model-compose 0.4.64__tar.gz → 0.4.66__tar.gz

{model_compose-0.4.64 → model_compose-0.4.66}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-compose
-Version: 0.4.64
+Version: 0.4.66
 Summary: model-compose: Declarative AI Workflow Orchestrator
 Author-email: Hanyeol Cho <hanyeol.cho@gmail.com>
 License-Expression: MIT
@@ -61,14 +61,47 @@ Inspired by `docker-compose` — one YAML file defines your entire AI system.
 
 ---
 
+## Philosophy
+
+AI systems should not be locked into a single provider, runtime, or cloud. They should remain portable, inspectable, and able to run anywhere.
+
+Today, many AI applications are tightly coupled to provider-specific APIs, managed runtimes, and closed ecosystems. While convenient at first, this coupling introduces vendor lock-in — components can't be swapped without rewriting, systems can't move between environments, and teams are forced to choose between cloud convenience and local control.
+
+**model-compose** takes a fundamentally different approach based on three core principles:
+
+* **Composable** — Models, agents, workflows, tools, memory, and protocols are treated as modular, interchangeable building blocks.
+
+* **Portable** — Define your AI system once, then deploy it locally, in containers, or across distributed production environments without re-engineering the core architecture.
+
+* **Hybrid-First** — Bridge cloud APIs and local models on your own terms. Swap infrastructure layers seamlessly to optimize for privacy, latency, or cost without changing how your system behaves.
+
+The goal of model-compose is not to build another closed platform, but to restore architectural autonomy to developers.
+
+---
+
+## Why model-compose?
+
+| Feature | Managed APIs (OpenAI, etc.) | Code Frameworks (LangChain, etc.) | **model-compose** |
+|---|---|---|---|
+| **Provider Coupling** | Single provider per SDK | Multi-provider via abstractions | **Multi-provider via config** |
+| **Code Coupling** | Application code required | Framework-specific code required | **Declarative YAML — no application code** |
+| **Infrastructure Control** | Provider-controlled | Heavy Abstraction | **Full Sovereignty** |
+| **Runtime Flexibility** | Cloud Only | Complex to customize | **Hybrid-First (Local + Cloud)** |
+| **Protocol Support** | Provider-specific | Limited | **HTTP / WebSocket / MCP** |
+| **Deployment** | Provider-managed | Manual integration | **Docker / Native / Process** |
+
+---
+
 ## Highlights
 
-- **Any model, anywhere** — run models locally via HuggingFace, vLLM, and llama.cpp, or connect to OpenAI, Anthropic, Google, and more
-- **20+ components ready** — models, agents, HTTP clients, vector/graph stores, shell commands, and more
-- **Built-in data stores** — Chroma, FAISS, Milvus, Qdrant, Neo4j, ArangoDB, Redis
-- **Deploy as container** — Docker, native containers, or standalone process with one config
+- **Any model, anywhere** — run models locally via HuggingFace, vLLM, or llama.cpp for privacy, offline use, or zero API cost — or connect to OpenAI, Anthropic, Google, and more
+- **AI agents in YAML** — build autonomous agents with tool use, planning, and multi-step reasoning — all declarative
+- **Human-in-the-loop** — workflows can pause for approval gates, user input, or manual review before continuing
+- **Real-time streaming** — built-in SSE streaming for live AI responses from any provider or local model
+- **20+ components ready** — models, agents, HTTP/WebSocket clients, vector/graph stores, shell commands, and more
+- **Deploy as container** — same YAML runs as a Docker container, native process, or standalone service — switch runtime with one line
 - **Serve any protocol** — HTTP REST, WebSocket, or MCP with one line change
-- **Distributed execution** — scale across machines with Redis-backed queue dispatch
+- **Distributed execution** — dispatch workflows to remote workers via Redis queues — scale horizontally by adding servers
 - **Instant Web UI** — add a Gradio-powered interface with 2 lines of YAML
 
 ---

{model_compose-0.4.64 → model_compose-0.4.66}/README.md

@@ -30,14 +30,47 @@ Inspired by `docker-compose` — one YAML file defines your entire AI system.
 
