nexaai 1.0.23__cp313-cp313-win_arm64.whl → 1.0.24__cp313-cp313-win_arm64.whl

nexaai/__init__.py

@@ -28,7 +28,7 @@ from .log import set_logger, get_error_message
 from .runtime_error import (
     NexaRuntimeError,
     ContextLengthExceededError,
-    GenerationError
+    GenerationError,
 )
 
 # Create alias for PluginID to be accessible as plugin_id
@@ -41,30 +41,34 @@ from .vlm import VLM
 from .asr import ASR, ASRConfig, ASRResult
 from .cv import CVModel, CVModelConfig, CVResult, CVResults, CVCapabilities, BoundingBox
 from .rerank import Reranker, RerankConfig
-from .image_gen import ImageGen, ImageGenerationConfig, ImageSamplerConfig, SchedulerConfig, Image
+from .image_gen import (
+    ImageGen,
+    ImageGenerationConfig,
+    ImageSamplerConfig,
+    SchedulerConfig,
+    Image,
+)
 from .tts import TTS, TTSConfig, TTSSamplerConfig, TTSResult
+from .diarize import Diarize, DiarizeConfig, DiarizeResult, SpeechSegment
 
 # Build __all__ list dynamically
 __all__ = [
     "__version__",
     # Common configurations (always available)
     "ModelConfig",
-    "GenerationConfig", 
+    "GenerationConfig",
     "ChatMessage",
     "SamplerConfig",
     "EmbeddingConfig",
     "PluginID",
     "plugin_id",
-    
     # Logging functionality
     "set_logger",
     "get_error_message",
-    
     # Runtime errors
     "NexaRuntimeError",
     "ContextLengthExceededError",
     "GenerationError",
-
     "LLM",
     "Embedder",
     "VLM",
@@ -73,12 +77,12 @@ __all__ = [
     "Reranker",
     "ImageGen",
     "TTS",
-
+    "Diarize",
     "ASRConfig",
     "ASRResult",
     "CVModelConfig",
     "CVResult",
-    "CVResults", 
+    "CVResults",
     "CVCapabilities",
     "BoundingBox",
     "RerankConfig",
@@ -89,7 +93,7 @@ __all__ = [
     "TTSConfig",
     "TTSSamplerConfig",
     "TTSResult",
+    "DiarizeConfig",
+    "DiarizeResult",
+    "SpeechSegment",
 ]
-
-
-

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__ = "1.0.23" 
+__version__ = "1.0.24" 

nexaai/binds/__init__.py

@@ -4,3 +4,4 @@ from .embedder_bind import *
 from .vlm_bind import *
 from .rerank_bind import *
 from .asr_bind import *
+from .diarize_bind import *

nexaai/common.py

@@ -57,6 +57,7 @@ class ModelConfig:
     n_gpu_layers: int = 999
     chat_template_path: str = None
     chat_template_content: str = None
+    system_prompt: str = None  # For NPU plugin: system prompt must be set at model creation time
 
 
 @dataclass(frozen=True) # Read-only

nexaai/diarize.py

@@ -0,0 +1,80 @@
+from typing import List, Optional, Sequence, Union
+from abc import abstractmethod
+from dataclasses import dataclass
+
+from nexaai.base import BaseModel
+from nexaai.common import PluginID, ModelConfig
+
+
+@dataclass
+class DiarizeConfig:
+    """Configuration for speaker diarization."""
+
+    min_speakers: int = 0  # Minimum number of speakers (0 = auto-detect)
+    max_speakers: int = 0  # Maximum number of speakers (0 = no limit)
+
+
+@dataclass
+class SpeechSegment:
+    """Speech segment with speaker label and timestamps."""
+
+    start_time: float  # Segment start time in seconds
+    end_time: float  # Segment end time in seconds
+    speaker_label: str  # Speaker label (e.g., "SPEAKER_00")
+
+
+@dataclass
+class DiarizeResult:
+    """Result from speaker diarization."""
+
+    segments: Sequence[SpeechSegment]  # Array of speech segments
+    segment_count: int  # Number of segments
+    num_speakers: int  # Total unique speakers detected
+    duration: float  # Total audio duration in seconds
+
+
+class Diarize(BaseModel):
+    """Abstract base class for speaker diarization models."""
+
+    def __init__(self, m_cfg: ModelConfig = ModelConfig()):
+        """Initialize base Diarize class."""
+        self._m_cfg = m_cfg
+
+    @classmethod
+    def _load_from(
+        cls,
+        model_path: str,
+        model_name: Optional[str] = None,
+        m_cfg: ModelConfig = ModelConfig(),
+        plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
+        device_id: Optional[str] = None,
+        **kwargs
+    ) -> "Diarize":
+        """Load diarization model from local path using PyBind backend."""
+        from nexaai.diarize_impl.pybind_diarize_impl import PyBindDiarizeImpl
+
+        # There is no MLX implementation for diarization as of now.
+        return PyBindDiarizeImpl._load_from(
+            model_path, model_name, m_cfg, plugin_id, device_id
+        )
+
+    @abstractmethod
+    def infer(
+        self,
+        audio_path: str,
+        config: Optional[DiarizeConfig] = None,
+    ) -> DiarizeResult:
+        """
+        Perform speaker diarization on audio file.
+
+        Determines "who spoke when" in the audio recording, producing time-stamped segments
+        with speaker labels. Segments are time-ordered and non-overlapping.
+
+        Args:
+            audio_path: Path to audio file
+            config: Optional diarization configuration
+
+        Returns:
+            DiarizeResult with segments, speaker count, and duration
+        """
+        pass

