agent-hitch 0.2.9 → 0.2.10

package/docs/schemas/inference-execution-evidence.schema.json

@@ -0,0 +1,487 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://agent-hitch.local/schemas/inference-execution-evidence.schema.json",
+  "title": "Hitch inference execution evidence v1/v2",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema_version",
+    "run_id",
+    "service",
+    "doctor",
+    "prepared_at"
+  ],
+  "properties": {
+    "schema_version": {
+      "enum": [
+        "1",
+        "2"
+      ]
+    },
+    "run_id": {
+      "type": "string",
+      "pattern": "^run_[a-f0-9]{32}$"
+    },
+    "eval_id": {
+      "type": "string",
+      "pattern": "^eval_[a-f0-9]{32}$"
+    },
+    "rerun_id": {
+      "type": "string",
+      "pattern": "^rerun_[a-f0-9]{32}$"
+    },
+    "service": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "service_id",
+        "epoch",
+        "isolation_key"
+      ],
+      "properties": {
+        "service_id": {
+          "type": "string",
+          "pattern": "^inference_[a-f0-9]{32}$"
+        },
+        "epoch": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "isolation_key": {
+          "$ref": "#/$defs/sha256"
+        }
+      }
+    },
+    "doctor": {
+      "oneOf": [
+        {
+          "type": "null"
+        },
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "schema_version",
+            "backend",
+            "ready",
+            "checks"
+          ],
+          "properties": {
+            "schema_version": {
+              "const": "1"
+            },
+            "backend": {
+              "enum": [
+                "cpu",
+                "cuda",
+                "metal"
+              ]
+            },
+            "ready": {
+              "type": "boolean"
+            },
+            "checks": {
+              "type": "object",
+              "additionalProperties": {
+                "$ref": "#/$defs/check"
+              }
+            },
+            "gpu": {
+              "type": "object",
+              "additionalProperties": false,
+              "required": [
+                "uuid",
+                "name",
+                "memory_mib",
+                "driver_version"
+              ],
+              "properties": {
+                "uuid": {
+                  "type": "string",
+                  "minLength": 1
+                },
+                "name": {
+                  "type": "string",
+                  "minLength": 1
+                },
+                "memory_mib": {
+                  "type": "number",
+                  "exclusiveMinimum": 0
+                },
+                "driver_version": {
+                  "type": "string",
+                  "minLength": 1
+                },
+                "free_memory_mib": {
+                  "type": "number",
+                  "minimum": 0
+                },
+                "compute_capability": {
+                  "type": "string",
+                  "minLength": 1
+                }
+              }
+            }
+          }
+        }
+      ]
+    },
+    "prepared_at": {
+      "type": "string",
+      "format": "date-time"
+    },
+    "observation": {
+      "oneOf": [
+        {
+          "type": "null"
+        },
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "observed_at",
+            "version",
+            "dtype",
+            "kv_cache_dtype",
+            "attention_backend",
+            "sampling_backend",
+            "container_image_id",
+            "schema_version",
+            "device",
+            "gpu_uuid",
+            "context_length",
+            "max_total_num_tokens",
+            "max_running_requests",
+            "probe"
+          ],
+          "properties": {
+            "observed_at": {
+              "type": "string",
+              "minLength": 1
+            },
+            "version": {
+              "type": "string",
+              "minLength": 1
+            },
+            "dtype": {
+              "type": "string",
+              "minLength": 1
+            },
+            "kv_cache_dtype": {
+              "type": "string",
+              "minLength": 1
+            },
+            "attention_backend": {
+              "type": "string",
+              "minLength": 1
+            },
+            "sampling_backend": {
+              "type": "string",
+              "minLength": 1
+            },
+            "container_image_id": {
+              "type": "string",
+              "minLength": 1
+            },
+            "schema_version": {
+              "const": "1"
+            },
+            "device": {
+              "enum": [
+                "cpu",
+                "cuda"
+              ]
+            },
+            "gpu_uuid": {
+              "type": [
+                "string",
+                "null"
+              ]
+            },
+            "context_length": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "max_total_num_tokens": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "max_running_requests": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "probe": {
+              "type": "object",
+              "additionalProperties": false,
+              "required": [
+                "api",
+                "max_output_tokens",
+                "streaming"
+              ],
+              "properties": {
+                "api": {
+                  "enum": [
+                    "responses",
+                    "chat-completions"
+                  ]
+                },
+                "max_output_tokens": {
+                  "type": "integer",
+                  "minimum": 3
+                },
+                "streaming": {
+                  "type": "boolean"
+                }
+              }
+            }
+          }
+        },
