nexaai 1.0.17rc9__cp310-cp310-win_amd64.whl → 1.0.19__cp310-cp310-win_amd64.whl

nexaai/_version.py

@@ -1,4 +1,4 @@
 # This file is generated by CMake from _version.py.in
 # Do not modify this file manually - it will be overwritten
 
-__version__ = "1.0.17-rc9" 
+__version__ = "1.0.19" 

nexaai/asr.py

@@ -35,7 +35,8 @@ class ASR(BaseModel):
                    tokenizer_path: Optional[str] = None,
                    language: Optional[str] = None,
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'ASR':
         """Load ASR model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/cv.py

@@ -73,7 +73,8 @@ class CVModel(BaseModel):
                    _: 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
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'CVModel':
         """Load CV model from configuration, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/embedder.py

@@ -22,7 +22,7 @@ class Embedder(BaseModel):
         pass
 
     @classmethod
-    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP):
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP, **kwargs):
         """
         Load an embedder from model files, routing to appropriate implementation.
 

nexaai/image_gen.py

@@ -71,7 +71,8 @@ class ImageGen(BaseModel):
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None,
                    float16: bool = True,
-                   quantize: bool = False
+                   quantize: bool = False,
+                   **kwargs
         ) -> 'ImageGen':
         """Load image generation model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/llm.py

@@ -15,10 +15,12 @@ class LLM(BaseModel):
     @classmethod
     def _load_from(cls,
                    local_path: str,
+                   model_name: Optional[str] = None,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'LLM':
         """Load model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string
@@ -26,10 +28,10 @@ class LLM(BaseModel):
         
         if plugin_value == "mlx":
             from nexaai.llm_impl.mlx_llm_impl import MLXLLMImpl
-            return MLXLLMImpl._load_from(local_path, tokenizer_path, m_cfg, plugin_id, device_id)
+            return MLXLLMImpl._load_from(local_path, model_name, tokenizer_path, m_cfg, plugin_id, device_id)
         else:
             from nexaai.llm_impl.pybind_llm_impl import PyBindLLMImpl
-            return PyBindLLMImpl._load_from(local_path, tokenizer_path, m_cfg, plugin_id, device_id)
+            return PyBindLLMImpl._load_from(local_path, model_name, tokenizer_path, m_cfg, plugin_id, device_id)
 
     def cancel_generation(self):
         """Signal to cancel any ongoing stream generation."""

nexaai/llm_impl/mlx_llm_impl.py

@@ -16,6 +16,7 @@ class MLXLLMImpl(LLM):
     @classmethod
     def _load_from(cls,
                    local_path: str,
+                   model_name: Optional[str] = None,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.MLX,
@@ -40,6 +41,7 @@ class MLXLLMImpl(LLM):
             instance = cls(m_cfg)
             instance._mlx_llm = MLXLLMInterface(
                 model_path=local_path,
+                # model_name=model_name, # FIXME: For MLX LLM, model_name is not used
                 tokenizer_path=tokenizer_path or local_path,
                 config=mlx_config,
                 device=device_id

nexaai/llm_impl/pybind_llm_impl.py

@@ -19,6 +19,7 @@ class PyBindLLMImpl(LLM):
     @classmethod
     def _load_from(cls,
                    local_path: str,
+                   model_name: Optional[str] = None,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
@@ -55,6 +56,7 @@ class PyBindLLMImpl(LLM):
         plugin_id_str = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
         handle = llm_bind.ml_llm_create(
             model_path=local_path,
+            model_name=model_name,
             tokenizer_path=tokenizer_path,
             model_config=config,
             plugin_id=plugin_id_str,

nexaai/rerank.py

@@ -26,7 +26,8 @@ class Reranker(BaseModel):
                    model_path: str,
                    tokenizer_file: str = "tokenizer.json",
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'Reranker':
         """Load reranker model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/tts.py

@@ -45,7 +45,8 @@ class TTS(BaseModel):
                    model_path: str,
                    vocoder_path: str,
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'TTS':
         """Load TTS model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/utils/manifest_utils.py

@@ -22,6 +22,11 @@ from .model_types import (
     MODEL_TYPE_TO_PIPELINE
 )
 
+MODEL_FILE_TYPE_TO_PLUGIN_ID_MAPPING = {
+    'npu': 'npu',
+    'mlx': 'mlx',
+    'gguf': 'llama_cpp'
+}
 
 def process_manifest_metadata(manifest: Dict[str, Any], repo_id: str) -> Dict[str, Any]:
     """Process manifest metadata to handle null/missing fields."""
@@ -94,12 +99,20 @@ def save_download_metadata(directory_path: str, metadata: Dict[str, Any]) -> Non
         pass
 
 
-def create_gguf_manifest(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None) -> Dict[str, Any]:
+def _get_plugin_id_from_model_file_type(model_file_type: Optional[str], default: str = "llama_cpp") -> str:
+    """Map model file type to PluginId."""
+    return MODEL_FILE_TYPE_TO_PLUGIN_ID_MAPPING.get(model_file_type, default)
+
+
+def create_gguf_manifest(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> Dict[str, Any]:
     """Create GGUF format manifest."""
     
     # Load existing manifest to merge GGUF files if it exists
     existing_manifest = load_nexa_manifest(directory_path)
     
+    # Check if there's a downloaded nexa.manifest from the repo
+    downloaded_manifest = old_metadata.get('downloaded_manifest', {})
+    
     model_files = {}
     if existing_manifest and "ModelFile" in existing_manifest:
         model_files = existing_manifest["ModelFile"].copy()
@@ -151,10 +164,41 @@ def create_gguf_manifest(repo_id: str, files: List[str], directory_path: str, ol
                     "Size": file_size
                 }
     
