agent-hitch 0.2.9 → 0.2.10

package/docs/schemas/inference-lock.schema.json

@@ -0,0 +1,487 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://agent-hitch.local/schemas/inference-lock.schema.json",
+  "title": "Hitch inference lock v1/v2",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema_version",
+    "engine",
+    "model_id",
+    "runtime_id",
+    "inference_id",
+    "profile",
+    "execution",
+    "generation",
+    "protocol",
+    "resources"
+  ],
+  "properties": {
+    "schema_version": {
+      "enum": [
+        "1",
+        "2"
+      ]
+    },
+    "engine": {
+      "const": "sglang"
+    },
+    "model_id": {
+      "$ref": "#/$defs/sha256"
+    },
+    "runtime_id": {
+      "$ref": "#/$defs/sha256"
+    },
+    "inference_id": {
+      "$ref": "#/$defs/sha256"
+    },
+    "profile": {
+      "enum": [
+        "baseline",
+        "throughput"
+      ]
+    },
+    "execution": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "platform",
+        "load_format",
+        "dtype",
+        "quantization",
+        "tensor_parallel_size",
+        "data_parallel_size",
+        "pipeline_parallel_size",
+        "context_tokens_per_request",
+        "max_running_requests",
+        "max_queued_requests",
+        "max_total_tokens",
+        "chunked_prefill_size",
+        "max_prefill_tokens",
+        "kv_cache_dtype",
+        "attention_backend",
+        "sampling_backend",
+        "deterministic_inference",
+        "prefix_cache",
+        "hicache",
+        "speculative_decoding",
+        "cpu_offload_gb",
+        "startup_timeout_ms",
+        "queue_timeout_ms",
+        "request_timeout_ms",
+        "idle_ttl_ms"
+      ],
+      "properties": {
+        "platform": {
+          "$ref": "#/$defs/platform"
+        },
+        "load_format": {
+          "const": "safetensors"
+        },
+        "dtype": {
+          "type": "string",
+          "minLength": 1
+        },
+        "quantization": {
+          "type": [
+            "string",
+            "null"
+          ]
+        },
+        "tensor_parallel_size": {
+          "const": 1
+        },
+        "data_parallel_size": {
+          "const": 1
+        },
+        "pipeline_parallel_size": {
+          "const": 1
+        },
+        "context_tokens_per_request": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "max_running_requests": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "max_queued_requests": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "max_total_tokens": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "chunked_prefill_size": {
+          "type": "integer",
+          "minimum": -1
+        },
+        "max_prefill_tokens": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "kv_cache_dtype": {
+          "type": "string",
+          "minLength": 1
+        },
+        "attention_backend": {
+          "type": "string",
+          "minLength": 1
+        },
+        "sampling_backend": {
+          "type": "string",
+          "minLength": 1
+        },
+        "deterministic_inference": {
+          "type": "boolean"
+        },
+        "prefix_cache": {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "mode",
+            "scope",
+            "initial_state"
+          ],
+          "properties": {
+            "mode": {
+              "enum": [
+                "disabled",
+                "radix"
+              ]
+            },
+            "scope": {
+              "enum": [
+                "run",
+                "eval"
+              ]
+            },
+            "initial_state": {
+              "const": "empty"
+            }
+          }
+        },
+        "hicache": {
+          "const": false
+        },
+        "speculative_decoding": {
+          "const": false
+        },
+        "cpu_offload_gb": {
+          "const": 0
+        },
+        "startup_timeout_ms": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "queue_timeout_ms": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "request_timeout_ms": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "idle_ttl_ms": {
+          "type": "integer",
+          "minimum": 0
+        }
+      }
+    },
+    "generation": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "temperature",
+        "top_p",
+        "top_k",
+        "min_p",
+        "repetition_penalty",
+        "seed",
+        "max_output_tokens",
+        "override_policy"
+      ],
+      "properties": {
+        "temperature": {
+          "type": "number",
+          "minimum": 0
+        },
+        "top_p": {
+          "type": "number",
