cortexgrid 0.2.95__tar.gz → 0.2.97__tar.gz

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: cortexgrid
-Version: 0.2.95
+Version: 0.2.97
 Summary: Connect your ML code to the RoboLab compute cluster — Ray, MLflow, and S3
 Project-URL: Homepage, https://github.com/robodatalab/cortexgrid
 Project-URL: Repository, https://github.com/robodatalab/cortexgrid
@@ -165,20 +165,24 @@ app = FastAPI()
 
 @serve.ingress(app)
 class MyServeApp:
-    num_gpus = 1
-
     def __init__(self, family: str, suffix: str, run_name: str) -> None:
         self._weights_dir = cortexgrid.load_model(family, suffix, run_name)
 
     @app.post("/complete")
     async def complete(self, body: dict): ...
 
-saved = cortexgrid.save_model(weights_dir, MyServeApp, family="qwen", suffix="instruct")
+saved = cortexgrid.save_model(
+    weights_dir, MyServeApp, family="qwen", suffix="instruct",
+    # What one replica needs; the model is deployed only on a host that has it.
+    requirements=cortexgrid.ModelRequirements(num_gpus=1, ram_gb=8, vram_gb=16),
+)
 deployed = cortexgrid.deploy_model("qwen", "instruct", saved.run_name, wait=True)
 print(deployed.url)
 ```
 
-`save_model` saves a new copy under every run - meant for weights the run produced (e.g. a fine-tune). For a model produced elsewhere (e.g. a pretrained base model), `cortexgrid.import_model(source, MyServeApp, family, suffix)` uploads it once under `run_name=cortexgrid.IMPORTED` and is a no-op on later runs; deploy it with `deploy_model(family, suffix, cortexgrid.IMPORTED)`.
+The requirements are part of the model, not of the serve-app class: GPUs, RAM and VRAM (GiB, 0 meaning no requirement) are stored with it and matched against what the cluster's hosts have free. Correct them later with `cortexgrid.set_model_requirements(family, suffix, run_name, requirements)` or on the model card in the dashboard; `cortexgrid.deploy_model(..., num_replicas=2)` chooses how many copies to run.
+
+`save_model` saves a new copy under every run - meant for weights the run produced (e.g. a fine-tune). For a model produced elsewhere (e.g. a pretrained base model), `cortexgrid.import_model(source, MyServeApp, family, suffix)` uploads it once under `run_name=cortexgrid.IMPORTED` and on later runs only re-bundles `MyServeApp` if its code changed; deploy it with `deploy_model(family, suffix, cortexgrid.IMPORTED)`.
 
 `save_model` is synchronous (registry lifecycle: `uploading` -> `ready`); `deploy_model` schedules the serving lifecycle (`deploying` -> `running`). With `wait=True` a failed deploy raises `cortexgrid.ModelDeployFailed`; `cortexgrid.wait_for_model_serving(family, suffix, run_name, timeout=...)` waits on a deploy started elsewhere, and re-deploying a failed model retries it from scratch. See [model-serving.md](https://github.com/robodatalab/cortexgrid/blob/main/docs/cortexgrid/model-serving.md) for both lifecycles end to end - upload/deploy/undeploy/delete, status queries (`model_registry_status`, `model_serving_status`), and error handling.
 

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/cortexgrid/__init__.py

@@ -80,12 +80,14 @@ from cortexgrid.model_storage import (
     list_models,
     load_model,
     model_registry_status,
+    set_model_requirements,
 )
 from cortexgrid.model_storage import import_model as _import_model_storage
 from cortexgrid.model_storage import save_model as _save_model_storage
 from cortexgrid.model_serving import (
     Deployment,
     ModelDeployFailed,
+    ModelRequirements,
     ServingStatus,
     deploy_model,
     list_deployed_models,
@@ -118,10 +120,15 @@ def remote(
 
 
 def save_model(
-    weights_dir: str | Path, serve_app: type, family: str, suffix: str
+    weights_dir: str | Path,
+    serve_app: type,
+    family: str,
+    suffix: str,
+    requirements: ModelRequirements | None = None,
 ) -> SavedModel:
     """Persist a weights directory under the current Experiment's run, paired
-    with the serve-app class that will front it at deploy time.
+    with the serve-app class that will front it at deploy time and the
+    hardware one replica of it needs.
 
     Every run saves a new copy under its own run_name - meant for weights the
     run produced (e.g. a fine-tune). For a model produced elsewhere that should
@@ -134,6 +141,7 @@ def save_model(
         family,
         run_id=experiment.run_id,
         run_name=experiment.run_name(),
+        requirements=requirements,
     )
 
 
@@ -142,6 +150,7 @@ def import_model(
     serve_app: type,
     family: str,
     suffix: str,
+    requirements: ModelRequirements | None = None,
 ) -> SavedModel:
     """Register a model produced elsewhere once, reuse it on every later call,
     and record on the current Experiment's run which imported model it used.
