model-compose 0.4.4__tar.gz → 0.4.6__tar.gz

{model_compose-0.4.4 → model_compose-0.4.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-compose
-Version: 0.4.4
+Version: 0.4.6
 Summary: model-compose: Declarative AI Model and Workflow Orchestrator
 Author-email: Hanyeol Cho <hanyeol.cho@gmail.com>
 License-Expression: MIT

{model_compose-0.4.4 → model_compose-0.4.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "model-compose"
-version = "0.4.4"
+version = "0.4.6"
 description = "model-compose: Declarative AI Model and Workflow Orchestrator"
 authors = [
     { name = "Hanyeol Cho", email = "hanyeol.cho@gmail.com" }

{model_compose-0.4.4 → model_compose-0.4.6}/src/mindor/core/component/services/mcp_server.py

@@ -5,6 +5,7 @@ from mindor.core.utils.mcp_client import McpClient, ContentBlock, TextContent, I
 from mindor.core.utils.shell import run_command_streaming
 from ..base import ComponentService, ComponentType, ComponentGlobalConfigs, register_component
 from ..context import ComponentActionContext
+import asyncio
 
 class McpServerAction:
     def __init__(self, config: McpServerActionConfig):
@@ -61,10 +62,19 @@ class McpServerComponent(ComponentService):
 
     async def _serve(self) -> None:
         if self.config.manage.scripts.start:
-            await run_command_streaming(self.config.manage.scripts.start, self.config.manage.working_dir, self.config.manage.env, block=False)
+            await run_command_streaming(self.config.manage.scripts.start, self.config.manage.working_dir, self.config.manage.env)
 
     async def _shutdown(self) -> None:
         pass
 
+    async def _is_ready(self) -> bool:
+        try:
+            _, writer = await asyncio.open_connection("localhost", self.config.port)
+            writer.close()
+            await writer.wait_closed()
+            return True
+        except (ConnectionRefusedError, OSError):
+            return False
+
     async def _run(self, action: ActionConfig, context: ComponentActionContext) -> Any:
         return await McpServerAction(action).run(context, self.client)

model_compose-0.4.6/src/mindor/core/component/services/model/base/__init__.py

@@ -0,0 +1,3 @@
+from .common import *
+from .huggingface import *
+from .unsloth import *

model_compose-0.4.6/src/mindor/core/component/services/model/base/common.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING
+
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Callable, Mapping, Any
+from abc import ABC, abstractmethod
+from mindor.dsl.schema.component import ModelComponentConfig, ModelTaskType, ModelDriver, HuggingfaceModelConfig, LocalModelConfig
+from mindor.dsl.schema.action import ModelActionConfig
+from mindor.core.services import AsyncService
+from mindor.core.logger import logging
+from ....context import ComponentActionContext
+import asyncio
+
+if TYPE_CHECKING:
+    import torch
+
+class ModelTaskService(AsyncService):
+    def __init__(self, id: str, config: ModelComponentConfig, daemon: bool):
+        super().__init__(daemon)
+
+        self.id: str = id
+        self.config: ModelComponentConfig = config
+
+    def get_setup_requirements(self) -> Optional[List[str]]:
+        return [ "torch" ]
+
+    async def run(self, action: ModelActionConfig, context: ComponentActionContext) -> Any:
+        loop: asyncio.AbstractEventLoop = asyncio.get_running_loop()
+
+        async def _run():
+            return await self._run(action, context, loop)
+
+        return await self.run_in_thread(_run)
+
+    @abstractmethod
+    async def _run(self, action: ModelActionConfig, context: ComponentActionContext, loop: asyncio.AbstractEventLoop) -> Any:
+        pass
+
+    def _get_model_path(self) -> str:
+        if isinstance(self.config.model, HuggingfaceModelConfig):
+            from .huggingface_hub import get_model_path
+            return get_model_path(self.config.model)
+
+        if isinstance(self.config.model, LocalModelConfig):
+            return self.config.model.path
+        
+        if isinstance(self.config.model, str):
+            return self.config.model
+
+        raise ValueError(f"Unknown model config type: {type(self.config.model)}")
+
+    def _resolve_device(self) -> torch.device:
+        import torch
+
+        try:
+            return torch.device(self.config.device)
+        except:
+            logging.warning(f"Invalid device '{self.config.device}', falling back to 'cpu'")
+        
+        return torch.device("cpu")
+
+    def _load_model_checkpoint(self, model: torch.nn.Module, model_path: str) -> None:
+        checkpoint = torch.load(model_path, map_location="cpu")
+        state_dict = self._get_state_dict_from_checkpoint(checkpoint)
+
+        model.load_state_dict(state_dict, strict=True)
+
+    def _get_state_dict_from_checkpoint(self, checkpoint: Any) -> Mapping[str, Any]:
+        for key in [ "params", "state_dict" ]:
+            if key in checkpoint:
+                return checkpoint[key]
+        return checkpoint
+
+def register_model_task_service(type: ModelTaskType, driver: ModelDriver):
+    def decorator(cls: Type[ModelTaskService]) -> Type[ModelTaskService]:
+        if type not in ModelTaskServiceRegistry:
+            ModelTaskServiceRegistry[type] = {}
+        ModelTaskServiceRegistry[type][driver] = cls
+        return cls
+    return decorator
+
+ModelTaskServiceRegistry: Dict[ModelTaskType, Dict[ModelDriver, Type[ModelTaskService]]] = {}

