| { |
| "_attn_implementation_autoset": true, |
| "architectures": [ |
| "Qwen3ForCausalLM" |
| ], |
| "attention_bias": false, |
| "attention_dropout": 0.0, |
| "attn_impl": "rg_attn_flex", |
| "base_loss": "fwkl_ntp", |
| "base_model": "Qwen/Qwen3-1.7B", |
| "bos_token_id": 151643, |
| "buffer_size": 128, |
| "compress_memory": true, |
| "compress_strategy": "alpha", |
| "eos_token_id": 151645, |
| "head_dim": 128, |
| "hidden_act": "silu", |
| "hidden_size": 2048, |
| "initializer_range": 0.02, |
| "intermediate_size": 6144, |
| "logit_block_size": 8192, |
| "max_position_embeddings": 40960, |
| "max_seq_len": 32768, |
| "max_window_layers": 28, |
| "memory_size": 512.0, |
| "model_type": "qwen3", |
| "num_attention_heads": 16, |
| "num_hidden_layers": 28, |
| "num_key_value_heads": 8, |
| "rms_norm_eps": 1e-06, |
| "rope_scaling": null, |
| "rope_theta": 1000000, |
| "skip_layers": 0, |
| "sliding_window": null, |
| "tie_word_embeddings": true, |
| "torch_dtype": "bfloat16", |
| "trainable_params": "self_attn.retention_gate", |
| "transformers_version": "4.51.0", |
| "use_cache": true, |
| "use_sliding_window": false, |
| "vocab_size": 151936, |
| "retention_gate": "rg", |
| "retention_gate_bias_init": 8.0, |
| "retention_gate_intermediate_size": 512, |
| "retention_weight": 1.0, |
| "rg_dropout": 0.0 |
| } |
