nexaai 1.0.16rc13__cp310-cp310-macosx_15_0_universal2.whl

nexaai/binds/nexa_mlx/py-lib/profiling.py

@@ -0,0 +1,239 @@
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass, field
+from typing import Any, Optional
+from enum import IntEnum
+
+# --------------------------------------------------------------------------------------
+# Stop reason constants matching profile.h
+# --------------------------------------------------------------------------------------
+
+class StopReason(IntEnum):
+    """Stop reason constants matching profile.h"""
+    ML_STOP_REASON_UNKNOWN = 0
+    ML_STOP_REASON_EOS = 1
+    ML_STOP_REASON_LENGTH = 2
+    ML_STOP_REASON_USER = 3
+    ML_STOP_REASON_STOP_SEQUENCE = 4
+    ML_STOP_REASON_COMPLETED = 5
+
+# --------------------------------------------------------------------------------------
+# Profiling data structure
+# --------------------------------------------------------------------------------------
+
+@dataclass
+class ProfilingData:
+    """Profiling data for performance metrics."""
+    ttft_us: int = 0             # Time to first token (us)
+    total_time_us: int = 0       # Total generation time (us) 
+    prompt_time_us: int = 0      # Prompt processing time (us)
+    decode_time_us: int = 0      # Token generation time (us)
+    tokens_per_second: float = 0.0  # Decoding speed (tokens/sec)
+    total_tokens: int = 0        # Total tokens generated
+    prompt_tokens: int = 0       # Number of prompt tokens
+    generated_tokens: int = 0    # Number of generated tokens
+    stop_reason: int = StopReason.ML_STOP_REASON_UNKNOWN  # Stop reason (numeric)
+    
+    def reset(self):
+        """Reset all profiling data."""
+        self.ttft_us = 0
+        self.total_time_us = 0
+        self.prompt_time_us = 0
+        self.decode_time_us = 0
+        self.tokens_per_second = 0.0
+        self.total_tokens = 0
+        self.prompt_tokens = 0
+        self.generated_tokens = 0
+        self.stop_reason = StopReason.ML_STOP_REASON_UNKNOWN
+
+# --------------------------------------------------------------------------------------
+# Profiling context (similar to ml_ProfilingContext in profile.h)
+# --------------------------------------------------------------------------------------
+
+@dataclass
+class ProfilingContext:
+    """Profiling context for tracking timing and state."""
+    start_time: Optional[float] = None
+    prompt_start_time: Optional[float] = None
+    prompt_end_time: Optional[float] = None
+    decode_start_time: Optional[float] = None
+    decode_end_time: Optional[float] = None
+    first_token_time: Optional[float] = None
+    end_time: Optional[float] = None
+    
+    ttft_recorded: bool = False
+    stop_reason: int = StopReason.ML_STOP_REASON_UNKNOWN
+    prompt_tokens: int = 0
+    generated_tokens: int = 0
+    
+    def reset(self):
+        """Reset profiling context."""
+        self.start_time = None
+        self.prompt_start_time = None
+        self.prompt_end_time = None
+        self.decode_start_time = None
+        self.decode_end_time = None
+        self.first_token_time = None
+        self.end_time = None
+        self.ttft_recorded = False
+        self.stop_reason = StopReason.ML_STOP_REASON_UNKNOWN
+        self.prompt_tokens = 0
+        self.generated_tokens = 0
+
+# --------------------------------------------------------------------------------------
+# Profiling functions (similar to profile.h functions)
+# --------------------------------------------------------------------------------------
+
+def profiling_reset(ctx: ProfilingContext) -> None:
+    """Reset profiling context (ml_profiling_reset)."""
+    ctx.reset()
+
+def profiling_start(ctx: ProfilingContext) -> None:
+    """Start profiling (ml_profiling_start)."""
+    ctx.start_time = time.perf_counter()
+    ctx.prompt_start_time = ctx.start_time
+
+def profiling_prompt_start(ctx: ProfilingContext) -> None:
+    """Start prompt processing timing (ml_profiling_prompt_start)."""
+    ctx.prompt_start_time = time.perf_counter()
+
+def profiling_prompt_end(ctx: ProfilingContext) -> None:
+    """End prompt processing timing (ml_profiling_prompt_end)."""
+    ctx.prompt_end_time = time.perf_counter()
+
+def profiling_decode_start(ctx: ProfilingContext) -> None:
+    """Start decode timing (ml_profiling_decode_start)."""
+    ctx.decode_start_time = time.perf_counter()
+
+def profiling_decode_end(ctx: ProfilingContext) -> None:
+    """End decode timing (ml_profiling_decode_end)."""
+    ctx.decode_end_time = time.perf_counter()
+
+def profiling_record_ttft(ctx: ProfilingContext) -> None:
+    """Record time to first token (ml_profiling_record_ttft)."""
+    if not ctx.ttft_recorded and ctx.start_time is not None:
+        ctx.first_token_time = time.perf_counter()
+        ctx.ttft_recorded = True
+
