nexaai 1.0.16rc13__cp310-cp310-macosx_15_0_universal2.whl

nexaai/embedder_impl/mlx_embedder_impl.py

@@ -0,0 +1,116 @@
+from typing import List, Union
+import numpy as np
+
+from nexaai.common import PluginID
+from nexaai.embedder import Embedder, EmbeddingConfig
+from nexaai.mlx_backend.embedding.interface import create_embedder
+from nexaai.mlx_backend.ml import ModelConfig as MLXModelConfig, SamplerConfig as MLXSamplerConfig, GenerationConfig as MLXGenerationConfig, EmbeddingConfig
+
+
+class MLXEmbedderImpl(Embedder):
+    def __init__(self):
+        """Initialize MLX Embedder implementation."""
+        super().__init__()
+        self._mlx_embedder = None
+
+    @classmethod
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.MLX):
+        """
+        Load an embedder from model files using MLX backend.
+
+        Args:
+            model_path: Path to the model file
+            tokenizer_file: Path to the tokenizer file (default: "tokenizer.json")
+            plugin_id: Plugin ID to use for the model (default: PluginID.MLX)
+
+        Returns:
+            MLXEmbedderImpl instance
+        """
+        try:
+            # Create instance
+            instance = cls()
+            
+            # Use the factory function to create the appropriate embedder based on model type
+            # This will automatically detect if it's JinaV2 or generic model and route correctly
+            instance._mlx_embedder = create_embedder(
+                model_path=model_path,
+                tokenizer_path=tokenizer_file
+            )
+            
+            # Load the model
+            success = instance._mlx_embedder.load_model(model_path)
+            if not success:
+                raise RuntimeError("Failed to load MLX embedder model")
+            
+            return instance
+        except Exception as e:
+            raise RuntimeError(f"Failed to load MLX Embedder: {str(e)}")
+
+    def eject(self):
+        """
+        Clean up resources and destroy the embedder
+        """
+        if self._mlx_embedder:
+            self._mlx_embedder.destroy()
+            self._mlx_embedder = None
+
+    def generate(self, texts: Union[List[str], str] = None, config: EmbeddingConfig = EmbeddingConfig(), input_ids: Union[List[int], List[List[int]]] = None) -> np.ndarray:
+        """
+        Generate embeddings for the given texts or input_ids.
+
+        Args:
+            texts: List of strings or single string to embed
+            input_ids: Pre-tokenized input as:
+                      - Single sequence: list of integers [1, 2, 3, 4]
+                      - Multiple sequences: list of lists [[1, 2, 3], [4, 5, 6]]
+            config: Configuration for embedding generation
+
+        Returns:
+            numpy array of embeddings with shape (num_sequences, embedding_dim)
+        """
+        if not self._mlx_embedder:
+            raise RuntimeError("MLX Embedder not loaded")
+        
+        if texts is None and input_ids is None:
+            raise ValueError("Either texts or input_ids must be provided")
+        
+        # MLX embedder currently only supports text input, not pre-tokenized input_ids
+        if input_ids is not None:
+            raise NotImplementedError("MLX embedder does not support input_ids, only text input")
+        
+        try:
+            # Convert single string to list if needed
+            if isinstance(texts, str):
+                texts = [texts]
+            
+            # MLX config classes are already imported
+            
+            # Convert our config to MLX config
+            mlx_config = EmbeddingConfig()
+            mlx_config.batch_size = config.batch_size
+            mlx_config.normalize = config.normalize
+            mlx_config.normalize_method = config.normalize_method
+            
+            # Generate embeddings using MLX
+            embeddings = self._mlx_embedder.embed(texts, mlx_config)
+            
+            # Convert to numpy array
+            return np.array(embeddings, dtype=np.float32)
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate embeddings: {str(e)}")
+
+    def get_embedding_dim(self) -> int:
+        """
+        Get the embedding dimension of the model
+
+        Returns:
+            The embedding dimension in int
+        """
+        if not self._mlx_embedder:
+            raise RuntimeError("MLX Embedder not loaded")
+        
+        try:
+            return self._mlx_embedder.embedding_dim()
