metrale_model_layers/forward/
mod.rs

1// SPDX-License-Identifier: MIT OR Apache-2.0
2
3//! 2026-09-25: Per-layer decode forwards written against `GpuBackend` and the
4//! `QuantWeights` trait, so a driver runs a layer without knowing the weight
5//! format.
6//!
7//! Owner: model-layers.
8//! Invariants: none beyond the types.
9//!
10//! The only caller is the model-arch example `metal_qwen35_inference`. The
11//! served decode path does not use this module; it lives in model-engine
12//! (`model/trait_impl/decode_a.rs`).
13
14pub mod quant_weights;
15pub mod qwen3_5;
16
17pub use quant_weights::QuantWeights;