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "observed_at",
+            "version",
+            "dtype",
+            "kv_cache_dtype",
+            "attention_backend",
+            "sampling_backend",
+            "schema_version",
+            "device",
+            "gpu_uuid",
+            "context_length",
+            "max_total_num_tokens",
+            "max_running_requests",
+            "probe",
+            "runtime"
+          ],
+          "properties": {
+            "observed_at": {
+              "type": "string",
+              "minLength": 1
+            },
+            "version": {
+              "type": "string",
+              "minLength": 1
+            },
+            "dtype": {
+              "type": "string",
+              "minLength": 1
+            },
+            "kv_cache_dtype": {
+              "type": "string",
+              "minLength": 1
+            },
+            "attention_backend": {
+              "type": "string",
+              "minLength": 1
+            },
+            "sampling_backend": {
+              "type": "string",
+              "minLength": 1
+            },
+            "schema_version": {
+              "const": "2"
+            },
+            "device": {
+              "enum": [
+                "cpu",
+                "cuda"
+              ]
+            },
+            "gpu_uuid": {
+              "type": [
+                "string",
+                "null"
+              ]
+            },
+            "context_length": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "max_total_num_tokens": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "max_running_requests": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "probe": {
+              "type": "object",
+              "additionalProperties": false,
+              "required": [
+                "api",
+                "max_output_tokens",
+                "streaming"
+              ],
+              "properties": {
+                "api": {
+                  "enum": [
+                    "responses",
+                    "chat-completions"
+                  ]
+                },
+                "max_output_tokens": {
+                  "type": "integer",
+                  "minimum": 3
+                },
+                "streaming": {
+                  "type": "boolean"
+                }
+              }
+            },
+            "runtime": {
+              "type": "object",
+              "additionalProperties": false,
+              "required": [
+                "kind",
+                "node_id",
+                "generation",
+                "environment_digest",
+                "python_version",
+                "packages_digest",
+                "outer_image_digest"
+              ],
+              "properties": {
+                "kind": {
+                  "const": "python-env"
+                },
+                "node_id": {
+                  "type": "string",
+                  "pattern": "^[a-zA-Z0-9][a-zA-Z0-9_.-]{0,127}$"
+                },
+                "generation": {
+                  "type": "string",
+                  "pattern": "^[a-zA-Z0-9][a-zA-Z0-9_.-]{0,127}$"
+                },
+                "environment_digest": {
+                  "$ref": "#/$defs/sha256"
+                },
+                "python_version": {
+                  "type": "string",
+                  "minLength": 1
+                },
+                "packages_digest": {
+                  "$ref": "#/$defs/sha256"
+                },
+                "outer_image_digest": {
+                  "oneOf": [
+                    {
+                      "type": "null"
+                    },
+                    {
+                      "$ref": "#/$defs/sha256"
+                    }
+                  ]
+                }
+              }
+            }
+          }
+        }
+      ]
+    },
+    "model_node_observation": {
+      "type": "object",
+      "required": [
+        "nodeId",
+        "generation",
+        "runtimeDigest",
+        "runtime",
+        "gpuUuids",
+        "launchers",
+        "capabilities"
+      ],
+      "properties": {
+        "nodeId": {
+          "type": "string"
+        },
+        "generation": {
+          "type": "string"
+        },
+        "runtimeDigest": {
+          "type": "string",
+          "pattern": "^sha256:[a-f0-9]{64}$"
+        },
+        "runtime": {
+          "type": "object"
+        },
+        "gpuUuids": {
+          "type": "array",
+          "items": {
+            "type": "string"
+          }
+        },
+        "launchers": {
+          "type": "array",
+          "items": {
+            "type": "string"
+          }
+        },
+        "capabilities": {
+          "type": "object"
+        }
+      }
+    }
+  },
+  "allOf": [
+    {
+      "if": {
+        "required": [
+          "rerun_id"
+        ]
+      },
+      "then": {
+        "required": [
+          "eval_id"
+        ]
+      }
+    },
+    {
+      "if": {
+        "properties": {
+          "schema_version": {
+            "const": "2"
+          }
+        }
+      },
+      "then": {
+        "required": [
+          "model_node_observation"
+        ]
+      },
+      "else": {
+        "not": {
+          "required": [
+            "model_node_observation"
+          ]
+        }
+      }
+    }
+  ],
+  "$defs": {
+    "sha256": {
+      "type": "string",
+      "pattern": "^sha256:[a-f0-9]{64}$"
+    },
+    "check": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "status",
+        "message"
+      ],
+      "properties": {
+        "status": {
+          "enum": [
+            "pass",
+            "fail"
+          ]
+        },
+        "message": {
+          "type": "string"
+        }
+      }
+    }
+  }
+}