nexaai 1.0.21rc16__cp312-cp312-win_arm64.whl

nexaai/utils/model_types.py

@@ -0,0 +1,49 @@
+"""
+Model type mappings for HuggingFace pipeline tags to our internal model types.
+
+This module provides centralized model type mapping functionality to avoid
+circular imports between other utility modules.
+"""
+
+from enum import Enum
+from typing import Dict
+
+
+class ModelTypeMapping(Enum):
+    """Enum for mapping HuggingFace pipeline_tag to our ModelType."""
+    TEXT_GENERATION = ("text-generation", "llm")
+    IMAGE_TEXT_TO_TEXT = ("image-text-to-text", "vlm")
+    ANY_TO_ANY = ("any-to-any", "ata")
+    AUTOMATIC_SPEECH_RECOGNITION = ("automatic-speech-recognition", "asr")
+    
+    def __init__(self, pipeline_tag: str, model_type: str):
+        self.pipeline_tag = pipeline_tag
+        self.model_type = model_type
+
+
+# Create mapping dictionaries from the enum
+PIPELINE_TO_MODEL_TYPE: Dict[str, str] = {
+    mapping.pipeline_tag: mapping.model_type 
+    for mapping in ModelTypeMapping
+}
+
+MODEL_TYPE_TO_PIPELINE: Dict[str, str] = {
+    mapping.model_type: mapping.pipeline_tag 
+    for mapping in ModelTypeMapping
+}
+
+
+def map_pipeline_tag_to_model_type(pipeline_tag: str) -> str:
+    """Map HuggingFace pipeline_tag to our ModelType."""
+    if not pipeline_tag:
+        return "other"
+    
+    return PIPELINE_TO_MODEL_TYPE.get(pipeline_tag, "other")
+
+
+def map_model_type_to_pipeline_tag(model_type: str) -> str:
+    """Reverse map ModelType back to HuggingFace pipeline_tag."""
+    if not model_type:
+        return None
+    
+    return MODEL_TYPE_TO_PIPELINE.get(model_type)

nexaai/utils/progress_tracker.py

@@ -0,0 +1,385 @@
+"""
+Progress tracking utilities for downloads with tqdm integration.
+
+This module provides custom progress tracking classes that can monitor
+download progress with callback support and customizable display options.
+"""
+
+import os
+import sys
+import time
+from typing import Optional, Callable, Dict, Any
+from tqdm.auto import tqdm
+
+
+class CustomProgressTqdm(tqdm):
+    """Custom tqdm that tracks progress but completely hides terminal output."""
+    
+    def __init__(self, *args, **kwargs):
+        # Redirect output to devnull to completely suppress terminal output
+        kwargs['file'] = open(os.devnull, 'w')
+        kwargs['disable'] = False  # Keep enabled for tracking
+        kwargs['leave'] = False  # Don't leave progress bar
+        super().__init__(*args, **kwargs)
+    
+    def display(self, msg=None, pos=None):
+        # Override display to show nothing
+        pass
+    
+    def write(self, s, file=None, end="\n", nolock=False):
+        # Override write to prevent any output
+        pass
+    
+    def close(self):
+        # Override close to avoid printing and properly close devnull
+        if hasattr(self, 'fp') and self.fp and self.fp != sys.stdout and self.fp != sys.stderr:
+            try:
+                self.fp.close()
+            except:
+                pass
+        self.disable = True
+        super(tqdm, self).close()
+
+
+class DownloadProgressTracker:
+    """Progress tracker for HuggingFace downloads with callback support."""
+    
+    def __init__(self, progress_callback: Optional[Callable[[Dict[str, Any]], None]] = None, show_progress: bool = True):
+        self.progress_data: Dict[str, Dict[str, Any]] = {}
+        self.total_repo_size = 0
+        self.repo_file_count = 0
+        self.original_tqdm_update = None
+        self.original_tqdm_init = None
+        self.original_tqdm_display = None
+        self.original_tqdm_write = None
+        self.is_tracking = False
+        
+        # Callback function
+        self.progress_callback = progress_callback
+        
+        # Progress display
+        self.show_progress = show_progress
+        self.last_display_length = 0
+        
+        # Speed tracking
+        self.last_downloaded = None  # Use None to indicate no previous measurement
+        self.last_time = None  # Use None to indicate no previous time measurement
+        self.speed_history = []
+        self.max_speed_history = 10
+        
+        # Download status
+        self.download_status = "idle"  # idle, downloading, completed, error
+        self.error_message = None
+        self.download_start_time = None
+    
+    def set_repo_info(self, total_size: int, file_count: int):
+        """Set the total repository size and file count before download."""