nexaai/diarize_impl/__init__.py

@@ -0,0 +1 @@
+

nexaai/diarize_impl/pybind_diarize_impl.py

@@ -0,0 +1,125 @@
+from typing import Any, Optional, Union
+
+from nexaai.common import PluginID, ModelConfig
+from nexaai.diarize import Diarize, DiarizeConfig, DiarizeResult, SpeechSegment
+from nexaai.binds import diarize_bind, common_bind
+from nexaai.runtime import _ensure_runtime
+
+
+class PyBindDiarizeImpl(Diarize):
+    def __init__(self, handle: Any, m_cfg: ModelConfig = ModelConfig()):
+        """Private constructor, should not be called directly."""
+        super().__init__(m_cfg)
+        self._handle = handle  # This is a py::capsule
+        self._model_config = None
+
+    @classmethod
+    def _load_from(
+        cls,
+        model_path: str,
+        model_name: Optional[str] = None,
+        m_cfg: ModelConfig = ModelConfig(),
+        plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
+        device_id: Optional[str] = None,
+    ) -> "PyBindDiarizeImpl":
+        """Load diarization model from local path using PyBind backend."""
+        _ensure_runtime()
+
+        # Create model config
+        config = common_bind.ModelConfig()
+
+        config.n_ctx = m_cfg.n_ctx
+        if m_cfg.n_threads is not None:
+            config.n_threads = m_cfg.n_threads
+        if m_cfg.n_threads_batch is not None:
+            config.n_threads_batch = m_cfg.n_threads_batch
+        if m_cfg.n_batch is not None:
+            config.n_batch = m_cfg.n_batch
+        if m_cfg.n_ubatch is not None:
+            config.n_ubatch = m_cfg.n_ubatch
+        if m_cfg.n_seq_max is not None:
+            config.n_seq_max = m_cfg.n_seq_max
+        config.n_gpu_layers = m_cfg.n_gpu_layers
+
+        # handle chat template strings (if needed for diarization)
+        if m_cfg.chat_template_path:
+            config.chat_template_path = m_cfg.chat_template_path
+
+        if m_cfg.chat_template_content:
+            config.chat_template_content = m_cfg.chat_template_content
+
+        # Convert plugin_id to string
+        plugin_id_str = (
+            plugin_id.value if isinstance(plugin_id, PluginID) else str(plugin_id)
+        )
+
+        # Create Diarize handle using the binding
+        handle = diarize_bind.ml_diarize_create(
+            model_path=model_path,
+            model_name=model_name,
+            model_config=config,
+            plugin_id=plugin_id_str,
+            device_id=device_id,
+            license_id=None,  # Optional
+            license_key=None,  # Optional
+        )
+
+        return cls(handle, m_cfg)
+
+    def eject(self):
+        """Release the model from memory."""
+        # py::capsule handles cleanup automatically
+        if hasattr(self, "_handle") and self._handle is not None:
+            del self._handle
+            self._handle = None
+
+    def infer(
+        self,
+        audio_path: str,
+        config: Optional[DiarizeConfig] = None,
+    ) -> DiarizeResult:
+        """
+        Perform speaker diarization on audio file.
+
+        Determines "who spoke when" in the audio recording, producing time-stamped segments
+        with speaker labels. Segments are time-ordered and non-overlapping.
+
+        Args:
+            audio_path: Path to audio file
+            config: Optional diarization configuration
+
+        Returns:
+            DiarizeResult with segments, speaker count, and duration
+        """
+        if self._handle is None:
+            raise RuntimeError("Diarization model not loaded. Call _load_from first.")
+
+        # Convert DiarizeConfig to binding format if provided
+        diarize_config = None
+        if config:
+            diarize_config = diarize_bind.DiarizeConfig()
+            diarize_config.min_speakers = config.min_speakers
+            diarize_config.max_speakers = config.max_speakers
+
+        # Perform diarization using the binding
+        result_dict = diarize_bind.ml_diarize_infer(
+            handle=self._handle, audio_path=audio_path, config=diarize_config
+        )
+
+        # Convert result to DiarizeResult
+        segments = []
+        for segment_dict in result_dict.get("segments", []):
+            segments.append(
+                SpeechSegment(
+                    start_time=float(segment_dict["start_time"]),
+                    end_time=float(segment_dict["end_time"]),
+                    speaker_label=segment_dict["speaker_label"],
+                )
+            )
+
+        return DiarizeResult(
+            segments=segments,
+            segment_count=result_dict.get("segment_count", 0),
+            num_speakers=result_dict.get("num_speakers", 0),
+            duration=result_dict.get("duration", 0.0),
+        )

nexaai/llm_impl/pybind_llm_impl.py

@@ -17,14 +17,15 @@ class PyBindLLMImpl(LLM):
         self._profiling_data = None
 
