model-compose 0.3.13__tar.gz → 0.3.15__tar.gz

{model_compose-0.3.13 → model_compose-0.3.15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-compose
-Version: 0.3.13
+Version: 0.3.15
 Summary: model-compose: Declarative AI Model and Workflow Orchestrator
 Author-email: Hanyeol Cho <hanyeol.cho@gmail.com>
 License-Expression: MIT
@@ -45,6 +45,7 @@ Dynamic: license-file
 - **Multi-Workflow Support**: Define multiple named workflows in one project. Run them by name or set a default for quick execution.
 - **Modular Components**: Break down logic into reusable components and jobs. Easily plug, swap, and extend them across workflows.
 - **Flexible I/O Routing**: Connect inputs and outputs between jobs using clean, scoped variables—no glue code needed.
+- **Streaming Mode Support**: Stream real-time outputs from models and APIs, enabling interactive applications and faster feedback loops.
 - **Run Locally, Serve Remotely**: Execute workflows from the CLI or expose them as HTTP or MCP endpoints with an optional Web UI.
 - **Docker Deployment**: Build and deploy your workflow controller as a Docker container for consistent and portable execution environments.
 - **Environment Variable Support**: Easily inject secrets and configuration via `.env` files or environment variables to keep your YAML clean and secure.

{model_compose-0.3.13 → model_compose-0.3.15}/README.md

@@ -16,6 +16,7 @@
 - **Multi-Workflow Support**: Define multiple named workflows in one project. Run them by name or set a default for quick execution.
 - **Modular Components**: Break down logic into reusable components and jobs. Easily plug, swap, and extend them across workflows.
 - **Flexible I/O Routing**: Connect inputs and outputs between jobs using clean, scoped variables—no glue code needed.
+- **Streaming Mode Support**: Stream real-time outputs from models and APIs, enabling interactive applications and faster feedback loops.
 - **Run Locally, Serve Remotely**: Execute workflows from the CLI or expose them as HTTP or MCP endpoints with an optional Web UI.
 - **Docker Deployment**: Build and deploy your workflow controller as a Docker container for consistent and portable execution environments.
 - **Environment Variable Support**: Easily inject secrets and configuration via `.env` files or environment variables to keep your YAML clean and secure.

{model_compose-0.3.13 → model_compose-0.3.15}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "model-compose"
-version = "0.3.13"
+version = "0.3.15"
 description = "model-compose: Declarative AI Model and Workflow Orchestrator"
 authors = [
     { name = "Hanyeol Cho", email = "hanyeol.cho@gmail.com" }

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/base.py

@@ -64,6 +64,10 @@ class ComponentService(AsyncService):
     async def teardown(self) -> None:
         await self._teardown()
 
+    async def start(self, background: bool = False) -> None:
+        await super().start(background)
+        await self.wait_until_ready()
+
     async def run(self, action_id: Union[str, None], run_id: str, input: Dict[str, Any]) -> Dict[str, Any]:
         _, action = ActionResolver(self.config.actions).resolve(action_id)
         context = ComponentActionContext(run_id, input)
@@ -97,6 +101,9 @@ class ComponentService(AsyncService):
     async def _shutdown(self) -> None:
         pass
 
+    async def _is_ready(self) -> bool:
+        return True
+
     @abstractmethod
     async def _run(self, action: ActionConfig, context: ComponentActionContext) -> Any:
         pass

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/__init__.py

@@ -3,6 +3,7 @@ from .http_client import *
 from .mcp_server import *
 from .mcp_client import *
 from .model import *
+from .vector_store import *
 from .workflow import *
 from .shell import *
 from .text_splitter import *

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/http_server.py

@@ -161,10 +161,19 @@ class HttpServerComponent(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 HttpServerAction(action).run(context, self.client)

