Whisper medium

ENCODER-DECODER. ENCODER: 24 layers, d_model 1024, 16 heads, encoder_ffn_dim 4096, gelu, num_mel_bins 80, max_source_positions 1500. DECODER: 24 layers, 16 heads, decoder_ffn_dim 4096, max_target_positions 448. KV CACHE: decoder self-attention 24 * 16 * 64 * 2 * 2 = 98,304 B/token = 96.0 KiB/token. Cross-attention over 1500 encoder frames: 24 * 16 * 64 * 1500 * 2 * 2 = 117,964,800 B = 112.5 MiB per in-flight request. STREAMING: not natively supported (offline/batch model). KV CACHE CORRECTED (2026-10-05): kv_cache_bytes_per_token was 131,072 B/token and overstated the decoder self-attention cache by 1.33x. Correct value 98,304 B/token = 96.0 KiB/token, from 2 (K and V) x 24 decoder_layers x 16 decoder_attention_heads x 64 head_dim x 2 bytes (fp16). head_dim 64 is CORRECT here and was not changed: d_model 1024 / decoder_attention_heads 16 = 64 for every Whisper variant, so 64 holds across the whole family and is not the source of the error. What was wrong was the multiplication itself - this record's own derivation line read '24 * 16 * 64 * 2 * 2 = 131,072', and that expression evaluates to 98,304, not 131,072. The formula was right and the result beside it was not, which is why the error survived: nothing in validate.py evaluates a notes expression against the field it claims to derive. Cross-attention over the fixed 1500-frame encoder span was already correct and is unchanged. See gotcha mdl-gotcha-derived-figures-field-disagrees-with-its-own-note.