+    # Determine PluginId with priority: kwargs > downloaded_manifest > model_file_type > default
+    plugin_id = kwargs.get('plugin_id')
+    if not plugin_id:
+        model_file_type = old_metadata.get('model_file_type')
+        if downloaded_manifest.get('PluginId'):
+            plugin_id = downloaded_manifest.get('PluginId')
+        elif model_file_type:
+            plugin_id = _get_plugin_id_from_model_file_type(model_file_type)
+        else:
+            plugin_id = "llama_cpp"
+    
+    # Determine ModelType with priority: kwargs > downloaded_manifest > pipeline_tag mapping
+    model_type = kwargs.get('model_type')
+    if not model_type:
+        if downloaded_manifest.get('ModelType'):
+            model_type = downloaded_manifest.get('ModelType')
+        else:
+            model_type = PIPELINE_TO_MODEL_TYPE.get(old_metadata.get('pipeline_tag'), "other")
+    
+    # Determine ModelName with priority: kwargs > downloaded_manifest > empty string
+    model_name = kwargs.get('model_name')
+    if not model_name:
+        model_name = downloaded_manifest.get('ModelName', '')
+    
+    # Get DeviceId and MinSDKVersion from kwargs or default to empty string
+    device_id = kwargs.get('device_id', '')
+    min_sdk_version = kwargs.get('min_sdk_version', '')
+    
     manifest = {
         "Name": repo_id,
-        "ModelType": PIPELINE_TO_MODEL_TYPE.get(old_metadata.get('pipeline_tag'), "other"),
-        "PluginId": "llama_cpp",
+        "ModelName": model_name,
+        "ModelType": model_type,
+        "PluginId": plugin_id,
+        "DeviceId": device_id,
+        "MinSDKVersion": min_sdk_version,
         "ModelFile": model_files,
         "MMProjFile": mmproj_file,
         "TokenizerFile": {
@@ -172,12 +216,15 @@ def create_gguf_manifest(repo_id: str, files: List[str], directory_path: str, ol
     return manifest
 
 
-def create_mlx_manifest(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None) -> Dict[str, Any]:
+def create_mlx_manifest(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> Dict[str, Any]:
     """Create MLX format manifest."""
     
     # Load existing manifest to merge MLX files if it exists
     existing_manifest = load_nexa_manifest(directory_path)
     
+    # Check if there's a downloaded nexa.manifest from the repo
+    downloaded_manifest = old_metadata.get('downloaded_manifest', {})
+    
     model_files = {}
     extra_files = []
     
@@ -233,10 +280,153 @@ def create_mlx_manifest(repo_id: str, files: List[str], directory_path: str, old
                 "Size": file_size
             })
     
+    # Determine PluginId with priority: kwargs > downloaded_manifest > model_file_type > default
+    plugin_id = kwargs.get('plugin_id')
+    if not plugin_id:
+        model_file_type = old_metadata.get('model_file_type')
+        if downloaded_manifest.get('PluginId'):
+            plugin_id = downloaded_manifest.get('PluginId')
+        elif model_file_type:
+            plugin_id = _get_plugin_id_from_model_file_type(model_file_type)
+        else:
+            plugin_id = "mlx"
+    
+    # Determine ModelType with priority: kwargs > downloaded_manifest > pipeline_tag mapping
+    model_type = kwargs.get('model_type')
+    if not model_type:
+        if downloaded_manifest.get('ModelType'):
+            model_type = downloaded_manifest.get('ModelType')
+        else:
+            model_type = PIPELINE_TO_MODEL_TYPE.get(old_metadata.get('pipeline_tag'), "other")
+    
+    # Determine ModelName with priority: kwargs > downloaded_manifest > empty string
+    model_name = kwargs.get('model_name')
+    if not model_name:
+        model_name = downloaded_manifest.get('ModelName', '')
+    
+    # Get DeviceId and MinSDKVersion from kwargs or default to empty string
+    device_id = kwargs.get('device_id', '')
+    min_sdk_version = kwargs.get('min_sdk_version', '')
+    
+    manifest = {
+        "Name": repo_id,
+        "ModelName": model_name,
+        "ModelType": model_type,
+        "PluginId": plugin_id,
+        "DeviceId": device_id,
+        "MinSDKVersion": min_sdk_version,
+        "ModelFile": model_files,
+        "MMProjFile": mmproj_file,
+        "TokenizerFile": {
+            "Name": "",
+            "Downloaded": False,
+            "Size": 0
+        },
+        "ExtraFiles": extra_files if extra_files else None,
+        # Preserve old metadata fields
+        "pipeline_tag": old_metadata.get('pipeline_tag') if old_metadata.get('pipeline_tag') else existing_manifest.get('pipeline_tag'),
+        "download_time": old_metadata.get('download_time') if old_metadata.get('download_time') else existing_manifest.get('download_time'),
+        "avatar_url": old_metadata.get('avatar_url') if old_metadata.get('avatar_url') else existing_manifest.get('avatar_url')
+    }
+    
+    return manifest
+
+
+def create_npu_manifest(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> Dict[str, Any]:
+    """Create NPU format manifest."""
+    
+    # Load existing manifest to merge NPU files if it exists
+    existing_manifest = load_nexa_manifest(directory_path)
+    
+    # Check if there's a downloaded nexa.manifest from the repo
+    downloaded_manifest = old_metadata.get('downloaded_manifest', {})
+    
+    model_files = {}
+    extra_files = []
+    
+    # Initialize MMProjFile
+    mmproj_file = {
+        "Name": "",
+        "Downloaded": False,
+        "Size": 0
+    }
+    
+    for current_file_name in files:
+        file_path = os.path.join(directory_path, current_file_name)
+        file_size = 0
+        if os.path.exists(file_path):
+            try:
+                file_size = os.path.getsize(file_path)
+            except (OSError, IOError):
+                pass
+        
+        # Check if this file is an mmproj file
+        is_current_mmproj = 'mmproj' in current_file_name.lower()
+        
+        # If we're downloading specific files and this is marked as mmproj, respect that
+        if is_mmproj and file_name is not None:
+            filenames_to_check = file_name if isinstance(file_name, list) else [file_name]
+            is_current_mmproj = current_file_name in filenames_to_check
+        
+        if is_current_mmproj:
+            # This is an mmproj file, put it in MMProjFile
+            mmproj_file = {
+                "Name": current_file_name,
+                "Downloaded": True,
+                "Size": file_size
+            }
+        else:
+            # For NPU, all non-mmproj files go to extra_files
+            extra_files.append({
+                "Name": current_file_name,
+                "Downloaded": True,
+                "Size": file_size
+            })
+    
+    # Pick the first file from extra_files and add it to ModelFile with key "N/A"
+    if extra_files:
+        first_file = extra_files[0]
+        model_files["N/A"] = {
+            "Name": first_file["Name"],
+            "Downloaded": first_file["Downloaded"],
+            "Size": first_file["Size"]
+        }
+    
+    # Determine PluginId with priority: kwargs > downloaded_manifest > model_file_type > default
+    plugin_id = kwargs.get('plugin_id')
+    if not plugin_id:
+        model_file_type = old_metadata.get('model_file_type')
+        if downloaded_manifest.get('PluginId'):
+            plugin_id = downloaded_manifest.get('PluginId')
+        elif model_file_type:
+            plugin_id = _get_plugin_id_from_model_file_type(model_file_type)
+        else:
+            plugin_id = "npu"
+    
+    # Determine ModelType with priority: kwargs > downloaded_manifest > pipeline_tag mapping
+    model_type = kwargs.get('model_type')
+    if not model_type:
+        if downloaded_manifest.get('ModelType'):
+            model_type = downloaded_manifest.get('ModelType')
+        else:
+            model_type = PIPELINE_TO_MODEL_TYPE.get(old_metadata.get('pipeline_tag'), "other")
+    
+    # Determine ModelName with priority: kwargs > downloaded_manifest > empty string
+    model_name = kwargs.get('model_name')
+    if not model_name:
+        model_name = downloaded_manifest.get('ModelName', '')
+    
+    # Get DeviceId and MinSDKVersion from kwargs or default to empty string
+    device_id = kwargs.get('device_id', '')
+    min_sdk_version = kwargs.get('min_sdk_version', '')
+    
     manifest = {
         "Name": repo_id,
-        "ModelType": PIPELINE_TO_MODEL_TYPE.get(old_metadata.get('pipeline_tag'), "other"),
-        "PluginId": "mlx",
+        "ModelName": model_name,
+        "ModelType": model_type,
+        "PluginId": plugin_id,
+        "DeviceId": device_id,
+        "MinSDKVersion": min_sdk_version,
         "ModelFile": model_files,
         "MMProjFile": mmproj_file,
         "TokenizerFile": {
@@ -254,8 +444,21 @@ def create_mlx_manifest(repo_id: str, files: List[str], directory_path: str, old
     return manifest
 
