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>,
}

Nemotron-H MoE layer weights.

Fields

gate: DenseWeight

Router gate: [num_experts, hidden_size] F32→BF16.

e_score_correction_bias: DenseWeight

Expert score correction bias: [num_experts] F32.

experts: Vec<NemotronExpertWeight>

Per-expert weights (routed): NVFP4.

shared_up: QuantizedWeight

Shared expert up_proj: [shared_inter, hidden_size] NVFP4.

shared_up_fp8: Option<Fp8Weight>

Shared expert up_proj kept as NATIVE FP8 when the checkpoint ships it that way (ModelOpt MIXED_PRECISION), instead of the FP8→BF16→NVFP4 requant. Some only under ATLAS_NEMOTRON_NATIVE_FP8_SSM; decode prefers it via w8a16_gemv. Measured on Puzzle-75B: with the SSM projections already native, a 977-token story went from calling the dog “Rover”/“Rex” to using the given name “Rufus” 8 times with no substitutions — proper-noun retrieval is what the requant was destroying.

shared_down: QuantizedWeight

Shared expert down_proj: [hidden_size, shared_inter] NVFP4.

shared_down_fp8: Option<Fp8Weight>

Shared expert down_proj kept as NATIVE FP8, same rationale as shared_up_fp8. Consumed by relu_squared_inplace + w8a16_gemv instead of the fused moe_expert_relu2_down_shared kernel, which only speaks NVFP4 — the fused launch simply drops its shared slot (grid.y = top_k) when this is present.

fc1_latent_proj: Option<DenseWeight>

LatentMoE: fc1 [moe_latent_size, hidden_size] BF16 (dequant from FP8 at load). Present only for Super 120B (moe_latent_size > 0).

fc2_latent_proj: Option<DenseWeight>

LatentMoE: fc2 [hidden_size, moe_latent_size] BF16. Present only for Super 120B (moe_latent_size > 0).