Module loaders_fp8 

2026-09-25: FP8 block-scaled weight loading, BF16-to-NVFP4 quantization, and the attention, SSM and MoE layer loaders.

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

• Every Fp8Weight returned here holds its block scale as an FP32 device buffer.

Functions

load_attention

2026-09-25: Load the {layer_prefix}.self_attn weights of a full-attention layer.

load_fp8_block_scaled_as_fp8weight

2026-09-25: Load an FP8 E4M3 block-scaled checkpoint weight as a native Fp8Weight.

load_moe

2026-09-25: Load the {layer_prefix}.mlp MoE weights of a layer.

load_moe_skip_experts

2026-09-25: load_moe with every routed expert left as ExpertWeight::null(). The native-FP8 qwen3 loader uses it and loads the routed experts separately.

load_ssm

2026-09-25: Load the {layer_prefix}.linear_attn weights of a linear-attention layer.

quantize_to_nvfp4

2026-09-25: Quantize a BF16 dense [n, k] weight to NVFP4 on the GPU.