 ---
 
+## Philosophy
+
+AI systems should not be locked into a single provider, runtime, or cloud. They should remain portable, inspectable, and able to run anywhere.
+
+Today, many AI applications are tightly coupled to provider-specific APIs, managed runtimes, and closed ecosystems. While convenient at first, this coupling introduces vendor lock-in — components can't be swapped without rewriting, systems can't move between environments, and teams are forced to choose between cloud convenience and local control.
+
+**model-compose** takes a fundamentally different approach based on three core principles:
+
+* **Composable** — Models, agents, workflows, tools, memory, and protocols are treated as modular, interchangeable building blocks.
+
+* **Portable** — Define your AI system once, then deploy it locally, in containers, or across distributed production environments without re-engineering the core architecture.
+
+* **Hybrid-First** — Bridge cloud APIs and local models on your own terms. Swap infrastructure layers seamlessly to optimize for privacy, latency, or cost without changing how your system behaves.
+
+The goal of model-compose is not to build another closed platform, but to restore architectural autonomy to developers.
+
+---
+
+## Why model-compose?
+
+| Feature | Managed APIs (OpenAI, etc.) | Code Frameworks (LangChain, etc.) | **model-compose** |
+|---|---|---|---|
+| **Provider Coupling** | Single provider per SDK | Multi-provider via abstractions | **Multi-provider via config** |
+| **Code Coupling** | Application code required | Framework-specific code required | **Declarative YAML — no application code** |
+| **Infrastructure Control** | Provider-controlled | Heavy Abstraction | **Full Sovereignty** |
+| **Runtime Flexibility** | Cloud Only | Complex to customize | **Hybrid-First (Local + Cloud)** |
+| **Protocol Support** | Provider-specific | Limited | **HTTP / WebSocket / MCP** |
+| **Deployment** | Provider-managed | Manual integration | **Docker / Native / Process** |
+
+---
+
 ## Highlights
 
-- **Any model, anywhere** — run models locally via HuggingFace, vLLM, and llama.cpp, or connect to OpenAI, Anthropic, Google, and more
-- **20+ components ready** — models, agents, HTTP clients, vector/graph stores, shell commands, and more
-- **Built-in data stores** — Chroma, FAISS, Milvus, Qdrant, Neo4j, ArangoDB, Redis
-- **Deploy as container** — Docker, native containers, or standalone process with one config
+- **Any model, anywhere** — run models locally via HuggingFace, vLLM, or llama.cpp for privacy, offline use, or zero API cost — or connect to OpenAI, Anthropic, Google, and more
+- **AI agents in YAML** — build autonomous agents with tool use, planning, and multi-step reasoning — all declarative
+- **Human-in-the-loop** — workflows can pause for approval gates, user input, or manual review before continuing
+- **Real-time streaming** — built-in SSE streaming for live AI responses from any provider or local model
+- **20+ components ready** — models, agents, HTTP/WebSocket clients, vector/graph stores, shell commands, and more
+- **Deploy as container** — same YAML runs as a Docker container, native process, or standalone service — switch runtime with one line
 - **Serve any protocol** — HTTP REST, WebSocket, or MCP with one line change
-- **Distributed execution** — scale across machines with Redis-backed queue dispatch
+- **Distributed execution** — dispatch workflows to remote workers via Redis queues — scale horizontally by adding servers
 - **Instant Web UI** — add a Gradio-powered interface with 2 lines of YAML
 
 ---

{model_compose-0.4.64 → model_compose-0.4.66}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "model-compose"
-version = "0.4.64"
+version = "0.4.66"
 description = "model-compose: Declarative AI Workflow Orchestrator"
 authors = [
     { name = "Hanyeol Cho", email = "hanyeol.cho@gmail.com" }

{model_compose-0.4.64 → model_compose-0.4.66}/src/mindor/cli/compose.py

@@ -1,5 +1,4 @@
 from typing import Optional, List
-import importlib.metadata
 from pathlib import Path
 import click
 import asyncio
@@ -16,10 +15,9 @@ def _load_compose_config(config_files: List[Path], env_files: List[Path], env_da
     return load_compose_config(".", config_files, env)
 
 def _get_version() -> str:
-    try:
-        return importlib.metadata.version("model-compose")
-    except importlib.metadata.PackageNotFoundError:
-        return "latest"
+    from mindor.version import __version__
+
+    return __version__
 
 @click.group()
 @click.option(

{model_compose-0.4.64 → model_compose-0.4.66}/src/mindor/core/component/services/__init__.py