     @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
-        ) -> 'PyBindLLMImpl':
+    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,
+    ) -> "PyBindLLMImpl":
         """Load model from local path."""
         _ensure_runtime()
 
@@ -51,16 +52,22 @@ class PyBindLLMImpl(LLM):
         if m_cfg.chat_template_content:
             config.chat_template_content = m_cfg.chat_template_content
 
+        # handle system prompt (required for NPU plugin)
+        if m_cfg.system_prompt:
+            config.system_prompt = m_cfg.system_prompt
+
         # Create handle : returns py::capsule with automatic cleanup
         # Convert enum to string for C++ binding
-        plugin_id_str = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        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,
-            device_id=device_id
+            device_id=device_id,
         )
         return cls(handle, m_cfg)
 
@@ -70,16 +77,23 @@ class PyBindLLMImpl(LLM):
         del self._handle
         self._handle = None
 
-    def apply_chat_template(self, messages: list[ChatMessage], tools: Optional[str] = None, enable_thinking: bool = True, add_generation_prompt: bool = True) -> str:
+    def apply_chat_template(
+        self,
+        messages: list[ChatMessage],
+        tools: Optional[str] = None,
+        enable_thinking: bool = True,
+        add_generation_prompt: bool = True,
+    ) -> str:
         """Apply the chat template to messages."""
         # Convert TypedDict to list of dicts for binding
-        message_dicts = [
-            {"role": m["role"], "content": m["content"]}
-            for m in messages
-        ]
-        return llm_bind.ml_llm_apply_chat_template(self._handle, message_dicts, tools, enable_thinking)
+        message_dicts = [{"role": m["role"], "content": m["content"]} for m in messages]
+        return llm_bind.ml_llm_apply_chat_template(
+            self._handle, message_dicts, tools, enable_thinking
+        )
 
-    def generate_stream(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> Generator[str, None, None]:
+    def generate_stream(
+        self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()
+    ) -> Generator[str, None, None]:
         """Generate text with streaming."""
         token_queue = queue.Queue()
         exception_container = [None]
@@ -87,10 +101,10 @@ class PyBindLLMImpl(LLM):
 
         def on_token(token: str, user_data) -> bool:
             if self._cancel_event.is_set():
-                token_queue.put(('end', None))
+                token_queue.put(("end", None))
                 return False  # Stop generation
             try:
-                token_queue.put(('token', token))
+                token_queue.put(("token", token))
                 return True  # Continue generation
             except Exception as e:
                 exception_container[0] = e
@@ -106,13 +120,15 @@ class PyBindLLMImpl(LLM):
                     prompt=prompt,
                     config=config,
                     on_token=on_token,
-                    user_data=None
+                    user_data=None,
+                )
+                self._profiling_data = ProfilingData.from_dict(
+                    result.get("profile_data", {})
                 )
-                self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
             except Exception as e:
                 exception_container[0] = e
             finally:
-                token_queue.put(('end', None))
+                token_queue.put(("end", None))
 
         thread = threading.Thread(target=generate)
         thread.start()
@@ -121,9 +137,9 @@ class PyBindLLMImpl(LLM):
         try:
             while True:
                 msg_type, token = token_queue.get()
-                if msg_type == 'token':
+                if msg_type == "token":
                     yield token
-                elif msg_type in ('error', 'end'):
+                elif msg_type in ("error", "end"):
                     break
         finally:
             thread.join()
@@ -131,7 +147,9 @@ class PyBindLLMImpl(LLM):
         if exception_container[0]:
             raise exception_container[0]
 
-    def generate(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> str:
+    def generate(
+        self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()
+    ) -> str:
         """
         Generate text without streaming.
 
@@ -148,7 +166,7 @@ class PyBindLLMImpl(LLM):
             prompt=prompt,
             config=config,
             on_token=None,  # No callback for non-streaming
-            user_data=None
+            user_data=None,
         )
 
         self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))

nexaai/vlm_impl/pybind_vlm_impl.py

@@ -23,23 +23,24 @@ class PyBindVLMImpl(VLM):
         self._profiling_data = None
 
     @classmethod
-    def _load_from(cls,
-                   local_path: str,
-                   mmproj_path: str = None,
-                   model_name: Optional[str] = None,
-                   m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
-        ) -> 'PyBindVLMImpl':
+    def _load_from(
+        cls,
+        local_path: str,
+        mmproj_path: str = None,
+        model_name: Optional[str] = None,
+        m_cfg: ModelConfig = ModelConfig(),
+        plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
+        device_id: Optional[str] = None,
+    ) -> "PyBindVLMImpl":
         """Load VLM model from local path.
-        
+
         Args:
             local_path: Path to the main model file
             mmproj_path: Path to the multimodal projection file
             m_cfg: Model configuration
             plugin_id: Plugin identifier
             device_id: Optional device ID (not used in current binding)
-            
+
         Returns:
             PyBindVLMImpl instance
         """
@@ -67,16 +68,22 @@ class PyBindVLMImpl(VLM):
         if m_cfg.chat_template_content:
             config.chat_template_content = m_cfg.chat_template_content
 
