prefer-inference-core 0.0.0-g0be6ef9

package/schemas/prefer-model-catalog-extension.schema.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://github.com/cvalusek/PreFer/blob/develop/catalog/model-catalog-extension.schema.json",
+  "title": "PreFer runtime Hugging Face catalog extension",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema_version", "generated_at", "repositories", "models"],
+  "properties": {
+    "schema_version": {"const": "prefer.model-catalog-extension.v1"},
+    "generated_at": {"type": "string", "format": "date-time"},
+    "repositories": {"type": "object", "additionalProperties": {"type": "object"}},
+    "models": {
+      "type": "object",
+      "additionalProperties": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["repository", "resolved_revision"],
+        "properties": {
+          "repository": {"type": "string", "pattern": "^[^/]+/[^/]+$"},
+          "resolved_revision": {"type": "string", "pattern": "^[0-9a-f]{40}$"}
+        }
+      }
+    }
+  }
+}

package/schemas/prefer-model-catalog.schema.json

@@ -0,0 +1,54 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://github.com/cvalusek/PreFer/blob/develop/catalog/model-catalog.schema.json",
+  "title": "PreFer materialized model catalog",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema_version", "generated_at", "source_fingerprint", "catalog_fingerprint", "defaults", "models", "repositories"],
+  "properties": {
+    "schema_version": {"const": "prefer.model-catalog.v1"},
+    "generated_at": {"type": "string", "format": "date-time"},
+    "source_fingerprint": {"type": "string", "pattern": "^[0-9a-f]{64}$"},
+    "catalog_fingerprint": {"type": "string", "pattern": "^[0-9a-f]{64}$"},
+    "defaults": {"type": "object"},
+    "models": {"type": "object", "additionalProperties": {"type": "object"}},
+    "repositories": {"type": "object", "additionalProperties": {"$ref": "#/$defs/repository"}}
+  },
+  "$defs": {
+    "repository": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["repository", "requested_revision", "resolved_revision", "base_models", "gated", "private", "tags", "files_complete", "files", "retrieved_at"],
+      "properties": {
+        "repository": {"type": "string", "pattern": "^[^/]+/[^/]+$"},
+        "requested_revision": {"type": "string", "minLength": 1},
+        "resolved_revision": {"type": "string", "pattern": "^[0-9a-f]{40}$"},
+        "base_models": {"type": "array", "uniqueItems": true, "items": {"type": "string"}},
+        "pipeline_tag": {"type": "string"},
+        "library_name": {"type": "string"},
+        "license": {"type": "string"},
+        "gated": {"oneOf": [{"type": "boolean"}, {"type": "string"}]},
+        "private": {"type": "boolean"},
+        "tags": {"type": "array", "uniqueItems": true, "items": {"type": "string"}},
+        "files_complete": {"type": "boolean"},
+        "include_patterns": {"type": "array", "uniqueItems": true, "items": {"type": "string", "minLength": 1}},
+        "files": {
+          "type": "array",
+          "items": {
+            "type": "object",
+            "additionalProperties": false,
+            "required": ["path", "size"],
+            "properties": {
+              "path": {"type": "string", "minLength": 1},
+              "size": {"type": "integer", "minimum": 0},
+              "blob_oid": {"type": "string"},
+              "lfs_sha256": {"type": "string", "pattern": "^[0-9a-f]{64}$"},
+              "xet_hash": {"type": "string"}
+            }
+          }
+        },
+        "retrieved_at": {"type": "string", "format": "date-time"}
+      }
+    }
+  }
+}

