[WIP] Add support for converting to OCI artifacts

docs/ramalama-convert.1.md

@@ -39,14 +39,15 @@ Image to use when converting to GGUF format (when then `--gguf` option has been
 executable and available in the `PATH`. The script is available from the `llama.cpp` GitHub repo. Defaults to the current
 `quay.io/ramalama/ramalama-rag` image.
 
-#### **--type**=*raw* | *car*
+#### **--type**="artifact" | *raw* | *car*
 
-type of OCI Model Image to convert.
+Convert the MODEL to the specified OCI Object
 
-| Type | Description                                                   |
-| ---- | ------------------------------------------------------------- |
-| car  | Includes base image with the model stored in a /models subdir |
-| raw  | Only the model and a link file model.file to it stored at /   |
+| Type     | Description                                                   |
+| -------- | ------------------------------------------------------------- |
+| artifact | Store AI Models as artifacts                                  |
+| car      | Traditional OCI image including base image with the model stored in a /models subdir |
+| raw      | Traditional OCI image including only the model and a link file `model.file` pointed at it stored at /   |
 
 ## EXAMPLE
 

docs/ramalama.conf

@@ -32,6 +32,14 @@
 #
 #carimage = "registry.access.redhat.com/ubi10-micro:latest"
 
+# Convert the MODEL to the specified OCI Object
+# Options: artifact, car, raw
+#
+# artifact: Store AI Models as artifacts
+# car:      Traditional OCI image including base image with the model stored in a /models subdir
+# raw:      Traditional OCI image including only the model and a link file `model.file` pointed at it stored at /
+#convert_type = "raw"
+
 # Run RamaLama in the default container.
 #
 #container = true

docs/ramalama.conf.5.md

@@ -84,6 +84,18 @@ Min chunk size to attempt reusing from the cache via KV shifting
 Run RamaLama in the default container.
 RAMALAMA_IN_CONTAINER environment variable overrides this field.
 
+#convert_type = "raw"
+
+Convert the MODEL to the specified OCI Object
+Options: artifact, car, raw
+
+| Type     | Description                                                   |
+| -------- | ------------------------------------------------------------- |
+| artifact | Store AI Models as artifacts                                  |
+| car      | Traditional OCI image including base image with the model stored in a /models subdir |
+| raw      | Traditional OCI image including only the model and a link file `model.file` pointed at it stored at /   |
+
+
 **ctx_size**=0
 
 Size of the prompt context (0 = loaded from model)

ramalama/cli.py

@@ -719,11 +719,12 @@ def convert_parser(subparsers):
     )
     parser.add_argument(
         "--type",
-        default="raw",
-        choices=["car", "raw"],
+        default=CONFIG.convert_type,
+        choices=["artifact", "car", "raw"],
         help="""\
 type of OCI Model Image to push.
 
+Model "artifact" stores the AI Model as an OCI Artifact.
 Model "car" includes base image with the model stored in a /models subdir.
 Model "raw" contains the model and a link file model.file to it stored at /.""",
     )
@@ -762,11 +763,12 @@ def push_parser(subparsers):
     add_network_argument(parser)
     parser.add_argument(
         "--type",
-        default="raw",
-        choices=["car", "raw"],
+        default=CONFIG.convert_type,
+        choices=["artifact", "car", "raw"],
         help="""\
 type of OCI Model Image to push.
 
+Model "artifact" stores the AI Model as an OCI Artifact.
 Model "car" includes base image with the model stored in a /models subdir.
 Model "raw" contains the model and a link file model.file to it stored at /.""",
     )
@@ -1424,7 +1426,7 @@ def _rm_model(models, args):
         try:
             m = New(model, args)
             m.remove(args)
-        except KeyError as e:
+        except (KeyError, subprocess.CalledProcessError) as e:
             for prefix in MODEL_TYPES:
                 if model.startswith(prefix + "://"):
                     if not args.ignore:

ramalama/config.py

@@ -136,6 +136,7 @@ class BaseConfig:
     carimage: str = "registry.access.redhat.com/ubi10-micro:latest"
     container: bool = None  # type: ignore
     ctx_size: int = 0
+    convert_type: Literal["artifact", "car", "raw"] = "raw"
     default_image: str = DEFAULT_IMAGE
     default_rag_image: str = DEFAULT_RAG_IMAGE
     dryrun: bool = False

ramalama/oci_tools.py

@@ -8,7 +8,66 @@
 ocilabeltype = "org.containers.type"
 