+        # handle system prompt (required for NPU plugin)
+        if m_cfg.system_prompt:
+            config.system_prompt = m_cfg.system_prompt
+
         # Create handle : returns py::capsule with automatic cleanup
         # Convert enum to string for C++ binding
-        plugin_id_str = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        plugin_id_str = (
+            plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        )
         handle = vlm_bind.create_vlm(
             model_path=local_path,
             mmproj_path=mmproj_path,
             model_name=model_name,
             model_config=config,
             plugin_id=plugin_id_str,
-            device_id=device_id
+            device_id=device_id,
         )
         return cls(handle, m_cfg)
 
@@ -97,18 +104,18 @@ class PyBindVLMImpl(VLM):
         self,
         messages: List[MultiModalMessage],
         tools: Optional[List[Dict[str, Any]]] = None,
-        enable_thinking: bool = True
+        enable_thinking: bool = True,
     ) -> str:
         """Apply the chat template to multimodal messages."""
         payload = []
         for msg in messages:
-            role   = msg["role"]
+            role = msg["role"]
             blocks = []
 
             for c in msg["content"]:
                 t = c["type"]
                 if t == "text":
-                    blocks.append({"type": "text", "text": c.get("text","") or ""})
+                    blocks.append({"type": "text", "text": c.get("text", "") or ""})
                 else:
                     # Pass through the original structure for image, audio, and any other types
                     # Let vlm-bind.cpp handle field extraction (text/url/path)
@@ -116,10 +123,14 @@ class PyBindVLMImpl(VLM):
 
             payload.append({"role": role, "content": blocks})
 
-        result = vlm_bind.ml_vlm_apply_chat_template(self._handle, payload, tools, enable_thinking)
+        result = vlm_bind.ml_vlm_apply_chat_template(
+            self._handle, payload, tools, enable_thinking
+        )
         return result
 
-    def generate_stream(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> Generator[str, None, None]:
+    def generate_stream(
+        self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()
+    ) -> Generator[str, None, None]:
         """Generate text with streaming."""
         token_queue = queue.Queue()
         exception_container = [None]
@@ -127,10 +138,10 @@ class PyBindVLMImpl(VLM):
 
         def on_token(token: str, user_data) -> bool:
             if self._cancel_event.is_set():
-                token_queue.put(('end', None))
+                token_queue.put(("end", None))
                 return False  # Stop generation
             try:
-                token_queue.put(('token', token))
+                token_queue.put(("token", token))
                 return True  # Continue generation
             except Exception as e:
                 exception_container[0] = e
@@ -146,25 +157,31 @@ class PyBindVLMImpl(VLM):
                     prompt=prompt,
                     config=config,
                     on_token=on_token,
-                    user_data=None
+                    user_data=None,
                 )
-                
+
                 # Check for errors in result
                 error_code = result.get("error_code", ML_SUCCESS)
                 if error_code != ML_SUCCESS:
                     error_message = result.get("error_message", "Unknown error")
                     if error_code == ML_ERROR_LLM_TOKENIZATION_CONTEXT_LENGTH:
-                        exception_container[0] = ContextLengthExceededError(error_message, error_code)
+                        exception_container[0] = ContextLengthExceededError(
+                            error_message, error_code
+                        )
                     else:
-                        exception_container[0] = GenerationError(error_message, error_code)
-                    token_queue.put(('end', None))
+                        exception_container[0] = GenerationError(
+                            error_message, error_code
+                        )
+                    token_queue.put(("end", None))
                     return
-                
-                self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
+
+                self._profiling_data = ProfilingData.from_dict(
+                    result.get("profile_data", {})
+                )
             except Exception as e:
                 exception_container[0] = e
             finally:
-                token_queue.put(('end', None))
+                token_queue.put(("end", None))
 
         thread = threading.Thread(target=generate)
         thread.start()
@@ -173,9 +190,9 @@ class PyBindVLMImpl(VLM):
         try:
             while True:
                 msg_type, token = token_queue.get()
-                if msg_type == 'token':
+                if msg_type == "token":
                     yield token
-                elif msg_type in ('error', 'end'):
+                elif msg_type in ("error", "end"):
                     break
         finally:
             thread.join()
@@ -183,7 +200,9 @@ class PyBindVLMImpl(VLM):
         if exception_container[0]:
             raise exception_container[0]
 
-    def generate(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> str:
+    def generate(
+        self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()
+    ) -> str:
         """
         Generate text without streaming.
 