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/base.py

@@ -1,6 +1,6 @@
 from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Callable, Any
 from abc import ABC, abstractmethod
-from mindor.dsl.schema.component import ModelComponentConfig, ModelTaskType, DeviceMode
+from mindor.dsl.schema.component import ModelComponentConfig, ModelTaskType, ModelSourceConfig, DeviceMode
 from mindor.dsl.schema.action import ModelActionConfig
 from mindor.core.services import AsyncService
 from ...context import ComponentActionContext
@@ -26,7 +26,7 @@ class ModelTaskService(AsyncService):
     async def _run(self, action: ModelActionConfig, context: ComponentActionContext, loop: asyncio.AbstractEventLoop) -> Any:
         pass
 
-    def _load_pretrained_model(self, extra_params: Optional[Dict[str, Any]] = None) -> PreTrainedModel:
+    def _load_pretrained_model(self, extra_params: Optional[Dict[str, Any]] = None) -> Union[PreTrainedModel, torch.nn.Module]:
         params = self._get_common_model_params()
 
         if extra_params:
@@ -42,6 +42,10 @@ class ModelTaskService(AsyncService):
     def _get_common_model_params(self) -> Dict[str, Any]:
         params: Dict[str, Any] = {}
 
+        if isinstance(self.config.model, ModelSourceConfig):
+            if self.config.model.revision:
+                params["revision"] = self.config.model.revision
+
         if self.config.device_mode != DeviceMode.SINGLE:
             params["device_map"] = self.config.device_mode.value
     
@@ -51,9 +55,6 @@ class ModelTaskService(AsyncService):
         if self.config.low_cpu_mem_usage:
             params["low_cpu_mem_usage"] = True
 
-        if self.config.revision:
-            params["revision"] = self.config.revision
-
         if self.config.cache_dir:
             params["cache_dir"] = self.config.cache_dir
 
@@ -75,12 +76,13 @@ class ModelTaskService(AsyncService):
 
     def _get_common_tokenizer_params(self) -> Dict[str, Any]:
         params: Dict[str, Any] = {}
+
+        if isinstance(self.config.model, ModelSourceConfig):
+            if self.config.model.revision:
+                params["revision"] = self.config.model.revision
     
         if not self.config.fast_tokenizer:
             params["use_fast"] = False
-    
-        if self.config.revision:
-            params["revision"] = self.config.revision
 
         if self.config.cache_dir:
             params["cache_dir"] = self.config.cache_dir
@@ -104,9 +106,10 @@ class ModelTaskService(AsyncService):
     def _get_common_processor_params(self) -> Dict[str, Any]:
         params: Dict[str, Any] = {}
 
-        if self.config.revision:
-            params["revision"] = self.config.revision
-
+        if isinstance(self.config.model, ModelSourceConfig):
+            if self.config.model.revision:
+                params["revision"] = self.config.model.revision
+ 
         if self.config.cache_dir:
             params["cache_dir"] = self.config.cache_dir
 

model_compose-0.3.15/src/mindor/core/component/services/model/loader.py

@@ -0,0 +1,25 @@
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, AsyncIterator, Any
+from transformers import PreTrainedModel
+import torch
+    
+async def load_model(
+    provider: str,
+    model_id: str,
+    revision: Optional[str],
+    filename: Optional[str],
+    model_class: Type[PreTrainedModel],
+    *,
+    cache_dir: Optional[str] = None,
+    local_files_only: bool = False
+) -> Union[PreTrainedModel, torch.nn.Module]:
+    if provider in [ "huggingface.co", "huggingface.com", "huggingface" ]:
+        pass
+
+async def load_model_from_uri(
+    uri: str,
+    model_class: Type[PreTrainedModel],
+    *,
+    cache_dir: Optional[str] = None,
+    local_files_only: bool = False
+) -> Union[PreTrainedModel, torch.nn.Module]:
+    pass