 
-def engine_supports_manifest_attributes(engine):
+def convert_from_human_readable_size(input) -> float:
+    sizes = [("KB", 1024), ("MB", 1024**2), ("GB", 1024**3), ("TB", 1024**4), ("B", 1)]
+    for unit, size in sizes:
+        if input.endswith(unit) or input.endswith(unit.lower()):
+            return float(input[: -len(unit)]) * size
+
+    return float(input)
+
+
+def list_artifacts(args: EngineArgType):
+    if args.engine == "docker":
+        return []
+
+    conman_args = [
+        args.engine,
+        "artifact",
+        "ls",
+        "--format",
+        (
+            '{"name":"oci://{{ .Repository }}:{{ .Tag }}",\
+            "created":"{{ .CreatedAt }}", \
+            "size":"{{ .Size }}", \
+            "ID":"{{ .Digest }}"},'
+        ),
+    ]
+    output = run_cmd(conman_args).stdout.decode("utf-8").strip()
+    if output == "":
+        return []
+
+    artifacts = json.loads(f"[{output[:-1]}]")
+    models = []
+    for artifact in artifacts:
+        conman_args = [
+            args.engine,
+            "artifact",
+            "inspect",
+            artifact["ID"],
+        ]
+        output = run_cmd(conman_args).stdout.decode("utf-8").strip()
+
+        if output == "":
+            continue
+        inspect = json.loads(output)
+        if "Manifest" not in inspect:
+            continue
+        if "artifactType" not in inspect["Manifest"]:
+            continue
+        if inspect["Manifest"]['artifactType'] != annotations.ArtifactTypeModelManifest:
+            continue
+        models += [
+            {
+                "name": artifact["name"],
+                "modified": artifact["created"],
+                "size": convert_from_human_readable_size(artifact["size"]),
+            }
+        ]
+    return models
+
+
+def engine_supports_manifest_attributes(engine) -> bool:
     if not engine or engine == "" or engine == "docker":
         return False
     if engine == "podman" and engine_version(engine) < "5":
@@ -91,26 +150,26 @@ def list_models(args: EngineArgType):
         "--format",
         formatLine,
     ]
+    models = []
     output = run_cmd(conman_args, env={"TZ": "UTC"}).stdout.decode("utf-8").strip()
-    if output == "":
-        return []
-
-    models = json.loads(f"[{output[:-1]}]")
-    # exclude dangling images having no tag (i.e. <none>:<none>)
-    models = [model for model in models if model["name"] != "oci://<none>:<none>"]
-
-    # Grab the size from the inspect command
-    if conman == "docker":
-        # grab the size from the inspect command
-        for model in models:
-            conman_args = [conman, "image", "inspect", model["id"], "--format", "{{.Size}}"]
-            output = run_cmd(conman_args).stdout.decode("utf-8").strip()
-            # convert the number value from the string output
-            model["size"] = int(output)
-            # drop the id from the model
-            del model["id"]
+    if output != "":
+        models += json.loads(f"[{output[:-1]}]")
+        # exclude dangling images having no tag (i.e. <none>:<none>)
+        models = [model for model in models if model["name"] != "oci://<none>:<none>"]
+
+        # Grab the size from the inspect command
+        if conman == "docker":
+            # grab the size from the inspect command
+            for model in models:
+                conman_args = [conman, "image", "inspect", model["id"], "--format", "{{.Size}}"]
+                output = run_cmd(conman_args).stdout.decode("utf-8").strip()
+                # convert the number value from the string output
+                model["size"] = int(output)
+                # drop the id from the model
+                del model["id"]
 