+        except Exception as e:
+            raise RuntimeError(f"Failed to get embedding dimension: {str(e)}")

nexaai/embedder_impl/pybind_embedder_impl.py

@@ -0,0 +1,95 @@
+from typing import List, Union
+import numpy as np
+
+from nexaai.common import PluginID
+from nexaai.embedder import Embedder, EmbeddingConfig
+from nexaai.binds import embedder_bind
+from nexaai.runtime import _ensure_runtime
+
+
+class PyBindEmbedderImpl(Embedder):
+    def __init__(self, _handle_ptr):
+        """
+        Internal initializer
+        """
+        super().__init__()
+        self._handle = _handle_ptr
+
+    @classmethod
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP):
+        """
+        Load an embedder from model files
+
+        Args:
+            model_path: Path to the model file
+            tokenizer_file: Path to the tokenizer file (default: "tokenizer.json")
+            plugin_id: Plugin ID to use for the model (default: PluginID.LLAMA_CPP)
+
+        Returns:
+            PyBindEmbedderImpl instance
+        """
+        _ensure_runtime()
+        # Convert enum to string for C++ binding
+        plugin_id_str = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        # New parameter order: model_path, plugin_id, tokenizer_path (optional)
+        handle = embedder_bind.ml_embedder_create(model_path, plugin_id_str, tokenizer_file)
+        return cls(handle)
+
+    def eject(self):
+        """
+        Clean up resources and destroy the embedder
+        """
+        # Destructor of the handle will unload the model correctly
+        del self._handle
+        self._handle = None
+
+    def generate(self, texts: Union[List[str], str] = None, config: EmbeddingConfig = EmbeddingConfig(), input_ids: Union[List[int], List[List[int]]] = None) -> np.ndarray:
+        """
+        Generate embeddings for the given texts or input_ids.
+
+        Args:
+            texts: List of strings or single string to embed
+            input_ids: Pre-tokenized input as:
+                      - Single sequence: list of integers [1, 2, 3, 4]
+                      - Multiple sequences: list of lists [[1, 2, 3], [4, 5, 6]]
+            config: Configuration for embedding generation
+
+        Returns:
+            numpy array of embeddings with shape (num_sequences, embedding_dim)
+        """
+        if texts is None and input_ids is None:
+            raise ValueError("Either texts or input_ids must be provided")
+
+        # Create bind config
+        bind_config = embedder_bind.EmbeddingConfig()
+        bind_config.batch_size = config.batch_size
+        bind_config.normalize = config.normalize
+        bind_config.normalize_method = config.normalize_method
+
+        # Convert single string to list if needed
+        if isinstance(texts, str):
+            texts = [texts]
+
+        # Convert input_ids to 2D format if needed
+        processed_input_ids = None
+        if input_ids is not None:
+            if len(input_ids) > 0 and isinstance(input_ids[0], int):
+                # Single sequence: convert [1, 2, 3] to [[1, 2, 3]]
+                processed_input_ids = [input_ids]
+            else:
+                # Multiple sequences: already in correct format [[1, 2], [3, 4]]
+                processed_input_ids = input_ids
+
+        # Pass both parameters, let the ABI handle validation
+        embeddings = embedder_bind.ml_embedder_embed(self._handle, bind_config, texts, processed_input_ids)
+
+        return embeddings
+
+    def get_embedding_dim(self) -> int:
+        """
+        Get the embedding dimension of the model
+
+        Returns:
+            The embedding dimension in int
+        """
+        return embedder_bind.ml_embedder_embedding_dim(self._handle)

nexaai/image_gen.py

@@ -0,0 +1,140 @@
+from typing import List, Optional, Union
+from abc import abstractmethod
+from dataclasses import dataclass
+
+from nexaai.base import BaseModel
+from nexaai.common import PluginID
+
+
+@dataclass
+class Image:
+    """Image data structure."""
+    data: List[float]  # width × height × channels
+    width: int
+    height: int
+    channels: int  # 3 = RGB, 4 = RGBA
+
+
+@dataclass
+class ImageSamplerConfig:
+    """Configuration for image sampling."""
+    method: str = "ddim"
+    steps: int = 20
+    guidance_scale: float = 7.5
+    eta: float = 0.0
+    seed: int = -1  # –1 for random
+
+
+@dataclass
+class ImageGenerationConfig:
+    """Configuration for image generation."""
