Add pre-converted MLX weights (converter_version 1)

60d7136 verified 3 days ago

1.22 kB

license: mit
library_name: mlx
tags:
  - depth-estimation
  - mlx
  - apple-silicon
base_model: robbyant/lingbot-depth-pretrain-vitl-14-v0.5

lingbot-depth-mlx

MLX weights for LingBot-Depth (mdm), pre-converted for the mlx-native backend of lingbot-depth-viz.

These are the original robbyant/lingbot-depth-pretrain-vitl-14-v0.5 weights, converted to MLX safetensors with the position embeddings precomputed, so the whole model runs on Apple Silicon through mlx.core alone. Hosting them lets the torch-free install run without downloading the torch checkpoint or running the conversion:

uv sync --extra mlx
uv run lingbot-depth-viz --backend mlx-native --mode benchmark
# weights.safetensors is fetched from here on first run

Files

weights.safetensors — MLX arrays (bf16 encoder, fp32 decoder) plus precomputed position embeddings.
config.json — model dims and preprocessing constants read by the runtime.

Provenance

Produced by --mode convert-mlx (converter_version 1). Parity vs torch-MPS is 0.27–0.48% AbsRel end-to-end. See the repo's results/mlx-backend.md.