attn_nvfp4_bytes

spark_model::weight_loader::fp8_residency

Function attn_nvfp4_bytes 

Source 
pub fn attn_nvfp4_bytes(
    q_n: usize,
    kv_n: usize,
    o_k: usize,
    hidden: usize,
) -> usize
Expand description

What the pre-#915 loader spent per full-attention layer on NVFP4: q/k/v/o, each with a transposed twin, plus the fused [q|k|v] transposed twin (transpose_concat_for_gemm).

q_n is num_attention_heads * head_dim, doubled when attn_gated; kv_n is num_key_value_heads * head_dim; o_k is the o_proj contraction width num_attention_heads * head_dim.