+          "minimum": 0,
+          "maximum": 1
+        },
+        "top_k": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "min_p": {
+          "type": "number",
+          "minimum": 0,
+          "maximum": 1
+        },
+        "repetition_penalty": {
+          "type": "number",
+          "exclusiveMinimum": 0
+        },
+        "seed": {
+          "type": "integer"
+        },
+        "max_output_tokens": {
+          "type": "integer",
+          "minimum": 1
+        },
+        "override_policy": {
+          "const": "reject-conflicts"
+        }
+      }
+    },
+    "protocol": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "api",
+        "streaming",
+        "tool_calls",
+        "parallel_tool_calls",
+        "input_modalities",
+        "tool_call_parser",
+        "reasoning_parser",
+        "compatibility_profile_id"
+      ],
+      "properties": {
+        "api": {
+          "enum": [
+            "responses",
+            "chat-completions"
+          ]
+        },
+        "streaming": {
+          "type": "boolean"
+        },
+        "tool_calls": {
+          "type": "boolean"
+        },
+        "parallel_tool_calls": {
+          "type": "boolean"
+        },
+        "input_modalities": {
+          "const": [
+            "text"
+          ]
+        },
+        "tool_call_parser": {
+          "type": [
+            "string",
+            "null"
+          ]
+        },
+        "reasoning_parser": {
+          "type": [
+            "string",
+            "null"
+          ]
+        },
+        "compatibility_profile_id": {
+          "$ref": "#/$defs/sha256"
+        }
+      }
+    },
+    "resources": {
+      "$ref": "#/$defs/resources"
+    },
+    "model_node": {
+      "$ref": "#/$defs/modelNodeBinding"
+    }
+  },
+  "$defs": {
+    "sha256": {
+      "type": "string",
+      "pattern": "^sha256:[a-f0-9]{64}$"
+    },
+    "platform": {
+      "oneOf": [
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "backend",
+            "cpu_threads",
+            "numa_policy",
+            "cpu_feature_requirement",
+            "overlap_schedule"
+          ],
+          "properties": {
+            "backend": {
+              "const": "cpu"
+            },
+            "cpu_threads": {
+              "type": "integer",
+              "minimum": 1
+            },
+            "numa_policy": {
+              "const": "single-node"
+            },
+            "cpu_feature_requirement": {
+              "const": "amx"
+            },
+            "overlap_schedule": {
+              "const": false
+            }
+          }
+        },
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "backend",
+            "mem_fraction_static",
+            "overlap_schedule",
+            "cuda_graph"
+          ],
+          "properties": {
+            "backend": {
+              "const": "cuda"
+            },
+            "device_constraint": {
+              "type": "string",
+              "minLength": 1
+            },
+            "mem_fraction_static": {
+              "type": "number",
+              "exclusiveMinimum": 0,
+              "maximum": 1
+            },
+            "overlap_schedule": {
+              "type": "boolean"
+            },
+            "cuda_graph": {
+              "enum": [
+                "enabled",
+                "disabled"
+              ]
+            }
+          }
+        },
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "backend",
+            "mlx_quantization",
+            "overlap_schedule"
+          ],
+          "properties": {
+            "backend": {
+              "const": "metal"
+            },
+            "mlx_quantization": {
+              "enum": [
+                "none",
+                "prequantized"
+              ]
+            },
+            "overlap_schedule": {
+              "type": "boolean"
+            }
+          }
+        }
+      ]
+    },
+    "resources": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "cpu_millis",
+        "memory_bytes",
+        "container_slots",
+        "build_slots"
+      ],
+      "properties": {
+        "cpu_millis": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "memory_bytes": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "container_slots": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "build_slots": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "gpu_count": {
+          "type": "integer",
+          "minimum": 0
+        },
+        "ephemeral_disk_bytes": {
+          "type": "integer",
+          "minimum": 0
+        }
+      }
+    },
