{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "$id": "https://www.schemastore.org/litertlm_manifest.json",
  "$comment": "Mirror of manifest/litertlm_manifest.schema.json in the hf-to-litertlm repository, for the manifest_schema 0.1 line. Keep the two in sync.",
  "title": "litertlm deployment manifest",
  "description": "litertlm_manifest.json sits at the root of a Hugging Face model repo that ships .litertlm bundles for the LiteRT-LM runtime. It describes every bundle in the repo: the backends each file is verified on, which file a device should pick, what it requires, and how fast it measured.\nhttps://github.com/john-rocky/hf-to-litertlm/blob/main/manifest/SCHEMA.md",
  "type": "object",
  "additionalProperties": true,
  "required": ["manifest_schema", "repo", "generated", "model", "variants"],
  "properties": {
    "manifest_schema": {
      "description": "Version of the manifest format. The 0.1 line is the compatibility line: a 0.1.x release only adds optional fields, so a 0.2 manifest is rejected rather than half-parsed.",
      "type": "string",
      "pattern": "^0\\.1\\.[0-9]+$",
      "examples": ["0.1.0", "0.1.1", "0.1.2"]
    },
    "repo": {
      "description": "Hugging Face repo id (owner/name) the manifest lives in. Together with a variant's file name this is the download address.",
      "type": "string",
      "pattern": "^[^/]+/[^/]+$"
    },
    "generated": {
      "description": "Date the manifest was generated (YYYY-MM-DD).",
      "type": "string",
      "format": "date"
    },
    "generator": {
      "description": "Tool that produced the manifest.",
      "type": "string"
    },
    "model": {
      "$ref": "#/definitions/model"
    },
    "variants": {
      "description": "One entry per .litertlm file in the repo.",
      "type": "array",
      "minItems": 1,
      "items": {
        "$ref": "#/definitions/variant"
      }
    }
  },
  "definitions": {
    "backend": {
      "description": "LiteRT-LM execution backend.",
      "type": "string",
      "enum": ["cpu", "gpu", "npu"]
    },
    "platform": {
      "description": "Target platform of a recommendation.",
      "type": "string",
      "enum": ["android", "ios", "macos", "windows", "linux"]
    },
    "speed": {
      "description": "A measured value: a number, or a \"lo-hi\" range string when the runs spread (CPU rows on phones spread with thermal throttling).",
      "anyOf": [
        {
          "type": "number"
        },
        {
          "type": "string"
        }
      ]
    },
    "channelMarkers": {
      "description": "Exact marker strings of a channel, whitespace included.",
      "type": "object",
      "properties": {
        "start": {
          "description": "Opening marker, e.g. \"<think>\".",
          "type": "string"
        },
        "end": {
          "description": "Closing marker, e.g. \"</think>\".",
          "type": "string"
        }
      }
    },
    "declaredChannel": {
      "description": "One entry of the bundle's declared channel set (manifest 0.1.1 and later).",
      "type": "object",
      "required": ["name", "start", "end"],
      "properties": {
        "name": {
          "description": "Channel name as declared in the bundle header.",
          "type": "string"
        },
        "start": {
          "description": "Opening marker, whitespace included.",
          "type": "string"
        },
        "end": {
          "description": "Closing marker, whitespace included.",
          "type": "string"
        },
        "is_reasoning": {
          "description": "True when the bundle marks this channel as its reasoning channel.",
          "type": "boolean"
        }
      }
    },
    "capabilities": {
      "description": "Capability flags derived from the bundle header, so they are readable before download.",
      "type": "object",
      "properties": {
        "vision": {
          "description": "Derived: the bundle declares image input.",
          "type": "boolean"
        },
        "audio": {
          "description": "Derived: the bundle declares audio input.",
          "type": "boolean"
        },
        "thinking": {
          "description": "Derived: whether the bundle declares a thinking channel, and its markers. Mirrors the first declared channel.",
          "type": "object",
          "properties": {
            "declared": {
              "description": "True when the bundle declares a channel.",
              "type": "boolean"
            },
            "channel": {
              "$ref": "#/definitions/channelMarkers"
            }
          }
        },
        "channels": {
          "description": "Derived (manifest 0.1.1 and later): the bundle's full declared channel set, whether thinking, tool-call, or anything else the model declares.",
          "type": "array",
          "items": {
            "$ref": "#/definitions/declaredChannel"