+        self.total_repo_size = total_size
+        self.repo_file_count = file_count
+    
+    def register_tqdm(self, tqdm_instance):
+        """Register a tqdm instance for monitoring."""
+        tqdm_id = str(id(tqdm_instance))
+        self.progress_data[tqdm_id] = {
+            'current': 0,
+            'total': getattr(tqdm_instance, 'total', 0) or 0,
+            'desc': getattr(tqdm_instance, 'desc', 'Unknown'),
+            'tqdm_obj': tqdm_instance
+        }
+        # Trigger callback when new file is registered
+        self._trigger_callback()
+    
+    def update_progress(self, tqdm_instance, n=1):
+        """Update progress for a tqdm instance."""
+        tqdm_id = str(id(tqdm_instance))
+        if tqdm_id in self.progress_data:
+            self.progress_data[tqdm_id]['current'] = getattr(tqdm_instance, 'n', 0)
+            self.progress_data[tqdm_id]['total'] = getattr(tqdm_instance, 'total', 0) or 0
+            # Trigger callback on every progress update
+            self._trigger_callback()
+    
+    def calculate_speed(self, current_downloaded: int) -> float:
+        """Calculate download speed in bytes per second."""
+        current_time = time.time()
+        
+        # Check if we have a previous measurement to compare against
+        if self.last_time is not None and self.last_downloaded is not None:
+            time_diff = current_time - self.last_time
+            
+            # Only calculate if we have a meaningful time difference (avoid division by very small numbers)
+            if time_diff > 0.1:  # At least 100ms between measurements
+                bytes_diff = current_downloaded - self.last_downloaded
+                
+                # Only calculate speed if bytes actually changed
+                if bytes_diff >= 0:  # Allow 0 for periods with no progress
+                    speed = bytes_diff / time_diff
+                    
+                    # Add to speed history for smoothing
+                    self.speed_history.append(speed)
+                    if len(self.speed_history) > self.max_speed_history:
+                        self.speed_history.pop(0)
+                    
+                    # Update tracking variables when we actually calculate speed
+                    self.last_downloaded = current_downloaded
+                    self.last_time = current_time
+        else:
+            # First measurement - initialize tracking variables
+            self.last_downloaded = current_downloaded
+            self.last_time = current_time
+        
+        # Return the average of historical speeds if we have any
+        # This ensures we show the last known speed even when skipping updates
+        if self.speed_history:
+            return sum(self.speed_history) / len(self.speed_history)
+        
+        return 0.0
+    
+    def format_bytes(self, bytes_value: int) -> str:
+        """Format bytes to human readable string."""
+        for unit in ['B', 'KB', 'MB', 'GB']:
+            if bytes_value < 1024.0:
+                return f"{bytes_value:.1f} {unit}"
+            bytes_value /= 1024.0
+        return f"{bytes_value:.1f} TB"
+    
+    def format_speed(self, speed: float) -> str:
+        """Format speed to human readable string."""
+        if speed == 0:
+            return "0 B/s"
+        
+        for unit in ['B/s', 'KB/s', 'MB/s', 'GB/s']:
+            if speed < 1024.0:
+                return f"{speed:.1f} {unit}"
+            speed /= 1024.0
+        return f"{speed:.1f} TB/s"
+    
+    def get_progress_data(self) -> Dict[str, Any]:
+        """Get current progress data."""
