nexaai 1.0.4rc15__cp310-cp310-macosx_14_0_universal2.whl → 1.0.5__cp310-cp310-macosx_14_0_universal2.whl

nexaai/__init__.py

@@ -19,7 +19,10 @@ except ImportError:
     __version__ = "0.0.1"
 
 # Import common configuration classes first (no external dependencies)
-from .common import ModelConfig, GenerationConfig, ChatMessage, SamplerConfig
+from .common import ModelConfig, GenerationConfig, ChatMessage, SamplerConfig, PluginID
+
+# Create alias for PluginID to be accessible as plugin_id
+plugin_id = PluginID
 
 # Import new feature classes (no external dependencies in base classes)
 from .llm import LLM
@@ -40,6 +43,8 @@ __all__ = [
     "ChatMessage",
     "SamplerConfig",
     "EmbeddingConfig",
+    "PluginID",
+    "plugin_id",
 
     "LLM",
     "Embedder",

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__ = "1.0.4-rc15" 
+__version__ = "1.0.5" 

nexaai/asr.py

@@ -1,8 +1,9 @@
-from typing import List, Optional, Sequence, Tuple
+from typing import List, Optional, Sequence, Tuple, Union
 from abc import abstractmethod
 from dataclasses import dataclass
 
 from nexaai.base import BaseModel
+from nexaai.common import PluginID
 
 
 @dataclass
@@ -33,11 +34,14 @@ class ASR(BaseModel):
                    model_path: str,
                    tokenizer_path: Optional[str] = None,
                    language: Optional[str] = None,
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'ASR':
         """Load ASR model from local path, routing to appropriate implementation."""
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.asr_impl.mlx_asr_impl import MLXASRImpl
             return MLXASRImpl._load_from(model_path, tokenizer_path, language, plugin_id, device_id)
         else:

nexaai/asr_impl/mlx_asr_impl.py

@@ -1,7 +1,8 @@
 # Note: This code is generated by Cursor, not tested yet.
 
-from typing import List, Optional
+from typing import List, Optional, Union
 
+from nexaai.common import PluginID
 from nexaai.asr import ASR, ASRConfig, ASRResult
 from nexaai.mlx_backend.asr.interface import MlxAsr as MLXASRInterface
 from nexaai.mlx_backend.ml import ModelConfig as MLXModelConfig, SamplerConfig as MLXSamplerConfig, GenerationConfig as MLXGenerationConfig, EmbeddingConfig
@@ -18,7 +19,7 @@ class MLXASRImpl(ASR):
                    model_path: str,
                    tokenizer_path: Optional[str] = None,
                    language: Optional[str] = None,
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None
         ) -> 'MLXASRImpl':
         """Load ASR model from local path using MLX backend."""

nexaai/asr_impl/pybind_asr_impl.py

@@ -1,5 +1,6 @@
-from typing import List, Optional
+from typing import List, Optional, Union
 
+from nexaai.common import PluginID
 from nexaai.asr import ASR, ASRConfig, ASRResult
 
 
@@ -14,7 +15,7 @@ class PyBindASRImpl(ASR):
                    model_path: str,
                    tokenizer_path: Optional[str] = None,
                    language: Optional[str] = None,
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'PyBindASRImpl':
         """Load ASR model from local path using PyBind backend."""

nexaai/common.py

@@ -1,5 +1,12 @@
 from dataclasses import dataclass
 from typing import TypedDict, Literal, Optional, List
+from enum import Enum
+
+
+class PluginID(str, Enum):
+    """Enum for plugin identifiers."""
+    MLX = "mlx"
+    LLAMA_CPP = "llama_cpp"
 
 
 class ChatMessage(TypedDict):
@@ -52,10 +59,45 @@ class ModelConfig:
 
 @dataclass(frozen=True) # Read-only
 class ProfilingData:
-    start_time: int
-    end_time: int
-    prompt_start_time: int = None
-    prompt_end_time: int = None
-    decode_start_time: int = None
-    decode_ent_time: int = None
-    first_token_time: int = None
+    """Profiling data structure for LLM/VLM performance metrics."""
+    ttft: int = 0             # Time to first token (us)
+    prompt_time: int = 0      # Prompt processing time (us)
+    decode_time: int = 0      # Token generation time (us)
+    prompt_tokens: int = 0    # Number of prompt tokens
+    generated_tokens: int = 0  # Number of generated tokens
+    audio_duration: int = 0   # Audio duration (us)
+    prefill_speed: float = 0.0  # Prefill speed (tokens/sec)
+    decoding_speed: float = 0.0 # Decoding speed (tokens/sec)
+    real_time_factor: float = 0.0 # Real-Time Factor (RTF)
+    stop_reason: str = ""     # Stop reason: "eos", "length", "user", "stop_sequence"
+    
+    @classmethod
+    def from_dict(cls, data: dict) -> "ProfilingData":
+        """Create ProfilingData from dictionary."""
+        return cls(
+            ttft=data.get("ttft", 0),
+            prompt_time=data.get("prompt_time", 0),
+            decode_time=data.get("decode_time", 0),
+            prompt_tokens=data.get("prompt_tokens", 0),
+            generated_tokens=data.get("generated_tokens", 0),
+            audio_duration=data.get("audio_duration", 0),
+            prefill_speed=data.get("prefill_speed", 0.0),
+            decoding_speed=data.get("decoding_speed", 0.0),
+            real_time_factor=data.get("real_time_factor", 0.0),
+            stop_reason=data.get("stop_reason", "")
+        )
+    
+    def to_dict(self) -> dict:
+        """Convert to dictionary."""
+        return {
+            "ttft": self.ttft,
+            "prompt_time": self.prompt_time,
+            "decode_time": self.decode_time,
+            "prompt_tokens": self.prompt_tokens,
+            "generated_tokens": self.generated_tokens,
+            "audio_duration": self.audio_duration,
+            "prefill_speed": self.prefill_speed,
+            "decoding_speed": self.decoding_speed,
+            "real_time_factor": self.real_time_factor,
+            "stop_reason": self.stop_reason
+        }

nexaai/cv.py

@@ -1,8 +1,9 @@
-from typing import List, Optional
+from typing import List, Optional, Union
 from abc import abstractmethod
 from dataclasses import dataclass
 
 from nexaai.base import BaseModel
+from nexaai.common import PluginID
 
 
 @dataclass
@@ -71,11 +72,14 @@ class CVModel(BaseModel):
     def _load_from(cls,
                    _: str,  # TODO: remove this argument, this is a hack to make api design happy
                    config: CVModelConfig,
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'CVModel':
         """Load CV model from configuration, routing to appropriate implementation."""
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.cv_impl.mlx_cv_impl import MLXCVImpl
             return MLXCVImpl._load_from(config, plugin_id, device_id)
         else:

nexaai/cv_impl/mlx_cv_impl.py

@@ -1,8 +1,9 @@
 # Note: This code is generated by Cursor, not tested yet.
 