model_compose-0.4.6/src/mindor/core/component/services/model/base/huggingface.py

@@ -0,0 +1,171 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING
+
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Callable, Any
+from mindor.dsl.schema.component import ModelComponentConfig, HuggingfaceModelConfig, DeviceMode
+from mindor.core.logger import logging
+from .common import ModelTaskService
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
+    import torch
+
+class HuggingfaceModelTaskService(ModelTaskService):
+    def __init__(self, id: str, config: ModelComponentConfig, daemon: bool):
+        super().__init__(id, config, daemon)
+
+    def _load_pretrained_model(self) -> PreTrainedModel:
+        model_cls = self._get_model_class()
+        model = model_cls.from_pretrained(self.config.model, **self._get_model_params())
+
+        if self.config.device_mode == DeviceMode.SINGLE:
+            model = model.to(torch.device(self.config.device))
+
+        return model
+
+    def _get_model_class(self) -> Type[PreTrainedModel]:
+        raise NotImplementedError("Model class loader not implemented.")
+
+    def _get_model_params(self) -> Dict[str, Any]:
+        params: Dict[str, Any] = {}
+
+        if isinstance(self.config.model, HuggingfaceModelConfig):
+            if self.config.model.revision:
+                params["revision"] = self.config.model.revision
+            
+            if self.config.model.cache_dir:
+                params["cache_dir"] = self.config.model.cache_dir
+
+            if self.config.model.local_files_only:
+                params["local_files_only"] = True
+
+        if self.config.device_mode != DeviceMode.SINGLE:
+            params["device_map"] = self.config.device_mode.value
+    
+        if self.config.precision is not None:
+            params["torch_dtype"] = getattr(torch, self.config.precision.value)
+    
+        if self.config.low_cpu_mem_usage:
+            params["low_cpu_mem_usage"] = True
+
+        return params
+
+    def _get_model_device(self, model: PreTrainedModel) -> torch.device:
+        return next(model.parameters()).device
+
+class HuggingfaceLanguageModelTaskService(HuggingfaceModelTaskService):
+    def __init__(self, id: str, config: ModelComponentConfig, daemon: bool):
+        super().__init__(id, config, daemon)
+
+        self.model: Optional[PreTrainedModel] = None
+        self.tokenizer: Optional[PreTrainedTokenizer] = None
+        self.device: Optional[torch.device] = None
+
+    def get_setup_requirements(self) -> Optional[List[str]]:
+        return [ 
+            "transformers>=4.21.0",
+            "torch",
+            "sentencepiece",
+            "accelerate"
+        ]
+
+    async def _serve(self) -> None:
+        try:
+            self.model = self._load_pretrained_model()
+            self.tokenizer = self._load_pretrained_tokenizer()
+            self.device = self._get_model_device(self.model)
+            logging.info(f"Model and tokenizer loaded successfully on device '{self.device}': {self.config.model}")
+        except Exception as e:
+            logging.error(f"Failed to load model '{self.config.model}': {e}")
+            raise
+
+    async def _shutdown(self) -> None:
+        self.model = None
+        self.tokenizer = None
+        self.device = None
+
+    def _load_pretrained_tokenizer(self) -> Optional[PreTrainedTokenizer]:
+        tokenizer_cls = self._get_tokenizer_class()
+
+        if not tokenizer_cls:
+            return None
+
+        return tokenizer_cls.from_pretrained(self.config.model, **self._get_tokenizer_params())
+
+    def _get_tokenizer_class(self) -> Optional[Type[PreTrainedTokenizer]]:
+        return None
+
+    def _get_tokenizer_params(self) -> Dict[str, Any]:
+        params: Dict[str, Any] = {}
+
+        if isinstance(self.config.model, HuggingfaceModelConfig):
+            if self.config.model.revision:
+                params["revision"] = self.config.model.revision
+
+            if self.config.model.cache_dir:
+                params["cache_dir"] = self.config.model.cache_dir
+
+            if self.config.model.local_files_only:
+                params["local_files_only"] = True
+
+        if not self.config.fast_tokenizer:
+            params["use_fast"] = False
+
+        return params
+
+class HuggingfaceMultimodalModelTaskService(HuggingfaceModelTaskService):
+    def __init__(self, id: str, config: ModelComponentConfig, daemon: bool):
+        super().__init__(id, config, daemon)
+
+        self.model: Optional[PreTrainedModel] = None
+        self.processor: Optional[ProcessorMixin] = None
+        self.device: Optional[torch.device] = None
+
+    def get_setup_requirements(self) -> Optional[List[str]]:
+        return [ 
+            "transformers>=4.21.0",
+            "torch",
+            "sentencepiece",
+            "accelerate"
+        ]
+
+    async def _serve(self) -> None:
+        try:
+            self.model = self._load_pretrained_model()
+            self.processor = self._load_pretrained_processor()
+            self.device = self._get_model_device(self.model)
+            logging.info(f"Model and processor loaded successfully on device '{self.device}': {self.config.model}")
+        except Exception as e:
+            logging.error(f"Failed to load model '{self.config.model}': {e}")
+            raise
+
+    async def _shutdown(self) -> None:
+        self.model = None
+        self.processor = None
+        self.device = None
+
+    def _load_pretrained_processor(self) -> Optional[ProcessorMixin]:
+        processor_cls = self._get_processor_class()
+
+        if not processor_cls:
+            return None
+
+        return processor_cls.from_pretrained(self.config.model, **self._get_processor_params())
+
+    def _get_processor_class(self) -> Optional[Type[ProcessorMixin]]:
+        return None
+
+    def _get_processor_params(self) -> Dict[str, Any]:
+        params: Dict[str, Any] = {}
+
+        if isinstance(self.config.model, HuggingfaceModelConfig):
+            if self.config.model.revision:
+                params["revision"] = self.config.model.revision
+ 
+            if self.config.model.cache_dir:
+                params["cache_dir"] = self.config.model.cache_dir
+
+            if self.config.model.local_files_only:
+                params["local_files_only"] = True
+
+        return params

