ngxson's picture
ngxson HF Staff
tiny slice of d3-nano for testing
88ecebc verified
|
Raw History Blame Contribute Delete
1.79 kB
metadata
license: apache-2.0
base_model: vllm-sr/d3-nano
tags:
  - testing

tiny-d3-nano-for-testing

This is a modified, damaged copy of vllm-sr/d3-nano, made only to test software. Its output is gibberish. Do not use it for anything else.

It keeps the architecture, tokenizer, chat template, decision_config.json (codes, token ids, temperature, attention mode) and readout of d3-nano, so it runs the same code paths as the real model in llama.cpp (vllm-d3 decision type: non-causal attention, readout of 255 option codes stored as an LM head, entropy confidence, image input), but it is small enough for CI.

GGUF version: ggml-org-dev/tiny-d3-nano-for-testing-GGUF

Changes from the original model

Sliced from the safetensors with slice_qwen35_decision.py (python3 slice_qwen35_decision.py vllm-sr/d3-nano <out dir>), which follows the slice-model skill of llama.cpp:

  • Text model: 4 of 24 layers (0-3: 3 linear attention + 1 full attention), hidden size 2048 -> 32, FFN 6144 -> 64, heads 8 -> 2, KV heads 2 -> 1 (head_dim 256 kept), linear attention key heads 16 -> 2, value heads 16 -> 2 (head dims 128 kept)
  • Vision encoder: 2 of 24 blocks, hidden size 1024 -> 128, heads 16 -> 2 (head_dim 64 kept), FFN 4096 -> 256, output 2048 -> 32
  • readout.safetensors: 255 x 2048 -> 255 x 32 (first columns)
  • The Python code of the original repo (d3_*.py, modeling_d3.py, pipeline_d3.py), MODEL_MANIFEST.json, assets and eval results are not included

License

Derived from d3-nano by the vLLM Semantic Router Team (itself built on Qwen/Qwen3.5-2B), under the Apache License 2.0. See NOTICE.