@@ -6,6 +6,7 @@ from .mcp_server import *
 from .mcp_client import *
 from .model import *
 from .model_memory import *
+from .model_tokenizer import *
 from .datasets import *
 from .vector_store import *
 from .workflow import *
@@ -21,4 +22,4 @@ from .audio_extractor import *
 from .audio_converter import *
 from .key_value_store import *
 from .graph_store import *
-from .tokenizer import *
+from .search_engine import *

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/text_generation → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/chat_completion}/__init__.py

@@ -1,2 +1,3 @@
 from .huggingface import *
 from .llamacpp import *
+from .custom import *

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/chat_completion/custom/custom.py

@@ -0,0 +1,7 @@
+from mindor.dsl.schema.component import ModelComponentConfig, CustomChatCompletionModelFamily
+from ....base import ModelTaskType, ModelDriver, register_model_task_service
+
+@register_model_task_service(ModelTaskType.CHAT_COMPLETION, ModelDriver.CUSTOM)
+class CustomChatCompletionTaskService:
+    def __new__(cls, id: str, config: ModelComponentConfig, daemon: bool):
+        raise ValueError(f"Unknown family: {config.family}")

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/face_embedding → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/face_embedding/custom}/custom.py

@@ -1,11 +1,10 @@
-from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, TypeAlias, Any
-from mindor.dsl.schema.component import ModelComponentConfig, FaceEmbeddingModelFamily
-from ...base import ModelTaskType, ModelDriver, register_model_task_service
+from mindor.dsl.schema.component import ModelComponentConfig, CustomFaceEmbeddingModelFamily
+from ....base import ModelTaskType, ModelDriver, register_model_task_service
 
 @register_model_task_service(ModelTaskType.FACE_EMBEDDING, ModelDriver.CUSTOM)
 class CustomFaceEmbeddingTaskService:
     def __new__(cls, id: str, config: ModelComponentConfig, daemon: bool):
-        if config.family == FaceEmbeddingModelFamily.INSIGHTFACE:
+        if config.family == CustomFaceEmbeddingModelFamily.INSIGHTFACE:
             from .insightface import InsightfaceFaceEmbeddingTaskService
             return InsightfaceFaceEmbeddingTaskService(id, config, daemon)
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/face_embedding → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/face_embedding/custom}/insightface.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig, LocalModelConfig
 from mindor.dsl.schema.action import ModelActionConfig, InsightfaceFaceEmbeddingModelActionConfig
 from mindor.core.logger import logging
-from .common import FaceEmbeddingTaskService, FaceEmbeddingTaskAction
-from ...base import ComponentActionContext
+from ..common import FaceEmbeddingTaskService, FaceEmbeddingTaskAction
+from ....base import ComponentActionContext
 from PIL import Image as PILImage
 import asyncio, os, shutil
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_generation → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_generation/custom}/custom.py

@@ -1,11 +1,10 @@
-from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, TypeAlias, Any
-from mindor.dsl.schema.component import ModelComponentConfig, ImageGenerationModelFamily
-from ...base import ModelTaskType, ModelDriver, register_model_task_service
+from mindor.dsl.schema.component import ModelComponentConfig, CustomImageGenerationModelFamily
+from ....base import ModelTaskType, ModelDriver, register_model_task_service
 
 @register_model_task_service(ModelTaskType.IMAGE_GENERATION, ModelDriver.CUSTOM)
 class CustomImageGenerationTaskService:
     def __new__(cls, id: str, config: ModelComponentConfig, daemon: bool):
-        if config.family == ImageGenerationModelFamily.HUNYUAN_IMAGE:
+        if config.family == CustomImageGenerationModelFamily.HUNYUAN_IMAGE:
             from .hunyuan_image import HunyuanImageGenerationTaskService
             return HunyuanImageGenerationTaskService(id, config, daemon)
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_generation → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_generation/custom}/flux.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, FluxImageGenerationModelActionConfig
 from mindor.core.logger import logging
-from ...base import ComponentActionContext
-from .common import ImageGenerationTaskService, ImageGenerationTaskAction
+from ....base import ComponentActionContext
+from ..common import ImageGenerationTaskService, ImageGenerationTaskAction
 from PIL import Image as PILImage
 import asyncio
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_generation → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_generation/custom}/hunyuan_image.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, HunyuanImageGenerationModelActionConfig
 from mindor.core.logger import logging
-from ...base import ComponentActionContext
-from .common import ImageGenerationTaskService, ImageGenerationTaskAction
+from ....base import ComponentActionContext
+from ..common import ImageGenerationTaskService, ImageGenerationTaskAction
 from PIL import Image as PILImage
 import asyncio
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_generation → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_generation/custom}/sdxl.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, SdxlImageGenerationModelActionConfig
 from mindor.core.logger import logging
-from ...base import ComponentActionContext
-from .common import ImageGenerationTaskService, ImageGenerationTaskAction
+from ....base import ComponentActionContext
+from ..common import ImageGenerationTaskService, ImageGenerationTaskAction
 from PIL import Image as PILImage
 import asyncio
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/speech_to_text → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_to_text}/__init__.py