@@ -200,7 +219,7 @@ class PyBindVLMImpl(VLM):
             prompt=prompt,
             config=config,
             on_token=None,  # No callback for non-streaming
-            user_data=None
+            user_data=None,
         )
 
         # Check for errors in result

{nexaai-1.0.23.dist-info → nexaai-1.0.24.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.23
+Version: 1.0.24
 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

{nexaai-1.0.23.dist-info → nexaai-1.0.24.dist-info}/RECORD

@@ -1,10 +1,11 @@
-nexaai/__init__.py,sha256=gOd7sNsqEESopw_24xgnOSkIRENrk4Fa-RMtmVv62eA,2421
-nexaai/_stub.cp313-win_arm64.pyd,sha256=9PfCTx57-dNmLuotZ4vwQx-QymdYzVRvqaI8tWAPtiU,10240
-nexaai/_version.py,sha256=jvuw8FsRzBuy2gWtiNcMHIovBrvEB6BUEhxloSiR5bE,143
+nexaai/__init__.py,sha256=S6GrUD-_rZhDY2tcIY4aJvNIZa6KOk6TfV26Jb6x0pE,2582
+nexaai/_stub.cp313-win_arm64.pyd,sha256=s04H3HEwpabF09TmfNFTdN8-CW_yZSyE3_AhkqqrU38,10240
+nexaai/_version.py,sha256=OKw2jr_ECtbEM3MNGRIR8lheCGGhID-ysEdmnHkH_pA,143
 nexaai/asr.py,sha256=wqtq71cxIMGE4KvOIYZebHdWik8dy4LyKrDI98PDvzQ,2294
 nexaai/base.py,sha256=N8PRgDFA-XPku2vWnQIofQ7ipz3pPlO6f8YZGnuhquE,982
-nexaai/common.py,sha256=MRWZ6a7pnci_OUHxZRm3YqgKLAtZFD7b88STYDfeIF8,3460
+nexaai/common.py,sha256=LuOMGtDp_HQa4cpW84Cezqc3cYBgFySi4qYlEQsL9J4,3558
 nexaai/cv.py,sha256=WbyFzvByx6hOoR2ZrGSm6i4uG0ubxqeMZ9x71kG6Des,3338
+nexaai/diarize.py,sha256=WmwHr6VfN1SbRX45KdSvFS44t8Ca9R8R1WkudKswDC0,2461
 nexaai/embedder.py,sha256=lXOT16PEvd_hT23d77dZH38VHNOAk-3JvoOUdQTEaGI,2552
 nexaai/image_gen.py,sha256=MkGw1HXqqv8cJzbiGERNPKFXfq9vMOlvuq0pgekXw68,4385
 nexaai/llm.py,sha256=-agVJuj0FOaDvDiT-fFSOpoyVt-MpNudBucsod3Vp1M,3673
@@ -17,35 +18,36 @@ nexaai/vlm.py,sha256=LUrd1_SGHOsYpWyUymX93oEIsNJv7XzHIHo4hBZOhQA,4800
 nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/asr_impl/mlx_asr_impl.py,sha256=eosd8-TIWAOwV0HltmoFrLwzXHcU4jyxtncvuZE9pgA,3257
 nexaai/asr_impl/pybind_asr_impl.py,sha256=FLOWIph37q_nIiNx8xYi-VnhQ6CrPuc4HFAJZQKc42w,4680
-nexaai/binds/__init__.py,sha256=2-Rr0NwyWygqwS8Xlxq0BJ2ltyID-WbGuzEYNlSanCI,155
-nexaai/binds/asr_bind.cp313-win_arm64.pyd,sha256=XqDvZcckuk_wzRH6yEszY7I_rjdJ0YU-8y87fnwu96o,267776
-nexaai/binds/common_bind.cp313-win_arm64.pyd,sha256=fgJXE6sjXcvEvhSpO2G8HIwjGsbqnZJyK10drcOm0zw,291328
-nexaai/binds/cv_bind.cp313-win_arm64.pyd,sha256=YHiVojWcBwxewA5CSudqWeUudGg5HKbijkH0g3ubLgc,328192
-nexaai/binds/embedder_bind.cp313-win_arm64.pyd,sha256=I5YAWe4JF_1pEhsVXxliT1ZVsQmm_MOkl7Vf-fDE0sQ,243712
+nexaai/binds/__init__.py,sha256=-dSUMEA_-SghPgdI0qE2RWF2Fng_Klu2rGpP5zqA8_o,183
+nexaai/binds/asr_bind.cp313-win_arm64.pyd,sha256=X8-9Ro_-7g_wrCpZylgkTyaBu_R8suIgcUihPRfYvn8,267776
+nexaai/binds/common_bind.cp313-win_arm64.pyd,sha256=_7ZnfjzG7XG5-TMMtUiteY4OrHZ7kKu3H8v5nxkVdIk,293888
+nexaai/binds/cv_bind.cp313-win_arm64.pyd,sha256=QE8WAbLZKxPIPOonaLM2OaWdMhwZgjlvLKExVcIB3tw,328192
+nexaai/binds/diarize_bind.cp313-win_arm64.pyd,sha256=ZjXL6i6pGhcHCqx5w9FCJAYpg_Q7WGo0NMrfr1Dtllw,232448