+def profiling_update_prompt_tokens(ctx: ProfilingContext, prompt_tokens: int) -> None:
+    """Update prompt token count (ml_profiling_update_prompt_tokens)."""
+    ctx.prompt_tokens = prompt_tokens
+
+def profiling_update_generated_tokens(ctx: ProfilingContext, generated_tokens: int) -> None:
+    """Update generated token count (ml_profiling_update_generated_tokens)."""
+    ctx.generated_tokens = generated_tokens
+
+def profiling_stop_reason(ctx: ProfilingContext, stop_reason: int) -> None:
+    """Set stop reason (ml_profiling_stop_reason)."""
+    ctx.stop_reason = stop_reason
+
+def profiling_end(ctx: ProfilingContext) -> None:
+    """End profiling (ml_profiling_end)."""
+    ctx.end_time = time.perf_counter()
+
+def profiling_gen_data(ctx: ProfilingContext) -> ProfilingData:
+    """Generate profiling data from context (ml_profiling_gen_data)."""
+    data = ProfilingData()
+    
+    if ctx.start_time is None or ctx.end_time is None:
+        return data
+    
+    # Calculate total time
+    data.total_time_us = int((ctx.end_time - ctx.start_time) * 1_000_000)
+    
+    # Calculate prompt time
+    if ctx.prompt_start_time is not None and ctx.prompt_end_time is not None:
+        data.prompt_time_us = int((ctx.prompt_end_time - ctx.prompt_start_time) * 1_000_000)
+    
+    # Calculate decode time
+    if ctx.decode_start_time is not None and ctx.decode_end_time is not None:
+        data.decode_time_us = int((ctx.decode_end_time - ctx.decode_start_time) * 1_000_000)
+    
+    # Calculate TTFT
+    if ctx.first_token_time is not None and ctx.start_time is not None:
+        data.ttft_us = int((ctx.first_token_time - ctx.start_time) * 1_000_000)
+    
+    # Set token counts
+    data.prompt_tokens = ctx.prompt_tokens
+    data.generated_tokens = ctx.generated_tokens
+    data.total_tokens = ctx.prompt_tokens + ctx.generated_tokens
+    
+    # Calculate tokens per second
+    if data.decode_time_us > 0:
+        data.tokens_per_second = (data.generated_tokens * 1_000_000.0) / data.decode_time_us
+    
+    # Set stop reason
+    data.stop_reason = ctx.stop_reason
+    
+    return data
+
+def stop_reason_to_string(reason: int) -> str:
+    """Convert stop reason to string (stop_reason_to_string)."""
+    try:
+        return StopReason(reason).name
+    except ValueError:
+        return f"UNKNOWN({reason})"
+
+# --------------------------------------------------------------------------------------
+# Profiling mixin for model classes
+# --------------------------------------------------------------------------------------
+
+class ProfilingMixin:
+    """Mixin class to add profiling capabilities to model classes."""
+    
+    def __init__(self):
+        """Initialize profiling mixin."""
+        self._profiling_context = ProfilingContext()
+        self._profiling_data = ProfilingData()
+    
+    def _start_profiling(self) -> None:
+        """Start profiling for an operation."""
+        profiling_reset(self._profiling_context)
+        profiling_start(self._profiling_context)
+    
+    def _prompt_start(self) -> None:
+        """Start prompt processing timing."""
+        profiling_prompt_start(self._profiling_context)
+    
+    def _prompt_end(self) -> None:
+        """End prompt processing timing."""
+        profiling_prompt_end(self._profiling_context)
+    
+    def _decode_start(self) -> None:
+        """Start decode timing."""
+        profiling_decode_start(self._profiling_context)
+    
+    def _decode_end(self) -> None:
+        """End decode timing."""
+        profiling_decode_end(self._profiling_context)
+    
+    def _record_ttft(self) -> None:
+        """Record time to first token."""
+        profiling_record_ttft(self._profiling_context)
+    
+    def _update_prompt_tokens(self, prompt_tokens: int) -> None:
+        """Update prompt token count."""
+        profiling_update_prompt_tokens(self._profiling_context, prompt_tokens)
+    
+    def _update_generated_tokens(self, generated_tokens: int) -> None:
+        """Update generated token count."""
+        profiling_update_generated_tokens(self._profiling_context, generated_tokens)
+    
+    def _set_stop_reason(self, stop_reason: int) -> None:
+        """Set stop reason."""
+        profiling_stop_reason(self._profiling_context, stop_reason)
+    
+    def _end_profiling(self) -> ProfilingData:
+        """End profiling and return data."""
+        profiling_end(self._profiling_context)
+        self._profiling_data = profiling_gen_data(self._profiling_context)
+        return self._profiling_data
+    
+    def get_profiling_data(self) -> ProfilingData:
+        """Get profiling data for the last operation."""
+        return self._profiling_data
+    
+    def reset_profiling(self) -> None:
+        """Reset profiling data."""
+        self._profiling_data.reset()