 
-def detect_model_type(files: List[str]) -> str:
-    """Detect if this is a GGUF or MLX model based on file extensions."""
+def detect_model_type(files: List[str], old_metadata: Dict[str, Any] = None) -> str:
+    """Detect if this is a GGUF, MLX, or NPU model based on file extensions and metadata.
+    
+    Args:
+        files: List of files in the model directory
+        old_metadata: Metadata dict that may contain 'model_file_type'
+        
+    Returns:
+        Model type string: 'gguf', 'mlx', or 'npu'
+    """
+    # Check if model_file_type is explicitly set to NPU
+    if old_metadata and old_metadata.get('model_file_type') == 'npu':
+        return "npu"
+    
+    # Otherwise, detect based on file extensions
     has_gguf = any(f.endswith('.gguf') for f in files)
     has_safetensors = any(f.endswith('.safetensors') or 'safetensors' in f for f in files)
     
@@ -268,7 +471,7 @@ def detect_model_type(files: List[str]) -> str:
         return "mlx"
 
 
-def create_manifest_from_files(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None) -> Dict[str, Any]:
+def create_manifest_from_files(repo_id: str, files: List[str], directory_path: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> Dict[str, Any]:
     """
     Create appropriate manifest format based on detected model type.
     
@@ -276,22 +479,25 @@ def create_manifest_from_files(repo_id: str, files: List[str], directory_path: s
         repo_id: Repository ID
         files: List of files in the model directory
         directory_path: Path to the model directory
-        old_metadata: Existing metadata (pipeline_tag, download_time, avatar_url)
+        old_metadata: Existing metadata (pipeline_tag, download_time, avatar_url, model_file_type)
         is_mmproj: Whether the downloaded file is an mmproj file
         file_name: The specific file(s) that were downloaded (None if entire repo was downloaded)
+        **kwargs: Additional metadata including plugin_id, model_name, model_type, device_id, min_sdk_version
         
     Returns:
         Dict containing the appropriate manifest format
     """
-    model_type = detect_model_type(files)
+    model_type = detect_model_type(files, old_metadata)
     
     if model_type == "gguf":
-        return create_gguf_manifest(repo_id, files, directory_path, old_metadata, is_mmproj, file_name)
+        return create_gguf_manifest(repo_id, files, directory_path, old_metadata, is_mmproj, file_name, **kwargs)
+    elif model_type == "npu":
+        return create_npu_manifest(repo_id, files, directory_path, old_metadata, is_mmproj, file_name, **kwargs)
     else:  # mlx or other
-        return create_mlx_manifest(repo_id, files, directory_path, old_metadata, is_mmproj, file_name)
+        return create_mlx_manifest(repo_id, files, directory_path, old_metadata, is_mmproj, file_name, **kwargs)
 
 
-def save_manifest_with_files_metadata(repo_id: str, local_dir: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None) -> None:
+def save_manifest_with_files_metadata(repo_id: str, local_dir: str, old_metadata: Dict[str, Any], is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> None:
     """
     Create and save manifest based on files found in the directory.
     
@@ -301,6 +507,7 @@ def save_manifest_with_files_metadata(repo_id: str, local_dir: str, old_metadata
         old_metadata: Existing metadata to preserve
         is_mmproj: Whether the downloaded file is an mmproj file
         file_name: The specific file(s) that were downloaded (None if entire repo was downloaded)
+        **kwargs: Additional metadata including plugin_id, model_name, model_type, device_id, min_sdk_version
     """
     # Get list of files in the directory
     files = []
@@ -314,7 +521,7 @@ def save_manifest_with_files_metadata(repo_id: str, local_dir: str, old_metadata
         pass
     
     # Create appropriate manifest
-    manifest = create_manifest_from_files(repo_id, files, local_dir, old_metadata, is_mmproj, file_name)
+    manifest = create_manifest_from_files(repo_id, files, local_dir, old_metadata, is_mmproj, file_name, **kwargs)
     
     # Save manifest
     save_download_metadata(local_dir, manifest)

nexaai/utils/model_manager.py

@@ -595,6 +595,7 @@ class HuggingFaceDownloader:
         self.enable_transfer = enable_transfer
         self.original_hf_transfer = None
         self.endpoint = endpoint  # Store endpoint for avatar fetching
+        self._model_info_cache: Dict[str, Any] = {}  # Cache for model_info results
         
     def _create_repo_directory(self, local_dir: str, repo_id: str) -> str:
         """Create a directory structure for the repository following HF convention."""
@@ -618,6 +619,32 @@ class HuggingFaceDownloader:
         os.makedirs(local_dir, exist_ok=True)
         return local_dir
     
