support model config

Author: wsmlby

cli/homl_cli/daemon_pb2.py

@@ -65,6 +65,21 @@ def __init__(self, channel):
                 request_serializer=daemon__pb2.VersionRequest.SerializeToString,
                 response_deserializer=daemon__pb2.VersionResponse.FromString,
                 _registered_method=True)
+        self.ConfigModel = channel.unary_unary(
+                '/homl.Daemon/ConfigModel',
+                request_serializer=daemon__pb2.ModelSettingsRequest.SerializeToString,
+                response_deserializer=daemon__pb2.ModelConfigResponse.FromString,
+                _registered_method=True)
+        self.ConfigModelParam = channel.unary_unary(
+                '/homl.Daemon/ConfigModelParam',
+                request_serializer=daemon__pb2.ModelParamRequest.SerializeToString,
+                response_deserializer=daemon__pb2.ModelConfigResponse.FromString,
+                _registered_method=True)
+        self.GetModelConfig = channel.unary_unary(
+                '/homl.Daemon/GetModelConfig',
+                request_serializer=daemon__pb2.ModelInfoRequest.SerializeToString,
+                response_deserializer=daemon__pb2.ModelConfigResponse.FromString,
+                _registered_method=True)
 
 
 class DaemonServicer(object):
@@ -112,6 +127,24 @@ def Version(self, request, context):
         context.set_details('Method not implemented!')
         raise NotImplementedError('Method not implemented!')
 
+    def ConfigModel(self, request, context):
+        """Missing associated documentation comment in .proto file."""
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details('Method not implemented!')
+        raise NotImplementedError('Method not implemented!')
+
+    def ConfigModelParam(self, request, context):
+        """Missing associated documentation comment in .proto file."""
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details('Method not implemented!')
+        raise NotImplementedError('Method not implemented!')
+
+    def GetModelConfig(self, request, context):
+        """Missing associated documentation comment in .proto file."""
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details('Method not implemented!')
+        raise NotImplementedError('Method not implemented!')
+
 
 def add_DaemonServicer_to_server(servicer, server):
     rpc_method_handlers = {
@@ -145,6 +178,21 @@ def add_DaemonServicer_to_server(servicer, server):
                     request_deserializer=daemon__pb2.VersionRequest.FromString,
                     response_serializer=daemon__pb2.VersionResponse.SerializeToString,
             ),
+            'ConfigModel': grpc.unary_unary_rpc_method_handler(
+                    servicer.ConfigModel,
+                    request_deserializer=daemon__pb2.ModelSettingsRequest.FromString,
+                    response_serializer=daemon__pb2.ModelConfigResponse.SerializeToString,
+            ),
+            'ConfigModelParam': grpc.unary_unary_rpc_method_handler(
+                    servicer.ConfigModelParam,
+                    request_deserializer=daemon__pb2.ModelParamRequest.FromString,
+                    response_serializer=daemon__pb2.ModelConfigResponse.SerializeToString,
+            ),
+            'GetModelConfig': grpc.unary_unary_rpc_method_handler(
+                    servicer.GetModelConfig,
+                    request_deserializer=daemon__pb2.ModelInfoRequest.FromString,
+                    response_serializer=daemon__pb2.ModelConfigResponse.SerializeToString,
+            ),
     }
     generic_handler = grpc.method_handlers_generic_handler(
             'homl.Daemon', rpc_method_handlers)
@@ -318,3 +366,84 @@ def Version(request,
             timeout,
             metadata,
             _registered_method=True)
+
+    @staticmethod
+    def ConfigModel(request,
+            target,
+            options=(),
+            channel_credentials=None,
+            call_credentials=None,
+            insecure=False,
+            compression=None,
+            wait_for_ready=None,
+            timeout=None,
+            metadata=None):
+        return grpc.experimental.unary_unary(
+            request,
+            target,
+            '/homl.Daemon/ConfigModel',
+            daemon__pb2.ModelSettingsRequest.SerializeToString,
+            daemon__pb2.ModelConfigResponse.FromString,
+            options,
+            channel_credentials,
+            insecure,
+            call_credentials,
+            compression,
+            wait_for_ready,
+            timeout,
+            metadata,
+            _registered_method=True)
+
+    @staticmethod
+    def ConfigModelParam(request,
+            target,
+            options=(),
+            channel_credentials=None,
+            call_credentials=None,
+            insecure=False,
+            compression=None,
+            wait_for_ready=None,
+            timeout=None,
+            metadata=None):
+        return grpc.experimental.unary_unary(
+            request,
+            target,
+            '/homl.Daemon/ConfigModelParam',
+            daemon__pb2.ModelParamRequest.SerializeToString,
+            daemon__pb2.ModelConfigResponse.FromString,
+            options,
+            channel_credentials,
+            insecure,
+            call_credentials,
+            compression,
+            wait_for_ready,
+            timeout,
+            metadata,
+            _registered_method=True)
+
+    @staticmethod
+    def GetModelConfig(request,
+            target,
+            options=(),
+            channel_credentials=None,
+            call_credentials=None,
+            insecure=False,
+            compression=None,
+            wait_for_ready=None,
+            timeout=None,
+            metadata=None):
+        return grpc.experimental.unary_unary(
+            request,
+            target,
+            '/homl.Daemon/GetModelConfig',
+            daemon__pb2.ModelInfoRequest.SerializeToString,
+            daemon__pb2.ModelConfigResponse.FromString,
+            options,
+            channel_credentials,
+            insecure,
+            call_credentials,
+            compression,
+            wait_for_ready,
+            timeout,
+            metadata,
+            _registered_method=True)