@@ -151,7 +160,7 @@ def import_model(
     `imported_model/<family>/<suffix>` holds the version's `created_at`, set
     whether this call uploaded the model or reused it."""
     experiment = Experiment.get_instance()
-    model = _import_model_storage(source, serve_app, family, suffix)
+    model = _import_model_storage(source, serve_app, family, suffix, requirements)
     get_mlflow_client().set_tag(
         experiment.run_id, f"imported_model/{family}/{suffix}", model.created_at
     )
@@ -213,6 +222,8 @@ __all__ = [
     "load_model",
     "list_models",
     "model_registry_status",
+    "ModelRequirements",
+    "set_model_requirements",
     "delete_model",
     # Model serving
     "Deployment",

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/cortexgrid/_bundle.py

@@ -12,6 +12,8 @@ Bundles of several seeds combine with `BundleDesc.merge`.
 
 `stage(files, dest)` lays a bundle out under `dest` at each file's import path,
 so `dest` on sys.path (e.g. a Ray working_dir) makes every module importable.
+`digest(files)` hashes that layout, so bundles that stage identically compare
+equal wherever their files live.
 
 `BundleDesc.pip_requirements(worker_provides())` pins the third-party
 distributions the Ray worker image does not already have, for a Ray `pip`
@@ -24,6 +26,7 @@ import ast
 from collections.abc import Iterable, Iterator
 from dataclasses import dataclass
 import functools
+import hashlib
 import importlib.machinery
 import importlib.metadata
 import importlib.util
@@ -105,11 +108,23 @@ def stage(files: set[Path], dest: Path) -> None:
     in installed distributions)."""
     dest.mkdir(parents=True, exist_ok=True)
     for file in files:
-        target = dest / file.relative_to(_sys_path_root(file))
+        target = dest / _import_path(file)
         target.parent.mkdir(parents=True, exist_ok=True)
         shutil.copy2(file, target)
 
 
+def digest(files: set[Path]) -> str:
+    """SHA-256 of `files` as `stage` lays them out: each file's import path and
+    contents, in import-path order. Files that stage identically digest
+    identically, wherever they live on disk."""
+    sha = hashlib.sha256()
+    for path, file in sorted((_import_path(file).as_posix(), file) for file in files):
+        content = file.read_bytes()
+        sha.update(f"{path}\0{len(content)}\0".encode())
+        sha.update(content)
+    return sha.hexdigest()
+
+
 # What the Ray worker image pip-installs, as the Dockerfile spells it. They and
 # their dependency trees are on the worker already, so they are never installed
 # there again -- a second copy in the job's virtualenv would shadow the image's.
@@ -335,6 +350,11 @@ def _package(file: Path) -> str:
     return ".".join(parts)
 
 
+def _import_path(file: Path) -> Path:
+    """`file` relative to the sys.path entry it is imported from."""
+    return file.relative_to(_sys_path_root(file))
+
+
 def _sys_path_root(file: Path) -> Path:
     """The sys.path entry `file` is imported from: its first ancestor directory
     without an __init__.py."""

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/cortexgrid/_serve_entry.py

@@ -4,9 +4,9 @@ Ray Serve's REST `import_path` resolves to `cortexgrid._serve_entry:build`.
 On the cluster replica, `build` imports the serve-app class bundled at
 `save_model` time (its import path was stored as an MLflow tag), applies Ray's
 ingress with the app it was marked with by `cortexgrid.serve.ingress` (again on
-each replica, see `_IngressOnReplica`), reads its
-`num_gpus`/`num_replicas` class attributes for actor placement, wraps it as a
-Ray Serve deployment, and binds it with the (family, suffix, run_name)
+each replica, see `_IngressOnReplica`), wraps it as a Ray Serve deployment
+with the replica count and Ray resource requests `deploy_model` derived from
+the model's requirements, and binds it with the (family, suffix, run_name)
 identifiers.
 
 The serve-app owns everything about traffic: its own routes, request schemas,
@@ -14,11 +14,9 @@ streaming, and timeouts. cortexgrid does not interpose a request/response
 contract - it only schedules the app and hands it the identifiers it needs to
 fetch its own weights via `cortexgrid.load_model`.
 
-Design note: resource needs (`num_gpus`/`num_replicas`) are read from plain
-class attributes rather than a cortexgrid decorator or base class. This is a
-deliberate, provisional choice - kept minimal until we see how serve-apps
-declare resources in practice; revisit if plain class attributes prove too
-limited.
+The serve-app declares no resources: the hardware a replica needs belongs to
+the model and is stored in the registry (`cortexgrid.ModelRequirements`), and
+the replica count is chosen per `deploy_model`.
