Module ssm_qwen35 

2026-09-25: Qwen3.5-family loaders: the separate-projection linear-attention weights and three MoE loaders (NVFP4, native FP8 experts, no shared expert).

Owner: model-layers (weight loading). Invariants: none beyond the types.

Structs

Nvfp4MoeCopies

2026-09-27: Which NVFP4 copies load_moe_qwen35 builds. A projection it does not build is NULL (QuantizedWeight::null()).

SsmWeightsQwen35

2026-09-25: Qwen3.5 linear-attention weights with separate projections. The projections, conv1d and norm are BF16; a_log and dt_bias are FP32.

Functions

load_moe_no_shared

2026-09-25: Load a MoE block that has no shared expert. A zero-filled shared expert of real expert size stands in: its packed codes, scales and weight_scale_2 are all 0, so it contributes 0. Routed experts go through quantized_auto, which panics for Fp8Dequanted and Bf16Raw.

load_moe_qwen35

2026-09-25: Load a Qwen3.5 MoE block ({layer_prefix}.mlp) as NVFP4.

load_moe_qwen35_fp8_experts

2026-09-25: Load the routed experts of {layer_prefix}.mlp as native FP8 block-scaled tables, one entry per expert id. Experts this rank does not hold are NULL entries tagged Fp8BlockScaled.

load_ssm_qwen35

2026-09-25: Load the Qwen3.5 {layer_prefix}.linear_attn weights.