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/tasks/image_to_text.py

@@ -5,7 +5,7 @@ from mindor.core.utils.streamer import AsyncStreamer
 from mindor.core.logger import logging
 from ..base import ModelTaskService, ModelTaskType, register_model_task_service
 from ..base import ComponentActionContext
-from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin, GenerationMixin, TextIteratorStreamer
+from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin, GenerationMixin, TextIteratorStreamer, StopStringCriteria
 from PIL import Image as PILImage
 from threading import Thread
 from torch import Tensor
@@ -25,10 +25,22 @@ class ImageToTextTaskAction:
         image: Union[PILImage.Image, List[PILImage.Image]] = await context.render_image(self.config.image)
         prompt: Optional[Union[str, List[str]]] = await context.render_variable(self.config.prompt)
 
-        batch_size = await context.render_variable(self.config.params.batch_size)
-        stream     = await context.render_variable(self.config.stream)
-
-        is_single_input: bool = True if not isinstance(images, list) else False
+        max_input_length     = await context.render_variable(self.config.params.max_input_length)
+        max_output_length    = await context.render_variable(self.config.params.max_output_length)
+        min_output_length    = await context.render_variable(self.config.params.min_output_length)
+        num_return_sequences = await context.render_variable(self.config.params.num_return_sequences)
+        do_sample            = await context.render_variable(self.config.params.do_sample)
+        temperature          = await context.render_variable(self.config.params.temperature) if do_sample else None
+        top_k                = await context.render_variable(self.config.params.top_k) if do_sample else None
+        top_p                = await context.render_variable(self.config.params.top_p) if do_sample else None
+        num_beams            = await context.render_variable(self.config.params.num_beams)
+        length_penalty       = await context.render_variable(self.config.params.length_penalty) if num_beams > 1 else None
+        early_stopping       = await context.render_variable(self.config.params.early_stopping) if num_beams > 1 else False
+        stop_sequences       = await context.render_variable(self.config.params.stop_sequences)
+        batch_size           = await context.render_variable(self.config.params.batch_size)
+        stream               = await context.render_variable(self.config.stream)
+
+        is_single_input: bool = bool(not isinstance(images, list))
         images: List[PILImage.Image] = [ image ] if is_single_input else image
         prompts: Optional[List[str]] = [ prompt ] if is_single_input else prompt
         results = []
@@ -37,17 +49,31 @@ class ImageToTextTaskAction:
             raise ValueError("Streaming mode only supports a single input image with batch size of 1.")
 
         streamer = TextIteratorStreamer(self.processor.tokenizer, skip_prompt=True, skip_special_tokens=True) if stream 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_prompts = prompts[index:index + batch_size] if prompts else None
             
-            inputs: Tensor = self.processor(images=batch_images, text=batch_prompts, return_tensors="pt")
+            inputs: Tensor = self.processor(images=batch_images, text=batch_prompts, max_length=max_input_length, return_tensors="pt")
             inputs = inputs.to(self.device)
 
             def _generate():
                 with torch.inference_mode():
                     outputs = self.model.generate(
                         **inputs,
+                        max_new_tokens=max_output_length,
+                        min_length=min_output_length,
+                        num_return_sequences=num_return_sequences,
+                        do_sample=do_sample,
+                        temperature=temperature,
+                        top_k=top_k,
+                        top_p=top_p,
+                        num_beams=num_beams,
+                        length_penalty=length_penalty,
+                        early_stopping=early_stopping,
+                        pad_token_id=getattr(self.processor.tokenizer, "pad_token_id", None),
+                        eos_token_id=getattr(self.processor.tokenizer, "eos_token_id", None),
+                        stopping_criteria=stopping_criteria,
                         streamer=streamer
                     )
 
@@ -61,13 +87,13 @@ class ImageToTextTaskAction:
                 _generate()
 
         if stream:
-            async def _stream_generator():
+            async def _stream_output_generator():
                 async for chunk in AsyncStreamer(streamer, loop):
                     if chunk:
                         context.register_source("result[]", chunk)
                         yield (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else chunk
 
-            return _stream_generator()
+            return _stream_output_generator()
         else:
             result = results[0] if is_single_input else results
             context.register_source("result", result)