     models += list_manifests(args)
+    models += list_artifacts(args)
     for model in models:
         # Convert to ISO 8601 format
         parsed_date = datetime.fromisoformat(

ramalama/transports/base.py

@@ -145,6 +145,7 @@ def __init__(self, model: str, model_store_path: str):
         self._model_type: str
         self._model_name, self._model_tag, self._model_organization = self.extract_model_identifiers()
         self._model_type = type(self).__name__.lower()
+        self.artifact = False
 
         self._model_store_path: str = model_store_path
         self._model_store: Optional[ModelStore] = None
@@ -200,6 +201,8 @@ def _get_entry_model_path(self, use_container: bool, should_generate: bool, dry_
 
         if self.model_type == 'oci':
             if use_container or should_generate:
+                if getattr(self, "artifact", False):
+                    return os.path.join(MNT_DIR, self.artifact_name())
                 return os.path.join(MNT_DIR, 'model.file')
             else:
                 return f"oci://{self.model}"
@@ -345,9 +348,10 @@ def exec_model_in_container(self, cmd_args, args):
     def setup_mounts(self, args):
         if args.dryrun:
             return
+
         if self.model_type == 'oci':
             if self.engine.use_podman:
-                mount_cmd = f"--mount=type=image,src={self.model},destination={MNT_DIR},subpath=/models,rw=false"
+                mount_cmd = self.mount_cmd()
             elif self.engine.use_docker:
                 output_filename = self._get_entry_model_path(args.container, True, args.dryrun)
                 volume = populate_volume_from_image(self, os.path.basename(output_filename))
@@ -651,40 +655,48 @@ def inspect(
         as_json: bool = False,
         dryrun: bool = False,
     ) -> None:
+        print(self.get_inspect(show_all, show_all_metadata, get_field, dryrun, as_json))
+
+    def get_inspect(
+        self,
+        show_all: bool = False,
+        show_all_metadata: bool = False,
+        get_field: str = "",
+        dryrun: bool = False,
+        as_json: bool = False,
+    ) -> Any:
         model_name = self.filename
         model_registry = self.type.lower()
         model_path = self._get_inspect_model_path(dryrun)
-
         if GGUFInfoParser.is_model_gguf(model_path):
             if not show_all_metadata and get_field == "":
                 gguf_info: GGUFModelInfo = GGUFInfoParser.parse(model_name, model_registry, model_path)
-                print(gguf_info.serialize(json=as_json, all=show_all))
-                return
+                return gguf_info.serialize(json=as_json, all=show_all)
 
             metadata = GGUFInfoParser.parse_metadata(model_path)
             if show_all_metadata:
-                print(metadata.serialize(json=as_json))
-                return
+                return metadata.serialize(json=as_json)
             elif get_field != "":  # If a specific field is requested, print only that field
                 field_value = metadata.get(get_field)
                 if field_value is None:
                     raise KeyError(f"Field '{get_field}' not found in GGUF model metadata")
-                print(field_value)
-                return
+                return field_value
 
         if SafetensorInfoParser.is_model_safetensor(model_name):
             safetensor_info: SafetensorModelInfo = SafetensorInfoParser.parse(model_name, model_registry, model_path)
-            print(safetensor_info.serialize(json=as_json, all=show_all))
-            return
+            return safetensor_info.serialize(json=as_json, all=show_all)
 
-        print(ModelInfoBase(model_name, model_registry, model_path).serialize(json=as_json))
+        return ModelInfoBase(model_name, model_registry, model_path).serialize(json=as_json)
 
-    def print_pull_message(self, model_name):
+    def print_pull_message(self, model_name) -> None:
         model_name = trim_model_name(model_name)
         # Write messages to stderr
         perror(f"Downloading {model_name} ...")
         perror(f"Trying to pull {model_name} ...")
 
+    def is_artifact(self) -> bool:
+        return False
+
 
 def compute_ports(exclude: list[str] | None = None) -> list[int]:
     exclude = exclude and set(map(int, exclude)) or set()