package/schemas/prefer-model-plan.schema.json

@@ -0,0 +1,115 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://github.com/cvalusek/PreFer/blob/develop/catalog/model-plan.schema.json",
+  "title": "PreFer resource-aware model plan",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema_version", "engine", "resources", "selected", "skipped", "capabilities", "staged_artifact_bytes", "complete"],
+  "properties": {
+    "schema_version": {"const": "prefer.model-plan.v1"},
+    "engine": {"type": "string", "minLength": 1},
+    "resources": {"$ref": "prefer-resource-profile.schema.json"},
+    "selected": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["model_id", "preferred_quant", "selected", "fit", "quant_changed", "reason"],
+        "properties": {
+          "model_id": {"type": "string", "minLength": 1},
+          "preferred_quant": {"type": "string", "minLength": 1},
+          "requested_quant": {"type": "string", "minLength": 1},
+          "selected": {"type": "object"},
+          "fit": {"$ref": "#/$defs/fit"},
+          "quant_changed": {"type": "boolean"},
+          "reason": {"type": "string", "minLength": 1},
+          "workload_request": {"$ref": "#/$defs/workloadRequest"},
+          "workload": {"$ref": "#/$defs/workloadResult"}
+        }
+      }
+    },
+    "skipped": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["model_id", "required", "reason", "details"],
+        "properties": {
+          "model_id": {"type": "string", "minLength": 1},
+          "required": {"type": "boolean"},
+          "reason": {"enum": ["no-compatible-variant", "missing-capability", "insufficient-memory", "storage-budget", "model-limit"]},
+          "details": {"type": "array", "items": {"type": "string"}}
+        }
+      }
+    },
+    "capabilities": {"type": "array", "uniqueItems": true, "items": {"type": "string"}},
+    "staged_artifact_bytes": {"type": "integer", "minimum": 0},
+    "complete": {"type": "boolean"},
+    "hints": {"$ref": "#/$defs/hints"}
+  },
+  "$defs": {
+    "hints": {
+      "type": "object",
+      "additionalProperties": false,
+      "properties": {
+        "required_capabilities": {"$ref": "#/$defs/stringList"},
+        "preferred_capabilities": {"$ref": "#/$defs/stringList"},
+        "preferred_roles": {"$ref": "#/$defs/stringList"},
+        "speed_importance": {"type": "number", "minimum": 0, "maximum": 1},
+        "quality_importance": {"type": "number", "minimum": 0, "maximum": 1},
+        "quant_bias": {"enum": ["quality", "balanced", "capacity"]}
+      }
+    },
+    "stringList": {
+      "type": "array",
+      "uniqueItems": true,
+      "items": {"type": "string", "minLength": 1}
+    },
+    "workloadRequest": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["desired_context_tokens", "desired_concurrency", "bytes_per_token"],
+      "properties": {
+        "desired_context_tokens": {"type": "integer", "minimum": 1},
+        "desired_concurrency": {"type": "integer", "minimum": 1},
+        "minimum_context_tokens": {"type": "integer", "minimum": 1},
+        "minimum_concurrency": {"type": "integer", "minimum": 1},
+        "priority": {"enum": ["context", "concurrency"]},
+        "bytes_per_token": {"type": "number", "exclusiveMinimum": 0},
+        "fixed_device_bytes": {"type": "integer", "minimum": 0},
+        "source": {"enum": ["measured", "architecture", "operator", "estimate"]}
+      }
+    },
+    "workloadResult": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["fits", "context_tokens", "concurrency", "aggregate_token_capacity", "changed", "reason"],
+      "properties": {
+        "fits": {"type": "boolean"},
+        "context_tokens": {"type": "integer", "minimum": 0},
+        "concurrency": {"type": "integer", "minimum": 0},
+        "aggregate_token_capacity": {"type": "integer", "minimum": 0},
+        "changed": {"type": "boolean"},
+        "reason": {"type": "string", "minLength": 1}
+      }
+    },
+    "fit": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["status", "confidence", "weight_bytes", "runtime_overhead_bytes", "workload_bytes", "required_device_bytes", "reasons"],
+      "properties": {
+        "status": {"enum": ["fits", "tight", "host-offload", "does-not-fit", "unknown"]},
+        "confidence": {"enum": ["measured", "architecture", "artifact-only", "unknown"]},
+        "capacity_bytes": {"type": "integer", "minimum": 0},
+        "usable_capacity_bytes": {"type": "integer", "minimum": 0},
+        "weight_bytes": {"type": "integer", "minimum": 0},
+        "runtime_overhead_bytes": {"type": "integer", "minimum": 0},
+        "workload_bytes": {"type": "integer", "minimum": 0},
+        "required_device_bytes": {"type": "integer", "minimum": 0},
+        "remaining_device_bytes": {"type": "integer"},
+        "host_offload_bytes": {"type": "integer", "minimum": 0},
+        "reasons": {"type": "array", "items": {"type": "string"}}
+      }
+    }
+  }
+}

package/schemas/prefer-resource-profile.schema.json

@@ -0,0 +1,70 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://github.com/cvalusek/PreFer/blob/develop/catalog/resource-profile.schema.json",
+  "title": "PreFer normalized runtime resources",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema_version", "source", "memory_topology", "accelerators", "capabilities"],
+  "properties": {
+    "schema_version": {"const": "prefer.resources.v1"},
+    "source": {"enum": ["deployment", "runtime", "merged"]},
+    "provider": {"type": "string", "minLength": 1},
+    "provider_sku": {"type": "string", "minLength": 1},
+    "memory_topology": {"enum": ["discrete", "unified", "host", "unknown"]},
+    "accelerators": {"type": "array", "items": {"$ref": "#/$defs/accelerator"}},
+    "unified_memory": {"$ref": "#/$defs/memory"},
+    "host_memory": {"$ref": "#/$defs/memory"},
+    "cpu": {
+      "type": "object",
+      "additionalProperties": false,
+      "properties": {
+        "architecture": {"type": "string"},
+        "logical_cores": {"type": "integer", "minimum": 1},
+        "model": {"type": "string"},
+        "capabilities": {"$ref": "#/$defs/stringList"}
+      }
+    },
+    "storage": {
+      "type": "object",
+      "additionalProperties": false,
+      "properties": {
+        "total_bytes": {"type": "integer", "minimum": 0},
+        "available_bytes": {"type": "integer", "minimum": 0},
+        "kind": {"enum": ["local", "network", "unknown"]}
+      }
+    },
+    "capabilities": {"$ref": "#/$defs/stringList"}
+  },
+  "$defs": {
+    "memory": {
+      "type": "object",
+      "additionalProperties": false,
+      "properties": {
+        "total_bytes": {"type": "integer", "minimum": 0},
+        "available_bytes": {"type": "integer", "minimum": 0}
+      }
+    },
+    "accelerator": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["capabilities"],
+      "properties": {
+        "id": {"type": "string"},
+        "slug": {"type": "string"},
+        "canonical_id": {"type": "string"},
+        "provider_id": {"type": "string"},
+        "name": {"type": "string"},
+        "architecture": {"type": "string"},
+        "compute_capability": {"type": "string"},
+        "total_bytes": {"type": "integer", "minimum": 0},
+        "available_bytes": {"type": "integer", "minimum": 0},
+        "capabilities": {"$ref": "#/$defs/stringList"}
+      }
+    },
+    "stringList": {
+      "type": "array",
+      "uniqueItems": true,
+      "items": {"type": "string", "minLength": 1}
+    }
+  }
+}