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/tasks/summarization.py

@@ -5,7 +5,7 @@ from mindor.core.utils.streamer import AsyncStreamer
 from mindor.core.logger import logging
 from ..base import ModelTaskService, ModelTaskType, register_model_task_service
 from ..base import ComponentActionContext
-from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer, GenerationMixin, TextIteratorStreamer
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer, GenerationMixin, TextIteratorStreamer, StopStringCriteria
 from threading import Thread
 from torch import Tensor
 import torch, asyncio
@@ -23,16 +23,18 @@ class SummarizationTaskAction:
         max_input_length  = await context.render_variable(self.config.params.max_input_length)
         max_output_length = await context.render_variable(self.config.params.max_output_length)
         min_output_length = await context.render_variable(self.config.params.min_output_length)
-        num_beams         = await context.render_variable(self.config.params.num_beams)
-        length_penalty    = await context.render_variable(self.config.params.length_penalty) if num_beams > 1 else None
-        early_stopping    = await context.render_variable(self.config.params.early_stopping) if num_beams > 1 else False
         do_sample         = await context.render_variable(self.config.params.do_sample)
+        temperature       = await context.render_variable(self.config.params.temperature) if do_sample else None
         top_k             = await context.render_variable(self.config.params.top_k) if do_sample else None
         top_p             = await context.render_variable(self.config.params.top_p) if do_sample else None
+        num_beams         = await context.render_variable(self.config.params.num_beams)
+        length_penalty    = await context.render_variable(self.config.params.length_penalty) if num_beams > 1 else None
+        early_stopping    = await context.render_variable(self.config.params.early_stopping) if num_beams > 1 else False
+        stop_sequences    = await context.render_variable(self.config.params.stop_sequences)
         batch_size        = await context.render_variable(self.config.params.batch_size)
         stream            = await context.render_variable(self.config.stream)
 
-        is_single_input: bool = True if not isinstance(text, list) else False
+        is_single_input: bool = bool(not isinstance(text, list))
         texts: List[str] = [ text ] if is_single_input else text
         results = []
 
@@ -40,6 +42,7 @@ class SummarizationTaskAction:
             raise ValueError("Streaming mode only supports a single input text with batch size of 1.")
 
         streamer = TextIteratorStreamer(self.tokenizer, skip_prompt=True, skip_special_tokens=True) if stream else None
+        stopping_criteria = [ StopStringCriteria(self.tokenizer, stop_sequences) ] if stop_sequences else None
         for index in range(0, len(texts), batch_size):
             batch_texts = texts[index:index + batch_size]
             inputs: Dict[str, Tensor] = self.tokenizer(batch_texts, return_tensors="pt", max_length=max_input_length, padding=True, truncation=True)
@@ -49,16 +52,18 @@ class SummarizationTaskAction:
                 with torch.inference_mode():
                     outputs = self.model.generate(
                         **inputs,
-                        max_length=max_output_length,
+                        max_new_tokens=max_output_length,
                         min_length=min_output_length,
-                        num_beams=num_beams,
-                        length_penalty=length_penalty,
-                        early_stopping=early_stopping,
                         do_sample=do_sample,
+                        temperature=temperature,
                         top_k=top_k,
                         top_p=top_p,
+                        num_beams=num_beams,
+                        length_penalty=length_penalty,
+                        early_stopping=early_stopping,
                         pad_token_id=getattr(self.tokenizer, "pad_token_id", None),
                         eos_token_id=getattr(self.tokenizer, "eos_token_id", None),
+                        stopping_criteria=stopping_criteria,
                         streamer=streamer
                     )
 
@@ -73,13 +78,13 @@ class SummarizationTaskAction:
                 _generate()
 
         if stream:
-            async def _stream_generator():
+            async def _stream_output_generator():
                 async for chunk in AsyncStreamer(streamer, loop):
                     if chunk:
                         context.register_source("result[]", chunk)
                         yield (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else chunk
 
-            return _stream_generator()
+            return _stream_output_generator()
         else:
             result = results[0] if is_single_input else results
             context.register_source("result", result)