+    def _get_model_info_cached(self, repo_id: str, files_metadata: bool = False):
+        """Get model info with caching to avoid rate limiting.
+        
+        Args:
+            repo_id: Repository ID
+            files_metadata: Whether to include files metadata
+            
+        Returns:
+            Model info object from HuggingFace API
+        """
+        # Create cache key based on repo_id and files_metadata flag
+        cache_key = f"{repo_id}:files={files_metadata}"
+        
+        # Return cached result if available
+        if cache_key in self._model_info_cache:
+            return self._model_info_cache[cache_key]
+        
+        # Fetch from API and cache the result
+        try:
+            info = self.api.model_info(repo_id, files_metadata=files_metadata, token=self.token)
+            self._model_info_cache[cache_key] = info
+            return info
+        except Exception:
+            # Don't cache errors, re-raise
+            raise
+    
     def _get_repo_info_for_progress(
         self,
         repo_id: str,
@@ -625,7 +652,7 @@ class HuggingFaceDownloader:
     ) -> tuple[int, int]:
         """Get total repository size and file count for progress tracking."""
         try:
-            info = self.api.model_info(repo_id, files_metadata=True, token=self.token)
+            info = self._get_model_info_cached(repo_id, files_metadata=True)
             
             total_size = 0
             file_count = 0
@@ -720,7 +747,7 @@ class HuggingFaceDownloader:
     ):
         """Validate repository exists and get info."""
         try:
-            info = self.api.model_info(repo_id, token=self.token)
+            info = self._get_model_info_cached(repo_id, files_metadata=False)
             return info
         except RepositoryNotFoundError:
             error_msg = f"Repository '{repo_id}' not found. Please check the repository ID."
@@ -789,6 +816,36 @@ class HuggingFaceDownloader:
         # If no expected size, just check that file is not empty
         return os.path.getsize(file_path) > 0
     
+    def _extract_model_file_type_from_tags(self, repo_id: str) -> Optional[str]:
+        """Extract model file type from repo tags with priority: NPU > MLX > GGUF."""
+        try:
+            info = self._get_model_info_cached(repo_id, files_metadata=False)
+            if hasattr(info, 'tags') and info.tags:
+                # Convert tags to lowercase for case-insensitive matching
+                tags_lower = [tag.lower() for tag in info.tags]
+                
+                # Check with priority: NPU > MLX > GGUF
+                if 'npu' in tags_lower:
+                    return 'npu'
+                elif 'mlx' in tags_lower:
+                    return 'mlx'
+                elif 'gguf' in tags_lower:
+                    return 'gguf'
+        except Exception:
+            pass
+        return None
+    
+    def _load_downloaded_manifest(self, local_dir: str) -> Dict[str, Any]:
+        """Load nexa.manifest from the downloaded repository if it exists."""
+        manifest_path = os.path.join(local_dir, 'nexa.manifest')
+        if os.path.exists(manifest_path):
+            try:
+                with open(manifest_path, 'r', encoding='utf-8') as f:
+                    return json.load(f)
+            except (json.JSONDecodeError, IOError):
+                pass
+        return {}
+    
     def _fetch_and_save_metadata(self, repo_id: str, local_dir: str, is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> None:
         """Fetch model info and save metadata after successful download."""
         # Initialize metadata with defaults to ensure manifest is always created
@@ -800,8 +857,8 @@ class HuggingFaceDownloader:
         
         # Try to fetch additional metadata, but don't let failures prevent manifest creation
         try:
-            # Fetch model info to get pipeline_tag
-            info = self.api.model_info(repo_id, token=self.token)
+            # Fetch model info to get pipeline_tag (using cache)
+            info = self._get_model_info_cached(repo_id, files_metadata=False)
             if hasattr(info, 'pipeline_tag') and info.pipeline_tag:
                 old_metadata['pipeline_tag'] = info.pipeline_tag
         except Exception as e:
@@ -810,11 +867,21 @@ class HuggingFaceDownloader:
         
         # Use input avater url if provided
         old_metadata['avatar_url'] = kwargs.get('avatar_url')
+        
+        # Extract model file type from tags
+        model_file_type = self._extract_model_file_type_from_tags(repo_id)
+        if model_file_type:
+            old_metadata['model_file_type'] = model_file_type
+        
+        # Load existing nexa.manifest from downloaded repo (if exists)
+        downloaded_manifest = self._load_downloaded_manifest(local_dir)
+        if downloaded_manifest:
+            old_metadata['downloaded_manifest'] = downloaded_manifest
 
         
         # CRITICAL: Always create the manifest file, regardless of metadata fetch failures
         try:
-            save_manifest_with_files_metadata(repo_id, local_dir, old_metadata, is_mmproj, file_name)
+            save_manifest_with_files_metadata(repo_id, local_dir, old_metadata, is_mmproj, file_name, **kwargs)
             print(f"[OK] Successfully created nexa.manifest for {repo_id}")
         except Exception as e:
             # This is critical - if manifest creation fails, we should know about it
@@ -823,8 +890,11 @@ class HuggingFaceDownloader:
             try:
                 minimal_manifest = {
                     "Name": repo_id,
-                    "ModelType": "other",
-                    "PluginId": "unknown",
+                    "ModelName": kwargs.get('model_name', ''),
+                    "ModelType": kwargs.get('model_type', 'other'),
+                    "PluginId": kwargs.get('plugin_id', 'unknown'),
+                    "DeviceId": kwargs.get('device_id', ''),
+                    "MinSDKVersion": kwargs.get('min_sdk_version', ''),
                     "ModelFile": {},
                     "MMProjFile": {"Name": "", "Downloaded": False, "Size": 0},
                     "TokenizerFile": {"Name": "", "Downloaded": False, "Size": 0},
@@ -1136,6 +1206,12 @@ def download_from_huggingface(
         is_mmproj (bool, optional): Whether the file being downloaded is an mmproj file. Only used when
                                    file_name is not None. If None, defaults to True if 'mmproj' is in 
                                    the filename, False otherwise.
+        **kwargs: Additional parameters including:
+            - plugin_id (str): Override PluginId in nexa.manifest (highest priority)
+            - model_name (str): Override ModelName in nexa.manifest (highest priority)
+            - model_type (str): Override ModelType in nexa.manifest (highest priority)
+            - device_id (str): Set DeviceId in nexa.manifest (highest priority)
+            - min_sdk_version (str): Set MinSDKVersion in nexa.manifest (highest priority)
     
     Returns:
         str: Path to the downloaded file or directory
@@ -1213,7 +1289,7 @@ def _download_model_if_needed(
     token: Union[bool, str, None] = None,
     is_mmproj: bool = False,
     **kwargs
-) -> str:
+) -> tuple[str, Optional[str]]:
     """
     Helper function to download a model from HuggingFace if it doesn't exist locally.
     