+nexaai/binds/embedder_bind.cp313-win_arm64.pyd,sha256=SEkNn8kVLRODVlb-3j4DI2l7fuK9hEmDHEbOHyKPEfQ,243712
 nexaai/binds/libcrypto-3-arm64.dll,sha256=jl9ZGYpUKEfkqU2HosjSJQ4Mg0tQdkEka7mKI371NcQ,7297536
 nexaai/binds/libssl-3-arm64.dll,sha256=GeAhDqjiuxhc792cpJIzjEx03noAWUM6R3w2ekHKnJw,1722368
-nexaai/binds/llm_bind.cp313-win_arm64.pyd,sha256=1J_Xw90zAHOAUkcrrP4w0s49VElyQzRQvg62M8Q38O0,226304
-nexaai/binds/nexa_bridge.dll,sha256=T6-0QPCk3622hsTRpqfG6sxOuYZZyEMTuIv2XrL3M78,231936
-nexaai/binds/rerank_bind.cp313-win_arm64.pyd,sha256=rclsjzOxknQ3-jot1z6c9GrgtVRwsna3FPDcl61WqJg,239616
-nexaai/binds/vlm_bind.cp313-win_arm64.pyd,sha256=vRCIGzKdAM0SISTLb4y-ps_TofkuevV8DElGnDpWZGc,237568
-nexaai/binds/cpu_gpu/ggml-base.dll,sha256=FMsS34AC9cG-zBmVQPD7TAAuZxFQBPVYx6eZDpTqL6I,534016
-nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=7oYrOG_GtLyMzUIEeJn76IIzL3cuYT9m-qyGu3PsJYs,555520
-nexaai/binds/cpu_gpu/ggml-opencl.dll,sha256=8eGkRxrHhdh2juBAWKM5iZeGvivdeEgCI-m0-hbK1ag,620544
-nexaai/binds/cpu_gpu/ggml.dll,sha256=eGqguBEG6LvgnMO140-PuR0JOL5pDkeQeeOROkhP1JE,70656
+nexaai/binds/llm_bind.cp313-win_arm64.pyd,sha256=jqbpyuJfJTiYHmlKpHt7Ofg5yWo4hdjs9_48RjcWjG0,226304
+nexaai/binds/nexa_bridge.dll,sha256=_7VuTbWqKV-tI2dpBcW1QnutbmavubywSSMm12jIHsA,231936
+nexaai/binds/rerank_bind.cp313-win_arm64.pyd,sha256=JG_K6fwSDcqx9WHnpLDcM2RbcddLR903UAnoDAfRTGo,239616
+nexaai/binds/vlm_bind.cp313-win_arm64.pyd,sha256=4K8imx11I_CHQY5Qwfkc-uLKvx0jma2J3zfCbB0TTow,237568
+nexaai/binds/cpu_gpu/ggml-base.dll,sha256=B6tz8_Gj5bE8LaDOlv4g8ggS3zE02rifrcVdL-NaSKQ,534016
+nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=I8MuXSSYaVirOaDCcuylzcrVMdeDIlMenkaRlz9V9XM,555520
+nexaai/binds/cpu_gpu/ggml-opencl.dll,sha256=Iwj6PCJhLolKkNWUDmCigYIgzL5cfN8HiqU5F8rX0lI,620544
+nexaai/binds/cpu_gpu/ggml.dll,sha256=I3iEGHmXbFIVzeBoIfdPgmo6lFWcKC8MsnTtBpVDdac,70656
 nexaai/binds/cpu_gpu/libomp140.aarch64.dll,sha256=0mSWmPxoRm7ojPWK46sPqxBvKTJug6qqVkPTBlGpYuM,599504
-nexaai/binds/cpu_gpu/mtmd.dll,sha256=_g6n8iwEt-ylyHi3D_PPYZ_ZT2rjgI-3M1Ee1eHOJNM,551936
-nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=Y5KRSWKMlV_57zX451q1PXHfkVqmn5KqBE0BYDaYzcE,1581056
-nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=P_9bapuF-vvEFo13gwwQ7yEbdw0cNS5dJ9Lyv4LtGdo,2177024
+nexaai/binds/cpu_gpu/mtmd.dll,sha256=gEAKzury4q1Y32C0ZUGJENqiahPypkV4n3N4chiw8GQ,551936
+nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=8jKFNoAjW6art7Cxxp11-dqgJFine_nGuF36RrB7ZJc,1581056
+nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=-LBW7eP3hPqoGGUldfxOAQU363-I0Gwjfip4U4WKnac,2177024
 nexaai/binds/npu/FLAC.dll,sha256=KieFm07q-TkzZ-NoOzaLutkwlMoYNnMD3IuI1iMAt9s,224256
 nexaai/binds/npu/convnext-sdk.dll,sha256=90dv6IKyGUnsW61V8byEVxn9sOkRT_bAoOjVr2QE7rU,1210368
 nexaai/binds/npu/embed-gemma-sdk.dll,sha256=SIOpKs_9_XxvoflWR6KQtkUqwYa-fZt4wtPxIoifXJ8,4683264
 nexaai/binds/npu/fftw3.dll,sha256=UHNc9Fdpo50qWP6jOIVftyGJeoFbExBV4taNrI-rcG8,569856