+    "modelNodeBinding": {
+      "$ref": "model-node-binding.schema.json"
+    }
+  },
+  "allOf": [
+    {
+      "if": {
+        "properties": {
+          "schema_version": {
+            "const": "2"
+          }
+        }
+      },
+      "then": {
+        "required": [
+          "model_node"
+        ],
+        "properties": {
+          "resources": {
+            "properties": {
+              "container_slots": {
+                "const": 0
+              }
+            }
+          },
+          "execution": {
+            "properties": {
+              "platform": {
+                "required": [
+                  "device_constraint"
+                ],
+                "properties": {
+                  "backend": {
+                    "const": "cuda"
+                  }
+                }
+              }
+            }
+          }
+        }
+      },
+      "else": {
+        "not": {
+          "required": [
+            "model_node"
+          ]
+        },
+        "properties": {
+          "resources": {
+            "properties": {
+              "container_slots": {
+                "minimum": 1
+              }
+            }
+          }
+        }
+      }
+    }
+  ]
+}

package/docs/schemas/inference-runtime-manifest.schema.json

@@ -0,0 +1,156 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://agent-hitch.local/schemas/inference-runtime-manifest.schema.json",
+  "title": "Hitch SGLang runtime manifest v1/v2",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema_version",
+    "runtime_id",
+    "engine",
+    "sglang_version",
+    "sglang_commit",
+    "backend",
+    "package",
+    "compatibility_profile"
+  ],
+  "properties": {
+    "schema_version": {
+      "enum": [
+        "1",
+        "2"
+      ]
+    },
+    "runtime_id": {
+      "$ref": "#/$defs/sha256"
+    },
+    "engine": {
+      "const": "sglang"
+    },
+    "sglang_version": {
+      "type": "string",
+      "pattern": "^v?[0-9]+\\.[0-9]+\\.[0-9]+"
+    },
+    "sglang_commit": {
+      "type": ["string", "null"],
+      "pattern": "^[a-f0-9]{40}$"
+    },
+    "backend": {
+      "enum": [
+        "cpu",
+        "cuda",
+        "metal"
+      ]
+    },
+    "package": {
+      "oneOf": [
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "kind",
+            "image",
+            "image_digest",
+            "platform"
+          ],
+          "properties": {
+            "kind": {
+              "const": "oci"
+            },
+            "image": {
+              "type": "string",
+              "minLength": 1
+            },
+            "image_digest": {
+              "$ref": "#/$defs/sha256"
+            },
+            "platform": {
+              "const": "linux/amd64"
+            }
+          }
+        },
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "required": [
+            "kind",
+            "environment_digest",
+            "python_version",
+            "packages_digest"
+          ],
+          "properties": {
+            "kind": {
+              "const": "python-env"
+            },
+            "environment_digest": {
+              "$ref": "#/$defs/sha256"
+            },
+            "python_version": {
+              "type": "string",
+              "minLength": 1
+            },
+            "packages_digest": {
+              "$ref": "#/$defs/sha256"
+            }
+          }
+        }
+      ]
+    },
+    "compatibility_profile": {
+      "type": "string",
+      "minLength": 1
+    }
+  },
+  "$defs": {
+    "sha256": {
+      "type": "string",
+      "pattern": "^sha256:[a-f0-9]{64}$"
+    }
+  },
+  "allOf": [
+    {
+      "if": { "properties": { "sglang_commit": { "type": "null" } } },
+      "then": { "properties": { "schema_version": { "const": "2" }, "package": { "properties": { "kind": { "const": "python-env" } } } } }
+    },
+    {
+      "if": {
+        "properties": {
+          "schema_version": {
+            "const": "1"
+          }
+        }
+      },
+      "then": {
+        "if": {
+          "properties": {
+            "backend": {
+              "const": "metal"
+            }
+          }
+        },
+        "then": {
+          "properties": {
+            "package": {
+              "properties": {
+                "kind": {
+                  "const": "python-env"
+                }
+              }
+            }
+          }
+        },
+        "else": {
+          "properties": {
+            "package": {
+              "properties": {
+                "kind": {
+                  "const": "oci"
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  ]
+}