nexaai/common.py

@@ -0,0 +1,104 @@
+from dataclasses import dataclass
+from typing import TypedDict, Literal, Optional, List
+from enum import Enum
+
+
+class PluginID(str, Enum):
+    """Enum for plugin identifiers."""
+    MLX = "mlx"
+    LLAMA_CPP = "llama_cpp"
+    NEXAML = "nexaml"
+
+
+class ChatMessage(TypedDict):
+    role: Literal["user", "assistant", "system"]
+    content: str
+
+class MultiModalMessageContent(TypedDict):
+    type: Literal["text", "image", "audio", "video"]
+    text: Optional[str]
+    url: Optional[str]
+    path: Optional[str]
+
+class MultiModalMessage(TypedDict):
+    role: Literal["user", "assistant", "system"]
+    content: List[MultiModalMessageContent] 
+
+
+@dataclass
+class SamplerConfig:
+    temperature: float = 0.8
+    top_p: float = 0.95
+    top_k: int = 40
+    repetition_penalty: float = 1.0
+    presence_penalty: float = 0.0
+    frequency_penalty: float = 0.0
+    seed: int = -1
+    grammar_path: str = None
+    grammar_string: str = None
+
+@dataclass
+class GenerationConfig:
+    max_tokens: int = 1024
+    stop_words: list[str] = None
+    sampler_config: SamplerConfig = None
+    image_paths: list[str] = None
+    audio_paths: list[str] = None
+
+@dataclass
+class ModelConfig:
+    n_ctx: int = 4096
+    n_threads: int = None
+    n_threads_batch: int = None
+    n_batch: int = 512
+    n_ubatch: int = 512
+    n_seq_max: int = 1
+    n_gpu_layers: int = 999
+    chat_template_path: str = None
+    chat_template_content: str = None
+
+
+@dataclass(frozen=True) # Read-only
+class ProfilingData:
+    """Profiling data structure for LLM/VLM performance metrics."""
+    ttft: int = 0             # Time to first token (us)
+    prompt_time: int = 0      # Prompt processing time (us)
+    decode_time: int = 0      # Token generation time (us)
+    prompt_tokens: int = 0    # Number of prompt tokens
+    generated_tokens: int = 0  # Number of generated tokens
+    audio_duration: int = 0   # Audio duration (us)
+    prefill_speed: float = 0.0  # Prefill speed (tokens/sec)
+    decoding_speed: float = 0.0 # Decoding speed (tokens/sec)
+    real_time_factor: float = 0.0 # Real-Time Factor (RTF)
+    stop_reason: str = ""     # Stop reason: "eos", "length", "user", "stop_sequence"
+    
+    @classmethod
+    def from_dict(cls, data: dict) -> "ProfilingData":
+        """Create ProfilingData from dictionary."""
+        return cls(
+            ttft=data.get("ttft", 0),
+            prompt_time=data.get("prompt_time", 0),
+            decode_time=data.get("decode_time", 0),
+            prompt_tokens=data.get("prompt_tokens", 0),
+            generated_tokens=data.get("generated_tokens", 0),
+            audio_duration=data.get("audio_duration", 0),
+            prefill_speed=data.get("prefill_speed", 0.0),
+            decoding_speed=data.get("decoding_speed", 0.0),
+            real_time_factor=data.get("real_time_factor", 0.0),
+            stop_reason=data.get("stop_reason", "")
+        )
+    
+    def to_dict(self) -> dict:
+        """Convert to dictionary."""
+        return {
+            "ttft": self.ttft,
+            "prompt_time": self.prompt_time,
+            "decode_time": self.decode_time,
+            "prompt_tokens": self.prompt_tokens,
+            "generated_tokens": self.generated_tokens,
+            "audio_duration": self.audio_duration,
+            "prefill_speed": self.prefill_speed,
+            "decoding_speed": self.decoding_speed,
+            "real_time_factor": self.real_time_factor,
+            "stop_reason": self.stop_reason
+        }