-from typing import Optional
+from typing import Optional, Union
 import os
 
+from nexaai.common import PluginID
 from nexaai.cv import CVModel, CVModelConfig, CVResults
 from nexaai.mlx_backend.cv.interface import CVModel as MLXCVInterface, create_cv_model
 
@@ -16,7 +17,7 @@ class MLXCVImpl(CVModel):
     @classmethod
     def _load_from(cls,
                    config: CVModelConfig,
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None
         ) -> 'MLXCVImpl':
         """Load CV model from configuration using MLX backend."""

nexaai/cv_impl/pybind_cv_impl.py

@@ -1,5 +1,6 @@
-from typing import Optional
+from typing import Optional, Union
 
+from nexaai.common import PluginID
 from nexaai.cv import CVModel, CVModelConfig, CVResults
 
 
@@ -12,7 +13,7 @@ class PyBindCVImpl(CVModel):
     @classmethod
     def _load_from(cls,
                    config: CVModelConfig,
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'PyBindCVImpl':
         """Load CV model from configuration using PyBind backend."""

nexaai/embedder.py

@@ -4,6 +4,7 @@ from abc import abstractmethod
 import numpy as np
 
 from nexaai.base import BaseModel
+from nexaai.common import PluginID
 
 
 @dataclass
@@ -21,19 +22,22 @@ class Embedder(BaseModel):
         pass
 
     @classmethod
-    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: str = "llama_cpp"):
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP):
         """
         Load an embedder from model files, routing to appropriate implementation.
 
         Args:
             model_path: Path to the model file
             tokenizer_file: Path to the tokenizer file (default: "tokenizer.json")
-            plugin_id: Plugin ID to use for the model (default: "llama_cpp")
+            plugin_id: Plugin ID to use for the model (default: PluginID.LLAMA_CPP)
 
         Returns:
             Embedder instance
         """
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.embedder_impl.mlx_embedder_impl import MLXEmbedderImpl
             return MLXEmbedderImpl._load_from(model_path, tokenizer_file, plugin_id)
         else:

nexaai/embedder_impl/mlx_embedder_impl.py

@@ -1,6 +1,7 @@
 from typing import List, Union
 import numpy as np
 
+from nexaai.common import PluginID
 from nexaai.embedder import Embedder, EmbeddingConfig
 from nexaai.mlx_backend.embedding.interface import Embedder as MLXEmbedderInterface
 from nexaai.mlx_backend.ml import ModelConfig as MLXModelConfig, SamplerConfig as MLXSamplerConfig, GenerationConfig as MLXGenerationConfig, EmbeddingConfig
@@ -13,14 +14,14 @@ class MLXEmbedderImpl(Embedder):
         self._mlx_embedder = None
 
     @classmethod
-    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: str = "mlx"):
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.MLX):
         """
         Load an embedder from model files using MLX backend.
 
         Args:
             model_path: Path to the model file
             tokenizer_file: Path to the tokenizer file (default: "tokenizer.json")
-            plugin_id: Plugin ID to use for the model (default: "mlx")
+            plugin_id: Plugin ID to use for the model (default: PluginID.MLX)
 
         Returns:
             MLXEmbedderImpl instance

nexaai/embedder_impl/pybind_embedder_impl.py

@@ -1,6 +1,7 @@
 from typing import List, Union
 import numpy as np
 
+from nexaai.common import PluginID
 from nexaai.embedder import Embedder, EmbeddingConfig
 from nexaai.binds import embedder_bind
 from nexaai.runtime import _ensure_runtime
@@ -15,20 +16,22 @@ class PyBindEmbedderImpl(Embedder):
         self._handle = _handle_ptr
 
     @classmethod
-    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: str = "llama_cpp"):
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP):
         """
         Load an embedder from model files
 
         Args:
             model_path: Path to the model file
             tokenizer_file: Path to the tokenizer file (default: "tokenizer.json")
-            plugin_id: Plugin ID to use for the model (default: "llama_cpp")
+            plugin_id: Plugin ID to use for the model (default: PluginID.LLAMA_CPP)
 
         Returns:
             PyBindEmbedderImpl instance
         """
         _ensure_runtime()
-        handle = embedder_bind.ml_embedder_create(model_path, tokenizer_file, plugin_id)
+        # Convert enum to string for C++ binding
+        plugin_id_str = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        handle = embedder_bind.ml_embedder_create(model_path, tokenizer_file, plugin_id_str)
         return cls(handle)
 
     def eject(self):

nexaai/image_gen.py

@@ -3,6 +3,7 @@ from abc import abstractmethod
 from dataclasses import dataclass
 
 from nexaai.base import BaseModel
+from nexaai.common import PluginID
 
 
 @dataclass
@@ -67,13 +68,16 @@ class ImageGen(BaseModel):
     def _load_from(cls,
                    model_path: str,
                    scheduler_config_path: str = "",
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None,
                    float16: bool = True,
                    quantize: bool = False
         ) -> 'ImageGen':
         """Load image generation model from local path, routing to appropriate implementation."""
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.image_gen_impl.mlx_image_gen_impl import MLXImageGenImpl
             return MLXImageGenImpl._load_from(model_path, scheduler_config_path, plugin_id, device_id, float16, quantize)
         else:

nexaai/image_gen_impl/mlx_image_gen_impl.py

@@ -1,8 +1,9 @@
 # Note: This code is generated by Cursor, not tested yet.
 
