Struct NemotronMoeWeights 

pub struct NemotronMoeWeights {
    pub gate: DenseWeight,
    pub e_score_correction_bias: DenseWeight,
    pub experts: Vec<NemotronExpertWeight>,
    pub shared_up: QuantizedWeight,
    pub shared_up_fp8: Option<Fp8Weight>,
    pub shared_down: QuantizedWeight,
    pub shared_down_fp8: Option<Fp8Weight>,
    pub fc1_latent_proj: Option<DenseWeight>,
    pub fc2_latent_proj: Option<DenseWeight>,
}

2026-09-25: Nemotron-H MoE layer weights.

Fields

gate: DenseWeight

2026-09-25: Router gate, BF16 (converted when the checkpoint stores FP32).

e_score_correction_bias: DenseWeight

2026-09-25: The store’s gate.e_score_correction_bias pointer, unconverted; moe_topk_sigmoid reads it as F32.

experts: Vec<NemotronExpertWeight>

2026-09-25: Routed experts as NVFP4; experts this rank does not hold are null.

shared_up: QuantizedWeight

2026-09-25: Shared-expert up_proj as NVFP4; NULL when shared_up_fp8 is Some.

shared_up_fp8: Option<Fp8Weight>

2026-09-25: Shared-expert up_proj as the checkpoint’s own FP8 bytes, in an owned copy. Some when METRALE_NEMOTRON_NATIVE_FP8_SSM is unset, 1, both or decode, the tensor has weight_scale but no weight_scale_2, and it loads.

shared_down: QuantizedWeight

2026-09-25: Shared-expert down_proj as NVFP4; NULL when shared_down_fp8 is Some.

shared_down_fp8: Option<Fp8Weight>

2026-09-25: Shared-expert down_proj as native FP8, under the same condition as shared_up_fp8. When its kernels are available, decode applies relu² in place and w8a16_gemv to it, and launches the fused NVFP4 kernel without its shared slot.

fc1_latent_proj: Option<DenseWeight>

2026-09-25: LatentMoE fc1, BF16 (FP8 is dequantized at load). Some when config.moe_latent_size > 0.

fc2_latent_proj: Option<DenseWeight>

2026-09-25: LatentMoE fc2, BF16 (FP8 is dequantized at load). Some when config.moe_latent_size > 0.