ahmedandaloes/d3-lite-mlx

d3-lite — MLX (Apple Silicon)

MLX conversion of vllm-sr/d3-lite, part of the Decision 3.0 collection — the System One code-readout decision model (Qwen3.5-VL backbone + a 255-way readout head).

Install

pip install mlx-vlm transformers

Use

from d3_mlx import D3MLX

m = D3MLX(".")
result = m.system_one(
    "state text the model should treat as data",
    {"route": {"type": "choice", "criteria": {"a": "first option", "b": "second option"}}},
)
print(result)

Files

• weights.safetensors — MLX backbone weights (remapped + transposed)
• readout.safetensors — the original 255-way readout head (unchanged)
• d3_mlx.py — MLX runtime: loader, prompt contract, decision path
• config.json, decision_config.json, tokenizer files — copied from the original checkpoint

What was changed to make it run on MLX

• key remap: language_model.* → language_model.model.*; language_model.visual.* → vision_tower.*
• linear_attn.conv1d.weight transposed [C,1,K] → [C,K,1]
• patch_embed.proj.weight transposed NCDHW → NDHWC
• decoder layer types taken from the checkpoint's explicit layer_types list (irregular patterns: the small models end with several full-attention layers; the 32/64-layer models use a uniform every-4 pattern — the stock MLX class derives this and gets it wrong)
• tied lm_head dropped (embeddings are tied; only hidden states are needed)
• prompt format, answer codes and the readout head are identical to the original

Numerics: BF16 backbone, FP32 readout + softmax — the same as the reference implementation.

Weights are unchanged from the original; only the format/layout was converted. Converted with mlx-vlm on an Apple M3 Max.