-from typing import List, Optional
+from typing import List, Optional, Union
 import os
 
+from nexaai.common import PluginID
 from nexaai.image_gen import ImageGen, ImageGenerationConfig, ImageSamplerConfig, SchedulerConfig, Image
 from nexaai.mlx_backend.sd.interface import ImageGen as MLXImageGenInterface
 
@@ -17,7 +18,7 @@ class MLXImageGenImpl(ImageGen):
     def _load_from(cls,
                    model_path: str,
                    scheduler_config_path: str = "",
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None,
                    float16: bool = True,
                    quantize: bool = False

nexaai/image_gen_impl/pybind_image_gen_impl.py

@@ -1,5 +1,6 @@
-from typing import List, Optional
+from typing import List, Optional, Union
 
+from nexaai.common import PluginID
 from nexaai.image_gen import ImageGen, ImageGenerationConfig, ImageSamplerConfig, SchedulerConfig, Image
 
 
@@ -13,7 +14,7 @@ class PyBindImageGenImpl(ImageGen):
     def _load_from(cls,
                    model_path: str,
                    scheduler_config_path: str = "",
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None,
                    float16: bool = True,
                    quantize: bool = False

nexaai/llm.py

@@ -1,10 +1,10 @@
-from typing import Generator, Optional
+from typing import Generator, Optional, Union
 from abc import abstractmethod
 import queue
 import threading
 
-from nexaai.common import ModelConfig, GenerationConfig, ChatMessage
-from nexaai.base import BaseModel
+from nexaai.common import ModelConfig, GenerationConfig, ChatMessage, PluginID
+from nexaai.base import BaseModel, ProfilingData
 
 class LLM(BaseModel):
     def __init__(self, m_cfg: ModelConfig = ModelConfig()):
@@ -17,11 +17,14 @@ class LLM(BaseModel):
                    local_path: str,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'LLM':
         """Load model from local path, routing to appropriate implementation."""
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.llm_impl.mlx_llm_impl import MLXLLMImpl
             return MLXLLMImpl._load_from(local_path, tokenizer_path, m_cfg, plugin_id, device_id)
         else:
@@ -37,7 +40,7 @@ class LLM(BaseModel):
         self._cancel_event.clear()
 
     @abstractmethod
-    def apply_chat_template(self, messages: list[ChatMessage]) -> 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."""
         pass
 
@@ -60,6 +63,10 @@ class LLM(BaseModel):
         """
         pass
 
+    def get_profiling_data(self) -> Optional[ProfilingData]:
+        """Get profiling data from the last generation."""  
+        pass
+
     @abstractmethod
     def save_kv_cache(self, path: str):
         """

nexaai/llm_impl/mlx_llm_impl.py

@@ -1,6 +1,7 @@
-from typing import Generator, Optional, Any
+from typing import Generator, Optional, Any, Sequence, Union
 
-from nexaai.common import ModelConfig, GenerationConfig, ChatMessage
+from nexaai.base import ProfilingData
+from nexaai.common import ModelConfig, GenerationConfig, ChatMessage, PluginID
 from nexaai.llm import LLM
 from nexaai.mlx_backend.llm.interface import LLM as MLXLLMInterface
 from nexaai.mlx_backend.ml import ModelConfig as MLXModelConfig, SamplerConfig as MLXSamplerConfig, GenerationConfig as MLXGenerationConfig, EmbeddingConfig
@@ -17,7 +18,7 @@ class MLXLLMImpl(LLM):
                    local_path: str,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None
         ) -> 'MLXLLMImpl':
         """Load model from local path using MLX backend."""
@@ -54,7 +55,13 @@ class MLXLLMImpl(LLM):
             self._mlx_llm.destroy()
             self._mlx_llm = None
 
-    def apply_chat_template(self, messages: list[ChatMessage]) -> str:
+    def apply_chat_template(
+        self,
+        messages: Sequence[ChatMessage], 
+        tools: Optional[str] = None,
+        enable_thinking: bool = True, 
+        add_generation_prompt: bool = True
+    ) -> str:
         """Apply the chat template to messages."""
         if not self._mlx_llm:
             raise RuntimeError("MLX LLM not loaded")
@@ -68,9 +75,16 @@ class MLXLLMImpl(LLM):
                     def __init__(self, role, content):
                         self.role = role
                         self.content = content
-                mlx_messages.append(MLXChatMessage(msg["role"], msg["content"]))
+                
+                # Handle both dict-style and attribute-style access
+                if hasattr(msg, 'role') and hasattr(msg, 'content'):
+                    # Message is already an object with attributes
+                    mlx_messages.append(MLXChatMessage(msg.role, msg.content))
+                else:
+                    # Message is a dict
+                    mlx_messages.append(MLXChatMessage(msg["role"], msg["content"]))
             
-            return self._mlx_llm.apply_chat_template(mlx_messages)
+            return self._mlx_llm.apply_chat_template(mlx_messages, tools=tools, enable_thinking=enable_thinking, add_generation_prompt=add_generation_prompt)
         except Exception as e:
             raise RuntimeError(f"Failed to apply chat template: {str(e)}")
 
@@ -202,6 +216,12 @@ class MLXLLMImpl(LLM):
         except Exception as e:
             raise RuntimeError(f"Failed to generate text: {str(e)}")
 
+    def get_profiling_data(self) -> Optional[ProfilingData]:
+        """Get profiling data from the last generation."""
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        return self._mlx_llm.get_profiling_data()
+
     def save_kv_cache(self, path: str):
         """
         Save the key-value cache to the file.

nexaai/llm_impl/pybind_llm_impl.py

@@ -1,8 +1,9 @@
-from typing import Generator, Optional
+from typing import Generator, Optional, Union
 import queue
 import threading
 
-from nexaai.common import ModelConfig, GenerationConfig, ChatMessage
+from nexaai.base import ProfilingData
+from nexaai.common import ModelConfig, GenerationConfig, ChatMessage, PluginID
 from nexaai.binds import llm_bind, common_bind
 from nexaai.runtime import _ensure_runtime
 from nexaai.llm import LLM
@@ -13,13 +14,14 @@ class PyBindLLMImpl(LLM):
         """Private constructor, should not be called directly."""
         super().__init__(m_cfg)
         self._handle = handle  # This is a py::capsule
+        self._profiling_data = None
 
     @classmethod
     def _load_from(cls,
                    local_path: str,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'PyBindLLMImpl':
         """Load model from local path."""
@@ -49,11 +51,13 @@ class PyBindLLMImpl(LLM):
             config.chat_template_content = m_cfg.chat_template_content
 