model_compose-0.4.6/src/mindor/core/component/services/model/base/huggingface_hub.py

@@ -0,0 +1,23 @@
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Callable, Any
+from mindor.dsl.schema.component import HuggingfaceModelConfig
+
+def get_model_path(model: HuggingfaceModelConfig) -> Optional[str]:
+    if model.filename:
+        from huggingface_hub import hf_hub_download
+        return hf_hub_download(
+            repo_id=model.repository,
+            filename=model.filename,
+            revision=model.revision,
+            cache_dir=model.cache_dir,
+            token=model.token,
+            local_files_only=model.local_files_only
+        )
+    else:
+        from huggingface_hub import snapshot_download
+        return snapshot_download(
+            repo_id=model.repository,
+            revision=model.revision,
+            cache_dir=model.cache_dir,
+            token=model.token,
+            local_files_only=model.local_files_only
+        )

model_compose-0.4.6/src/mindor/core/component/services/model/base/unsloth.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING
+
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Callable, Any
+from mindor.dsl.schema.component import ModelComponentConfig, DeviceMode
+from mindor.core.logger import logging
+from .common import ModelTaskService
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+class UnslothModelTaskService(ModelTaskService):
+    def __init__(self, id: str, config: ModelComponentConfig, daemon: bool):
+        super().__init__(id, config, daemon)
+
+        self.model: Optional[PreTrainedModel] = None
+        self.tokenizer: Optional[PreTrainedTokenizer] = None
+
+    def get_setup_requirements(self) -> Optional[List[str]]:
+        return [ 
+            "unsloth",
+            "torch"
+        ]
+
+    async def _serve(self) -> None:
+        try:
+            self.model, self.tokenizer = self._load_pretrained_model()
+            logging.info(f"Model and tokenizer loaded successfully: {self.config.model}")
+        except Exception as e:
+            logging.error(f"Failed to load model '{self.config.model}': {e}")
+            raise
+
+    async def _shutdown(self) -> None:
+        self.model = None
+        self.tokenizer = None
+        self.device = None
+
+    def _load_pretrained_model(self) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
+        from unsloth import FastLanguageModel
+        
+        model = FastLanguageModel.from_pretrained(self.config.model, **self._get_model_params())
+
+        return model
+
+    def _get_model_params(self) -> Dict[str, Any]:
+        params: Dict[str, Any] = {}
+
+        if self.config.device_mode != DeviceMode.SINGLE:
+            params["device_map"] = self.config.device_mode.value
+    
+        if self.config.precision is not None:
+            params["dtype"] = self.config.precision.value
+
+        return params

