load_nemotron_attention

metrale_model_layers::weight_map::nemotron

Function load_nemotron_attention 

Source 
pub fn load_nemotron_attention(
    store: &WeightStore,
    layer: usize,
    gpu: &dyn GpuBackend,
    layer_prefix: &str,
) -> Result<(AttentionWeights, Option<QuantizedWeight>, Option<QuantizedWeight>, Option<QuantizedWeight>, DenseWeight, bool)>
Expand description

2026-09-25: Load a Nemotron-H attention mixer.

Returns (attn, q, k, v, o_dense, is_nvfp4). When q_proj has a weight_scale_2, q/k/v are Some NVFP4 weights, attn.o_proj is NVFP4 and the dense fields are NULL. Otherwise q/k/v/o are BF16 in attn and o_dense (an FP8 projection with weight_scale is dequantized), q/k/v are None, and attn.o_proj is null. The q/k norms are always NULL.