         # 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
         handle = llm_bind.ml_llm_create(
             model_path=local_path,
             tokenizer_path=tokenizer_path,
             model_config=config,
-            plugin_id=plugin_id,
+            plugin_id=plugin_id_str,
             device_id=device_id
         )
         return cls(handle, m_cfg)
@@ -64,7 +68,7 @@ class PyBindLLMImpl(LLM):
         del self._handle
         self._handle = None
 
-    def apply_chat_template(self, messages: list[ChatMessage]) -> 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 = [
@@ -95,13 +99,14 @@ class PyBindLLMImpl(LLM):
         # Run generation in thread
         def generate():
             try:
-                llm_bind.ml_llm_generate(
+                result = llm_bind.ml_llm_generate(
                     handle=self._handle,
                     prompt=prompt,
                     config=config,
                     on_token=on_token,
                     user_data=None
                 )
+                self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
             except Exception as e:
                 exception_container[0] = e
             finally:
@@ -143,8 +148,14 @@ class PyBindLLMImpl(LLM):
             on_token=None,  # No callback for non-streaming
             user_data=None
         )
+
+        self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
         return result.get("text", "")
 
+    def get_profiling_data(self) -> Optional[ProfilingData]:
+        """Get profiling data."""
+        return self._profiling_data
+
     def save_kv_cache(self, path: str):
         """
         Save the key-value cache to the file.

nexaai/mlx_backend/llm/interface.py

@@ -371,19 +371,19 @@ class LLM(BaseLLM, ProfilingMixin):
         cached_tokens = 0
 
         # Only offset prefix kv-cache at first round
-        if is_first_round:
+        # if is_first_round:
         
-            # Handle KV cache prefix offset if available
-            if self.kv_cache is not None and len(self.kv_cache) > 0:
-                # Get the offset from the first cache layer
-                if hasattr(self.kv_cache[0], 'offset'):
-                    cached_tokens = self.kv_cache[0].offset - 1
+        #     # Handle KV cache prefix offset if available
+        #     if self.kv_cache is not None and len(self.kv_cache) > 0:
+        #         # Get the offset from the first cache layer
+        #         if hasattr(self.kv_cache[0], 'offset'):
+        #             cached_tokens = self.kv_cache[0].offset - 1
             
-            # Process only the non-cached tokens
-            incremental_tokens = incremental_tokens[cached_tokens:] if cached_tokens > 0 else incremental_tokens
+        #     # Process only the non-cached tokens
+        #     incremental_tokens = incremental_tokens[cached_tokens:] if cached_tokens > 0 else incremental_tokens
             
-            if len(incremental_tokens) == 0:
-                raise ValueError("No tokens to process, KV cache is too long.")
+        #     if len(incremental_tokens) == 0:
+        #         raise ValueError("No tokens to process, KV cache is too long.")
             
         # Since apply_chat_template now returns incremental prompts, we can use the prompt directly
         # The prompt is already the incremental part based on global_n_past
@@ -467,7 +467,7 @@ class LLM(BaseLLM, ProfilingMixin):
         # We'll ignore the argument for now.
         return self.tokenizer.chat_template
 
-    def apply_chat_template(self, messages: Sequence[ChatMessage], tools: Optional[str] = None, enable_thinking: bool = True) -> str:
+    def apply_chat_template(self, messages: Sequence[ChatMessage], tools: Optional[str] = None, enable_thinking: bool = True, add_generation_prompt: bool = True) -> str:
         """
         Apply chat template to messages with incremental prompt support and optional tools.
         
@@ -526,7 +526,7 @@ class LLM(BaseLLM, ProfilingMixin):
                 incremental_messages,
                 tokenize=False,
                 enable_thinking=enable_thinking,   
-                add_generation_prompt=True,
+                add_generation_prompt=add_generation_prompt,
                 tools=parsed_tools
             )
         except Exception as e:

nexaai/rerank.py

@@ -1,8 +1,9 @@
-from typing import List, Optional, Sequence
+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
 
 
 @dataclass
@@ -24,11 +25,14 @@ class Reranker(BaseModel):
     def _load_from(cls,
                    model_path: str,
                    tokenizer_file: str = "tokenizer.json",
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'Reranker':
         """Load reranker model from local path, routing to appropriate implementation."""
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.rerank_impl.mlx_rerank_impl import MLXRerankImpl
             return MLXRerankImpl._load_from(model_path, tokenizer_file, plugin_id, device_id)
         else:

nexaai/rerank_impl/mlx_rerank_impl.py

@@ -1,8 +1,9 @@
 # Note: This code is generated by Cursor, not tested yet.
 
-from typing import List, Optional, Sequence
+from typing import List, Optional, Sequence, Union
 import os
 
+from nexaai.common import PluginID
 from nexaai.rerank import Reranker, RerankConfig
 from nexaai.mlx_backend.rerank.interface import Reranker as MLXRerankInterface, create_reranker
 
@@ -17,7 +18,7 @@ class MLXRerankImpl(Reranker):
     def _load_from(cls,
                    model_path: str,
                    tokenizer_file: str = "tokenizer.json",
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None
         ) -> 'MLXRerankImpl':
         """Load reranker model from local path using MLX backend."""

nexaai/rerank_impl/pybind_rerank_impl.py

@@ -1,5 +1,6 @@
-from typing import List, Optional, Sequence
+from typing import List, Optional, Sequence, Union
 
+from nexaai.common import PluginID
 from nexaai.rerank import Reranker, RerankConfig
 
 
@@ -13,7 +14,7 @@ class PyBindRerankImpl(Reranker):
     def _load_from(cls,
                    model_path: str,
                    tokenizer_file: str = "tokenizer.json",
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'PyBindRerankImpl':
         """Load reranker model from local path using PyBind backend."""

nexaai/tts.py

@@ -1,8 +1,9 @@
-from typing import List, Optional
+from typing import List, Optional, Union
 from abc import abstractmethod
 from dataclasses import dataclass
 
 from nexaai.base import BaseModel
+from nexaai.common import PluginID
 
 
 @dataclass
@@ -43,11 +44,14 @@ class TTS(BaseModel):
     def _load_from(cls,
                    model_path: str,
                    vocoder_path: str,
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'TTS':
         """Load TTS model from local path, routing to appropriate implementation."""
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.tts_impl.mlx_tts_impl import MLXTTSImpl
             return MLXTTSImpl._load_from(model_path, vocoder_path, plugin_id, device_id)
         else:

nexaai/tts_impl/mlx_tts_impl.py

@@ -1,8 +1,9 @@
 # Note: This code is generated by Cursor, not tested yet.
 