 """
 
 from __future__ import annotations
@@ -72,12 +70,13 @@ def build(args: dict[str, Any]) -> Application:
             {"_serve_app": serve_app},
         )
         serve_app = serve.ingress(app)(on_replica)
-    num_gpus = getattr(serve_app, "num_gpus", 0)
-    num_replicas = getattr(serve_app, "num_replicas", 1)
     return serve.deployment(serve_app).options(
-        num_replicas=num_replicas,
+        # This builder ships in the bundle, frozen at save time, while `args`
+        # come from the cortexgrid that deploys it; one older than the bundle
+        # sends neither key.
+        num_replicas=args.get("num_replicas", 1),
         # Ray 2.32 lowered the default from 100 to 5; keep what serve-apps
         # had on Ray 2.9.
         max_ongoing_requests=_MAX_ONGOING_REQUESTS,
-        ray_actor_options={"num_gpus": num_gpus},
+        ray_actor_options=args.get("ray_actor_options", {}),
     ).bind(args["family"], args["suffix"], args["run_name"])

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/cortexgrid/model_serving.py

@@ -3,11 +3,13 @@
 Caller stays HTTP-only: deploy/undeploy/list talk to the Ray dashboard's
 declarative `/api/serve/applications/` endpoint via [cortexgrid.ray_util],
 never `ray.init`. The deployment class is bundled at `save_model` time, zipped,
-uploaded to MinIO under `serve-bundles/<run_name>/<family>__<suffix>.zip`, and
+uploaded to MinIO under
+`serve-bundles/<run_name>/<family>__<suffix>/<fingerprint>.zip`, and
 referenced via `runtime_env.working_dir` so Ray workers fetch it from there.
-The bundle URL, class import path, and pip list are persisted as MLflow tags
-on the ModelVersion so `deploy_model` can find them later without the caller
-holding the class object.
+The bundle URL, class import path, pip list, and fingerprint are persisted as
+MLflow tags on the ModelVersion so `deploy_model` can find them later without
+the caller holding the class object. So are the model's `ModelRequirements`,
+which `deploy_model` turns into the replica's Ray resource requests.
 
 Naming: the Ray Serve application is named "<family>__<suffix>__<run_name>".
 This relies on family/suffix/run_name not containing the literal "__".
@@ -26,13 +28,14 @@ import shutil
 import tempfile
 import time
 from dataclasses import dataclass, field
+import hashlib
 from pathlib import Path
 from typing import Any
 
 from mlflow.tracking import MlflowClient
 from ray.serve.schema import ApplicationStatus
 
-from cortexgrid._bundle import bundle, stage, worker_provides
+from cortexgrid._bundle import bundle, digest, stage, worker_provides
 from cortexgrid.infra import get_mlflow_tracking_uri, get_ray_serve_uri
 from cortexgrid.ray_util import (
     get_serve_details,
@@ -95,17 +98,28 @@ class BundleMetadata:
     # pinned third-party requirements the replica pip-installs (the bundle's
     # distributions the Ray image does not already provide)
     pip_requirements: list[str] = field(default_factory=list)
+    # ServeBundle.fingerprint of the uploaded bundle; empty for models saved
+    # before bundles were fingerprinted
+    fingerprint: str = ""
 
 
-def bundle_class(
-    cls: type, family: str, suffix: str, run_name: str
-) -> BundleMetadata:
-    """Bundle the serve-app class's code (and the serve entrypoint), zip it, and
-    upload to MinIO.
+@dataclass
+class ServeBundle:
+    """A serve-app's bundle, resolved locally but not uploaded yet: what
+    `build_bundle` finds and `upload_bundle` ships."""
 
-    Returns the metadata `deploy_model` needs later; callers (typically
-    `save_model`) persist it on the ModelVersion so the deploy step can run
-    without holding the class object.
+    files: set[Path]
+    class_import_path: str
+    pip_requirements: list[str]
+    # Hash of everything the replica runs: the staged files, the class it
+    # imports, and the requirements it installs. Equal fingerprints mean the
+    # same code, so an uploaded bundle can be reused.
+    fingerprint: str
+
+
+def build_bundle(cls: type) -> ServeBundle:
+    """Resolve the serve-app class's code (and the serve entrypoint) into a
+    bundle, without uploading it.