nexaai/cv.py

@@ -0,0 +1,92 @@
+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 BoundingBox:
+    """Generic bounding box structure."""
+    x: float  # X coordinate (normalized or pixel, depends on model)
+    y: float  # Y coordinate (normalized or pixel, depends on model) 
+    width: float  # Width
+    height: float  # Height
+
+
+@dataclass
+class CVResult:
+    """Generic detection/classification result."""
+    image_paths: Optional[List[str]] = None  # Output image paths
+    image_count: int = 0  # Number of output images
+    class_id: int = 0  # Class ID (example: ConvNext)
+    confidence: float = 0.0  # Confidence score [0.0-1.0]
+    bbox: Optional[BoundingBox] = None  # Bounding box (example: YOLO)
+    text: Optional[str] = None  # Text result (example: OCR)
+    embedding: Optional[List[float]] = None  # Feature embedding (example: CLIP embedding)
+    embedding_dim: int = 0  # Embedding dimension
+
+
+@dataclass
+class CVResults:
+    """Generic CV inference result."""
+    results: List[CVResult]  # Array of CV results
+    result_count: int  # Number of CV results
+
+
+class CVCapabilities:
+    """CV capabilities enum."""
+    OCR = 0  # OCR
+    CLASSIFICATION = 1  # Classification
+    SEGMENTATION = 2  # Segmentation
+    CUSTOM = 3  # Custom task
+
+
+@dataclass
+class CVModelConfig:
+    """CV model preprocessing configuration."""
+    capabilities: int  # CVCapabilities
+
+    # MLX-OCR
+    det_model_path: Optional[str] = None  # Detection model path
+    rec_model_path: Optional[str] = None  # Recognition model path
+
+    # QNN
+    model_path: Optional[str] = None  # Model path
+    system_library_path: Optional[str] = None  # System library path
+    backend_library_path: Optional[str] = None  # Backend library path
+    extension_library_path: Optional[str] = None  # Extension library path
+    config_file_path: Optional[str] = None  # Config file path
+    char_dict_path: Optional[str] = None  # Character dictionary path
+
+
+class CVModel(BaseModel):
+    """Abstract base class for generic computer vision models."""
+
+    def __init__(self):
+        """Initialize base CV model class."""
+        pass
+
+    @classmethod
+    def _load_from(cls,
+                   _: str,  # TODO: remove this argument, this is a hack to make api design happy
+                   config: CVModelConfig,
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
+                   device_id: Optional[str] = None
+        ) -> 'CVModel':
+        """Load CV model from configuration, 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.cv_impl.mlx_cv_impl import MLXCVImpl
+            return MLXCVImpl._load_from(config, plugin_id, device_id)
+        else:
+            from nexaai.cv_impl.pybind_cv_impl import PyBindCVImpl
+            return PyBindCVImpl._load_from(config, plugin_id, device_id)
+
+    @abstractmethod
+    def infer(self, input_image_path: str) -> CVResults:
+        """Perform inference on image."""
+        pass

