config.json

{
  "architectures": [
    "Xerxes2ForCausalLM"
  ],
  "attention_axis_name": "sp",
  "attn_mechanism": "vanilla",
  "axis_dims": [
    1,
    -1,
    1,
    1
  ],
  "axis_names": [
    "dp",
    "fsdp",
    "tp",
    "sp"
  ],
  "backend": null,
  "bits": null,
  "blocksize_b": 1,
  "blocksize_k": 128,
  "blocksize_q": 128,
  "bos_token_id": 2,
  "cache_implementation": "hybrid",
  "compute_blocksize": 64,
  "easy_method": "train",
  "eos_token_id": 1,
  "fcm_max_ratio": 0.0,
  "fcm_min_ratio": 0.0,
  "flash_attention_backward_pass_impl": "triton",
  "gradient_checkpointing": "",
  "hardware_abstraction": false,
  "head_dim": 64,
  "hidden_size": 2048,
  "initializer_range": 0.02,
  "intermediate_size": 8192,
  "kv_cache_quantization_blocksize": 64,
  "kv_cache_quantization_method": "None",
  "kv_cache_sharding_sequence_axis_name": "sp",
  "max_position_embeddings": 2048,
  "model_type": "xerxes2",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "pad_token_id": 0,
  "pallas_k_block_size": null,
  "pallas_m_block_size": null,
  "pallas_n_block_size": null,
  "partition_axis": [
    [
      "fsdp",
      "dp"
    ],
    "sp",
    "sp",
    "tp",
    "sp",
    "tp",
    null,
    null,
    null,
    null,
    "tp",
    "sp",
    null
  ],
  "platform": null,
  "pretraining_tp": 1,
  "quantization_blocksize": 64,
  "quantization_method": "None",
  "quantization_pattern": ".*",
  "rms_norm_eps": 1e-06,
  "rope_theta": 10000.0,
  "scan_attention_layers": false,
  "scan_layers": false,
  "scan_mlp_chunk_size": 1024,
  "scan_ring_attention": true,
  "shard_attention_computation": true,
  "tie_word_embeddings": false,
  "transformers_version": "4.48.0",
  "use_cache": true,
  "use_scan_mlp": false,
  "use_sharded_kv_caching": false,
  "use_sharding_constraint": false,
  "vocab_size": 152064
}