was tested with vLLM 0.30.0 on an 8 GB RTX 3080 Laptop GPU.
vLLM test
vllm serve . --runner pooling --convert embed \
--max-model-len 256 --gpu-memory-utilization 0.75 --enforce-eager
/v1/embeddings returned a normalized 2048-dimensional vector. The test used
the default model pooler; configure a 1024-dimensional pooler in the serving
wrapper if required.
For the original model and license, see the upstream model card.
Independent screening and fidelity status
The published artifact has a constrained vLLM smoke result: HTTP readiness and
a finite normalized 2048-dimensional embedding were observed with
max-model-len=256. No BF16 cosine-drift or held-out retrieval score is
claimed yet; those metrics must use the persisted BF16 reference fixture.
The completed 60-case BF16 comparison measured mean cosine -0.03569794, mean drift 1.03569794, nearest-neighbor agreement 10%, Recall@1 10%, Recall@5 50%, Recall@10 100%, and nDCG 0.46418892. This artifact is therefore not a quality-preserving replacement despite its constrained vLLM smoke pass.