+    prompts: Union[str, List[str]]
+    negative_prompts: Optional[Union[str, List[str]]] = None
+    height: int = 512
+    width: int = 512
+    sampler_config: Optional[ImageSamplerConfig] = None
+    lora_id: int = -1  # –1 for none
+    init_image: Optional[Image] = None
+    strength: float = 1.0
+    n_images: int = 1
+    n_rows: int = 1
+    decoding_batch_size: int = 1
+
+
+@dataclass
+class SchedulerConfig:
+    """Configuration for diffusion scheduler."""
+    type: str = "ddim"
+    num_train_timesteps: int = 1000
+    steps_offset: int = 0  # An offset added to the inference steps
+    beta_start: float = 0.00085
+    beta_end: float = 0.012
+    beta_schedule: str = "scaled_linear"
+    prediction_type: str = "epsilon"
+    timestep_type: str = "discrete"
+    timestep_spacing: str = "linspace"
+    interpolation_type: str = "linear"
+    config_path: Optional[str] = None
+
+
+class ImageGen(BaseModel):
+    """Abstract base class for image generation models."""
+
+    def __init__(self):
+        """Initialize base image generation class."""
+        pass
+
+    @classmethod
+    def _load_from(cls,
+                   model_path: str,
+                   scheduler_config_path: str = "",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
+                   device_id: Optional[str] = None,
+                   float16: bool = True,
+                   quantize: bool = False
+        ) -> 'ImageGen':
+        """Load image generation model from local path, routing to appropriate implementation."""
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
+            from nexaai.image_gen_impl.mlx_image_gen_impl import MLXImageGenImpl
+            return MLXImageGenImpl._load_from(model_path, scheduler_config_path, plugin_id, device_id, float16, quantize)
+        else:
+            from nexaai.image_gen_impl.pybind_image_gen_impl import PyBindImageGenImpl
+            return PyBindImageGenImpl._load_from(model_path, scheduler_config_path, plugin_id, device_id, float16, quantize)
+
+    @abstractmethod
+    def load_model(self, model_path: str, extra_data: Optional[str] = None) -> bool:
+        """Load model from path."""
+        pass
+
+    @abstractmethod
+    def set_scheduler(self, config: SchedulerConfig) -> None:
+        """Set scheduler configuration."""
+        pass
+
+    @abstractmethod
+    def set_sampler(self, config: ImageSamplerConfig) -> None:
+        """Set sampler configuration."""
+        pass
+
+    @abstractmethod
+    def reset_sampler(self) -> None:
+        """Reset sampler to default configuration."""
+        pass
+
+    @abstractmethod
+    def txt2img(self, prompt: str, config: ImageGenerationConfig) -> Image:
+        """Generate image from text prompt."""
+        pass
+
+    @abstractmethod
+    def img2img(self, init_image: Image, prompt: str, config: ImageGenerationConfig) -> Image:
+        """Generate image from initial image and text prompt."""
+        pass
+
+    @abstractmethod
+    def generate(self, config: ImageGenerationConfig) -> Image:
+        """Generate image from configuration."""
+        pass
+
+    @abstractmethod
+    def set_lora(self, lora_id: int) -> None:
+        """Set active LoRA adapter."""
+        pass
+
+    @abstractmethod
+    def add_lora(self, lora_path: str) -> int:
+        """Add LoRA adapter and return its ID."""
+        pass
+
+    @abstractmethod
+    def remove_lora(self, lora_id: int) -> None:
+        """Remove LoRA adapter."""
+        pass
+
+    @abstractmethod
+    def list_loras(self) -> List[int]:
+        """List available LoRA adapters."""
+        pass

nexaai/image_gen_impl/mlx_image_gen_impl.py

@@ -0,0 +1,292 @@
+# Note: This code is generated by Cursor, not tested yet.
+
+from typing import List, Optional, Union
+import os
+
+from nexaai.common import PluginID
+from nexaai.image_gen import ImageGen, ImageGenerationConfig, ImageSamplerConfig, SchedulerConfig, Image
+from nexaai.mlx_backend.sd.interface import ImageGen as MLXImageGenInterface
+
+
+class MLXImageGenImpl(ImageGen):
+    def __init__(self):
+        """Initialize MLX Image Generation implementation."""