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/tasks/text_classification.py

@@ -24,7 +24,7 @@ class TextClassificationTaskAction:
         batch_size           = await context.render_variable(self.config.params.batch_size)
         stream               = await context.render_variable(self.config.stream)
 
-        is_single_input: bool = True if not isinstance(text, list) else False
+        is_single_input: bool = bool(not isinstance(text, list))
         is_output_array_mode: bool = context.contains_variable_reference("result[]", self.config.output)
         texts: List[str] = [ text ] if is_single_input else text
         results = []
@@ -64,7 +64,7 @@ class TextClassificationTaskAction:
                     yield predictions
 
         if stream:
-            async def _stream_generator():
+            async def _stream_output_generator():
                 async for predictions in _predict():
                     if not is_output_array_mode:
                         for prediction in predictions:
@@ -74,7 +74,7 @@ class TextClassificationTaskAction:
                         for prediction in predictions:
                             yield prediction
 
-            return _stream_generator()
+            return _stream_output_generator()
         else:
             async for predictions in _predict():
                 results.extend(predictions)

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/tasks/text_embedding.py

@@ -25,7 +25,7 @@ class TextEmbeddingTaskAction:
         batch_size       = await context.render_variable(self.config.params.batch_size)
         stream           = await context.render_variable(self.config.stream)
 
-        is_single_input: bool = True if not isinstance(text, list) else False
+        is_single_input: bool = bool(not isinstance(text, list))
         is_output_array_mode: bool = context.contains_variable_reference("result[]", self.config.output)
         texts: List[str] = [ text ] if is_single_input else text
         results = []
@@ -59,7 +59,7 @@ class TextEmbeddingTaskAction:
                     yield embeddings
 
         if stream:
-            async def _stream_generator():
+            async def _stream_output_generator():
                 async for embeddings in _embed():
                     if not is_output_array_mode:
                         for embedding in embeddings:
@@ -69,7 +69,7 @@ class TextEmbeddingTaskAction:
                         for embedding in embeddings:
                             yield embedding
 
-            return _stream_generator()
+            return _stream_output_generator()
         else:
             async for embeddings in _embed():
                 results.extend(embeddings)

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/tasks/text_generation.py

@@ -5,7 +5,7 @@ from mindor.core.utils.streamer import AsyncStreamer
 from mindor.core.logger import logging
 from ..base import ModelTaskService, ModelTaskType, register_model_task_service
 from ..base import ComponentActionContext
-from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer, GenerationMixin, TextIteratorStreamer
+from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer, GenerationMixin, TextIteratorStreamer, StopStringCriteria
 from threading import Thread
 from torch import Tensor
 import torch, asyncio
@@ -21,14 +21,20 @@ class TextGenerationTaskAction:
         prompt: Union[str, List[str]] = await context.render_variable(self.config.prompt)
 
         max_output_length    = await context.render_variable(self.config.params.max_output_length)
+        min_output_length    = await context.render_variable(self.config.params.min_output_length)
         num_return_sequences = await context.render_variable(self.config.params.num_return_sequences)
+        do_sample            = await context.render_variable(self.config.params.do_sample)
         temperature          = await context.render_variable(self.config.params.temperature)
         top_k                = await context.render_variable(self.config.params.top_k)
         top_p                = await context.render_variable(self.config.params.top_p)
+        num_beams            = await context.render_variable(self.config.params.num_beams)
+        length_penalty       = await context.render_variable(self.config.params.length_penalty) if num_beams > 1 else None
+        early_stopping       = await context.render_variable(self.config.params.early_stopping) if num_beams > 1 else False
+        stop_sequences       = await context.render_variable(self.config.params.stop_sequences)
         batch_size           = await context.render_variable(self.config.params.batch_size)
         stream               = await context.render_variable(self.config.stream)
 
-        is_single_input: bool = True if not isinstance(prompt, list) else False
+        is_single_input: bool = bool(not isinstance(prompt, list))
         prompts: List[str] = [ prompt ] if is_single_input else prompt
         results = []
 
