VeriLoop E2 config.json

config.json fetched 2026-10-18 from tsinghua-sigs-robot-lab/VeriLoop-E2. architectures Qwen3_5ForConditionalGeneration, model_type qwen3_5, language_model_only false. text_config: hidden_size 5120, num_hidden_layers 64, layer_types 48 linear_attention + 16 full_attention at full_attention_interval 4 (indices 3,7,...,63), num_attention_heads 24, num_key_value_heads 4, head_dim 256, intermediate_size 17408, hidden_act silu, vocab_size 248320, tie_word_embeddings false, max_position_embeddings 262144. Linear block: linear_num_key_heads 16, linear_num_value_heads 48, key and value head dim 128, linear_conv_kernel_dim 4. RoPE: rope_type default, mrope_interleaved true, mrope_section [11,11,10], partial_rotary_factor 0.25, rope_theta 10000000. No num_experts or num_experts_per_tok keys: dense. Top-level key veriloop_conditional_memory is present and is the only config difference from Qwen/Qwen3.8-27B. Vision tower is separate: vision_config depth 27, hidden 1152.