lmsys/vicuna-33b-v1.3 config.json

LlamaForCausalLM config read 2026-10-18. hidden_size 6656, num_hidden_layers 60, num_attention_heads 52, max_position_embeddings 2048, torch_dtype float16, vocab_size 32000. No num_key_value_heads key, so this is MHA: 52 KV heads, head_dim 6656/52 = 128. Used only to derive the KV bytes behind the SGLang RadixAttention hit-rate record.