+        super().__init__()
+        self._mlx_image_gen = None
+
+    @classmethod
+    def _load_from(cls,
+                   model_path: str,
+                   scheduler_config_path: str = "",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
+                   device_id: Optional[str] = None,
+                   float16: bool = True,
+                   quantize: bool = False
+        ) -> 'MLXImageGenImpl':
+        """Load image generation model from local path using MLX backend."""
+        try:
+            # MLX Image Generation interface is already imported
+            
+            # Create instance and load MLX image generation model
+            instance = cls()
+            instance._mlx_image_gen = MLXImageGenInterface(
+                model_path=model_path,
+                scheduler_config_path=scheduler_config_path,
+                device=device_id,
+                float16=float16,
+                quantize=quantize
+            )
+            
+            return instance
+        except Exception as e:
+            raise RuntimeError(f"Failed to load MLX Image Generation: {str(e)}")
+
+    def eject(self):
+        """Destroy the model and free resources."""
+        if self._mlx_image_gen:
+            self._mlx_image_gen.destroy()
+            self._mlx_image_gen = None
+
+    def load_model(self, model_path: str, extra_data: Optional[str] = None) -> bool:
+        """Load model from path."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not initialized")
+        
+        try:
+            return self._mlx_image_gen.load_model(model_path, extra_data)
+        except Exception as e:
+            raise RuntimeError(f"Failed to load image generation model: {str(e)}")
+
+    def set_scheduler(self, config: SchedulerConfig) -> None:
+        """Set scheduler configuration."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            # Convert our config to MLX format
+            import sys
+            mlx_path = self._get_mlx_path()
+            if mlx_path not in sys.path:
+                sys.path.insert(0, mlx_path)
+            
+            # MLX ImageGen uses a simpler interface, we'll store the config for use in generation
+            self._scheduler_config = config
+            self._mlx_image_gen.set_scheduler(config)
+        except Exception as e:
+            raise RuntimeError(f"Failed to set scheduler: {str(e)}")
+
+    def set_sampler(self, config: ImageSamplerConfig) -> None:
+        """Set sampler configuration."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            # Convert our config to MLX format
+            import sys
+            mlx_path = self._get_mlx_path()
+            if mlx_path not in sys.path:
+                sys.path.insert(0, mlx_path)
+            from sd.interface import ImageSamplerConfig as MLXImageSamplerConfig
+            
+            mlx_config = MLXImageSamplerConfig(
+                method=config.method,
+                steps=config.steps,
+                guidance_scale=config.guidance_scale,
+                eta=config.eta,
+                seed=config.seed
+            )
+            
+            self._mlx_image_gen.set_sampler(mlx_config)
+        except Exception as e:
+            raise RuntimeError(f"Failed to set sampler: {str(e)}")
+
+    def reset_sampler(self) -> None:
+        """Reset sampler to default configuration."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            self._mlx_image_gen.reset_sampler()
+        except Exception as e:
+            raise RuntimeError(f"Failed to reset sampler: {str(e)}")
+
+    def txt2img(self, prompt: str, config: ImageGenerationConfig) -> Image:
+        """Generate image from text prompt."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            # Convert our config to MLX format
+            import sys
+            mlx_path = self._get_mlx_path()
+            if mlx_path not in sys.path:
+                sys.path.insert(0, mlx_path)
+            from sd.interface import ImageGenerationConfig as MLXImageGenerationConfig
+            
+            mlx_config = self._convert_generation_config(config)
+            
+            # Use MLX text-to-image generation
+            result = self._mlx_image_gen.txt2img(prompt, mlx_config)
+            
+            # Convert MLX result to our format
+            return self._convert_image_result(result)
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate image from text: {str(e)}")
+
+    def img2img(self, init_image: Image, prompt: str, config: ImageGenerationConfig) -> Image:
+        """Generate image from initial image and text prompt."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            # Convert our config to MLX format
+            mlx_config = self._convert_generation_config(config)
+            
+            # Convert our image to MLX format
+            import sys
+            mlx_path = self._get_mlx_path()
+            if mlx_path not in sys.path:
+                sys.path.insert(0, mlx_path)
+            from sd.interface import Image as MLXImage