 nexaai/binds/npu/fftw3f.dll,sha256=Qeeq1HxRdRReCBihEpEBz4AfcPrJZ994u1hy4OwSv5M,565760
-nexaai/binds/npu/ggml-base.dll,sha256=FMsS34AC9cG-zBmVQPD7TAAuZxFQBPVYx6eZDpTqL6I,534016
-nexaai/binds/npu/ggml-cpu.dll,sha256=7oYrOG_GtLyMzUIEeJn76IIzL3cuYT9m-qyGu3PsJYs,555520
-nexaai/binds/npu/ggml-opencl.dll,sha256=8eGkRxrHhdh2juBAWKM5iZeGvivdeEgCI-m0-hbK1ag,620544
-nexaai/binds/npu/ggml.dll,sha256=eGqguBEG6LvgnMO140-PuR0JOL5pDkeQeeOROkhP1JE,70656
-nexaai/binds/npu/granite-nano-sdk.dll,sha256=03FsvBYHRjd7ryR57Drf5PQ1R-A0qkmLtTIJJQGHfVo,4841984
+nexaai/binds/npu/ggml-base.dll,sha256=B6tz8_Gj5bE8LaDOlv4g8ggS3zE02rifrcVdL-NaSKQ,534016
+nexaai/binds/npu/ggml-cpu.dll,sha256=I8MuXSSYaVirOaDCcuylzcrVMdeDIlMenkaRlz9V9XM,555520
+nexaai/binds/npu/ggml-opencl.dll,sha256=Iwj6PCJhLolKkNWUDmCigYIgzL5cfN8HiqU5F8rX0lI,620544
+nexaai/binds/npu/ggml.dll,sha256=I3iEGHmXbFIVzeBoIfdPgmo6lFWcKC8MsnTtBpVDdac,70656
+nexaai/binds/npu/granite-nano-sdk.dll,sha256=gFON6w2dppuYc-T1zCz3Rqdyy8FkHDCpVTzYPKRC3HM,4841984
 nexaai/binds/npu/granite4-sdk.dll,sha256=aNlT_Db8CYO4-A6gegxINhU9VbiGAXkArenruym4Jvs,4735488
 nexaai/binds/npu/jina-rerank-sdk.dll,sha256=End5A_pSnaswcjsmifuXkc41zZdl8c1HtA69lqjUFaM,4632576
 nexaai/binds/npu/libcrypto-3-arm64.dll,sha256=Zdx-VmPlmvRbagtlOdhjOsmZSGTcRoUCm_K-_6XcDe8,7309432
@@ -55,9 +57,9 @@ nexaai/binds/npu/libssl-3-arm64.dll,sha256=YbBpztXYi15lGe445EyV9DL0Q4rI0k23gta1d
 nexaai/binds/npu/liquid-sdk.dll,sha256=la2-wwueKuq8gzNtb-PVbfXUXB7TkG04HFrPKrxwg9k,4849664
 nexaai/binds/npu/llama3-3b-sdk.dll,sha256=C9SS6tYpll8yAb_xBqBkmweuVTk2aA-9sPE6DPlBl4Y,4738048
 nexaai/binds/npu/mpg123.dll,sha256=1iTmvrgIVievsDXO1RaWLntnE-TIAVPpbqEV8C-yaC0,253952
-nexaai/binds/npu/nexa-mm-process.dll,sha256=rZvP8d8eVjSqKGTkeoNV7Ill_gSjSzJ4XJ1p-Tt92rk,5470208
+nexaai/binds/npu/nexa-mm-process.dll,sha256=w4azeu0O7XiTGS26Lz9snzHCLNrfFf6_I2y5hxZVHwE,5465600
 nexaai/binds/npu/nexa-sampling.dll,sha256=eYclqTHNOGn8bEzS2yarCULnh1PwXsPgQO8ePGeZKx8,3795968
-nexaai/binds/npu/nexa_plugin.dll,sha256=-hhjP9OW0W5x41qZF86b1qkh5MuVSNef28Abz9W-fRc,773632
+nexaai/binds/npu/nexa_plugin.dll,sha256=0HvsfaF7rHomJh5bFScDPFieSBibrqHmf8knkw9VpA0,773632
 nexaai/binds/npu/nexaproc.dll,sha256=uWp5ENzDwlavO06t4SfcKBnc7JlF1OmAGsJCrV39TCw,2561024
 nexaai/binds/npu/ogg.dll,sha256=bgpqQM-OLactFaqZYmZcIDjkaTRCinrny8_6SBUU8B0,30720
 nexaai/binds/npu/omni-neural-sdk.dll,sha256=jA8XFNirsiK5uAZbnoU01RohWj33Bf2uv8Y8g8GySDQ,5257728
@@ -68,7 +70,7 @@ nexaai/binds/npu/paddleocr-sdk.dll,sha256=dJer3VBQRbUIqRDe8DjpTFVjOm18h9IJQYjozn
 nexaai/binds/npu/parakeet-sdk.dll,sha256=REkIVi5YLasAOiDql2J60ipQJjT0_GcRpzGi5JsyNZs,1732608
 nexaai/binds/npu/phi3-5-sdk.dll,sha256=__Q5BgnYarK9yWBrRDO-UKvFQq9gfFVF5JQCThrJl6A,4737024
 nexaai/binds/npu/phi4-sdk.dll,sha256=uBJL70m0qOAAlV_hEGi-tB753A2dGFSpD8JsxTDW3HA,4737536
-nexaai/binds/npu/pyannote-sdk.dll,sha256=ptzE3gdX57P8AU8M1i-eoAbParNROPAr7jVstb14Wlc,1217536
+nexaai/binds/npu/pyannote-sdk.dll,sha256=VPAI9Pj29FOjL77u3ZL8bv0cTVNXXf2Wq3IijfpypeA,1195520