@@ -1224,15 +1300,38 @@ def _download_model_if_needed(
         token: HuggingFace authentication token for private repositories
         
     Returns:
-        str: Local path to the model (either existing or downloaded)
+        tuple[str, Optional[str]]: Tuple of (local_path, model_name)
+            - local_path: Local path to the model (either existing or downloaded)
+            - model_name: ModelName from nexa.manifest if available, None otherwise
         
     Raises:
         RuntimeError: If download fails
     """
+    # Helper function to extract model_name from manifest
+    def _extract_model_name_from_manifest(path: str) -> Optional[str]:
+        """Extract ModelName from nexa.manifest if it exists."""
+        # If path is a file, check its parent directory for manifest
+        if os.path.isfile(path):
+            manifest_dir = os.path.dirname(path)
+        else:
+            manifest_dir = path
+        
+        manifest_path = os.path.join(manifest_dir, 'nexa.manifest')
+        if not os.path.exists(manifest_path):
+            return None
+        
+        try:
+            with open(manifest_path, 'r', encoding='utf-8') as f:
+                manifest = json.load(f)
+                return manifest.get('ModelName')
+        except (json.JSONDecodeError, IOError):
+            return None
+    
     # Check if model_path exists locally (file or directory)
     if os.path.exists(model_path):
-        # Local path exists, return as-is
-        return model_path
+        # Local path exists, try to extract model_name
+        model_name = _extract_model_name_from_manifest(model_path)
+        return model_path, model_name
     
     # Model path doesn't exist locally, try to download from HuggingFace
     try:
@@ -1252,7 +1351,10 @@ def _download_model_if_needed(
             **kwargs
         )
         
-        return downloaded_path
+        # Extract model_name from the downloaded manifest
+        model_name = _extract_model_name_from_manifest(downloaded_path)
+        
+        return downloaded_path, model_name
         
     except Exception as e:
         # Only handle download-related errors
@@ -1321,7 +1423,7 @@ def auto_download_model(func: Callable) -> Callable:
         # Download name_or_path if needed
         if name_or_path is not None:
             try:
-                downloaded_name_path = _download_model_if_needed(
+                downloaded_name_path, model_name = _download_model_if_needed(
                     name_or_path, 'name_or_path', progress_callback, token, **kwargs
                 )
                 
@@ -1332,6 +1434,10 @@ def auto_download_model(func: Callable) -> Callable:
                         args = tuple(args_list)
                 else:
                     kwargs['name_or_path'] = downloaded_name_path
+                
+                # Add model_name to kwargs if it exists and not already set
+                if model_name is not None and 'model_name' not in kwargs:
+                    kwargs['model_name'] = model_name
                     
             except Exception as e:
                 raise e  # Re-raise the error from _download_model_if_needed
@@ -1339,7 +1445,7 @@ def auto_download_model(func: Callable) -> Callable:
         # Download mmproj_path if needed
         if mmproj_path is not None:
             try:
-                downloaded_mmproj_path = _download_model_if_needed(
+                downloaded_mmproj_path, _ = _download_model_if_needed(
                     mmproj_path, 'mmproj_path', progress_callback, token, is_mmproj=True, **kwargs
                 )
                 

nexaai/utils/model_types.py

@@ -13,6 +13,8 @@ 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

nexaai/vlm.py

@@ -22,7 +22,8 @@ class VLM(BaseModel):
                    model_name: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'VLM':
         """Load VLM model from local path, routing to appropriate implementation.
         

{nexaai-1.0.17rc9.dist-info → nexaai-1.0.19.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.17rc9
+Version: 1.0.19
 Summary: Python bindings for NexaSDK C-lib backend
 Author-email: "Nexa AI, Inc." <dev@nexa.ai>
 Project-URL: Homepage, https://github.com/NexaAI/nexasdk-bridge
@@ -14,7 +14,6 @@ Requires-Python: >=3.7
 Description-Content-Type: text/markdown
 Requires-Dist: huggingface_hub
 Requires-Dist: tqdm
-Requires-Dist: hf_xet
 Requires-Dist: numpy
 Requires-Dist: httpx
 Provides-Extra: mlx

