A hybrid model stacks n_csa CSA layers (one KV entry plus an indexer key per m tokens; queries attend to top_k entries) and n_hca HCA layers (one entry per m_prime tokens, all attended), each with a window of n_win uncompressed entries. Implement hybrid_attention_budget(seq_len, n_csa, n_hca, m, m_prime, top_k, n_win, head_dim, indexer_dim, n_heads, n_indexer_heads, kv_entry_bytes, indexer_entry_bytes) (all ints) returning a dict.
KV cache bytes for the sequence:
(seq_len // m) * (kv_entry_bytes + indexer_entry_bytes) + ((seq_len % m) + n_win) * kv_entry_bytes.(seq_len // m_prime + (seq_len % m_prime) + n_win) * kv_entry_bytes."hybrid_cache_bytes": sum over all layers."baseline_cache_bytes": seq_len * (n_csa + n_hca) * 2 * 8 * 128 * 2 (BF16 GQA-8, head dim 128)."cache_ratio": hybrid / baseline, rounded to 6 decimals.Attention FLOPs for one decoded token (2 per multiply-add, no projections):
2 * n_indexer_heads * (seq_len // m) * indexer_dim + 4 * n_heads * (top_k + n_win) * head_dim.4 * n_heads * (seq_len // m_prime + n_win) * head_dim."hybrid_flops": sum over all layers; "baseline_flops": (n_csa + n_hca) * 4 * n_heads * seq_len * 128; "flops_ratio": hybrid / baseline, rounded to 6 decimals.Counts are Python ints; only the ratios are floats.