 
     Raises ValueError for a class wrapped by `ray.serve.ingress`: that wrapper
     is a subclass Ray defines in its own module, and on older Ray (e.g. 2.9) it
@@ -120,17 +134,51 @@ def bundle_class(
     entry_file = Path(inspect.getfile(cls)).resolve()
     serve_entry = Path(__file__).with_name("_serve_entry.py")
     desc = bundle(entry_file).merge(bundle(serve_entry))
+    class_import_path = f"{cls.__module__}:{cls.__name__}"
     pip_requirements = desc.pip_requirements(worker_provides())
+    fingerprint = hashlib.sha256(
+        json.dumps(
+            [digest(desc.local_files), class_import_path, pip_requirements]
+        ).encode()
+    ).hexdigest()
+    return ServeBundle(
+        files=desc.local_files,
+        class_import_path=class_import_path,
+        pip_requirements=pip_requirements,
+        fingerprint=fingerprint,
+    )
+
+
+def bundle_class(
+    cls: type, family: str, suffix: str, run_name: str
+) -> BundleMetadata:
+    """Bundle the serve-app class's code (and the serve entrypoint), zip it, and
+    upload to MinIO: `build_bundle` followed by `upload_bundle`.
+
+    Returns the metadata `deploy_model` needs later; callers (typically
+    `save_model`) persist it on the ModelVersion so the deploy step can run
+    without holding the class object."""
+    return upload_bundle(build_bundle(cls), family, suffix, run_name)
+
+
+def upload_bundle(
+    serve_bundle: ServeBundle, family: str, suffix: str, run_name: str
+) -> BundleMetadata:
+    """Zip a built bundle and upload it under its fingerprint.
+
+    The fingerprint is part of the URL because Ray keeps a remote working_dir
+    it has downloaded and reuses it for the same URL: new code at an old URL
+    would never reach a replica."""
     with tempfile.TemporaryDirectory() as tmp:
         code_root = Path(tmp) / "code"
-        stage(desc.local_files, code_root)
+        stage(serve_bundle.files, code_root)
         log.info(
             "Serve bundle for %s/%s/%s: %d files, pip: %s",
             family,
             suffix,
             run_name,
-            len(desc.local_files),
-            pip_requirements,
+            len(serve_bundle.files),
+            serve_bundle.pip_requirements,
         )
         # Ray unpacks a remote (s3://) working_dir zip by stripping its
         # top-level directory when there is exactly one, so a bundle of a single
@@ -146,19 +194,54 @@ def bundle_class(
         )
         bundle_url = upload(
             archive,
-            dest_path=f"serve-bundles/{run_name}/{family}__{suffix}.zip",
+            dest_path=(
+                f"serve-bundles/{run_name}/{family}__{suffix}/"
+                f"{serve_bundle.fingerprint}.zip"
+            ),
         )
     return BundleMetadata(
         bundle_url=bundle_url,
-        class_import_path=f"{cls.__module__}:{cls.__name__}",
-        pip_requirements=pip_requirements,
+        class_import_path=serve_bundle.class_import_path,
+        pip_requirements=serve_bundle.pip_requirements,
+        fingerprint=serve_bundle.fingerprint,
     )
 
 
+# Custom Ray resource each GPU worker advertises: the MiB of memory its GPUs
+# have (see the ray-worker DaemonSet). A replica requests its vram_gb of it in
+# MiB, so Ray places it only on a node with that much VRAM left. MiB because
+# nvidia-smi reports MiB and a GPU's memory is not a whole number of GiB.
+_VRAM_RESOURCE = "vram_mib"
+
+_MIB_PER_GIB = 1024
+
+_GIB = 1024**3
+
+
+def _ray_actor_options(requirements: ModelRequirements) -> dict[str, Any]:
+    """Translate ModelRequirements into a replica's Ray actor resource requests.
+    Ray places the replica only on a node with that much free and reserves it
+    there; a zero requirement requests nothing."""
+    options: dict[str, Any] = {"num_gpus": requirements.num_gpus}
+    if requirements.ram_gb > 0:
+        options["memory"] = int(requirements.ram_gb * _GIB)
+    if requirements.vram_gb > 0:
+        options["resources"] = {
+            _VRAM_RESOURCE: round(requirements.vram_gb * _MIB_PER_GIB)
+        }
+    return options
+
+
 def _build_application_spec(
-    family: str, suffix: str, run_name: str, meta: BundleMetadata
+    family: str,
+    suffix: str,
+    run_name: str,
+    meta: BundleMetadata,
+    requirements: ModelRequirements,
+    num_replicas: int,
 ) -> dict[str, Any]:
-    """Assemble a Ray Serve application schema from pre-bundled metadata."""
+    """Assemble a Ray Serve application schema from pre-bundled metadata and
+    the model's requirements."""