+            
+            mlx_init_image = MLXImage(
+                data=init_image.data,
+                width=init_image.width,
+                height=init_image.height,
+                channels=init_image.channels
+            )
+            
+            # Use MLX image-to-image generation
+            result = self._mlx_image_gen.img2img(mlx_init_image, prompt, mlx_config)
+            
+            # Convert MLX result to our format
+            return self._convert_image_result(result)
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate image from image: {str(e)}")
+
+    def generate(self, config: ImageGenerationConfig) -> Image:
+        """Generate image from configuration."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            # Convert our config to MLX format
+            mlx_config = self._convert_generation_config(config)
+            
+            # Use MLX generation
+            result = self._mlx_image_gen.generate(mlx_config)
+            
+            # Convert MLX result to our format
+            return self._convert_image_result(result)
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate image: {str(e)}")
+
+    def set_lora(self, lora_id: int) -> None:
+        """Set active LoRA adapter."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            self._mlx_image_gen.set_lora(lora_id)
+        except Exception as e:
+            raise RuntimeError(f"Failed to set LoRA: {str(e)}")
+
+    def add_lora(self, lora_path: str) -> int:
+        """Add LoRA adapter and return its ID."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            return self._mlx_image_gen.add_lora(lora_path)
+        except Exception as e:
+            raise RuntimeError(f"Failed to add LoRA: {str(e)}")
+
+    def remove_lora(self, lora_id: int) -> None:
+        """Remove LoRA adapter."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            self._mlx_image_gen.remove_lora(lora_id)
+        except Exception as e:
+            raise RuntimeError(f"Failed to remove LoRA: {str(e)}")
+
+    def list_loras(self) -> List[int]:
+        """List available LoRA adapters."""
+        if not self._mlx_image_gen:
+            raise RuntimeError("MLX Image Generator not loaded")
+        
+        try:
+            return self._mlx_image_gen.list_loras()
+        except Exception as e:
+            raise RuntimeError(f"Failed to list LoRAs: {str(e)}")
+
+    def _convert_generation_config(self, config: ImageGenerationConfig):
+        """Convert our config to MLX format."""
+        import sys
+        mlx_path = self._get_mlx_path()
+        if mlx_path not in sys.path:
+            sys.path.insert(0, mlx_path)
+        from sd.interface import ImageGenerationConfig as MLXImageGenerationConfig, ImageSamplerConfig as MLXImageSamplerConfig
+        
+        mlx_sampler_config = None
+        if config.sampler_config:
+            mlx_sampler_config = MLXImageSamplerConfig(
+                method=config.sampler_config.method,
+                steps=config.sampler_config.steps,
+                guidance_scale=config.sampler_config.guidance_scale,
+                eta=config.sampler_config.eta,
+                seed=config.sampler_config.seed
+            )
+        
+        mlx_init_image = None
+        if config.init_image:
+            from sd.interface import Image as MLXImage
+            mlx_init_image = MLXImage(
+                data=config.init_image.data,
+                width=config.init_image.width,
+                height=config.init_image.height,
+                channels=config.init_image.channels
+            )
+        
+        return MLXImageGenerationConfig(
+            prompts=config.prompts,
+            negative_prompts=config.negative_prompts,
+            height=config.height,
+            width=config.width,
+            sampler_config=mlx_sampler_config,
+            lora_id=config.lora_id,
+            init_image=mlx_init_image,
+            strength=config.strength,
+            n_images=config.n_images,
+            n_rows=config.n_rows,
+            decoding_batch_size=config.decoding_batch_size
+        )
+
+    def _convert_image_result(self, mlx_image):
+        """Convert MLX image to our format."""
+        return Image(
+            data=mlx_image.data,
+            width=mlx_image.width,
+            height=mlx_image.height,
+            channels=mlx_image.channels
+        )
+
+    @staticmethod
+    def _get_mlx_path():
+        """Get the MLX py-lib path."""
+        from pathlib import Path
+        current_dir = Path(__file__).parent.parent.parent.parent  # Go back to nexa root
+        possible_paths = [
+            current_dir / "plugins" / "mlx" / "py-lib",
+            Path(os.environ.get("NEXA_MLX_PATH", "")) if os.environ.get("NEXA_MLX_PATH") else None,
+        ]
+        
+        for path in possible_paths:
+            if path and path.exists() and (path / "sd" / "interface.py").exists():
+                return str(path.absolute())
+        
+        raise ImportError("Could not find MLX py-lib directory")