cli/homl_cli/daemon_pb2_grpc.py

@@ -80,6 +80,31 @@ def config_list():
             # Do not print credentials or sensitive values
             click.echo(f"    Current value: {value}")
 
+@config_cli.command("model")
+@click.argument("model_name", required=True)
+@click.option("--params", help="launch params")
+@click.option("--get", is_flag=True, help="Get model configuration")
+def config_model(model_name, params, get):
+    """Manage model-specific configuration."""
+    stub = get_client_stub()
+    if not stub:
+        click.secho("HoML server is not running.", fg="red")
+        return
+    if get:
+        """Get model configuration."""
+        
+        response = stub.GetModelConfig(daemon_pb2.ModelInfoRequest(model_name=model_name))
+    elif params:
+        params = params.split(" ")
+        response = stub.ConfigModelParam(daemon_pb2.ModelParamRequest(
+            model_name=model_name,
+            params=daemon_pb2.ModelParam(params=params)
+        ))
+    click.echo(f"Model settings for '{model_name}': {response.params.params}: \"{' '.join(response.params.params)}\"")
+
+
+
+
 
 main.add_command(config_cli, name="config")
 
@@ -211,7 +236,8 @@ def stop(model_name):
 
 @main.command()
 @click.argument('model_name')
-def pull(model_name):
+@click.option('--config', 'refresh_config', is_flag=True, help="Refresh the model config. Local config will be overridden if conflict with the model default.")
+def pull(model_name, refresh_config):
     """Pulls a model from a registry."""
     stub = get_client_stub()
     if "gptoss" in model_name.lower():
@@ -235,7 +261,7 @@ def pull(model_name):
         click.echo(f"Pulling model '{model_name}'...")
         try:
             responses = stub.PullModel(daemon_pb2.PullModelRequest(
-                model_name=model_name, hf_token=hf_token))
+                model_name=model_name, hf_token=hf_token, refresh_config=refresh_config))
             for resp in responses:
                 msg = resp.message
                 percent = resp.percent

cli/homl_cli/main.py

@@ -24,11 +24,38 @@ service Daemon {
   rpc ListLocalModels (ListLocalModelsRequest) returns (ListLocalModelsResponse) {}
 
   rpc Version (VersionRequest) returns (VersionResponse) {}
+
+  rpc ConfigModel(ModelSettingsRequest) returns (ModelConfigResponse) {}
+  rpc ConfigModelParam(ModelParamRequest) returns (ModelConfigResponse) {}
+  rpc GetModelConfig(ModelInfoRequest) returns (ModelConfigResponse) {}
 }
 
 
 
 // --- Messages ---
+message ModelSettings {
+  map<string, string> settings = 1;
+}
+message ModelParam {
+  repeated string params = 1;
+}
+message ModelConfigResponse {
+  ModelSettings settings = 1;
+  ModelParam params = 2;
+}
+message ModelSettingsRequest {
+  string model_name = 1;
+  ModelSettings settings = 2;
+}
+message ModelParamRequest {
+  string model_name = 1;
+  ModelParam params = 2;
+}
+
+message ModelInfoRequest {
+  string model_name = 1;
+}
+
 message VersionRequest {
 }
 message VersionResponse {
@@ -38,6 +65,7 @@ message VersionResponse {
 message PullModelRequest {
   string model_name = 1;
   string hf_token = 2;
+  bool refresh_config = 3;
 }
 
 message PullModelResponse {
@@ -55,6 +83,8 @@ message PullModelProgress {
 message StartModelRequest {
   string model_name = 1;
   bool eager_mode = 2;
+  ModelParam params = 3;
+  bool use_params = 4;
 }
 
 message StartModelResponse {