nexaai-1.0.19.dist-info/RECORD

@@ -0,0 +1,76 @@
+nexaai/__init__.py,sha256=mbzzeXrEHHI_E3BQ0_OukD9wNajKJJVk0ykxT0rz8uM,2267
+nexaai/_stub.cp310-win_amd64.pyd,sha256=5dzEltIGjiUo03HnBGj-H8Im7YjeWrrpl8vUyj7gEQ4,10752
+nexaai/_version.py,sha256=1EqJuyxDoSkEvRoBL8vxC0Ww_e-0l8BLCs6fquyOolY,143
+nexaai/asr.py,sha256=eeFMFKsaEKiJ2PCGmgedRR4fTfrhaDcnZBuhCChMAio,2148
+nexaai/base.py,sha256=qQBCiQVNzgpkQjZX9aiFDEdbAAe56TROKC3WnWra2Zg,1021
+nexaai/common.py,sha256=muQqFY-WllwL5IO83tImexbuUcoEQsKg73u4gWL2lNs,3548
+nexaai/cv.py,sha256=PhgqK5hO8FlXLR6QxE_pHtXjdtzLSEPzkCda-qrNpOs,3395
+nexaai/embedder.py,sha256=NdYsmq--KGlLah11OJhhJe8rwAqO3mb-SxKEoRR9vXs,2534
+nexaai/image_gen.py,sha256=_kLLCbvhc30a-YGidkB13ZXYoPOS-sKMo8bKVRkUado,4526
+nexaai/llm.py,sha256=woB98NQmaNf3_BH0Ogig3niU1rcD8YtHPiSpzyrbVmo,3771
+nexaai/log.py,sha256=F_Qe169kLbnFV25WJmS_ZtmBfOdcGic8BYFIsYVoD_o,2720
+nexaai/rerank.py,sha256=26Fizy3YLoJmEswCU4IgG7Ka-NHcu79YZr4IBz1XDQ0,1957
+nexaai/runtime.py,sha256=_BoAtTUv5ZR7wtOlJL5TldR3AZTP0OnMWjB9p71k-8E,2135
+nexaai/tts.py,sha256=5lEzv50Q5iLTNHLg3kX-4J_mi1T_t-rVdZzOwd6sALY,2271
+nexaai/vlm.py,sha256=bmdl215iIFhpFi3zloRkdym4j_z5BDbU6c981I80k-0,4890
+nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/asr_impl/mlx_asr_impl.py,sha256=XwMX3LYMeulp8cDS0TCCYcjvttFHAyDWQ_oMvABwQmI,3349
+nexaai/asr_impl/pybind_asr_impl.py,sha256=20o5SOPzhF9x41ra8L_qIM7YxCkYeLb5csSrNde-dds,1560
+nexaai/binds/__init__.py,sha256=ENl-uoIF9-3XGIXitVgZ2QmJ6p7Yet4h1-X7nUDZ0Hk,108
+nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=N7qfVNi48wmqhKJf1TIMhY0W6GZT14Fo97bQh7OFG40,205824
+nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=mkx4LibQEUY_OP0cc333_Qv_JWHhn4vJc3z7mLGHD0o,182784
+nexaai/binds/libcrypto-3-x64.dll,sha256=X7hDuQPMn6eRe58dTGJxTmRPKFWgd5i6Gjoy9QPLmPg,7315968
+nexaai/binds/libssl-3-x64.dll,sha256=GRw0cOaGUaR2QHRvBFuU_9S9iNgkyVgm0NIpZMnReAw,1313792
+nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=lOusUjV5teUtAk-lq62Y8ha-IZQ__3cb7yA1Tx0BvbA,163840
+nexaai/binds/nexa_bridge.dll,sha256=4x7QWF0PdfvUxf_DBsgac3Z2jZqPKvsNCHHDu0A79pg,187904
+nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=ziwINHPBAwC2WkFOoU5EJOQfKEw3JUJeUZ-jDkkLQho,168960
+nexaai/binds/cpu_gpu/ggml-base.dll,sha256=JBjHN894qYk8Izk0Lsl_aK5OrtU6s6pkVf-mNbE3dSY,532480
+nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=ZoIgujpu9WwK0CJLnz4dlVuEuSKphj3wrR_w9AuCIp8,672768
+nexaai/binds/cpu_gpu/ggml-cuda.dll,sha256=EDQMReypE0cQKj_mVT8UQ0kWRlf5mJ7S5duOuJLQ98w,313528832
+nexaai/binds/cpu_gpu/ggml-vulkan.dll,sha256=UQSw_60Z-d3NtnQgJVORlm5BofWKafXKOT-L19Ayc78,36627968
+nexaai/binds/cpu_gpu/ggml.dll,sha256=hvS-ukn4PDTZhc09MZ7r9mZuIu9q6fYS-rXQXH9ZiXw,66560
+nexaai/binds/cpu_gpu/mtmd.dll,sha256=rLo4waw9B9UawaJUMhiDao-zH12m4QqIAxfW0gW9wC8,561152
+nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=vFFvQp0pqV5blDrPs18ZgpCQ-owSwT1p8HTtf4pGa4M,1611776
+nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=O_06Z508L4Ia0Cd-OCcrMJvfFWtn8zhrUv_hzOLp9qw,1411072
+nexaai/binds/nexaml/ggml-base.dll,sha256=JBjHN894qYk8Izk0Lsl_aK5OrtU6s6pkVf-mNbE3dSY,532480
+nexaai/binds/nexaml/ggml-cpu.dll,sha256=ZoIgujpu9WwK0CJLnz4dlVuEuSKphj3wrR_w9AuCIp8,672768
+nexaai/binds/nexaml/ggml-cuda.dll,sha256=EDQMReypE0cQKj_mVT8UQ0kWRlf5mJ7S5duOuJLQ98w,313528832
+nexaai/binds/nexaml/ggml-vulkan.dll,sha256=UQSw_60Z-d3NtnQgJVORlm5BofWKafXKOT-L19Ayc78,36627968
+nexaai/binds/nexaml/ggml.dll,sha256=hvS-ukn4PDTZhc09MZ7r9mZuIu9q6fYS-rXQXH9ZiXw,66560
+nexaai/binds/nexaml/nexa-mm-process.dll,sha256=DDD5qPbelZAIrRWGggOd3Pg9U_fh_ZEYIplwqiyD5LY,4643328
+nexaai/binds/nexaml/nexa-sampling.dll,sha256=Notkz287laSUG2_ED3oFXWVLJM3t7x_USkvX6wKVOLA,4265984
+nexaai/binds/nexaml/nexa_plugin.dll,sha256=DcIOfUjhXaOfB9SXIEJZMXdZWa12_VCswP3hWch7lVE,600576
+nexaai/binds/nexaml/nexaproc.dll,sha256=cVU4ngSdsQ0MttqOv5IUZhmUE8lZLq-OjC8ueaAsx1s,2668544
+nexaai/binds/nexaml/qwen3-vl.dll,sha256=say_RLwLsIq_svPlgqPl12gfrOCQtcCiwa_q91Ql7YM,5873152
+nexaai/binds/nexaml/qwen3vl-vision.dll,sha256=5AX4MtAOWTcXSAkS2qqvl2pFti4HR55dtkCaOO03Evc,1063424
+nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/cv_impl/mlx_cv_impl.py,sha256=QLd_8w90gtxH8kmssaDYatCTRvQNIJuUGKZNnYrmx6E,3317
+nexaai/cv_impl/pybind_cv_impl.py,sha256=aSOCAxmHrwJbEkSN6VX3Cykqlj_9RIpVrZXILul04GA,1096
+nexaai/embedder_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/embedder_impl/mlx_embedder_impl.py,sha256=Kzd-veLNl95FbI2oEJMtr6qKbjtPDDajzsGUVjJfTRA,4598
+nexaai/embedder_impl/pybind_embedder_impl.py,sha256=eH2L--L6BAl-46UOzm84pfjIkJtQKVd63-r62eb0Vg0,3670
+nexaai/image_gen_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/image_gen_impl/mlx_image_gen_impl.py,sha256=peUE9ue9ApaPlZVOICBWiHtd13sY40OWQbE8EjfIUMU,11511
+nexaai/image_gen_impl/pybind_image_gen_impl.py,sha256=514RFQMSO0Rhacq0IYzlEhEr6QfaprnGew0Rjz8HZI4,3777
+nexaai/llm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/llm_impl/mlx_llm_impl.py,sha256=qJY0zwb6jfVFQnTsp1W7BDnIMdgcJlqnU5UJgSoU-Vs,11567
+nexaai/llm_impl/pybind_llm_impl.py,sha256=zzZZ2FovnNn9EJxjovNhLI_SDjt3MAXMi8jsR6N57Ng,8359
+nexaai/rerank_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/rerank_impl/mlx_rerank_impl.py,sha256=x8L6zCccV2I4bq4V5zAOcrWFe5Ckle7z5J_00tao8bI,3335
+nexaai/rerank_impl/pybind_rerank_impl.py,sha256=1IW925bYv4FRwZDNzf9epEzdDqR6T3OONgTLBd-cOn8,1556
+nexaai/tts_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/tts_impl/mlx_tts_impl.py,sha256=LcH9bVdIl3Q6lOzSUB_X2s-_nWFmlCl1yL7XSUK0fYI,3195
+nexaai/tts_impl/pybind_tts_impl.py,sha256=n3z4zmPQayQJgAwcvETw0IBUCp8IYROuYFSg0tAy_8Y,1487
+nexaai/utils/decode.py,sha256=0Z9jDH4ICzw4YXj8nD4L-sMouDaev-TISGRQ4KzidWE,421
+nexaai/utils/manifest_utils.py,sha256=E07t_bNq-EDrqVYZW7uyX8zsKsFklvqqwSvic0in0tQ,21800
+nexaai/utils/model_manager.py,sha256=BBtScKRPXSSJDhoeHePB_bucjC4dXHjxnYHSR4EPyx4,62488
+nexaai/utils/model_types.py,sha256=q2m7diYLOpLvRl1ixL2eMq5_kdTj8KqPBGWX4p6Ob08,1532
+nexaai/utils/progress_tracker.py,sha256=BztrFqtjwNUmeREwZ5m7H6ZcrVzQEbpZfsxndWh4z0A,15778
+nexaai/utils/quantization_utils.py,sha256=FxnZ6-uAE_bl_vQ5jsRXlpU0NBn-U4Y8iN9_O6aCdPA,8070
+nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nexaai/vlm_impl/mlx_vlm_impl.py,sha256=MgqJO7OzuPd79gOZZKhSXXMNSP2eBuhhrdCX8XHn6aQ,11090
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=NuQ_Ep1TnjmGAkjJuUS0Lb6z7iPu3wroLVOx7kiAwlE,8827
+nexaai-1.0.19.dist-info/METADATA,sha256=0IRxAP_axX6ofVihQYA_qWLx8vF0zON-0OdoVbePPy4,1207
+nexaai-1.0.19.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
+nexaai-1.0.19.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.19.dist-info/RECORD,,