     # working_dir carries the serve-app's own source; Ray pip-installs the
     # third-party distributions the image lacks into a per-node cached
     # virtualenv layered on the image. No pip key when there are none, so Ray
@@ -179,6 +262,8 @@ def _build_application_spec(
             "family": family,
             "suffix": suffix,
             "run_name": run_name,
+            "num_replicas": num_replicas,
+            "ray_actor_options": _ray_actor_options(requirements),
         },
         "runtime_env": runtime_env,
     }
@@ -189,23 +274,94 @@ def _build_application_spec(
 _CLASS_IMPORT_PATH_TAG = "class_import_path"
 _BUNDLE_URL_TAG = "serve_bundle_url"
 _PIP_REQUIREMENTS_TAG = "serve_pip_requirements"
+_BUNDLE_FINGERPRINT_TAG = "serve_bundle_fingerprint"
 
 
 def metadata_to_tags(meta: BundleMetadata) -> dict[str, str]:
     """Serialise BundleMetadata to MLflow tags. The inverse of
-    `_load_bundle_metadata`; lives here next to the consumer so the tag schema
+    `metadata_from_tags`; lives here next to the consumer so the tag schema
     stays in one place."""
     return {
         _CLASS_IMPORT_PATH_TAG: meta.class_import_path,
         _BUNDLE_URL_TAG: meta.bundle_url,
         _PIP_REQUIREMENTS_TAG: json.dumps(meta.pip_requirements),
+        _BUNDLE_FINGERPRINT_TAG: meta.fingerprint,
+    }
+
+
+def metadata_from_tags(tags: dict[str, str]) -> BundleMetadata:
+    """Deserialise BundleMetadata from a ModelVersion's MLflow tags. Raises
+    KeyError for a missing bundle URL or class import path."""
+    return BundleMetadata(
+        bundle_url=tags[_BUNDLE_URL_TAG],
+        class_import_path=tags[_CLASS_IMPORT_PATH_TAG],
+        # Absent on models saved before dependencies were pip-installed.
+        pip_requirements=json.loads(tags.get(_PIP_REQUIREMENTS_TAG, "[]")),
+        # Absent on models saved before bundles were fingerprinted.
+        fingerprint=tags.get(_BUNDLE_FINGERPRINT_TAG, ""),
+    )
+
+
+@dataclass
+class ModelRequirements:
+    """Hardware one replica of a model needs to be served, in GiB. Persisted as
+    tags on the ModelVersion next to the bundle metadata, so it is read without
+    touching the weights or importing the serve-app class.
+
+    Zero means no requirement: a model with no requirements is served on any
+    node, CPU-only included. Models saved before requirements existed carry
+    no tags and read as the defaults."""
+
+    num_gpus: int = 0
+    ram_gb: float = 0.0
+    # GPU memory across the replica's num_gpus GPUs, so it needs num_gpus >= 1.
+    vram_gb: float = 0.0
+
+    def __post_init__(self) -> None:
+        if self.num_gpus < 0 or self.ram_gb < 0 or self.vram_gb < 0:
+            raise ValueError(f"Model requirements cannot be negative: {self}")
+        if self.vram_gb > 0 and self.num_gpus == 0:
+            raise ValueError(
+                f"vram_gb={self.vram_gb} needs a GPU; set num_gpus >= 1"
+            )
+
+
+# MLflow tag keys for the ModelRequirements.
+_NUM_GPUS_TAG = "num_gpus"
+_RAM_GB_TAG = "ram_gb"
+_VRAM_GB_TAG = "vram_gb"
+
+
+def has_requirement_tags(tags: dict[str, str]) -> bool:
+    """Whether requirements were ever stored on the ModelVersion."""
+    return any(key in tags for key in (_NUM_GPUS_TAG, _RAM_GB_TAG, _VRAM_GB_TAG))
+
+
+def requirements_to_tags(requirements: ModelRequirements) -> dict[str, str]:
+    """Serialise ModelRequirements to MLflow tags. The inverse of
+    `requirements_from_tags`."""
+    return {
+        _NUM_GPUS_TAG: str(requirements.num_gpus),
+        _RAM_GB_TAG: str(requirements.ram_gb),
+        _VRAM_GB_TAG: str(requirements.vram_gb),
     }
 
 
-def _load_bundle_metadata(
+def requirements_from_tags(tags: dict[str, str]) -> ModelRequirements:
+    """Deserialise ModelRequirements from a ModelVersion's MLflow tags; a
+    missing tag reads as no requirement."""
+    return ModelRequirements(
+        num_gpus=int(tags.get(_NUM_GPUS_TAG, "0")),
+        ram_gb=float(tags.get(_RAM_GB_TAG, "0")),
+        vram_gb=float(tags.get(_VRAM_GB_TAG, "0")),
+    )
+
+
+def _load_deploy_metadata(
     family: str, suffix: str, run_name: str
-) -> BundleMetadata:
-    """Read the bundle metadata `save_model` persisted on the ModelVersion."""