          }
        }
      }
    },
    "sessionDefaults": {
      "description": "Curated session knobs a wrapper should set that the engine cannot infer. An open object: readers take keys by name and ignore what they do not consume.",
      "type": "object",
      "properties": {
        "max_output_tokens_min": {
          "description": "A floor on the output-token budget (e.g. 2048 for reasoning models), never a cap.",
          "type": "integer",
          "minimum": 1
        },
        "notes": {
          "description": "Curated guidance worth surfacing to the app developer.",
          "type": "string"
        },
        "temperature": {
          "description": "Sampler hint.",
          "type": "number"
        },
        "top_k": {
          "description": "Sampler hint.",
          "type": "integer",
          "minimum": 1
        },
        "top_p": {
          "description": "Sampler hint.",
          "type": "number"
        }
      }
    },
    "model": {
      "description": "Model-level facts. Fields marked Derived are read out of the bundle header by the generator; the rest are curated.",
      "type": "object",
      "additionalProperties": true,
      "required": ["display_name"],
      "properties": {
        "display_name": {
          "description": "Human-readable model name.",
          "type": "string"
        },
        "base_model": {
          "description": "Hugging Face id of the source model.",
          "type": "string"
        },
        "architecture": {
          "description": "Free-text architecture label, e.g. \"lfm2-hybrid\" or \"qwen3-dense\".",
          "type": "string"
        },
        "parameters_b": {
          "description": "Parameter count in billions.",
          "type": "number"
        },
        "license": {
          "description": "SPDX identifier, or a pointer to the license file.",
          "type": "string"
        },
        "context_length": {
          "description": "Derived: the bundle's max_num_tokens.",
          "type": "integer",
          "minimum": 1
        },
        "capabilities": {
          "$ref": "#/definitions/capabilities"
        },
        "session_defaults": {
          "$ref": "#/definitions/sessionDefaults"
        }
      }
    },
    "recommendation": {
      "description": "The fastest verified backend for a platform, optionally narrowed to a device class.",
      "type": "object",
      "required": ["platform", "backend"],
      "properties": {
        "platform": {
          "$ref": "#/definitions/platform"
        },
        "device_class": {
          "description": "Free-text device class the recommendation applies to, e.g. \"midrange\" or \"flagship\".",
          "type": "string"
        },
        "backend": {
          "$ref": "#/definitions/backend",
          "description": "Backend to use on this platform. Must be one of the variant's verified backends; resolvers ignore a recommendation naming an unlisted backend."
        },
        "reason": {
          "description": "Why this backend wins on this platform, in words that carry the evidence.",
          "type": "string"
        }
      }
    },
    "requirements": {
      "description": "What the variant needs beyond the runtime version.",
      "type": "object",
      "properties": {
        "peak_ram_mb": {
          "description": "Peak resident memory observed while generating, in megabytes.",
          "type": "integer"
        },
        "platform_notes": {
          "description": "Platform caveats to surface to the app developer.",
          "type": "array",
          "items": {
            "type": "string"
          }
        }
      }
    },
    "measuredRow": {
      "description": "One measured performance row. Every row states its conditions and its provenance; rows come from generation-verified backends only.",
      "type": "object",
      "required": ["device", "backend", "runtime", "date", "source"],
      "properties": {
        "device": {
          "description": "Device the row was measured on, e.g. \"Pixel 8a (Tensor G3)\".",
          "type": "string"
        },
        "os": {
          "description": "Operating system of the device.",
          "type": "string"
        },
        "backend": {
          "$ref": "#/definitions/backend",
          "description": "Backend the row was measured on."
        },
        "runtime": {
          "description": "LiteRT-LM release and build the row was measured with.",
          "type": "string"
        },
        "prompt_tokens": {
          "description": "Prompt length in tokens.",
          "type": "integer"
        },
        "decode_tokens": {
          "description": "Number of generated tokens.",
          "type": "integer"
        },
        "prefill_tps": {
          "$ref": "#/definitions/speed",
          "description": "Prefill throughput in tokens per second."