-from typing import List, Optional
+from typing import List, Optional, Union
 import os
 
+from nexaai.common import PluginID
 from nexaai.tts import TTS, TTSConfig, TTSResult
 from nexaai.mlx_backend.tts.interface import MlxTts as MLXTTSInterface
 
@@ -17,7 +18,7 @@ class MLXTTSImpl(TTS):
     def _load_from(cls,
                    model_path: str,
                    vocoder_path: str,
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None
         ) -> 'MLXTTSImpl':
         """Load TTS model from local path using MLX backend."""

nexaai/tts_impl/pybind_tts_impl.py

@@ -1,5 +1,6 @@
-from typing import List, Optional
+from typing import List, Optional, Union
 
+from nexaai.common import PluginID
 from nexaai.tts import TTS, TTSConfig, TTSResult
 
 
@@ -13,7 +14,7 @@ class PyBindTTSImpl(TTS):
     def _load_from(cls,
                    model_path: str,
                    vocoder_path: str,
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'PyBindTTSImpl':
         """Load TTS model from local path using PyBind backend."""

nexaai/vlm.py

@@ -5,8 +5,8 @@ import threading
 import base64
 from pathlib import Path
 
-from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage
-from nexaai.base import BaseModel
+from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage, PluginID
+from nexaai.base import BaseModel, ProfilingData
 
 
 class VLM(BaseModel):
@@ -20,7 +20,7 @@ class VLM(BaseModel):
                    local_path: str,
                    mmproj_path: str,
                    m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'VLM':
         """Load VLM model from local path, routing to appropriate implementation.
@@ -35,7 +35,10 @@ class VLM(BaseModel):
         Returns:
             VLM instance
         """
-        if plugin_id == "mlx":
+        # Check plugin_id value for routing - handle both enum and string
+        plugin_value = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
+        
+        if plugin_value == "mlx":
             from nexaai.vlm_impl.mlx_vlm_impl import MlxVlmImpl
             return MlxVlmImpl._load_from(local_path, mmproj_path, m_cfg, plugin_id, device_id)
         else:
@@ -117,4 +120,8 @@ class VLM(BaseModel):
         Returns:
             str: The generated text.
         """
+        pass
+
+    def get_profiling_data(self) -> Optional[ProfilingData]:
+        """Get profiling data from the last generation."""
         pass

nexaai/vlm_impl/mlx_vlm_impl.py

@@ -1,6 +1,7 @@
-from typing import Generator, Optional, List, Dict, Any
+from typing import Generator, Optional, List, Dict, Any, Union
 
-from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage
+from nexaai.base import ProfilingData
+from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage, PluginID
 from nexaai.vlm import VLM
 from nexaai.mlx_backend.vlm.interface import VLM as MLXVLMInterface
 from nexaai.mlx_backend.ml import ModelConfig as MLXModelConfig, SamplerConfig as MLXSamplerConfig, GenerationConfig as MLXGenerationConfig, EmbeddingConfig
@@ -17,7 +18,7 @@ class MlxVlmImpl(VLM):
                    local_path: str,
                    mmproj_path: str,
                    m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: str = "mlx",
+                   plugin_id: Union[PluginID, str] = PluginID.MLX,
                    device_id: Optional[str] = None
         ) -> 'MlxVlmImpl':
         """Load VLM model from local path using MLX backend.
@@ -247,3 +248,9 @@ class MlxVlmImpl(VLM):
             
         except Exception as e:
             raise RuntimeError(f"Failed to generate text: {str(e)}")
+
+    def get_profiling_data(self) -> Optional[ProfilingData]:
+        """Get profiling data from the last generation."""
+        if not self._mlx_vlm:
+            raise RuntimeError("MLX VLM not loaded")
+        return self._mlx_vlm.get_profiling_data()

nexaai/vlm_impl/pybind_vlm_impl.py

@@ -4,10 +4,11 @@ import threading
 import base64
 from pathlib import Path
 
-from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage
+from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage, PluginID
 from nexaai.binds import vlm_bind, common_bind
 from nexaai.runtime import _ensure_runtime
 from nexaai.vlm import VLM
+from nexaai.base import ProfilingData
 
 
 class PyBindVLMImpl(VLM):
@@ -15,13 +16,14 @@ class PyBindVLMImpl(VLM):
         """Private constructor, should not be called directly."""
         super().__init__(m_cfg)
         self._handle = handle  # This is a py::capsule
+        self._profiling_data = None
 
     @classmethod
     def _load_from(cls,
                    local_path: str,
                    mmproj_path: str,
                    m_cfg: ModelConfig = ModelConfig(),
-                   plugin_id: str = "llama_cpp",
+                   plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None
         ) -> 'PyBindVLMImpl':
         """Load VLM model from local path.
@@ -61,11 +63,13 @@ class PyBindVLMImpl(VLM):
             config.chat_template_content = m_cfg.chat_template_content
 
         # 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
         handle = vlm_bind.create_vlm(
             model_path=local_path,
             mmproj_path=mmproj_path,
             model_config=config,
-            plugin_id=plugin_id,
+            plugin_id=plugin_id_str,
             device_id=device_id
         )
         return cls(handle, m_cfg)
@@ -141,13 +145,14 @@ class PyBindVLMImpl(VLM):
         # Run generation in thread
         def generate():
             try:
-                vlm_bind.ml_vlm_generate(
+                result = vlm_bind.ml_vlm_generate(
                     handle=self._handle,
                     prompt=prompt,
                     config=config,
                     on_token=on_token,
                     user_data=None
                 )
+                self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
             except Exception as e:
                 exception_container[0] = e
             finally:
@@ -189,8 +194,14 @@ class PyBindVLMImpl(VLM):
             on_token=None,  # No callback for non-streaming
             user_data=None
         )
+
+        self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
         return result.get("text", "")
 
+    def get_profiling_data(self) -> Optional[ProfilingData]:
+        """Get profiling data."""
+        return self._profiling_data
+
     def _convert_generation_config(self, g_cfg: GenerationConfig):
         """Convert GenerationConfig to binding format."""
         config = common_bind.GenerationConfig()