nexaai-1.0.17rc9.dist-info/RECORD

@@ -1,76 +0,0 @@
-nexaai/__init__.py,sha256=mbzzeXrEHHI_E3BQ0_OukD9wNajKJJVk0ykxT0rz8uM,2267
-nexaai/_stub.cp310-win_amd64.pyd,sha256=ceSj-sc3ptDA4NQBrP0EkcvUmCoqnyARwOxjyZdlVt0,10752
-nexaai/_version.py,sha256=WXmF66x0dmzVLa6Qm6qe9msuViXrfu6n7ujttHcRaQg,147
-nexaai/asr.py,sha256=_fsGaxpiU137bUtO5ujtFSYCI1RLsyeEm3Gf4GhHVRk,2118
-nexaai/base.py,sha256=qQBCiQVNzgpkQjZX9aiFDEdbAAe56TROKC3WnWra2Zg,1021
-nexaai/common.py,sha256=muQqFY-WllwL5IO83tImexbuUcoEQsKg73u4gWL2lNs,3548
-nexaai/cv.py,sha256=a6-csgYNDzPziJ0EojE9-BeM_xCny4UvWWbpnJ7GL-A,3365
-nexaai/embedder.py,sha256=3a81s7JapvYxCRbWPFKp_9EWBKW7WYqF03gk87YuGKU,2524
-nexaai/image_gen.py,sha256=4iASOKxJosMznLarTvOxJQDNaas251O81bfUWJTUBfE,4496
-nexaai/llm.py,sha256=Qwm1q_NStLfD-JYZQIvxniWnAmwNl1V6LUON3Me7w_I,3663
-nexaai/log.py,sha256=F_Qe169kLbnFV25WJmS_ZtmBfOdcGic8BYFIsYVoD_o,2720
-nexaai/rerank.py,sha256=_zGWmX6eDigY2kViMKCtNssp4JMEeVycZZfJH9eAZOY,1927
-nexaai/runtime.py,sha256=_BoAtTUv5ZR7wtOlJL5TldR3AZTP0OnMWjB9p71k-8E,2135
-nexaai/tts.py,sha256=afs6sx0w0Tvs_aJlyZRPm62qQpTrs-fW_jDHrMkc4AA,2241
-nexaai/vlm.py,sha256=wUMbRURluxvULbS7vgyHsWdEPgLGrrC_S79bhdEN7Vg,4860
-nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/asr_impl/mlx_asr_impl.py,sha256=XwMX3LYMeulp8cDS0TCCYcjvttFHAyDWQ_oMvABwQmI,3349
-nexaai/asr_impl/pybind_asr_impl.py,sha256=20o5SOPzhF9x41ra8L_qIM7YxCkYeLb5csSrNde-dds,1560
-nexaai/binds/__init__.py,sha256=ENl-uoIF9-3XGIXitVgZ2QmJ6p7Yet4h1-X7nUDZ0Hk,108
-nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=5Tix3pQanZtjtpKkP61ffawmNWewXcBYCjhFFhAqoCU,205824
-nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=BK-floyPQQuCg3Ixbi6Dl54iEWk6-R2RX5nfMJSfCdk,182784
-nexaai/binds/libcrypto-3-x64.dll,sha256=TAWcTdJrWviUingNyWaraOfz0ILUy7ybpI4VaiVHURg,7315968
-nexaai/binds/libssl-3-x64.dll,sha256=ZxawOWlONNM0Neg47w4JbU8kZJldj046H1zV3LiETa0,1313792
-nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=q_rFOC-8d2LWic4SCCkJtNOfOQVdMI3Q0Flj9zLSEeQ,162816
-nexaai/binds/nexa_bridge.dll,sha256=VDz1HILMgloBFa-C5S2JmLbNZL7O1buLvNpSTWyCFAo,171008
-nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=VDaTHpuV2S97XMFHmTqPlEI0Flq6u5s3g0bSwZ5ViPE,168960
-nexaai/binds/nexa_llama_cpp/ggml-base.dll,sha256=itL13J0fh7MIzE9ZTcR4UmDAS018r12PVYpop6i7PlA,532480
-nexaai/binds/nexa_llama_cpp/ggml-cpu.dll,sha256=3t1tCZad7gAvyKYaW9cbQ4CJb7P2U0_Kab1Smtd0P9s,672768
-nexaai/binds/nexa_llama_cpp/ggml-cuda.dll,sha256=-LHsflgA9GRIkbBinD8Kmlux7DeGpuhjitSLX1R2E9c,313528832
-nexaai/binds/nexa_llama_cpp/ggml-vulkan.dll,sha256=vlZ0b6LV8ZqE6CGPNBWk_XqkgqIZN4ch4OL5SRld0D8,36627456
-nexaai/binds/nexa_llama_cpp/ggml.dll,sha256=8cdUzL8ppC6iz0nkESnxoIMX6Mtu4jfdTU3cGjacfF8,66560
-nexaai/binds/nexa_llama_cpp/llama.dll,sha256=yqq5h0w26rTVIdhpeduB0F9mfJDl6uibFcFKKu_DIyo,1611776
-nexaai/binds/nexa_llama_cpp/mtmd.dll,sha256=dAs6RQwSYxWuQQW_a1dG7ugUC9kLWeP--_t-Dj3oe6o,561152
-nexaai/binds/nexa_llama_cpp/nexa_plugin.dll,sha256=HuKy-Z-rNbcxHgBK8jSELWTVDiXW0CIUZTjQ4BDy-nE,1413120
-nexaai/binds/nexa_nexaml/ggml-base.dll,sha256=itL13J0fh7MIzE9ZTcR4UmDAS018r12PVYpop6i7PlA,532480
-nexaai/binds/nexa_nexaml/ggml-cpu.dll,sha256=3t1tCZad7gAvyKYaW9cbQ4CJb7P2U0_Kab1Smtd0P9s,672768
-nexaai/binds/nexa_nexaml/ggml-cuda.dll,sha256=-LHsflgA9GRIkbBinD8Kmlux7DeGpuhjitSLX1R2E9c,313528832