@@ -36,6 +42,7 @@ class TextGenerationTaskAction:
             raise ValueError("Streaming mode only supports a single input prompt with batch size of 1.")
 
         streamer = TextIteratorStreamer(self.tokenizer, skip_prompt=True, skip_special_tokens=True) if stream else None
+        stopping_criteria = [ StopStringCriteria(self.tokenizer, stop_sequences) ] if stop_sequences else None
         for index in range(0, len(prompts), batch_size):
             batch_prompts = prompts[index:index + batch_size]
             inputs: Dict[str, Tensor] = self.tokenizer(batch_prompts, return_tensors="pt", padding=True, truncation=True)
@@ -46,11 +53,18 @@ class TextGenerationTaskAction:
                     outputs = self.model.generate(
                         **inputs,
                         max_new_tokens=max_output_length,
+                        min_length=min_output_length,
                         num_return_sequences=num_return_sequences,
+                        do_sample=do_sample,
                         temperature=temperature,
-                        do_sample=True,
                         top_k=top_k,
                         top_p=top_p,
+                        num_beams=num_beams,
+                        length_penalty=length_penalty,
+                        early_stopping=early_stopping,
+                        pad_token_id=getattr(self.tokenizer, "pad_token_id", None),
+                        eos_token_id=getattr(self.tokenizer, "eos_token_id", None),
+                        stopping_criteria=stopping_criteria,
                         streamer=streamer
                     )
 
@@ -65,13 +79,13 @@ class TextGenerationTaskAction:
                 _generate()
 
         if stream:
-            async def _stream_generator():
+            async def _stream_output_generator():
                 async for chunk in AsyncStreamer(streamer, loop):
                     if chunk:
                         context.register_source("result[]", chunk)
                         yield (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else chunk
 
-            return _stream_generator()
+            return _stream_output_generator()
         else:
             result = results[0] if is_single_input else results
             context.register_source("result", result)

{model_compose-0.3.13 → model_compose-0.3.15}/src/mindor/core/component/services/model/tasks/translation.py

@@ -5,7 +5,7 @@ from mindor.core.utils.streamer import AsyncStreamer
 from mindor.core.logger import logging
 from ..base import ModelTaskService, ModelTaskType, register_model_task_service
 from ..base import ComponentActionContext
-from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer, GenerationMixin, TextIteratorStreamer
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer, GenerationMixin, TextIteratorStreamer, StopStringCriteria
 from threading import Thread
 from torch import Tensor
 import torch, asyncio
@@ -23,16 +23,18 @@ class TranslationTaskAction:
         max_input_length  = await context.render_variable(self.config.params.max_input_length)
         max_output_length = await context.render_variable(self.config.params.max_output_length)
         min_output_length = await context.render_variable(self.config.params.min_output_length)
-        num_beams         = await context.render_variable(self.config.params.num_beams)
-        length_penalty    = await context.render_variable(self.config.params.length_penalty) if num_beams > 1 else None
-        early_stopping    = await context.render_variable(self.config.params.early_stopping) if num_beams > 1 else False
         do_sample         = await context.render_variable(self.config.params.do_sample)
+        temperature       = await context.render_variable(self.config.params.temperature) if do_sample else None
         top_k             = await context.render_variable(self.config.params.top_k) if do_sample else None
         top_p             = await context.render_variable(self.config.params.top_p) if do_sample else None
+        num_beams         = await context.render_variable(self.config.params.num_beams)
+        length_penalty    = await context.render_variable(self.config.params.length_penalty) if num_beams > 1 else None
+        early_stopping    = await context.render_variable(self.config.params.early_stopping) if num_beams > 1 else False
+        stop_sequences    = await context.render_variable(self.config.params.stop_sequences)
         batch_size        = await context.render_variable(self.config.params.batch_size)
         stream            = await context.render_variable(self.config.stream)
 
