| { |
| "allow_neg_eigval": false, |
| "architectures": [ |
| "KDAForCausalLM" |
| ], |
| "attn": null, |
| "attn_mode": "chunk", |
| "attnres_block_size": null, |
| "bos_token_id": 1, |
| "conv_size": 4, |
| "disable_recompute": true, |
| "dtype": "float32", |
| "eos_token_id": 2, |
| "expand_v": 2, |
| "fuse_cross_entropy": true, |
| "fuse_linear_cross_entropy": false, |
| "fuse_norm": true, |
| "fuse_swiglu": true, |
| "head_dim": 64, |
| "head_init_design": null, |
| "hidden_act": "swish", |
| "hidden_ratio": 4, |
| "hidden_size": 512, |
| "initialization_manifest": { |
| "A_log_rule": "per-head A_log copied exactly from GDN v3_layerdt3 targets", |
| "design": "v4 layerdt3: GDN v3_layerdt3 (PoST tau by head + low-high-low layerwise sensitivity) targets", |
| "design_config": "/root/linear-attention/flame/configs/gated_deltanet_100M_v3_layerdt3_a.json", |
| "dt_bias_rule": "per-head mean equals GDN v3_layerdt3 target dt_bias; 64 channels use deterministic centered Gaussian noise, std=0.03", |
| "experiment": "v4_layerdt3_b", |
| "non_ad_tensors": "same seeded base KDA initialization (seed 42) as v4_random_a / v4_post_a" |
| }, |
| "initializer_range": 0.02, |
| "intermediate_size": null, |
| "lower_bound": null, |
| "max_position_embeddings": 4096, |
| "model_type": "kda", |
| "norm_eps": 1e-06, |
| "num_heads": 6, |
| "num_hidden_layers": 18, |
| "num_v_heads": 6, |
| "pad_token_id": 0, |
| "safe_gate": false, |
| "tie_word_embeddings": false, |
| "transformers_version": "5.9.0", |
| "use_cache": true, |
| "use_l2warp": false, |
| "use_short_conv": true, |
| "vocab_size": 32000 |
| } |
|
|
