Module loaders_moe 

2026-09-25: MoE and MTP-head weight loaders: the mlp MoE block, the Mistral w1/w2/w3 MoE block, and the MTP head.

Owner: model-layers (weight loading). Invariants:

• MtpWeights::fp8_experts is Some only when the experts are unstacked, the variant is not Bf16Raw, and expert 0 is FP8 with weight_scale_inv.

Functions

load_moe_mistral

2026-09-25: Load a Mistral MoE block (w1/w2/w3 naming: w1 = gate_proj, w2 = down_proj, w3 = up_proj) through quantized_v2.

load_mtp

2026-09-25: Load the MTP head (mtp.* keys) as BF16 DenseWeights.