@@ -1 +1,2 @@
 from .huggingface import *
+from .custom import *

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_to_text/common.py

@@ -0,0 +1,67 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING
+
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Any, Iterator
+from abc import ABC, abstractmethod
+from mindor.dsl.schema.action import ImageToTextModelActionConfig
+from mindor.core.utils.streamer import AsyncStreamer
+from ...base import ModelTaskService, ComponentActionContext
+from PIL import Image as PILImage
+import asyncio
+
+class ImageToTextTaskAction:
+    def __init__(self, config: ImageToTextModelActionConfig):
+        self.config: ImageToTextModelActionConfig = config
+
+    async def run(self, context: ComponentActionContext, loop: asyncio.AbstractEventLoop) -> Any:
+        image, text = await self._prepare_input(context)
+        is_single_input: bool = bool(not isinstance(image, list))
+        images: List[PILImage.Image] = [ image ] if is_single_input else image
+        texts: Optional[List[str]] = [ text ] if is_single_input else text
+
+        batch_size = await context.render_variable(self.config.batch_size)
+        streaming  = await context.render_variable(self.config.streaming)
+
+        if streaming:
+            if batch_size != 1 or len(images) != 1:
+                raise ValueError("Streaming mode only supports a single input image with batch size of 1")
+
+            streamer = await self._generate(images, texts, context, streaming=True)
+
+            async def _stream_output_generator():
+                async for chunk in AsyncStreamer(streamer, loop):
+                    if chunk:
+                        yield await self._render_output_chunk(context, chunk)
+
+            return _stream_output_generator()
+        else:
+            results = []
+            for index in range(0, len(images), batch_size):
+                batch_images = images[index:index + batch_size]
+                batch_texts = texts[index:index + batch_size] if texts else None
+                outputs = await self._generate(batch_images, batch_texts, context, streaming=False)
+                results.extend(outputs)
+
+            result = results[0] if is_single_input else results
+            return await self._render_output(context, result)
+
+    async def _prepare_input(self, context: ComponentActionContext) -> Tuple[Union[PILImage.Image, List[PILImage.Image]], Optional[Union[str, List[str]]]]:
+        image = await context.render_image(self.config.image)
+        text  = await context.render_variable(self.config.text)
+
+        return image, text
+
+    @abstractmethod
+    async def _generate(self, images: List[PILImage.Image], texts: Optional[List[str]], context: ComponentActionContext, streaming: bool) -> Union[List[str], Iterator[str]]:
+        pass
+
+    async def _render_output_chunk(self, context: ComponentActionContext, chunk: str) -> Any:
+        context.register_source("result[]", chunk)
+        return (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else chunk
+
+    async def _render_output(self, context: ComponentActionContext, result: Union[str, List[str]]) -> Any:
+        context.register_source("result", result)
+        return (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else result
+
+class ImageToTextTaskService(ModelTaskService):
+    pass

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_to_text/custom/__init__.py

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

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_to_text/custom/custom.py

@@ -0,0 +1,7 @@
+from mindor.dsl.schema.component import ModelComponentConfig, CustomImageToTextModelFamily
+from ....base import ModelTaskType, ModelDriver, register_model_task_service
+
+@register_model_task_service(ModelTaskType.IMAGE_TO_TEXT, ModelDriver.CUSTOM)
+class CustomImageToTextTaskService:
+    def __new__(cls, id: str, config: ModelComponentConfig, daemon: bool):
+        raise ValueError(f"Unknown family: {config.family}")