+) -> tuple[BundleMetadata, ModelRequirements]:
+    """Read the bundle metadata and requirements `save_model` persisted on the
+    ModelVersion."""
     client = MlflowClient(tracking_uri=get_mlflow_tracking_uri())
     name = f"{family}__{suffix}"
     versions = client.search_model_versions(
@@ -217,12 +373,7 @@ def _load_bundle_metadata(
         )
     tags = versions[0].tags or {}
     try:
-        return BundleMetadata(
-            bundle_url=tags[_BUNDLE_URL_TAG],
-            class_import_path=tags[_CLASS_IMPORT_PATH_TAG],
-            # Absent on models saved before dependencies were pip-installed.
-            pip_requirements=json.loads(tags.get(_PIP_REQUIREMENTS_TAG, "[]")),
-        )
+        return metadata_from_tags(tags), requirements_from_tags(tags)
     except KeyError as exc:
         raise ValueError(
             f"Saved model {family}/{suffix}/{run_name} is missing the deployment "
@@ -338,6 +489,7 @@ def deploy_model(
     family: str,
     suffix: str,
     run_name: str,
+    num_replicas: int = 1,
     wait: bool = False,
     timeout: float | None = 300.0,
 ) -> Deployment:
@@ -348,6 +500,8 @@ def deploy_model(
 
     The serve-app class is pulled from the MLflow ModelVersion tags `save_model`
     wrote at save time; the caller does not need to hold the class object.
+    Each of the `num_replicas` replicas requests the model's `ModelRequirements`
+    from Ray, so it is placed only on a node that has them free.
 
     A DEPLOY_FAILED app left by an earlier attempt is undeployed first, and it,
     or an app still DELETING, is waited out before the new spec is PUT, so the
@@ -361,8 +515,10 @@ def deploy_model(
     stuck in DEPLOYING) will hang forever.
     """
     deadline = _deadline(timeout)
-    meta = _load_bundle_metadata(family, suffix, run_name)
-    spec = _build_application_spec(family, suffix, run_name, meta)
+    meta, requirements = _load_deploy_metadata(family, suffix, run_name)
+    spec = _build_application_spec(
+        family, suffix, run_name, meta, requirements, num_replicas
+    )
     _clear_failed_application(family, suffix, run_name, timeout, deadline)
     existing = [a for a in _current_application_specs() if a["name"] != spec["name"]]
     # The controller registers the app, sets it DEPLOYING and stamps

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/cortexgrid/model_storage.py

@@ -8,11 +8,13 @@ Mapping cortexgrid taxonomy <-> MLflow Registry:
                            "s3://<bucket>/models/<run_name>/<family>/<suffix>/weights/"
     run linkage         -> ModelVersion.run_id  (built-in MLflow field; unset
                            for imported models)
+    requirements        -> ModelVersion.tags["num_gpus"], ["ram_gb"], ["vram_gb"]
 
 Two ways in: `save_model` registers a fresh copy under the calling run's
 run_name every time it runs (fine-tuned output); `import_model` registers a
-model produced elsewhere once, under the fixed run_name IMPORTED, and is a
-no-op after that. Both write the same layout, so every
+model produced elsewhere once, under the fixed run_name IMPORTED, and after
+that only re-bundles the serve-app when its code changed. Both write the same
+layout, so every
 (family, suffix, run_name) consumer - load_model, deploy_model - handles both.
 
 storage.py is pure: it takes run_id/run_name as explicit args and never reads
@@ -35,8 +37,15 @@ from mlflow.tracking import MlflowClient
 from cortexgrid import s3_util
 from cortexgrid.infra import get_mlflow_tracking_uri, get_s3_bucket
 from cortexgrid.model_serving import (
+    ModelRequirements,
+    build_bundle,
     bundle_class,
+    has_requirement_tags,
+    metadata_from_tags,
     metadata_to_tags,
+    requirements_from_tags,
+    requirements_to_tags,
+    upload_bundle,
 )
 
 
@@ -77,6 +86,8 @@ class SavedModel:
     # _UPLOAD_DEADLINE (writer presumed dead). Versions written before this tag
     # existed report "ready".
     phase: str
+    # Hardware one replica needs; defaults for versions stored without it.
+    requirements: ModelRequirements
 
 
 def _phase_for(version: Any) -> str:
@@ -107,6 +118,7 @@ def _to_saved_model(version: Any) -> SavedModel:
         data_blob_path=version.source,
         size_bytes=int(version.tags.get("size_bytes", "0")),
         phase=_phase_for(version),
+        requirements=requirements_from_tags(version.tags),
     )
 
 
@@ -146,6 +158,7 @@ def save_model(
     family: str,
     run_id: str,
     run_name: str,
+    requirements: ModelRequirements | None = None,
 ) -> SavedModel:
     """Upload a weights directory to S3 and register a new MLflow ModelVersion
     paired with the serve-app that fronts it.
@@ -164,8 +177,13 @@ def save_model(
     weights.
     Its code is bundled and its import path, bundle URL, and pip list are
     stored as tags on the ModelVersion so `deploy_model` can bind it later
-    without the caller holding the class object."""