+        total_downloaded = 0
+        active_file_count = 0
+        total_file_sizes = 0
+        
+        for data in self.progress_data.values():
+            if data['total'] > 0:
+                total_downloaded += data['current']
+                total_file_sizes += data['total']
+                active_file_count += 1
+        
+        # Calculate speed (tracking variables are updated internally)
+        speed = self.calculate_speed(total_downloaded)
+        
+        # Determine total size - prioritize pre-fetched repo size, then aggregate file sizes
+        if self.total_repo_size > 0:
+            # Use pre-fetched repository info if available
+            total_size = self.total_repo_size
+        elif total_file_sizes > 0:
+            # Use sum of individual file sizes if available
+            total_size = total_file_sizes
+        else:
+            # Last resort - we don't know the total size yet
+            total_size = 0
+        
+        file_count = self.repo_file_count if self.repo_file_count > 0 else active_file_count
+        
+        # Calculate percentage - handle unknown total size gracefully
+        if total_size > 0:
+            percentage = min((total_downloaded / total_size * 100), 100.0)
+        else:
+            percentage = 0
+        
+        # Calculate ETA
+        eta_seconds = None
+        if speed > 0 and total_size > total_downloaded:
+            eta_seconds = (total_size - total_downloaded) / speed
+        
+        # Calculate elapsed time
+        elapsed_seconds = None
+        if self.download_start_time:
+            elapsed_seconds = time.time() - self.download_start_time
+        
+        return {
+            'status': self.download_status,
+            'error_message': self.error_message,
+            'progress': {
+                'total_downloaded': total_downloaded,
+                'total_size': total_size,
+                'percentage': round(percentage, 2),
+                'files_active': active_file_count,
+                'files_total': file_count,
+                'known_total': total_size > 0
+            },
+            'speed': {
+                'bytes_per_second': speed,
+                'formatted': self.format_speed(speed)
+            },
+            'formatting': {
+                'downloaded': self.format_bytes(total_downloaded),
+                'total_size': self.format_bytes(total_size)
+            },
+            'timing': {
+                'elapsed_seconds': elapsed_seconds,
+                'eta_seconds': eta_seconds,
+                'start_time': self.download_start_time
+            }
+        }
+    
+    def _display_progress_bar(self, progress_data: Dict[str, Any]):
+        """Display a custom unified progress bar."""
+        if not self.show_progress:
+            return
+            
+        # Clear previous line
+        if self.last_display_length > 0:
+            print('\r' + ' ' * self.last_display_length, end='\r')
+        
+        progress_info = progress_data.get('progress', {})
+        speed_info = progress_data.get('speed', {})
+        timing_info = progress_data.get('timing', {})
+        formatting_info = progress_data.get('formatting', {})
+        
+        percentage = progress_info.get('percentage', 0)
+        downloaded = formatting_info.get('downloaded', '0 B')
+        total_size_raw = progress_info.get('total_size', 0)
+        total_size = formatting_info.get('total_size', 'Unknown')
+        speed = speed_info.get('formatted', '0 B/s')
+        known_total = progress_info.get('known_total', False)
+        
+        # Create progress bar
+        bar_width = 30
+        if known_total and total_size_raw > 0:
+            # Known total size - show actual progress
+            filled_width = int(bar_width * min(percentage, 100) / 100)
+            bar = '#' * filled_width + '-' * (bar_width - filled_width)
+        else:
+            # Unknown total size - show animated progress
+            animation_pos = int(time.time() * 2) % bar_width
+            bar = '-' * animation_pos + '#' + '-' * (bar_width - animation_pos - 1)
+        
+        # Format the progress line
+        status = progress_data.get('status', 'unknown')
+        if status == 'downloading':
+            if known_total:
+                progress_line = f"[{bar}] {percentage:.1f}% | {downloaded}/{total_size} | {speed}"
+            else:
+                progress_line = f"[{bar}] {downloaded} | {speed} | Calculating size..."
+        elif status == 'completed':
+            progress_line = f"[{bar}] 100.0% | {downloaded} | Complete!"
+        elif status == 'error':
+            progress_line = f"Error: {progress_data.get('error_message', 'Unknown error')}"
+        else:
+            progress_line = f"Starting download..."
+        
+        # Display and track length for next clear
+        print(progress_line, end='', flush=True)
+        self.last_display_length = len(progress_line)
+    
+    def _clear_progress_bar(self):
+        """Clear the progress bar display."""
+        if self.show_progress and self.last_display_length > 0:
+            print('\r' + ' ' * self.last_display_length, end='\r')
+            print()  # Move to next line
+            self.last_display_length = 0
+    
+    def _trigger_callback(self):
+        """Trigger the progress callback if one is set."""