{model_compose-0.4.4 → model_compose-0.4.6}/src/mindor/core/component/services/model/model.py

@@ -1,5 +1,5 @@
 from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, AsyncIterator, Any
-from mindor.dsl.schema.component import ModelComponentConfig, ModelTaskType
+from mindor.dsl.schema.component import ModelComponentConfig, ModelTaskType, ModelDriver
 from mindor.dsl.schema.action import ActionConfig, ModelActionConfig
 from ...base import ComponentService, ComponentType, ComponentGlobalConfigs, register_component
 from ...context import ComponentActionContext
@@ -17,26 +17,18 @@ class ModelComponent(ComponentService):
     def __init__(self, id: str, config: ModelComponentConfig, global_configs: ComponentGlobalConfigs, daemon: bool):
         super().__init__(id, config, global_configs, daemon)
 
-        self.task_service: ModelTaskService = self._create_task_service(self.config.task)
+        self.task_service: ModelTaskService = self._create_task_service(self.config.task, self.config.driver)
 
-    def _create_task_service(self, type: ModelTaskType) -> ModelTaskService:
+    def _create_task_service(self, type: ModelTaskType, driver: ModelDriver) -> ModelTaskService:
         try:
             if not ModelTaskServiceRegistry:
                 from . import tasks
-            return ModelTaskServiceRegistry[type](self.id, self.config, self.daemon)
+            return ModelTaskServiceRegistry[type][driver](self.id, self.config, self.daemon)
         except KeyError:
-            raise ValueError(f"Unsupported model task type: {type}")
+            raise ValueError(f"Unsupported model task type: {type} on {driver}")
 
     def _get_setup_requirements(self) -> Optional[List[str]]:
-        task_dependencies = self.task_service.get_setup_requirements()
-
-        return [ 
-            "transformers>=4.21.0",
-            "torch",
-            "sentencepiece",
-            "accelerate",
-            *(task_dependencies or [])
-        ]
+        return self.task_service.get_setup_requirements()
 
     async def _serve(self) -> None:
         await self.task_service.start()

{model_compose-0.4.4 → model_compose-0.4.6}/src/mindor/core/component/services/model/tasks/__init__.py

@@ -1,6 +1,6 @@
 from .text_generation import *
-from .summarization import *
-from .translation import *
+from .chat_completion import *
 from .text_classification import *
 from .text_embedding import *
 from .image_to_text import *
+from .image_upscale import *

model_compose-0.4.6/src/mindor/core/component/services/model/tasks/chat_completion/__init__.py

@@ -0,0 +1 @@
+from .huggingface import *

model_compose-0.4.6/src/mindor/core/component/services/model/tasks/chat_completion/huggingface.py

@@ -0,0 +1,41 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING
+
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Any
+from mindor.dsl.schema.action import ModelActionConfig, ChatCompletionModelActionConfig
+from ...base import ModelTaskType, ModelDriver, register_model_task_service
+from ...base import HuggingfaceLanguageModelTaskService, ComponentActionContext
+from ..text_generation import HuggingfaceTextGenerationTaskAction
+import asyncio
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+    import torch
+
+class HuggingfaceChatCompletionTaskAction(HuggingfaceTextGenerationTaskAction):
+    def __init__(self, config: ChatCompletionModelActionConfig, model: PreTrainedModel, tokenizer: PreTrainedTokenizer, device: torch.device):
+        super().__init__(config, model, tokenizer, device)
+
+        self.config: ChatCompletionModelActionConfig = config # For type only
+
+    async def _prepare_input(self, context: ComponentActionContext) -> Union[str, List[str]]:
+        messages = await context.render_variable(self.config.messages)
+
+        return self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+
+@register_model_task_service(ModelTaskType.CHAT_COMPLETION, ModelDriver.HUGGINGFACE)
+class HuggingfaceChatCompletionTaskService(HuggingfaceLanguageModelTaskService):
+    async def _run(self, action: ModelActionConfig, context: ComponentActionContext, loop: asyncio.AbstractEventLoop) -> Any:
+        return await HuggingfaceChatCompletionTaskAction(action, self.model, self.tokenizer, self.device).run(context, loop)
+
+    def _get_model_class(self) -> Type[PreTrainedModel]:
+        from transformers import AutoModelForCausalLM
+        return AutoModelForCausalLM
+
+    def _get_tokenizer_class(self) -> Type[PreTrainedTokenizer]:
+        from transformers import AutoTokenizer
+        return AutoTokenizer

model_compose-0.4.6/src/mindor/core/component/services/model/tasks/image_to_text/__init__.py

@@ -0,0 +1 @@
+from .huggingface import *