{model_compose-0.4.64 → model_compose-0.4.66}/src/mindor/core/component/services/model/tasks/image_to_text/huggingface.py

@@ -1,13 +1,13 @@
 from __future__ import annotations
 from typing import TYPE_CHECKING
 
-from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Protocol, Any
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Protocol, Any, Iterator
 from mindor.dsl.schema.component import ImageToTextModelArchitecture
 from mindor.dsl.schema.action import ModelActionConfig, ImageToTextModelActionConfig
-from mindor.core.utils.streamer import AsyncStreamer
 from mindor.core.logger import logging
 from ...base import ModelTaskType, ModelDriver, register_model_task_service
 from ...base import HuggingfaceMultimodalModelTaskService, ComponentActionContext
+from .common import ImageToTextTaskAction
 from PIL import Image as PILImage
 from threading import Thread
 import asyncio
@@ -20,7 +20,7 @@ if TYPE_CHECKING:
 class WithTokenizer(Protocol):
     tokenizer: PreTrainedTokenizer
 
-class HuggingfaceImageToTextTaskAction:
+class HuggingfaceImageToTextTaskAction(ImageToTextTaskAction):
     def __init__(
         self,
         config: ImageToTextModelActionConfig,
@@ -28,82 +28,51 @@ class HuggingfaceImageToTextTaskAction:
         processor: ProcessorMixin,
         device: torch.device
     ):
-        self.config: ImageToTextModelActionConfig = config
+        super().__init__(config)
+
         self.model: Union[PreTrainedModel, GenerationMixin] = model
         self.processor: Union[ProcessorMixin, WithTokenizer] = processor
         self.device: torch.device = device
 
-    async def run(self, context: ComponentActionContext, loop: asyncio.AbstractEventLoop) -> Any:
-        from transformers import TextIteratorStreamer, StopStringCriteria, GenerationConfig
+    async def _generate(self, images: List[PILImage.Image], texts: Optional[List[str]], context: ComponentActionContext, streaming: bool) -> Union[List[str], Iterator[str]]:
+        from transformers import StopStringCriteria, GenerationConfig
         import torch
 
-        image, text = await self._prepare_input(context)
-        is_single_input: bool = bool(not isinstance(image, list))
-        images: List[PILImage.Image] = [ image ] if is_single_input else image
-        texts: Optional[List[str]] = [ text ] if is_single_input else text
-        results = []
-
-        batch_size        = await context.render_variable(self.config.batch_size)
-        streaming         = await context.render_variable(self.config.streaming)
         stop_sequences    = await context.render_variable(self.config.stop_sequences)
         processor_params  = await self._resolve_processor_params(context)
         generation_params = await self._resolve_generation_params(context)
 
-        if streaming and (batch_size != 1 or len(images) != 1):
-            raise ValueError("Streaming mode only supports a single input image with batch size of 1")
-
-        streamer = TextIteratorStreamer(self.processor.tokenizer, skip_prompt=True) if streaming else None
         stopping_criteria = [ StopStringCriteria(self.processor.tokenizer, stop_sequences) ] if stop_sequences else None
 
-        for index in range(0, len(images), batch_size):
-            batch_images = images[index:index + batch_size]
-            batch_texts = texts[index:index + batch_size] if texts else None
+        inputs: Tensor = self.processor(images=images, text=texts, **processor_params)
+        inputs = inputs.to(self.device)
 
-            inputs: Tensor = self.processor(images=batch_images, text=batch_texts, **processor_params)
-            inputs = inputs.to(self.device)
+        if streaming:
+            from transformers import TextIteratorStreamer
+
+            streamer = TextIteratorStreamer(self.processor.tokenizer, skip_prompt=True)
 
-            def _generate():
+            def _run():
                 with torch.inference_mode():
-                    outputs = self.model.generate(
-                        **inputs, 
+                    self.model.generate(
+                        **inputs,
                         generation_config=GenerationConfig(**generation_params),
                         stopping_criteria=stopping_criteria,
                         streamer=streamer
                     )
 
-                outputs = self.processor.tokenizer.batch_decode(outputs, skip_special_tokens=True)
-                results.extend(outputs)
-
-            if streaming:
-                thread = Thread(target=_generate)
-                thread.start()
-            else:
-                _generate()
+            Thread(target=_run, daemon=True).start()
 