{nexaai-1.0.4rc15.dist-info → nexaai-1.0.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.4rc15
+Version: 1.0.5
 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
@@ -17,11 +17,15 @@ Requires-Dist: tqdm
 Requires-Dist: hf_xet
 Requires-Dist: numpy
 Requires-Dist: httpx
-Requires-Dist: mlx
-Requires-Dist: mlx-lm
-Requires-Dist: scipy
-Requires-Dist: soundfile
-Requires-Dist: Pillow
-Requires-Dist: opencv-python
-Requires-Dist: shapely
-Requires-Dist: pyclipper
+Provides-Extra: mlx
+Requires-Dist: mlx; extra == "mlx"
+Requires-Dist: mlx-lm; extra == "mlx"
+Requires-Dist: mlx-vlm; extra == "mlx"
+Requires-Dist: tokenizers; extra == "mlx"
+Requires-Dist: safetensors; extra == "mlx"
+Requires-Dist: Pillow; extra == "mlx"
+Requires-Dist: scipy; extra == "mlx"
+Requires-Dist: soundfile; extra == "mlx"
+Requires-Dist: opencv-python; extra == "mlx"
+Requires-Dist: shapely; extra == "mlx"
+Requires-Dist: pyclipper; extra == "mlx"

{nexaai-1.0.4rc15.dist-info → nexaai-1.0.5.dist-info}/RECORD

@@ -1,33 +1,35 @@
-nexaai/__init__.py,sha256=JTjJWdiBXHZyc_91Oe-GNOcODFp9gbUQM43bzNY7S8Q,1906
-nexaai/_stub.cpython-310-darwin.so,sha256=6cDmmUXwfQIBTowXBsoMjB7kqSQskigaWJAHEB8aaTw,66768
-nexaai/_version.py,sha256=UxLv07_TC8sCUMr5KTEXolBn9DNXJx2RUjkBcGewdXw,143
-nexaai/asr.py,sha256=Yg8Yml_nklzJYl3C_lwvEApTdNjY2czAurDaoEjkiIU,1813
+nexaai/__init__.py,sha256=jXdC4vv6DBK1fVewYTYSUhOOYfvf_Mk81UIeMGGIKUg,2029
+nexaai/_stub.cpython-310-darwin.so,sha256=z7e7YOjkyjyalKr6u9iLGZ5YHgdvh5d5pLTwGvCCmtM,66768
+nexaai/_version.py,sha256=j-3XhaOQERPf3uculltVeo9djhrsUyWzi1EY_j4wPEc,138
+nexaai/asr.py,sha256=NljMXDErwPNMOPaRkJZMEDka9Nk8xyur7L8i924TStY,2054
 nexaai/base.py,sha256=N8PRgDFA-XPku2vWnQIofQ7ipz3pPlO6f8YZGnuhquE,982
-nexaai/common.py,sha256=VPM7NaUNaLTT7quW-u4D2uOeNrQqPjvfcgJlYGS3Qy8,1525
-nexaai/cv.py,sha256=KOaiRouiQ-YFP8FL20QuiieJfHN7DzASEi5_0m6H-E0,3032
-nexaai/embedder.py,sha256=VheiZEYBuuBjhQcvLawCz26jX0I169Xk4b9VP-ERjqU,2211
-nexaai/image_gen.py,sha256=IhLQLpmPkK9KcHteUdaQdxrnTIjk6xdyekRqeJtHfWw,4122
-nexaai/llm.py,sha256=egHa6YafNWyZy5qrmZRNZlFHO8LRUejc_gkOpK0nbnw,3105
-nexaai/rerank.py,sha256=7EEm96gpvd6kXO_Q8xSrQDlLZdAYTk0MODeNWDq70WA,1631
+nexaai/common.py,sha256=yBnIbqYaQYnfrl7IczOBh6MDibYZVxwaRJEglYcKgGs,3422
+nexaai/cv.py,sha256=RHCDo8gvBH8BkGZx7qVyp-OKxqi7E1GG9XzyaXehCNA,3273
+nexaai/embedder.py,sha256=Cw0tSHkPgd-RI62afCqQAcTHMnQhaI2CvfTMO-1JKOg,2452
+nexaai/image_gen.py,sha256=0C_5Tjj4BYmxLbmMmvwajp-yy2mmEEOKwBFnDQNPzx4,4356
+nexaai/llm.py,sha256=S1o_k2VQoF5w2wO25f142OO1R75TP89Ii69VZv8pIGo,3567
+nexaai/rerank.py,sha256=vWaBucoQ1wz-2iYnZqyFIcEjm-4Xcs1KDbFN5X8zzDQ,1872
 nexaai/runtime.py,sha256=mxxHYsb5iBUAm2K_u-XJWr_U-spJ9S4eApc8kf9myjw,1957
-nexaai/tts.py,sha256=4EbC0BfFh5TLrm_3Q5vx1sXdug5gvOi-owNeX7ekbdA,1926
-nexaai/vlm.py,sha256=g65S8ChMnp_wsz_O4szjR3Z8sD_46NHaxDlfdoZoQ0c,4291
+nexaai/tts.py,sha256=ZnBpWUxIfHhh7KfEjddtH7hHOTa91zg7ogGLakMIALo,2167
+nexaai/vlm.py,sha256=3voXmAVnGlXnOiwA3wcX4p0Lvmp0X1VKkQVPObJdwBY,4649
 nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/asr_impl/mlx_asr_impl.py,sha256=JuyxFzFbbgclK5_2Rq5pT278h0q8LztJX7Tggz0zkbM,3191
-nexaai/asr_impl/pybind_asr_impl.py,sha256=ybvthYgtVbH_JgpSsl0nxjZYvXyk8KGRSKdsJ-hLfZE,1450
+nexaai/asr_impl/mlx_asr_impl.py,sha256=eosd8-TIWAOwV0HltmoFrLwzXHcU4jyxtncvuZE9pgA,3257
+nexaai/asr_impl/pybind_asr_impl.py,sha256=pE9Hb_hMi5yAc4MF83bLVOb8zDtreCkB3_u7XED9YpA,1516
 nexaai/binds/__init__.py,sha256=T9Ua7SzHNglSeEqXlfH5ymYXRyXhNKkC9z_y_bWCNMo,80
 nexaai/binds/common_bind.cpython-310-darwin.so,sha256=hVxY76tn7hN6uHDIgM7LWNvgoudHgNZVoaygM9X1RWE,217232
 nexaai/binds/embedder_bind.cpython-310-darwin.so,sha256=FT8581RNciilskK89PhtnNSjw4Oh0-xk8QdbJVFmOd8,202064