-        is_single_input: bool = True if not isinstance(text, list) else False
+        is_single_input: bool = bool(not isinstance(text, list))
         texts: List[str] = [ text ] if is_single_input else text
         results = []
 
@@ -40,6 +42,7 @@ class TranslationTaskAction:
             raise ValueError("Streaming mode only supports a single input text with batch size of 1.")
 
         streamer = TextIteratorStreamer(self.tokenizer, skip_prompt=True, skip_special_tokens=True) if stream else None
+        stopping_criteria = [ StopStringCriteria(self.tokenizer, stop_sequences) ] if stop_sequences else None
         for index in range(0, len(texts), batch_size):
             batch_texts = texts[index:index + batch_size]
             inputs: Dict[str, Tensor] = self.tokenizer(batch_texts, return_tensors="pt", max_length=max_input_length, padding=True, truncation=True)
@@ -49,16 +52,18 @@ class TranslationTaskAction:
                 with torch.inference_mode():
                     outputs = self.model.generate(
                         **inputs,
-                        max_length=max_output_length,
+                        max_new_tokens=max_output_length,
                         min_length=min_output_length,
-                        num_beams=num_beams,
-                        length_penalty=length_penalty,
-                        early_stopping=early_stopping,
                         do_sample=do_sample,
+                        temperature=temperature,
                         top_k=top_k,
                         top_p=top_p,
+                        num_beams=num_beams,
+                        length_penalty=length_penalty,
+                        early_stopping=early_stopping,
                         pad_token_id=getattr(self.tokenizer, "pad_token_id", None),
                         eos_token_id=getattr(self.tokenizer, "eos_token_id", None),
+                        stopping_criteria=stopping_criteria,
                         streamer=streamer
                     )
 
@@ -73,13 +78,13 @@ class TranslationTaskAction:
                 _generate()
 
         if stream:
-            async def _stream_generator():
+            async def _stream_output_generator():
                 async for chunk in AsyncStreamer(streamer, loop):
                     if chunk:
                         context.register_source("result[]", chunk)
                         yield (await context.render_variable(self.config.output, ignore_files=True)) if self.config.output else chunk
 
-            return _stream_generator()
+            return _stream_output_generator()
         else:
             result = results[0] if is_single_input else results
             context.register_source("result", result)

model_compose-0.3.15/src/mindor/core/component/services/vector_store/__init__.py

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

model_compose-0.3.15/src/mindor/core/component/services/vector_store/base.py

@@ -0,0 +1,28 @@
+from typing import Type, Union, Literal, Optional, Dict, List, Tuple, Set, Annotated, Callable, Any
+from abc import ABC, abstractmethod
+from mindor.dsl.schema.component import VectorStoreComponentConfig, VectorStoreDriver
+from mindor.dsl.schema.action import VectorStoreActionConfig
+from mindor.core.services import AsyncService
+from ...context import ComponentActionContext
+
+class VectorStoreService(AsyncService):
+    def __init__(self, id: str, config: VectorStoreComponentConfig, daemon: bool):
+        super().__init__(daemon)
+
+        self.id: str = id
+        self.config: VectorStoreComponentConfig = config
+
+    async def run(self, action: VectorStoreActionConfig, context: ComponentActionContext) -> Any:
+        return await self._run(action, context)
+
+    @abstractmethod
+    async def _run(self, action: VectorStoreActionConfig, context: ComponentActionContext) -> Any:
+        pass
+
+def register_vector_store_service(driver: VectorStoreDriver):
+    def decorator(cls: Type[VectorStoreService]) -> Type[VectorStoreService]:
+        VectorStoreServiceRegistry[driver] = cls
+        return cls
+    return decorator
+
+VectorStoreServiceRegistry: Dict[VectorStoreDriver, Type[VectorStoreService]] = {}

model_compose-0.3.15/src/mindor/core/component/services/vector_store/drivers/__init__.py

@@ -0,0 +1,2 @@
+from .milvus import *
+from .chroma import *