{
  "version": 1,
  "hardware": {
    "hosts": 2,
    "gpus_per_host": 8,
    "hbm_bytes_per_gpu": 85899345920,
    "scale_up_fabric": "measure",
    "scale_out_fabric": "measure",
    "gpu_nic_affinity": []
  },
  "model": {
    "shardable_weight_bytes": 1600000000000,
    "replicated_bytes_per_rank": 8000000000,
    "kv_bytes_per_replica": 128000000000,
    "runtime_reserve_fraction": 0.15
  },
  "plan": {
    "tensor_parallel": 8,
    "pipeline_parallel": 2,
    "data_parallel": 1,
    "kv_shard_factor": 4,
    "phase": "unified"
  },
  "rank_to_device": [
    {
      "rank": 0,
      "host": 0,
      "local_gpu": 0,
      "pipeline_stage": 0,
      "tensor_group": 0,
      "expert_group": null,
      "nic": "measure"
    }
  ],
  "stage_ledger": [
    {
      "pipeline_stage": 0,
      "layer_start": null,
      "layer_end": null,
      "owned_weight_bytes": null,
      "largest_unsharded_tensor_bytes": null,
      "activation_peak_bytes": null
    }
  ],
  "rank_ledger": [
    {
      "rank": 0,
      "stage_owned_weight_bytes": null,
      "replicated_tensor_bytes": null,
      "kv_or_recurrent_state_bytes": null,
      "activation_peak_bytes": null,
      "collective_workspace_bytes": null,
      "allocator_reserve_bytes": null,
      "modeled_peak_bytes": null
    }
  ],
  "collectives": [
    {
      "kind": "all_reduce",
      "group": "tp",
      "frequency": "per transformer layer",
      "bytes": "measure"
    },
    {
      "kind": "send_receive",
      "group": "pipeline_boundary",
      "frequency": "per microbatch",
      "bytes": "activation bytes"
    }
  ],
  "objectives": {
    "ttft_ms_p95": null,
    "tpot_ms_p95": null,
    "request_goodput": null
  },
  "required_measurements": [
    "per-rank peak HBM",
    "NCCL collective matrix by message size",
    "GPU-to-NIC affinity",
    "KV block rounding",
    "prompt/output/concurrency distributions"
  ]
}