-nexaai/binds/libnexa_bridge.dylib,sha256=cdriv4BpSoRUlNmFCVuSAHiWeVjnC8KKihCTiTNrJno,251256
-nexaai/binds/llm_bind.cpython-310-darwin.so,sha256=Bv08rn9OBAHy01eAQeANiJSrCxskn1xSx4Gl1Vcrhm0,166064
-nexaai/binds/nexa_llama_cpp/libggml-base.dylib,sha256=D_mlY_PNMIFlm6mHERSLMoA7QfYHwEPVlb0UKMbl2N0,632048
+nexaai/binds/libcrypto.dylib,sha256=aWif9WhTKVQhmZL3DmtIpMkZY5JSb_Ny6CClmUBKYM4,4710416
+nexaai/binds/libnexa_bridge.dylib,sha256=g4nlxyGyVJ-LJV1cHMDg2m2pYF8fFTBBXGTPQV-lotg,251480
+nexaai/binds/libssl.dylib,sha256=Q2frAdhR729oKYuCjJOEr1Ott3idFWoFp98fwNqtIaU,881616
+nexaai/binds/llm_bind.cpython-310-darwin.so,sha256=3Bsq0_tGkM027-bORVeJUDl6CYZxAF9sbDIn1l31XTQ,182704
+nexaai/binds/nexa_llama_cpp/libggml-base.dylib,sha256=CzsTec_QHlvbBGzmx4MBQ4LUjG7aIqW1rP5p_A90Vds,632048
 nexaai/binds/nexa_llama_cpp/libggml-cpu.so,sha256=RiMhOv6IAWY1zkFTp0JCB7CYoPfOv54vBVQHvj1koBM,661120
 nexaai/binds/nexa_llama_cpp/libggml-metal.so,sha256=L4RQvaD0w4qBjexi4O05RMCH8842fof5QgBEvyx0RcA,673104
 nexaai/binds/nexa_llama_cpp/libggml.dylib,sha256=aOTj_6RrAMkfDO0ZI28_3nfcC-l4Y3dRCiS3C0d0_eI,58592
 nexaai/binds/nexa_llama_cpp/libllama.dylib,sha256=fDPnTG6EQ1JN6aRmnIFQzag_kmtyImRxKjMOOtaTY5Q,1746928
 nexaai/binds/nexa_llama_cpp/libmtmd.dylib,sha256=ccnBRsJNFGTCsjgW03N9PvX26wUirqpxljnxdVPINVc,587008
-nexaai/binds/nexa_llama_cpp/libnexa_plugin.dylib,sha256=1kAoSq1w8pPiNmceOdmAZ7ehfENauFoGq8mpIwGl-kk,1806696
-nexaai/binds/nexa_mlx/libnexa_plugin.dylib,sha256=aw8if8RwXjb02CehbqGPHOeEKRUcTpzeJZLOkjTBm8A,596328
+nexaai/binds/nexa_llama_cpp/libnexa_plugin.dylib,sha256=bZOa9K6tROVv7sfBU6JaA7WEPvuLfJljkXH37um3WSU,2368632
+nexaai/binds/nexa_mlx/libnexa_plugin.dylib,sha256=yjbdy0FpBE_RwgqvwGxd3czIfs3OYVoh--vWpn2H7RQ,1422888
 nexaai/binds/nexa_mlx/py-lib/ml.py,sha256=LafDM_TeXmuQkld2tdQxUBGgooT0JPMXngLam2TADqU,23179
 nexaai/binds/nexa_mlx/py-lib/profiling.py,sha256=Dc-mybFwBdCIKFWL7CbSHjkOJGAoYHG7r_e_XPhzwBU,9361
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -181,17 +183,17 @@ nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_convert.py,sha256=79ddUhtT
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_interpolate.py,sha256=9dNmH03C46HtxwesH2DpT2oTNEG1KCZWYEKq6UQ3vfk,3536
 nexaai/binds/nexa_mlx/py-lib/mlx_audio/tts/tests/test_models.py,sha256=12RiOfPtSZQj5g5JM-yCJk3uGQfM3OdmRiPt5uUDE4E,35096
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/cv_impl/mlx_cv_impl.py,sha256=mdK4DEffPe96AgDGDXtQeHlG958hf8FO1fBZ1qjZMEE,3162
-nexaai/cv_impl/pybind_cv_impl.py,sha256=yS4JKfRSaIjjVP7hJ-CizG76pIX85bpmGLk9B9cnL24,998
+nexaai/cv_impl/mlx_cv_impl.py,sha256=gKECQOv8iaWwG3bl7xeqVy2NN_9K7tYerIFzfn4eLo4,3228
+nexaai/cv_impl/pybind_cv_impl.py,sha256=uSmwBste4cT7c8DQmXzRLmzwDf773PAbXNYWW1UzVls,1064
 nexaai/embedder_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/embedder_impl/mlx_embedder_impl.py,sha256=dTOz34WGDnhsI9L7Ctv6fGPngvMAUc4FwEwRgBp_M9I,4317
-nexaai/embedder_impl/pybind_embedder_impl.py,sha256=AGGrOq4z0mDpQZInOvJsOIlQWflByhDjsihMu_Wjtbk,3286
+nexaai/embedder_impl/mlx_embedder_impl.py,sha256=OsDzsc_2wZkSoWu6yCOZadMkaYdBW3uyjF11hDKTaX8,4383
+nexaai/embedder_impl/pybind_embedder_impl.py,sha256=Ga1JYauVkRq6jwAGL7Xx5HDaIx483_v9gZVoTyd3xNU,3495
 nexaai/image_gen_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/image_gen_impl/mlx_image_gen_impl.py,sha256=OxSbk9zIDj7tTvsdM8bMJQDBhpn-mygBNktewd_wgtE,11153
-nexaai/image_gen_impl/pybind_image_gen_impl.py,sha256=XoSdWG5ID_g93WT9QB0qCP64a4rX-Rva0u4fQ8xpoqg,3626
+nexaai/image_gen_impl/mlx_image_gen_impl.py,sha256=BuDkksvXyb4J02GsdnbGAmYckfUU0Eah6BimoMD3QqY,11219
+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=HCi1uQBjccDDi64LbAgyH85lWx7qDZIW8i43dojGfF0,10210
-nexaai/llm_impl/pybind_llm_impl.py,sha256=8Us4N5KF6oi-0-K_5Dpf2rYe9smd89ZfWFrP_fWBsM4,7374
