metrale_model_layers/quant_format/
modelopt.rs

1// SPDX-License-Identifier: MIT OR Apache-2.0
2
3//! 2026-09-25: The ModelOpt NVFP4 layout (`.weight`, `.weight_scale`,
4//! `.weight_scale_2`, optional `.input_scale`), mapped to
5//! [`Nvfp4Variant::Standard`].
6//!
7//! Owner: model-layers (weight loading).
8//! Invariants: none beyond the types.
9
10use crate::quant_format::{IgnoreList, QuantFormat};
11use crate::weight_map::Nvfp4Variant;
12use metrale_config::precision_plan::IgnoreDialect;
13
14/// 2026-09-25: A ModelOpt checkpoint, and the fallback of
15/// [`super::detect_quant_format`] when no quantized weights are found.
16#[derive(Debug)]
17pub struct ModeloptFormat {
18    /// 2026-09-25: The config's `quant_algo`; nothing in this module reads it,
19    /// and [`QuantFormat::base_variant`] is `Standard` whatever it says.
20    pub algo: String,
21    /// 2026-09-30: The modules that ship unquantized, matched as the format specifies.
22    pub ignore: IgnoreList,
23}
24
25impl ModeloptFormat {
26    /// 2026-09-30: A malformed ignore entry is refused.
27    pub fn new(algo: String, ignore_modules: &[String]) -> anyhow::Result<Self> {
28        Ok(Self {
29            algo,
30            ignore: IgnoreList::new(IgnoreDialect::ModelOpt, ignore_modules)?,
31        })
32    }
33}
34
35impl QuantFormat for ModeloptFormat {
36    fn name(&self) -> &'static str {
37        "modelopt"
38    }
39
40    fn base_variant(&self) -> Nvfp4Variant {
41        Nvfp4Variant::Standard
42    }
43
44    fn is_ignored(&self, module_path: &str) -> bool {
45        self.ignore.matches(module_path)
46    }
47}
metrale_model_layers/quant_format/modelopt.rs

metrale_model_layers/quant_format/
modelopt.rs