-nexaai/binds/nexa_nexaml/ggml-vulkan.dll,sha256=vlZ0b6LV8ZqE6CGPNBWk_XqkgqIZN4ch4OL5SRld0D8,36627456
-nexaai/binds/nexa_nexaml/ggml.dll,sha256=8cdUzL8ppC6iz0nkESnxoIMX6Mtu4jfdTU3cGjacfF8,66560
-nexaai/binds/nexa_nexaml/nexa-mm-process.dll,sha256=oDnqTaAUeZrWJePQEDuIYKEOMP-ndu221mMFHht4bSY,4642816
-nexaai/binds/nexa_nexaml/nexa-sampling.dll,sha256=PNtidsRbGO8dTfbf7-VMdhQRbF328yvLnmDjMy2Tb7Y,4265984
-nexaai/binds/nexa_nexaml/nexa_plugin.dll,sha256=j0VR1s0Pz25CeK_XjbQc8FaUS994eNeigqmqryakis8,602112
-nexaai/binds/nexa_nexaml/nexaproc.dll,sha256=9X4TF6p7wOiIXi_SA7uHLoEmrvcTlH2WpfijK7HxZgY,2668544
-nexaai/binds/nexa_nexaml/qwen3-vl.dll,sha256=Es-etUDiMmK_BwOki8V79Nja-NY4kBbrkhtnesXfMMU,5870592
-nexaai/binds/nexa_nexaml/qwen3vl-vision.dll,sha256=HFAk0FxWlf_lko3AQL6uBXTIhm6AzDFi5PjxI0lxONI,1184256
-nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/cv_impl/mlx_cv_impl.py,sha256=QLd_8w90gtxH8kmssaDYatCTRvQNIJuUGKZNnYrmx6E,3317
-nexaai/cv_impl/pybind_cv_impl.py,sha256=aSOCAxmHrwJbEkSN6VX3Cykqlj_9RIpVrZXILul04GA,1096
-nexaai/embedder_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/embedder_impl/mlx_embedder_impl.py,sha256=Kzd-veLNl95FbI2oEJMtr6qKbjtPDDajzsGUVjJfTRA,4598
-nexaai/embedder_impl/pybind_embedder_impl.py,sha256=eH2L--L6BAl-46UOzm84pfjIkJtQKVd63-r62eb0Vg0,3670
-nexaai/image_gen_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/image_gen_impl/mlx_image_gen_impl.py,sha256=peUE9ue9ApaPlZVOICBWiHtd13sY40OWQbE8EjfIUMU,11511
-nexaai/image_gen_impl/pybind_image_gen_impl.py,sha256=514RFQMSO0Rhacq0IYzlEhEr6QfaprnGew0Rjz8HZI4,3777
-nexaai/llm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/llm_impl/mlx_llm_impl.py,sha256=r9Qa1ZuduRAcRjmqo2J_zlqhprVPe3ioZtUbi7yIvYY,11426
-nexaai/llm_impl/pybind_llm_impl.py,sha256=s2Cb035xDbh1ZhGNys0LRtAR6b6n4-YVKSC5voD4_Tk,8269
-nexaai/rerank_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/rerank_impl/mlx_rerank_impl.py,sha256=x8L6zCccV2I4bq4V5zAOcrWFe5Ckle7z5J_00tao8bI,3335
-nexaai/rerank_impl/pybind_rerank_impl.py,sha256=1IW925bYv4FRwZDNzf9epEzdDqR6T3OONgTLBd-cOn8,1556
-nexaai/tts_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/tts_impl/mlx_tts_impl.py,sha256=LcH9bVdIl3Q6lOzSUB_X2s-_nWFmlCl1yL7XSUK0fYI,3195
-nexaai/tts_impl/pybind_tts_impl.py,sha256=n3z4zmPQayQJgAwcvETw0IBUCp8IYROuYFSg0tAy_8Y,1487
-nexaai/utils/decode.py,sha256=0Z9jDH4ICzw4YXj8nD4L-sMouDaev-TISGRQ4KzidWE,421
-nexaai/utils/manifest_utils.py,sha256=OyLi9e4JrOBunswf4FG2KoVvkbv2FulkOf5LqRDVf50,12979
-nexaai/utils/model_manager.py,sha256=bWMVvcBu5dC3YmWtsJK4C3IaXfaDSByAYcH7GVLtoEg,57334
-nexaai/utils/model_types.py,sha256=arIyb9q-1uG0nyUGdWZaxxDJAxv0cfnJEpjCzyELL5Q,1416
-nexaai/utils/progress_tracker.py,sha256=BztrFqtjwNUmeREwZ5m7H6ZcrVzQEbpZfsxndWh4z0A,15778
-nexaai/utils/quantization_utils.py,sha256=FxnZ6-uAE_bl_vQ5jsRXlpU0NBn-U4Y8iN9_O6aCdPA,8070
-nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/vlm_impl/mlx_vlm_impl.py,sha256=MgqJO7OzuPd79gOZZKhSXXMNSP2eBuhhrdCX8XHn6aQ,11090
-nexaai/vlm_impl/pybind_vlm_impl.py,sha256=NuQ_Ep1TnjmGAkjJuUS0Lb6z7iPu3wroLVOx7kiAwlE,8827
-nexaai-1.0.17rc9.dist-info/METADATA,sha256=UDHEOOf9isWWOBY3kBn-K4D0mlx9Dilhm-5PWM2D1Jw,1233
-nexaai-1.0.17rc9.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
-nexaai-1.0.17rc9.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.17rc9.dist-info/RECORD,,