Module helpers

metrale_model_layers::layers::qwen3_attention

Module helpers 

Source
Expand description

2026-09-25: Qwen3AttentionLayer setters that model loaders call after construction, the YaRN mscale helper, and small per-layer helpers (apply_layer_scalar, effective_attn_scale, nvfp4_decode_gemv, qsa_seq_state).

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

Functionsยง

yarn_rope_mscale
2026-09-25: The YaRN mscale ratio get_mscale(factor, yarn_mscale) / get_mscale(factor, yarn_mscale_all_dim), or 1.0 when yarn_factor <= 1. The DeepSeek-V4 RoPE call sites pass it to their kernels.