 nexaai/binds/npu/qwen3-4b-sdk.dll,sha256=94FRAX6LrXGsp0BgwRR2TS7MCwUD-dbqVxbUeIzNGPo,4730368
 nexaai/binds/npu/qwen3vl-sdk.dll,sha256=pmI701PrrYAj3J0nJsEa9zmSR5iqsmDuLrstanshrm8,5152256
 nexaai/binds/npu/qwen3vl-vision.dll,sha256=kCJBlpWv_4YQCgCIL4b99a6eFd9Ok3_V8TWB_X77w5A,546816
@@ -125,6 +127,8 @@ nexaai/binds/npu/htp-files/libsnpehtpv73.cat,sha256=hNtz9-i_cZjR5xLVEI_56jjD8pQr
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/cv_impl/mlx_cv_impl.py,sha256=gKECQOv8iaWwG3bl7xeqVy2NN_9K7tYerIFzfn4eLo4,3228
 nexaai/cv_impl/pybind_cv_impl.py,sha256=YVIgNZe8FaUs6eH4IfirqhAyMy9RXqZL0eqVjGwSfO0,4761
+nexaai/diarize_impl/__init__.py,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
+nexaai/diarize_impl/pybind_diarize_impl.py,sha256=SgdfY_BC0lXF4Ard73ZwsHFALX03DRjEj2DNWYJuPc0,4482
 nexaai/embedder_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/embedder_impl/mlx_embedder_impl.py,sha256=pFPraUAjm9EVvVbwIp1cjbtXUysF5pqxEcK2CAFvcDw,4639
 nexaai/embedder_impl/pybind_embedder_impl.py,sha256=lFpf0wI2d7kfO2GUyUuUS1U2L_PyZMJVGmAvF8EuQ0g,3653
@@ -133,7 +137,7 @@ nexaai/image_gen_impl/mlx_image_gen_impl.py,sha256=BuDkksvXyb4J02GsdnbGAmYckfUU0
 nexaai/image_gen_impl/pybind_image_gen_impl.py,sha256=ms34VYoD5AxZFG6cIG0QAJDjCtfphaZ1bHzKzey1xF8,3692
 nexaai/llm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/llm_impl/mlx_llm_impl.py,sha256=dPtaEribluHZZY_f9M114glcQhtDEckukw4Sfd5zJos,11296
-nexaai/llm_impl/pybind_llm_impl.py,sha256=-OFruPYQlAQvy0py6JoBQapBLbRCSgwuZ84WymHtLT0,8163
+nexaai/llm_impl/pybind_llm_impl.py,sha256=0ZNFA5xnNGqzmWcO_Fg-Nhv-fE8YVLjrLFPCDc499r0,8376
 nexaai/rerank_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/rerank_impl/mlx_rerank_impl.py,sha256=3nbqCdzyAugc4P_6K9mowEgy4LFdfzhy7GUvn9GMpSE,3377
 nexaai/rerank_impl/pybind_rerank_impl.py,sha256=tmzrpRYCCV3ATxbE9G1Io6SUtgYPO8BFe48nTae6_xw,4490
@@ -148,8 +152,8 @@ nexaai/utils/progress_tracker.py,sha256=jdUqtmPqyhwC9uSKvQcJEYETwSt-OhP4oitdJ946
 nexaai/utils/quantization_utils.py,sha256=FYcNSAKGlBqFDUTx3jSKOr2lnq4nyiyC0ZG8oSxFwiU,7825
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/vlm_impl/mlx_vlm_impl.py,sha256=sgHqnX5OCSGLccCnTuRiktIbqThNn3AAIvYE2_Dy4TI,10833
-nexaai/vlm_impl/pybind_vlm_impl.py,sha256=stJKHdhYhBuWUQkky-nHgCv625qDB_1geI3v5BLNGpM,9765
-nexaai-1.0.23.dist-info/METADATA,sha256=ghgqII8_rVkGTKEww-GkKbxEtYSuGl5oWOlqoQXKqBQ,1215
-nexaai-1.0.23.dist-info/WHEEL,sha256=QL7uMKXoDJRpSwAf1VOVpjVXYPYll2YWTJ-omqdO8-4,101
-nexaai-1.0.23.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.23.dist-info/RECORD,,
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=MKDZpU_FDwEm6V2wxgP6WrqvyDe5TpACFwHSYF2XwrQ,10021
+nexaai-1.0.24.dist-info/METADATA,sha256=iY5HsgQZcBXjDJwElBoK6V_L-nB2gzyTFcopPlCY6ZM,1215
+nexaai-1.0.24.dist-info/WHEEL,sha256=QL7uMKXoDJRpSwAf1VOVpjVXYPYll2YWTJ-omqdO8-4,101
+nexaai-1.0.24.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.24.dist-info/RECORD,,