-        if streaming:
-            async def _stream_output_generator():
-                async for chunk in AsyncStreamer(streamer, loop):
-                    if chunk:
-                        yield await self._render_output_chunk(context, chunk)
-
-            return _stream_output_generator()
+            return streamer
         else:
-            result = results[0] if is_single_input else results
-            return await self._render_output(context, result)
-
-    async def _prepare_input(self, context: ComponentActionContext) -> Tuple[Union[PILImage.Image, List[PILImage.Image]], Optional[Union[str, List[str]]]]:
-        image = await context.render_image(self.config.image)
-        text  = await context.render_variable(self.config.text)
-        
-        return image, text
-
-    async def _render_output_chunk(self, context: ComponentActionContext, chunk: str) -> Any:
-        context.register_source("result[]", chunk)
-        return (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else chunk
-
-    async def _render_output(self, context: ComponentActionContext, result: Union[str, List[str]]) -> Any:
-        context.register_source("result", result)
-        return (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else result
+            with torch.inference_mode():
+                outputs = self.model.generate(
+                    **inputs,
+                    generation_config=GenerationConfig(**generation_params),
+                    stopping_criteria=stopping_criteria,
+                )
+
+            return self.processor.tokenizer.batch_decode(outputs, skip_special_tokens=True)
 
     async def _resolve_processor_params(self, context: ComponentActionContext) -> Dict[str, Any]:
         max_input_length = await context.render_variable(self.config.max_input_length)

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/__init__.py

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

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/custom/__init__.py

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

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_upscale → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/custom}/custom.py

@@ -1,23 +1,22 @@
-from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, TypeAlias, Any
-from mindor.dsl.schema.component import ModelComponentConfig, ImageUpscaleModelFamily
-from ...base import ModelTaskType, ModelDriver, register_model_task_service
+from mindor.dsl.schema.component import ModelComponentConfig, CustomImageUpscaleModelFamily
+from ....base import ModelTaskType, ModelDriver, register_model_task_service
 
 @register_model_task_service(ModelTaskType.IMAGE_UPSCALE, ModelDriver.CUSTOM)
 class CustomImageUpscaleTaskService:
     def __new__(cls, id: str, config: ModelComponentConfig, daemon: bool):
-        if config.family == ImageUpscaleModelFamily.ESRGAN:
+        if config.family == CustomImageUpscaleModelFamily.ESRGAN:
             from .esrgan import EsrganImageUpscaleTaskService
             return EsrganImageUpscaleTaskService(id, config, daemon)
 
-        if config.family == ImageUpscaleModelFamily.REAL_ESRGAN:
+        if config.family == CustomImageUpscaleModelFamily.REAL_ESRGAN:
             from .real_esrgan import RealEsrganImageUpscaleTaskService
             return RealEsrganImageUpscaleTaskService(id, config, daemon)
 
-        if config.family == ImageUpscaleModelFamily.LDSR:
+        if config.family == CustomImageUpscaleModelFamily.LDSR:
             from .ldsr import LdsrImageUpscaleTaskService
             return LdsrImageUpscaleTaskService(id, config, daemon)
 
-        if config.family == ImageUpscaleModelFamily.SWINIR:
+        if config.family == CustomImageUpscaleModelFamily.SWINIR:
             from .swinir import SwinIRImageUpscaleTaskService
             return SwinIRImageUpscaleTaskService(id, config, daemon)
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_upscale → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/custom}/esrgan.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, EsrganImageUpscaleModelActionConfig
 from mindor.core.logger import logging
-from ...base import ComponentActionContext
-from .common import ImageUpscaleTaskService, ImageUpscaleTaskAction
+from ....base import ComponentActionContext
+from ..common import ImageUpscaleTaskService, ImageUpscaleTaskAction
 from PIL import Image as PILImage
 import asyncio
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_upscale → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/custom}/ldsr.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, LdsrImageUpscaleModelActionConfig
 from mindor.core.logger import logging
-from ...base import ComponentActionContext
-from .common import ImageUpscaleTaskService, ImageUpscaleTaskAction
+from ....base import ComponentActionContext
+from ..common import ImageUpscaleTaskService, ImageUpscaleTaskAction
 from PIL import Image as PILImage
 import asyncio
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_upscale → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/custom}/real_esrgan.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, RealEsrganImageUpscaleModelActionConfig
 from mindor.core.logger import logging
-from .common import ImageUpscaleTaskService, ImageUpscaleTaskAction
-from ...base import ComponentActionContext
+from ..common import ImageUpscaleTaskService, ImageUpscaleTaskAction
+from ....base import ComponentActionContext
 from PIL import Image as PILImage
 import asyncio
 