+        progress_data = self.get_progress_data()
+        
+        if self.progress_callback:
+            try:
+                self.progress_callback(progress_data)
+            except Exception as e:
+                print(f"Error in progress callback: {e}")
+        
+        # Show custom progress bar only if callback is enabled and show_progress is True
+        if self.progress_callback and self.show_progress:
+            self._display_progress_bar(progress_data)
+    
+    def start_tracking(self):
+        """Start progress tracking (monkey patch tqdm)."""
+        if self.is_tracking:
+            return
+        
+        # Store original methods
+        self.original_tqdm_update = tqdm.update
+        self.original_tqdm_init = tqdm.__init__
+        self.original_tqdm_display = tqdm.display
+        self.original_tqdm_write = tqdm.write
+        
+        # Create references to self for the nested functions
+        tracker = self
+        
+        def patched_init(self_tqdm, *args, **kwargs):
+            # Suppress tqdm display by redirecting to devnull
+            kwargs['file'] = open(os.devnull, 'w')
+            kwargs['disable'] = False  # Keep enabled for tracking
+            kwargs['leave'] = False  # Don't leave progress bar
+            
+            result = tracker.original_tqdm_init(self_tqdm, *args, **kwargs)
+            tracker.register_tqdm(self_tqdm)
+            return result
+        
+        def patched_update(self_tqdm, n=1):
+            result = tracker.original_tqdm_update(self_tqdm, n)
+            tracker.update_progress(self_tqdm, n)
+            return result
+        
+        def patched_display(self_tqdm, msg=None, pos=None):
+            # Override display to show nothing
+            pass
+        
+        def patched_write(self_tqdm, s, file=None, end="\n", nolock=False):
+            # Override write to prevent any output
+            pass
+        
+        # Apply patches
+        tqdm.__init__ = patched_init
+        tqdm.update = patched_update
+        tqdm.display = patched_display
+        tqdm.write = patched_write
+        
+        self.is_tracking = True
+        self.download_status = "downloading"
+        self.download_start_time = time.time()
+        
+        # Trigger initial callback
+        self._trigger_callback()
+    
+    def stop_tracking(self):
+        """Stop progress tracking and restore original tqdm."""
+        if not self.is_tracking:
+            return
+        
+        # Restore original tqdm methods
+        if self.original_tqdm_update:
+            tqdm.update = self.original_tqdm_update
+        if self.original_tqdm_init:
+            tqdm.__init__ = self.original_tqdm_init
+        if hasattr(self, 'original_tqdm_display') and self.original_tqdm_display:
+            tqdm.display = self.original_tqdm_display
+        if hasattr(self, 'original_tqdm_write') and self.original_tqdm_write:
+            tqdm.write = self.original_tqdm_write
+        
+        # Clean up any open devnull file handles from tqdm instances
+        for data in self.progress_data.values():
+            if 'tqdm_obj' in data and hasattr(data['tqdm_obj'], 'fp'):
+                try:
+                    fp = data['tqdm_obj'].fp
+                    if fp and fp != sys.stdout and fp != sys.stderr and not fp.closed:
+                        fp.close()
+                except:
+                    pass
+        
+        self.is_tracking = False
+        if self.download_status == "downloading":
+            self.download_status = "completed"
+        
+        # Trigger final callback and clear progress bar
+        self._trigger_callback()
+        self._clear_progress_bar()
+    
+    def set_error(self, error_message: str):
+        """Set error status and trigger callback."""
+        self.download_status = "error"
+        self.error_message = error_message
+        self._trigger_callback()
+        self._clear_progress_bar()

nexaai/utils/quantization_utils.py

@@ -0,0 +1,245 @@
+"""
+Quantization utilities for extracting quantization types from model files and configurations.
+
+This module provides utilities to extract quantization information from:
+- GGUF model filenames
+- MLX model repository IDs  
+- MLX model config.json files
+"""
+
+import os
+import json
+import re
+import logging
+from enum import Enum
+from typing import Optional
+
+# Set up logger
+logger = logging.getLogger(__name__)
+
+
+class QuantizationType(str, Enum):
+    """Enum for GGUF and MLX model quantization types."""
