{
  "_serialized_key": "NxDNeuronConfig",
  "batch_size": 1,
  "capacity_factor": null,
  "checkpoint_id": "ibm-granite/granite-3.3-8b-instruct",
  "checkpoint_revision": "51dd4bc2ade4059a6bd87649d68aa11e4fb2529b",
  "continuous_batching": false,
  "ep_degree": 1,
  "fused_qkv": true,
  "glu_mlp": true,
  "local_ranks_size": 8,
  "max_batch_size": 1,
  "max_context_length": 16384,
  "max_topk": 256,
  "n_active_tokens": 16384,
  "neuronxcc_version": "2.21.18209.0+043b1bf7",
  "on_device_sampling": true,
  "optimum_neuron_version": "0.4.2",
  "output_logits": false,
  "pp_degree": 1,
  "sequence_length": 16384,
  "speculation_length": 0,
  "start_rank_id": 0,
  "target": "trn1",
  "torch_dtype": "bfloat16",
  "tp_degree": 8
}