        },
        "decode_tps": {
          "$ref": "#/definitions/speed",
          "description": "Decode throughput in tokens per second."
        },
        "ttft_s": {
          "$ref": "#/definitions/speed",
          "description": "Time to first token in seconds."
        },
        "max_num_tokens": {
          "description": "Context budget (max_num_tokens) the run used.",
          "type": "integer"
        },
        "cache": {
          "description": "Whether a compiled-model cache was in use. \"no\" means a cold compile; caches mask load regressions and inflate disk cost.",
          "type": "string"
        },
        "runs": {
          "description": "Number of runs the row summarizes.",
          "type": "integer"
        },
        "load_s": {
          "description": "Engine load time in seconds under the row's cache condition (manifest 0.1.2 and later).",
          "type": "number",
          "minimum": 0
        },
        "peak_memory_mb": {
          "description": "Peak resident memory during the run, in megabytes (manifest 0.1.2 and later).",
          "type": "number",
          "minimum": 0
        },
        "date": {
          "description": "Date of the measurement (YYYY-MM-DD).",
          "type": "string",
          "format": "date"
        },
        "source": {
          "description": "Provenance of the numbers: which log or model-card table they come from.",
          "type": "string"
        },
        "evidence": {
          "description": "Pointer to the primary log. Stripped from published manifests and kept in the converter's own records.",
          "type": "string"
        }
      }
    },
    "section": {
      "description": "Derived: one row of the bundle's section table.",
      "type": "object",
      "properties": {
        "type": {
          "description": "Section data type as named in the bundle header, e.g. \"LlmMetadataProto\", \"HF_Tokenizer_Zlib\" or \"TFLiteModel\".",
          "type": "string"
        },
        "size_bytes": {
          "description": "Section size in bytes.",
          "type": "integer"
        },
        "model_type": {
          "description": "The section's model_type item when present, e.g. \"tf_lite_prefill_decode\" or \"tf_lite_embedder\".",
          "type": "string"
        },
        "backend_constraint": {
          "description": "The section's backend_constraint item when present: a comma-separated list of backends the engine will load it on.",
          "type": "string"
        }
      }
    },
    "variant": {
      "description": "One .litertlm file. Fields marked Derived are read from Hub metadata or the bundle header; the rest are curated and carry evidence.",
      "type": "object",
      "additionalProperties": true,
      "required": ["file", "quantization", "backends"],
      "properties": {
        "file": {
          "description": "File name inside the repo.",
          "type": "string",
          "pattern": "\\.litertlm$"
        },
        "sha256": {
          "description": "Derived: SHA-256 of the file from Hub LFS metadata. Verify after download.",
          "type": "string",
          "pattern": "^[0-9a-f]{64}$"
        },
        "size_bytes": {
          "description": "Derived: file size in bytes.",
          "type": "integer",
          "minimum": 1
        },
        "quantization": {
          "description": "Quantization recipe, stated honestly, e.g. \"int4 block-32 linears, fp32 activations\".",
          "type": "string"
        },
        "backends": {
          "description": "Backends this file is verified to generate on, not merely load. A resolver never returns a backend absent from this list.",
          "type": "array",
          "minItems": 1,
          "items": {
            "$ref": "#/definitions/backend"
          }
        },
        "default_backend": {
          "$ref": "#/definitions/backend",
          "description": "Backend to pick with no device knowledge."
        },
        "min_runtime_version": {
          "description": "Earliest LiteRT-LM release the file is verified on, e.g. \"0.15.0\".",
          "type": "string"
        },
        "recommended": {
          "description": "The fastest verified choice per platform and device class.",
          "type": "array",
          "items": {
            "$ref": "#/definitions/recommendation"
          }
        },
        "requirements": {
          "$ref": "#/definitions/requirements"
        },
        "measured": {
          "description": "Measured performance rows with conditions and provenance.",
          "type": "array",
          "items": {
            "$ref": "#/definitions/measuredRow"
          }
        },
        "known_issues": {
          "description": "Short, factual known issues, with upstream links where they exist.",
          "type": "array",
          "items": {
            "type": "string"
          }
        },
        "sections": {
          "description": "Derived: the bundle's section table.",
          "type": "array",
          "items": {
            "$ref": "#/definitions/section"
          }
        }
      }
    }
  }
}