+    # GGUF quantization types
+    BF16 = "BF16"
+    F16 = "F16"
+    Q2_K = "Q2_K"
+    Q2_K_L = "Q2_K_L"
+    Q3_K = "Q3_K"
+    Q3_K_M = "Q3_K_M"
+    Q3_K_S = "Q3_K_S"
+    Q4_0 = "Q4_0"
+    Q4_1 = "Q4_1"
+    Q4_K = "Q4_K"
+    Q4_K_M = "Q4_K_M"
+    Q4_K_S = "Q4_K_S"
+    Q5_K = "Q5_K"
+    Q5_K_M = "Q5_K_M"
+    Q5_K_S = "Q5_K_S"
+    Q6_K = "Q6_K"
+    Q8_0 = "Q8_0"
+    MXFP4 = "MXFP4"
+    MXFP8 = "MXFP8"
+    
+    # MLX bit-based quantization types
+    BIT_1 = "1BIT"
+    BIT_2 = "2BIT"
+    BIT_3 = "3BIT"
+    BIT_4 = "4BIT"
+    BIT_5 = "5BIT"
+    BIT_6 = "6BIT"
+    BIT_7 = "7BIT"
+    BIT_8 = "8BIT"
+    BIT_16 = "16BIT"
+
+
+def extract_quantization_from_filename(filename: str) -> Optional[QuantizationType]:
+    """
+    Extract quantization type from filename.
+    
+    Args:
+        filename: The filename to extract quantization from
+        
+    Returns:
+        QuantizationType enum value or None if not found
+    """
+    # Define mapping from lowercase patterns to enum values
+    # Include "." to ensure precise matching (e.g., "q4_0." not "q4_0_xl")
+    pattern_to_enum = {
+        'bf16.': QuantizationType.BF16,
+        'f16.': QuantizationType.F16,  # Add F16 support
+        'q2_k_l.': QuantizationType.Q2_K_L,  # Check Q2_K_L before Q2_K to avoid partial match
+        'q2_k.': QuantizationType.Q2_K,
+        'q3_k.': QuantizationType.Q3_K,
+        'q3_k_m.': QuantizationType.Q3_K_M,
+        'q3_k_s.': QuantizationType.Q3_K_S,
+        'q4_k_m.': QuantizationType.Q4_K_M,
+        'q4_k_s.': QuantizationType.Q4_K_S,
+        'q4_0.': QuantizationType.Q4_0,
+        'q4_1.': QuantizationType.Q4_1,
+        'q4_k.': QuantizationType.Q4_K,
+        'q5_k.': QuantizationType.Q5_K,
+        'q5_k_m.': QuantizationType.Q5_K_M,
+        'q5_k_s.': QuantizationType.Q5_K_S,
+        'q6_k.': QuantizationType.Q6_K,
+        'q8_0.': QuantizationType.Q8_0,
+        'mxfp4.': QuantizationType.MXFP4,
+        'mxfp8.': QuantizationType.MXFP8,
+    }
+    
+    filename_lower = filename.lower()
+    
+    # Check longer patterns first to avoid partial matches
+    # Sort by length descending to check q2_k_l before q2_k, q4_k_m before q4_0, etc.
+    for pattern in sorted(pattern_to_enum.keys(), key=len, reverse=True):
+        if pattern in filename_lower:
+            return pattern_to_enum[pattern]
+    
+    return None
+
+
+def extract_quantization_from_repo_id(repo_id: str) -> Optional[QuantizationType]:
+    """
+    Extract quantization type from repo_id for MLX models by looking for bit patterns.
