| { | |
| "_serialized_key": "NxDNeuronConfig", | |
| "async_mode": false, | |
| "attn_kernel_enabled": false, | |
| "batch_size": 1, | |
| "capacity_factor": null, | |
| "cc_pipeline_tiling_factor": 2, | |
| "checkpoint_id": "microsoft/Phi-3-mini-128k-instruct", | |
| "checkpoint_revision": "072cb7562cb8c4adf682a8e186aaafa49469eb5d", | |
| "continuous_batching": false, | |
| "enable_bucketing": false, | |
| "ep_degree": 1, | |
| "flash_decoding_enabled": false, | |
| "fused_qkv": true, | |
| "glu_mlp": true, | |
| "is_chunked_prefill": false, | |
| "local_ranks_size": 12, | |
| "logical_nc_config": 1, | |
| "max_batch_size": 1, | |
| "max_context_length": 128, | |
| "max_topk": 256, | |
| "mlp_kernel_enabled": false, | |
| "mlp_kernel_fuse_residual_add": false, | |
| "n_active_tokens": 128, | |
| "neuronxcc_version": "2.19.8089.0+8ab9f450", | |
| "num_cores_per_group": 1, | |
| "on_device_sampling": true, | |
| "optimum_neuron_version": "0.3.0", | |
| "output_logits": false, | |
| "padding_side": "right", | |
| "pp_degree": 1, | |
| "qk_layernorm": false, | |
| "qkv_kernel_enabled": false, | |
| "rpl_reduce_dtype": "bfloat16", | |
| "sequence_length": 128, | |
| "sequence_parallel_enabled": false, | |
| "speculation_length": 0, | |
| "start_rank_id": 0, | |
| "target": null, | |
| "torch_dtype": "bfloat16", | |
| "tp_degree": 12, | |
| "vocab_parallel": false | |
| } | |