nexaai/cv_impl/mlx_cv_impl.py

@@ -0,0 +1,89 @@
+# Note: This code is generated by Cursor, not tested yet.
+
+from typing import Optional, Union
+import os
+
+from nexaai.common import PluginID
+from nexaai.cv import CVModel, CVModelConfig, CVResults
+from nexaai.mlx_backend.cv.interface import CVModel as MLXCVInterface, create_cv_model
+
+
+class MLXCVImpl(CVModel):
+    def __init__(self):
+        """Initialize MLX CV implementation."""
+        super().__init__()
+        self._mlx_cv = None
+
+    @classmethod
+    def _load_from(cls,
+                   config: CVModelConfig,
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
+                   device_id: Optional[str] = None
+        ) -> 'MLXCVImpl':
+        """Load CV model from configuration using MLX backend."""
+        try:
+            # Get MLX config class
+            from nexaai.mlx_backend.ml import CVModelConfig as MLXCVModelConfig
+            
+            # Convert our config to MLX format
+            mlx_config = MLXCVModelConfig(
+                capabilities=config.capabilities,
+                det_model_path=config.det_model_path,
+                rec_model_path=config.rec_model_path,
+                model_path=config.model_path,
+                system_library_path=config.system_library_path,
+                backend_library_path=config.backend_library_path,
+                extension_library_path=config.extension_library_path,
+                config_file_path=config.config_file_path,
+                char_dict_path=config.char_dict_path
+            )
+            
+            # Create instance and load MLX CV model
+            instance = cls()
+            instance._mlx_cv = create_cv_model(mlx_config, device_id)
+            
+            return instance
+        except Exception as e:
+            raise RuntimeError(f"Failed to load MLX CV: {str(e)}")
+
+    def eject(self):
+        """Destroy the model and free resources."""
+        if self._mlx_cv:
+            self._mlx_cv.destroy()
+            self._mlx_cv = None
+
+    def infer(self, input_image_path: str) -> CVResults:
+        """Perform inference on image."""
+        if not self._mlx_cv:
+            raise RuntimeError("MLX CV not loaded")
+        
+        try:
+            # Use MLX CV inference
+            result = self._mlx_cv.infer(input_image_path)
+            
+            # Convert MLX result to our format
+            from nexaai.cv import CVResult
+            
+            our_results = []
+            for mlx_result in result.results:
+                our_result = CVResult(
+                    image_paths=mlx_result.image_paths,
+                    image_count=mlx_result.image_count,
+                    class_id=mlx_result.class_id,
+                    confidence=mlx_result.confidence,
+                    bbox=mlx_result.bbox,
+                    text=mlx_result.text,
+                    embedding=mlx_result.embedding,
+                    embedding_dim=mlx_result.embedding_dim
+                )
+                our_results.append(our_result)
+            
+            return CVResults(
+                results=our_results,
+                result_count=result.result_count
+            )
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to perform CV inference: {str(e)}")
+
+

nexaai/cv_impl/pybind_cv_impl.py

@@ -0,0 +1,32 @@
+from typing import Optional, Union
+
+from nexaai.common import PluginID
+from nexaai.cv import CVModel, CVModelConfig, CVResults
+
+
+class PyBindCVImpl(CVModel):
+    def __init__(self):
+        """Initialize PyBind CV implementation."""
+        super().__init__()
+        # TODO: Add PyBind-specific initialization
+
+    @classmethod
+    def _load_from(cls,
+                   config: CVModelConfig,
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
+                   device_id: Optional[str] = None
+        ) -> 'PyBindCVImpl':
+        """Load CV model from configuration using PyBind backend."""
+        # TODO: Implement PyBind CV loading
+        instance = cls()
+        return instance
+
+    def eject(self):
+        """Destroy the model and free resources."""
+        # TODO: Implement PyBind CV cleanup
+        pass
+
+    def infer(self, input_image_path: str) -> CVResults:
+        """Perform inference on image."""
+        # TODO: Implement PyBind CV inference
+        raise NotImplementedError("PyBind CV inference not yet implemented")

nexaai/embedder.py

@@ -0,0 +1,72 @@
+from typing import List, Union
+from dataclasses import dataclass
+from abc import abstractmethod
+import numpy as np
+
+from nexaai.base import BaseModel
+from nexaai.common import PluginID
+
+
+@dataclass
+class EmbeddingConfig:
+    batch_size: int = 32
+    normalize: bool = True
+    normalize_method: str = "l2"
+
+
+class Embedder(BaseModel):
+    def __init__(self):
+        """
+        Internal initializer
+        """
+        pass
+
+    @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, routing to appropriate implementation.
+
+        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:
+            Embedder instance
+        """
+        # 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.embedder_impl.mlx_embedder_impl import MLXEmbedderImpl
+            return MLXEmbedderImpl._load_from(model_path, tokenizer_file, plugin_id)
+        else:
+            from nexaai.embedder_impl.pybind_embedder_impl import PyBindEmbedderImpl
+            return PyBindEmbedderImpl._load_from(model_path, tokenizer_file, plugin_id)
+
+    @abstractmethod
+    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)
+        """
+        pass
+
+    @abstractmethod
+    def get_embedding_dim(self) -> int:
+        """
+        Get the embedding dimension of the model
+
+        Returns:
+            The embedding dimension in int
+        """
+        pass