+    
+    Args:
+        repo_id: The repository ID to extract quantization from
+        
+    Returns:
+        QuantizationType enum value or None if not found
+    """
+    # Define mapping from bit numbers to enum values
+    bit_to_enum = {
+        1: QuantizationType.BIT_1,
+        2: QuantizationType.BIT_2,
+        3: QuantizationType.BIT_3,
+        4: QuantizationType.BIT_4,
+        5: QuantizationType.BIT_5,
+        6: QuantizationType.BIT_6,
+        7: QuantizationType.BIT_7,
+        8: QuantizationType.BIT_8,
+        16: QuantizationType.BIT_16,
+    }
+    
+    # First check for patterns like "4bit", "8bit" etc. (case insensitive)
+    pattern = r'(\d+)bit'
+    matches = re.findall(pattern, repo_id.lower())
+    
+    for match in matches:
+        try:
+            bit_number = int(match)
+            if bit_number in bit_to_enum:
+                logger.debug(f"Found {bit_number}bit quantization in repo_id: {repo_id}")
+                return bit_to_enum[bit_number]
+        except ValueError:
+            continue
+    
+    # Also check for patterns like "-q8", "_Q4" etc.
+    q_pattern = r'[-_]q(\d+)'
+    q_matches = re.findall(q_pattern, repo_id.lower())
+    
+    for match in q_matches:
+        try:
+            bit_number = int(match)
+            if bit_number in bit_to_enum:
+                logger.debug(f"Found Q{bit_number} quantization in repo_id: {repo_id}")
+                return bit_to_enum[bit_number]
+        except ValueError:
+            continue
+    
+    return None
+
+
+def extract_quantization_from_mlx_config(mlx_folder_path: str) -> Optional[QuantizationType]:
+    """
+    Extract quantization type from MLX model's config.json file.
+    
+    Args:
+        mlx_folder_path: Path to the MLX model folder
+        
+    Returns:
+        QuantizationType enum value or None if not found
+    """
+    config_path = os.path.join(mlx_folder_path, "config.json")
+    
+    if not os.path.exists(config_path):
+        logger.debug(f"Config file not found: {config_path}")
+        return None
+    
+    try:
+        with open(config_path, 'r', encoding='utf-8') as f:
+            config = json.load(f)
+        
+        # Look for quantization.bits field
+        quantization_config = config.get("quantization", {})
+        if isinstance(quantization_config, dict):
+            bits = quantization_config.get("bits")
+            if isinstance(bits, int):
+                # Define mapping from bit numbers to enum values
+                bit_to_enum = {
+                    1: QuantizationType.BIT_1,
+                    2: QuantizationType.BIT_2,
+                    3: QuantizationType.BIT_3,
+                    4: QuantizationType.BIT_4,
+                    5: QuantizationType.BIT_5,
+                    6: QuantizationType.BIT_6,
+                    7: QuantizationType.BIT_7,
+                    8: QuantizationType.BIT_8,
+                    16: QuantizationType.BIT_16,
+                }
+                
+                if bits in bit_to_enum:
+                    logger.debug(f"Found {bits}bit quantization in config.json: {config_path}")
+                    return bit_to_enum[bits]
+                else:
+                    logger.debug(f"Unsupported quantization bits value: {bits}")
+        
+    except (json.JSONDecodeError, IOError) as e:
+        logger.warning(f"Error reading config.json from {config_path}: {e}")
+    except Exception as e:
+        logger.warning(f"Unexpected error reading config.json from {config_path}: {e}")
+    
+    return None
+
+
+def extract_gguf_quantization(filename: str) -> str:
+    """
+    Extract quantization level from GGUF filename using the enum-based approach.
+    
+    This function provides backward compatibility by returning a string representation
+    of the quantization type.
+    
+    Args:
+        filename: The GGUF filename
+        
+    Returns:
+        String representation of the quantization type or "UNKNOWN" if not found
+    """
+    quantization_type = extract_quantization_from_filename(filename)
+    if quantization_type:
+        return quantization_type.value
+    return "UNKNOWN"
+
+
+def detect_quantization_for_mlx(repo_id: str, directory_path: str) -> Optional[QuantizationType]:
+    """
+    Detect quantization for MLX models using multiple methods in priority order.
+    
+    Args:
+        repo_id: The repository ID
+        directory_path: Path to the model directory
+        
+    Returns:
+        QuantizationType enum value or None if not found
+    """
+    # Method 1: Extract from repo_id
+    quantization_type = extract_quantization_from_repo_id(repo_id)
+    if quantization_type:
+        return quantization_type
+    
+    # Method 2: Extract from config.json if available
+    quantization_type = extract_quantization_from_mlx_config(directory_path)
+    if quantization_type:
+        return quantization_type
+    
+    return None