-    return _upload_model(weights_dir, serve_app, suffix, family, run_id, run_name)
+    without the caller holding the class object.
+
+    `requirements` is the hardware one replica needs; None stores none, which
+    reads as no requirement. Change it later with `set_model_requirements`."""
+    return _upload_model(
+        weights_dir, serve_app, suffix, family, run_id, run_name, requirements
+    )
 
 
 def import_model(
@@ -173,6 +191,7 @@ def import_model(
     serve_app: type,
     family: str,
     suffix: str,
+    requirements: ModelRequirements | None = None,
 ) -> SavedModel:
     """Register a model produced elsewhere (e.g. a pretrained base model) under
     the fixed key (family, suffix, IMPORTED), once.
@@ -184,8 +203,12 @@ def import_model(
     while it downloads.
 
     If a version is already registered under the key:
-      - "ready": no-op, returns it. `source` and `serve_app` are ignored; to
-        replace the weights or the serve-app, `delete_model` it first.
+      - "ready": returns it without calling `source`. If `serve_app`'s code no
+        longer matches the stored bundle, it is re-bundled first and the
+        weights are kept (see `_refresh_bundle`). `requirements` are stored
+        only if the version has none yet, so values changed since with
+        `set_model_requirements` are kept. To replace the weights,
+        `delete_model` it first.
       - "uploading": raises RuntimeError - another process is importing it.
       - "upload_failed" / "broken": deleted and imported again.
 
@@ -195,6 +218,11 @@ def import_model(
     existing = model_registry_status(family, suffix, IMPORTED)
     if existing is not None:
         if existing.phase == _PHASE_READY:
+            _refresh_bundle(serve_app, family, suffix)
+            if requirements is not None:
+                existing.requirements = _set_missing_requirements(
+                    family, suffix, requirements
+                )
             return existing
         if existing.phase == _PHASE_UPLOADING:
             raise RuntimeError(
@@ -202,7 +230,47 @@ def import_model(
                 "another process"
             )
         delete_model(family, suffix, IMPORTED)
-    return _upload_model(source, serve_app, suffix, family, None, IMPORTED)
+    return _upload_model(
+        source, serve_app, suffix, family, None, IMPORTED, requirements
+    )
+
+
+def _refresh_bundle(serve_app: type, family: str, suffix: str) -> None:
+    """Re-bundle an imported model's serve-app when its fingerprint differs
+    from the bundle stored on the version, leaving the weights in place.
+
+    The new bundle is uploaded next to the old one and the version's tags are
+    pointed at it, so the next `deploy_model` runs the new code. An app that is
+    already running keeps the code it started with until it is deployed again;
+    the old bundle stays in storage so that app can still restart."""
+    name = f"{family}__{suffix}"
+    client = MlflowClient(tracking_uri=get_mlflow_tracking_uri())
+    version = client.search_model_versions(
+        f"name='{name}' and tags.run_name='{IMPORTED}'"
+    )[0]
+    serve_bundle = build_bundle(serve_app)
+    if metadata_from_tags(version.tags).fingerprint == serve_bundle.fingerprint:
+        return
+    meta = upload_bundle(serve_bundle, family, suffix, IMPORTED)
+    for key, value in metadata_to_tags(meta).items():
+        client.set_model_version_tag(name, version.version, key, value)
+
+
+def _set_missing_requirements(
+    family: str, suffix: str, requirements: ModelRequirements
+) -> ModelRequirements:
+    """Store `requirements` on an imported model whose version has none yet.
+    Returns the requirements the version holds afterwards."""
+    name = f"{family}__{suffix}"
+    client = MlflowClient(tracking_uri=get_mlflow_tracking_uri())
+    version = client.search_model_versions(
+        f"name='{name}' and tags.run_name='{IMPORTED}'"
+    )[0]
+    if has_requirement_tags(version.tags):
+        return requirements_from_tags(version.tags)
+    for key, value in requirements_to_tags(requirements).items():
+        client.set_model_version_tag(name, version.version, key, value)
+    return requirements
 
 
 def _upload_model(
@@ -212,6 +280,7 @@ def _upload_model(
     family: str,
     run_id: str | None,
     run_name: str,
+    requirements: ModelRequirements | None,
 ) -> SavedModel:
     """Register a ModelVersion in "uploading", resolve `weights` to a directory
     (calling it when it is a callable), upload the weights and the serve-app
@@ -227,7 +296,8 @@ def _upload_model(
     # fetched and uploaded, and a concurrent `import_model` sees the import in
     # flight for the whole download instead of starting one of its own. The
     # size is stamped once the directory exists; the bundle tags and the flip
-    # to "ready" happen only after the upload lands.
