Skip to content

dduf.md

Latest commit

 

History

History
42 lines (31 loc) · 1.45 KB

dduf.md

File metadata and controls

42 lines (31 loc) · 1.45 KB
Raw

DDUF weight loading

HuggingFace DDUF (.dduf) is a ZIP of nested *.safetensors plus JSON (model_index.json, per-component config.json). RLX reads it via rlx-dduf.

Name convention

Tensors are qualified as {component}/{tensor_name}, e.g. transformer/weight, vae/decoder.conv_in.weight. Root-level members use ./{tensor_name}.

CLI / APIs

rlx-pkg import-dduf model.dduf -o model.rlxp --no-graph
rlx-bake graph.json -o model.rlxp --weights model.dduf --weights-policy auto

DDUF has no MLX-style affine/mxfp packs. “Packed” for DDUF means keep native f16/bf16 bytes (rlx_dduf::load_native / bake --weights-policy packed|auto) instead of widening everything to f32.

use rlx_dduf::{DdufFile, load_native, load_shaped_f32, visit_f32_tensors};
let f = DdufFile::open("model.dduf")?;
let w = f.tensor_f32("transformer/weight")?;
let shaped = load_shaped_f32("model.dduf")?; // keeps on-disk dims
let native = load_native("model.dduf")?; // encoding = f16|bf16|f32|…
// Large packs: stream one safetensors ZIP member at a time
visit_f32_tensors("model.dduf", |t| { /* use t.shape + t.data */ Ok(()) })?;

import-dduf uses visit_f32_tensors so peak memory is roughly one member + the growing weight list (not the full ZIP decoded at once).

Dist URI: dduf://<path>#<component/tensor>.

v1 is read/import only (no DDUF writer).

License

MIT OR Apache-2.0.