+nexaai/llm_impl/mlx_llm_impl.py,sha256=4v7jUFzHfE7zw2uViekGQDaTROz8A6oaW31Z3iVe6tg,11157
+nexaai/llm_impl/pybind_llm_impl.py,sha256=aooqkcXZWhCo07wbSafGgBrA3WnijtnUADShjjgFsBQ,8051
 nexaai/mlx_backend/ml.py,sha256=LafDM_TeXmuQkld2tdQxUBGgooT0JPMXngLam2TADqU,23179
 nexaai/mlx_backend/profiling.py,sha256=Dc-mybFwBdCIKFWL7CbSHjkOJGAoYHG7r_e_XPhzwBU,9361
 nexaai/mlx_backend/asr/__init__.py,sha256=fuT_9_xpYJ28m4yjly5L2jChUrzlSQz-b_S7nujxkSM,451
@@ -211,7 +213,7 @@ nexaai/mlx_backend/embedding/modeling/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCe
 nexaai/mlx_backend/embedding/modeling/nexa_jina_v2.py,sha256=F9Z_9r-Dh0wNThiMp5W5hqE2dt5bf4ps5_c6h4BuWGw,15218
 nexaai/mlx_backend/llm/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/mlx_backend/llm/generate.py,sha256=Phes0tzxbbEWA2hDylQvD0LjorMaPwvcfZq9RKCAOt0,4399
-nexaai/mlx_backend/llm/interface.py,sha256=Fx28O2jCDPaEfr0xLffWnqGIU5Gspggxr-o54-fBWj4,29257
+nexaai/mlx_backend/llm/interface.py,sha256=SZFkuAUi2vxj_dSqj8RXf9vPTGMtpks_pZxxrF7iIe8,29330
 nexaai/mlx_backend/llm/main.py,sha256=gFDE4VZv_CLKMCTn0N521OfCKH_Ys26bHDh6g9VEFNc,1982
 nexaai/mlx_backend/mlx_audio/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/mlx_backend/mlx_audio/server.py,sha256=Pqy13Fafq4WX_cTuvRFz1jq89beQm2QQGpXmhK4b9jc,17547
@@ -502,19 +504,19 @@ nexaai/mlx_backend/vlm/modeling/trainer/lora.py,sha256=tGjvenjEQ8_1Az8Nz3smz5Mgv
 nexaai/mlx_backend/vlm/modeling/trainer/trainer.py,sha256=h16SaHt76JzFruXuidgXDx7_2evx4L0SecvzqLmhyZw,9081
 nexaai/mlx_backend/vlm/modeling/trainer/utils.py,sha256=29oHf_7946YeJKP_-Dt-NPeN4xJq8Fj7Yv4jZKO9RWA,4909
 nexaai/rerank_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/rerank_impl/mlx_rerank_impl.py,sha256=I-jumShLm1jAmKunRcDNUU4yjfWLoWClFMFONd88-Es,3177
-nexaai/rerank_impl/pybind_rerank_impl.py,sha256=FIIN96zCxXopqpqZdBd7OjuqqviFBY8HMZek1bCeoJw,1447
+nexaai/rerank_impl/mlx_rerank_impl.py,sha256=h37PKSIRBY8mwzVeLeP4ix9ui3waIsg4gorzelYLJbM,3243
+nexaai/rerank_impl/pybind_rerank_impl.py,sha256=CtwkG7YrW58GPMDERJSnISGTVCXWNju5__R2W837t7c,1513
 nexaai/tts_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/tts_impl/mlx_tts_impl.py,sha256=D71IFtIYWzrVdBS2y5vDBWjZ4ZAzRRjFHC0KO0pA5BU,3035
-nexaai/tts_impl/pybind_tts_impl.py,sha256=Be5QiXzDz6h1LTIQzUBd0ZyBs7rUpNA-pULCXFtt2Is,1378
+nexaai/tts_impl/mlx_tts_impl.py,sha256=i_uNPdvlXYtL3e01oKjDlP9jgkWCRt1bBHsExaaiJi8,3101
+nexaai/tts_impl/pybind_tts_impl.py,sha256=mpn44r6pfYLIl-NrEy2dXHjGtWtNCmM7HRyxiANxUI4,1444
 nexaai/utils/avatar_fetcher.py,sha256=bWy8ujgbOiTHFCjFxTwkn3uXbZ84PgEGUkXkR3MH4bI,3821
 nexaai/utils/decode.py,sha256=61n4Zf6c5QLyqGoctEitlI9BX3tPlP2a5aaKNHbw3T4,404
 nexaai/utils/model_manager.py,sha256=c07ocxxw1IHCQw6esbmYK0dX2R2OajfEIGsC_2teHXo,48572
 nexaai/utils/progress_tracker.py,sha256=76HlPkyN41IMHSsH56-qdlN_aY_oBfJz50J16Cx67R0,15102
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/vlm_impl/mlx_vlm_impl.py,sha256=7gm_tFNox3LC78DQEtlMQ-eBK55zDY0xWlJghUAOP5Y,10402
-nexaai/vlm_impl/pybind_vlm_impl.py,sha256=C-3fa0AIypI33OAGuGfVxo1V7zN0wjQMgruKlDIlW4Q,8333
-nexaai-1.0.4rc15.dist-info/METADATA,sha256=izOUOhvRNpO73EELnKolgU0Kn_PK79tsJkJr3RMWBzA,883
-nexaai-1.0.4rc15.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
-nexaai-1.0.4rc15.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.4rc15.dist-info/RECORD,,
+nexaai/vlm_impl/mlx_vlm_impl.py,sha256=od1R1mRoIgPG3NHC7JiDlcB_YJY8aklX8Em3ZkeHNpE,10734
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=5ZMFgDATthmMzjrd-vE5KX5ZAMoWPYbF_FTLz8DBKIk,8908
+nexaai-1.0.5.dist-info/METADATA,sha256=fxZuww9PtpWooTJNDcf04tzwKi7AEThWyD0Z9O5a0rY,1151
+nexaai-1.0.5.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
+nexaai-1.0.5.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.5.dist-info/RECORD,,