{
  "_serialized_key": "HloNeuronConfig",
  "all_reduce_dtype": null,
  "allow_flash_attention": true,
  "attention_layout": "HSB",
  "attn_output_transposed": false,
  "auto_cast_type": "bf16",
  "batch_size": 4,
  "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
  "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
  "collectives_layout": "HSB",
  "continuous_batching": true,
  "fuse_qkv": true,
  "group_query_attention": "shard-over-heads",
  "log_softmax_scores": false,
  "neuronxcc_version": "2.17.194.0+d312836f",
  "optimum_neuron_version": "0.2.0",
  "output_all_logits": false,
  "sequence_length": 4096,
  "tp_degree": 2
}