{ "architectures": [ "IQuestQ1ForCausalLM" ], "attention_bias": false, "attention_dropout": 0.0, "attn_out_scale": 1.0, "dense_intermediate_size": 12288, "dtype": "bfloat16", "embed_scale": 1.0, "enable_sink_attention": true, "eos_token_id": 0, "ffn_out_scale": 0.53881590608, "first_layer_attn_out_scale": 1.0, "first_layer_ffn_out_scale": 1.0, "first_layers_types": [ "full_attention" ], "fuse_sink_attention": true, "head_dim": 128, "hidden_act": "silu", "hidden_size": 3072, "hybrid_layers_types_block": [ "full_attention", "sliding_attention", "sliding_attention", "sliding_attention" ], "intermediate_size": 1536, "last_layers_types": [ "full_attention", "full_attention", "full_attention" ], "logit_scale": 1.0, "max_position_embeddings": 524288, "mlp_only_layers": [ 0 ], "model_type": "iquest_q1", "moe_router_dtype": "fp32", "enable_lm_head_fp32": true, "no_rope_layers": [], "num_attention_heads": 48, "num_experts": 256, "num_experts_per_tok": 8, "num_hidden_layers": 88, "num_hybrid_layers_block": 21, "num_key_value_heads": 8, "num_mtp_layers": 0, "rms_norm_eps": 1e-06, "rope_theta": 1000000, "rotary_dim": 32, "shared_kv_num_layers": 0, "shared_kv_source_begin": 0, "shared_kv_target_begin": 0, "sliding_window": 4096, "softmax_scale": null, "swa_rope_theta": 10000.0, "tie_word_embeddings": false, "transformers_version": "4.57.1", "use_cache": true, "use_hybrid_layers": true, "use_over_encoding": false, "use_sliding_window": true, "vocab_size": 160000 }