{model_compose-0.4.64/src/mindor/core/component/services/model/tasks/image_upscale → model_compose-0.4.66/src/mindor/core/component/services/model/tasks/image_upscale/custom}/swinir.py

@@ -5,8 +5,8 @@ from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annot
 from mindor.dsl.schema.component import ModelComponentConfig
 from mindor.dsl.schema.action import ModelActionConfig, SwinIRImageUpscaleModelActionConfig
 from mindor.core.logger import logging
-from .common import ImageUpscaleTaskService, ImageUpscaleTaskAction
-from ...base import ComponentActionContext
+from ..common import ImageUpscaleTaskService, ImageUpscaleTaskAction
+from ....base import ComponentActionContext
 from PIL import Image as PILImage
 import asyncio
 
@@ -148,7 +148,7 @@ class SwinIRImageUpscaleTaskAction(ImageUpscaleTaskAction):
         params = await super()._resolve_upscale_params(context)
 
         params["task"        ] = await context.render_variable(self.config.params.task)
-        params["tile"        ] = await context.render_variable(self.config.params.tile)
+        params["tile"        ] = await context.render_variable(self.config.params.tile_size)
         params["tile_overlap"] = await context.render_variable(self.config.params.tile_overlap)
         params["scale"       ] = await context.render_variable(self.config.params.scale)
         params["window_size" ] = await context.render_variable(self.config.params.window_size)

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/music_generation/__init__.py

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

{model_compose-0.4.64 → model_compose-0.4.66}/src/mindor/core/component/services/model/tasks/music_generation/common.py

@@ -1,9 +1,16 @@
-from typing import Optional, Dict, Any
+from __future__ import annotations
+from typing import TYPE_CHECKING
+
+from typing import Optional, Dict, Any, Tuple, Union, List
 from abc import ABC, abstractmethod
 from mindor.dsl.schema.action import MusicGenerationModelActionConfig
 from mindor.core.logger import logging
 from ...base import ModelTaskService, ComponentActionContext
 
+if TYPE_CHECKING:
+    import numpy as np
+    import torch
+
 class MusicGenerationTaskAction:
     def __init__(self, config: MusicGenerationModelActionConfig):
         self.config: MusicGenerationModelActionConfig = config
@@ -29,8 +36,34 @@ class MusicGenerationTaskAction:
         return params
 
     @abstractmethod
-    async def _generate(self, prompt: str, lyrics: Optional[str], params: Dict[str, Any]) -> bytes:
+    async def _generate(self, prompt: str, lyrics: Optional[str], params: Dict[str, Any]) -> Any:
         pass
 
+    def _encode_samples_to_pcm16(self, samples: Union[torch.Tensor, np.typing.ArrayLike]) -> Tuple[bytes, int]:
+        import numpy as np
+
+        if hasattr(samples, "detach"):
+            samples = samples.detach()
+        if hasattr(samples, "cpu"):
+            samples = samples.cpu()
+        if hasattr(samples, "numpy"):
+            samples = samples.numpy()
+
+        array = np.asarray(samples)
+        if array.ndim == 1:
+            array = array[:, None]
+        elif array.ndim == 2 and array.shape[0] <= 8 and array.shape[0] < array.shape[1]:
+            array = array.T
+        elif array.ndim != 2:
+            raise ValueError(f"Expected mono or stereo audio samples, got shape {array.shape}")
+
+        if np.issubdtype(array.dtype, np.floating):
+            array = np.clip(array, -1.0, 1.0)
+            array = (array * 32767.0).astype("<i2")
+        elif array.dtype != np.int16:
+            array = array.astype("<i2")
+
+        return array.tobytes(), int(array.shape[1])
+
 class MusicGenerationTaskService(ModelTaskService):
     pass

model_compose-0.4.66/src/mindor/core/component/services/model/tasks/music_generation/custom/__init__.py

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