{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://leadgreen.ai/api/v1/schema/measurement.schema.json",
  "title": "Lead Green contributed measurement, protocol v0.2",
  "description": "One measured serving configuration, submitted under the measurement protocol at leadgreen.ai/contribute/#protocol. Field names match the configurations in api/v1/models/<slug>.json where they exist. GPU energy only; host, facility and water figures are separate optional fields and are never folded into wh_per_million_output_tokens.",
  "type": "object",
  "additionalProperties": false,
  "required": [
    "protocol_version",
    "example",
    "model",
    "model_id",
    "gpu",
    "num_gpus",
    "tensor_parallel",
    "data_parallel",
    "expert_parallel",
    "precision",
    "max_num_seqs",
    "avg_running_batch",
    "joules_per_output_token",
    "wh_per_million_output_tokens",
    "median_itl_ms",
    "p90_itl_ms",
    "p99_itl_ms",
    "tokens_per_second",
    "gpu_power_watts_total",
    "avg_output_tokens",
    "avg_prompt_tokens",
    "window_seconds",
    "requests_in_window",
    "runs",
    "mig_profile",
    "slice_count",
    "slices_loaded",
    "engine",
    "engine_version",
    "driver_version",
    "cuda_version",
    "energy_source",
    "prompt_set",
    "measured_on",
    "contributor",
    "affiliation",
    "vendor_affiliation",
    "logs",
    "notes"
  ],
  "properties": {
    "protocol_version": {
      "type": "string",
      "enum": [
        "0.2"
      ]
    },
    "example": {
      "type": "boolean",
      "description": "true only for the example file; example rows are never published"
    },
    "model": {
      "type": "string",
      "description": "display name, as on the index where the model is on it"
    },
    "model_id": {
      "type": "string",
      "description": "Hugging Face id"
    },
    "gpu": {
      "type": "string",
      "description": "GPU model, e.g. H100 80GB HBM3 SXM"
    },
    "num_gpus": {
      "type": "integer",
      "minimum": 1
    },
    "tensor_parallel": {
      "type": "integer",
      "minimum": 1
    },
    "data_parallel": {
      "type": "integer",
      "minimum": 1
    },
    "expert_parallel": {
      "type": "integer",
      "minimum": 1
    },
    "precision": {
      "type": "string",
      "enum": [
        "BF16",
        "FP16",
        "FP8",
        "INT8",
        "INT4",
        "MXFP4",
        "other"
      ],
      "description": "weight precision; required, because it decides whether a sliced row is like-for-like"
    },
    "max_num_seqs": {
      "type": "integer",
      "minimum": 1,
      "description": "the configured cap"
    },
    "avg_running_batch": {
      "type": "number",
      "minimum": 0,
      "description": "mean of vllm:num_requests_running over the window, scraped every second; queued requests are not counted"
    },
    "joules_per_output_token": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "wh_per_million_output_tokens": {
      "type": "number",
      "exclusiveMinimum": 0,
      "description": "joules_per_output_token x 1,000,000 / 3,600"
    },
    "median_itl_ms": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "p90_itl_ms": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "p99_itl_ms": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "p95_itl_ms": {
      "type": [
        "number",
        "null"
      ]
    },
    "ttft_median_ms": {
      "type": [
        "number",
        "null"
      ],
      "description": "median time to first token, if recorded"
    },
    "tokens_per_second": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "gpu_power_watts_total": {
      "type": "number",
      "exclusiveMinimum": 0,
      "description": "average board power summed across every card, or every card holding a loaded slice"
    },
    "avg_output_tokens": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "avg_prompt_tokens": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "window_seconds": {
      "type": "number",
      "exclusiveMinimum": 0
    },
    "requests_in_window": {
      "type": "integer",
      "minimum": 1
    },
    "runs": {
      "type": "array",
      "minItems": 2,
      "items": {
        "type": "object",
        "additionalProperties": false,
        "required": [
          "joules_per_output_token"
        ],
        "properties": {
          "joules_per_output_token": {
            "type": "number",
            "exclusiveMinimum": 0
          }
        }
      },
      "description": "every run's joules per output token; the row carries their mean"
    },
    "mig_profile": {
      "type": [
        "string",
        "null"
      ],
      "description": "e.g. 3g.40gb, or null for a whole card"
    },
    "slice_count": {
      "type": [
        "integer",
        "null"
      ],
      "minimum": 1
    },
    "slices_loaded": {
      "type": [
        "integer",
        "null"
      ],
      "minimum": 1,
      "description": "must equal slice_count: every slice serves the workload at once"
    },
    "engine": {
      "type": "string"
    },
    "engine_version": {
      "type": "string"
    },
    "driver_version": {
      "type": "string"
    },
    "cuda_version": {
      "type": "string"
    },
    "energy_source": {
      "type": "string",
      "enum": [
        "dcgm_total_energy",
        "nvidia_smi_power_integrated",
        "other"
      ],
      "description": "how GPU energy was read"
    },
    "prompt_set": {
      "type": "string",
      "description": "URL or file name of the published prompt set"
    },
    "measured_on": {
      "type": "string",
      "format": "date"
    },
    "contributor": {
      "type": "string"
    },
    "affiliation": {
      "type": [
        "string",
        "null"
      ]
    },
    "vendor_affiliation": {
      "type": [
        "string",
        "null"
      ],
      "description": "set when a hardware, cloud or model vendor made or paid for the measurement"
    },
    "host_power_watts": {
      "type": [
        "number",
        "null"
      ],
      "description": "optional, separate; never folded in"
    },
    "site_pue": {
      "type": [
        "number",
        "null"
      ],
      "description": "optional, separate; ISO/IEC 30134-2"
    },
    "site_wue_l_per_kwh": {
      "type": [
        "number",
        "null"
      ],
      "description": "optional site water, ISO/IEC 30134-9; never folded into Wh/M"
    },
    "logs": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "power log and engine metrics log, request bodies stripped"
    },
    "notes": {
      "type": "string"
    }
  }
}