+    # to "ready" happen only after the upload lands. No requirements leaves
+    # their tags unset, so a later `import_model` can still store them.
     version = client.create_model_version(
         name=name,
         source=source,
@@ -237,6 +307,7 @@ def _upload_model(
             "suffix": suffix,
             "run_name": run_name,
             _LIFECYCLE_TAG: _PHASE_UPLOADING,
+            **(requirements_to_tags(requirements) if requirements is not None else {}),
         },
     )
     try:
@@ -305,6 +376,23 @@ def model_registry_status(
     return _to_saved_model(versions[0]) if versions else None
 
 
+def set_model_requirements(
+    family: str, suffix: str, run_name: str, requirements: ModelRequirements
+) -> None:
+    """Replace the hardware requirements stored on a model. Takes effect on
+    its next `deploy_model`; a replica already running keeps its placement.
+    Raises ValueError if the model was never registered."""
+    client = MlflowClient(tracking_uri=get_mlflow_tracking_uri())
+    name = f"{family}__{suffix}"
+    versions = client.search_model_versions(
+        f"name='{name}' and tags.run_name='{run_name}'"
+    )
+    if not versions:
+        raise ValueError(f"No model {family}/{suffix}/{run_name}")
+    for key, value in requirements_to_tags(requirements).items():
+        client.set_model_version_tag(name, versions[0].version, key, value)
+
+
 def delete_model(family: str, suffix: str, run_name: str) -> None:
     """Delete the ModelVersion in MLflow, its weights blob, and its serve bundle."""
     client = MlflowClient(tracking_uri=get_mlflow_tracking_uri())

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/docs/cortexgrid/README.md

@@ -137,20 +137,24 @@ app = FastAPI()
 
 @serve.ingress(app)
 class MyServeApp:
-    num_gpus = 1
-
     def __init__(self, family: str, suffix: str, run_name: str) -> None:
         self._weights_dir = cortexgrid.load_model(family, suffix, run_name)
 
     @app.post("/complete")
     async def complete(self, body: dict): ...
 
-saved = cortexgrid.save_model(weights_dir, MyServeApp, family="qwen", suffix="instruct")
+saved = cortexgrid.save_model(
+    weights_dir, MyServeApp, family="qwen", suffix="instruct",
+    # What one replica needs; the model is deployed only on a host that has it.
+    requirements=cortexgrid.ModelRequirements(num_gpus=1, ram_gb=8, vram_gb=16),
+)
 deployed = cortexgrid.deploy_model("qwen", "instruct", saved.run_name, wait=True)
 print(deployed.url)
 ```
 
-`save_model` saves a new copy under every run - meant for weights the run produced (e.g. a fine-tune). For a model produced elsewhere (e.g. a pretrained base model), `cortexgrid.import_model(source, MyServeApp, family, suffix)` uploads it once under `run_name=cortexgrid.IMPORTED` and is a no-op on later runs; deploy it with `deploy_model(family, suffix, cortexgrid.IMPORTED)`.
+The requirements are part of the model, not of the serve-app class: GPUs, RAM and VRAM (GiB, 0 meaning no requirement) are stored with it and matched against what the cluster's hosts have free. Correct them later with `cortexgrid.set_model_requirements(family, suffix, run_name, requirements)` or on the model card in the dashboard; `cortexgrid.deploy_model(..., num_replicas=2)` chooses how many copies to run.
+
+`save_model` saves a new copy under every run - meant for weights the run produced (e.g. a fine-tune). For a model produced elsewhere (e.g. a pretrained base model), `cortexgrid.import_model(source, MyServeApp, family, suffix)` uploads it once under `run_name=cortexgrid.IMPORTED` and on later runs only re-bundles `MyServeApp` if its code changed; deploy it with `deploy_model(family, suffix, cortexgrid.IMPORTED)`.
 
 `save_model` is synchronous (registry lifecycle: `uploading` -> `ready`); `deploy_model` schedules the serving lifecycle (`deploying` -> `running`). With `wait=True` a failed deploy raises `cortexgrid.ModelDeployFailed`; `cortexgrid.wait_for_model_serving(family, suffix, run_name, timeout=...)` waits on a deploy started elsewhere, and re-deploying a failed model retries it from scratch. See [model-serving.md](https://github.com/robodatalab/cortexgrid/blob/main/docs/cortexgrid/model-serving.md) for both lifecycles end to end - upload/deploy/undeploy/delete, status queries (`model_registry_status`, `model_serving_status`), and error handling.
 

{cortexgrid-0.2.95 → cortexgrid-0.2.97}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "cortexgrid"
-version = "0.2.95"
+version = "0.2.97"
 description = "Connect your ML code to the RoboLab compute cluster — Ray, MLflow, and S3"
 readme = "docs/cortexgrid/README.md"
 license = "Apache-2.0"