config.train.json

{
  "attn_impl": "sdpa",
  "attn_output_gate": true,
  "bqa_local_precision": "fp8",
  "bqa_nvfp4_block": 16,
  "bqa_remote_k_precision": "fp8",
  "bqa_remote_topk": -1,
  "bqa_remote_v_precision": "nvfp4",
  "bqa_rotate": true,
  "bqa_window": 512,
  "d_model": 2048,
  "ffn_mult": 2.6666666666666665,
  "ffn_multiple_of": 256,
  "gdn_head_dim": 128,
  "gdn_num_heads": 16,
  "gdn_num_v_heads": null,
  "head_dim": 128,
  "hidden_size": 2048,
  "intermediate_size": 6144,
  "kv_lora_rank": 384,
  "layer_mixers": [
    "gated_deltanet",
    "gated_deltanet",
    "gated_deltanet",
    "mla_fp8"
  ],
  "mamba2_chunk_size": 256,
  "mamba2_d_state": 128,
  "mamba2_expand": 2,
  "mamba2_headdim": null,
  "mamba2_ngroups": 1,
  "max_position_embeddings": 8192,
  "max_seq_len": 8192,
  "mixer": "gqa",
  "model_type": "bqalm",
  "n_heads": 16,
  "n_kv_heads": 16,
  "n_layers": 24,
  "nope": false,
  "norm_eps": 1e-06,
  "num_attention_heads": 16,
  "num_hidden_layers": 24,
  "partial_rotary_factor": 0.5,
  "qk_norm": true,
  "rms_norm_in_fp32": true,
  "rope_parameters": {
    "partial_rotary_factor": 0.5,
    "rope_theta": 10000000,
    "rope_type": "default"
  },
  "rope_theta": 10000000,
  "sb_impl": "triton",
  "tie_embeddings": true,
  "tie_word_embeddings": true,
  "transformers_version": "5.9.0",
  "vocab_size": 32000,
  "z_loss_weight": 0.0
}