nexaai 1.0.4rc10__py3-none-macosx_11_0_arm64.whl

nexaai/llm_impl/mlx_llm_impl.py

@@ -0,0 +1,249 @@
+from typing import Generator, Optional, Any
+
+from nexaai.common import ModelConfig, GenerationConfig, ChatMessage
+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
+
+
+class MLXLLMImpl(LLM):
+    def __init__(self, m_cfg: ModelConfig = ModelConfig()):
+        """Initialize MLX LLM implementation."""
+        super().__init__(m_cfg)
+        self._mlx_llm = None
+
+    @classmethod
+    def _load_from(cls,
+                   local_path: str,
+                   tokenizer_path: Optional[str] = None,
+                   m_cfg: ModelConfig = ModelConfig(),
+                   plugin_id: str = "mlx",
+                   device_id: Optional[str] = None
+        ) -> 'MLXLLMImpl':
+        """Load model from local path using MLX backend."""
+        try:
+            # MLX interface and configs are already imported
+            
+            # Convert our ModelConfig to MLX ModelConfig
+            mlx_config = MLXModelConfig()
+            mlx_config.n_ctx = m_cfg.n_ctx
+            mlx_config.n_threads = m_cfg.n_threads
+            mlx_config.n_threads_batch = m_cfg.n_threads_batch
+            mlx_config.n_batch = m_cfg.n_batch
+            mlx_config.n_ubatch = m_cfg.n_ubatch
+            mlx_config.n_seq_max = m_cfg.n_seq_max
+            mlx_config.chat_template_path = m_cfg.chat_template_path
+            mlx_config.chat_template_content = m_cfg.chat_template_content
+            
+            # Create instance and load MLX model
+            instance = cls(m_cfg)
+            instance._mlx_llm = MLXLLMInterface(
+                model_path=local_path,
+                tokenizer_path=tokenizer_path or local_path,
+                config=mlx_config,
+                device=device_id
+            )
+            
+            return instance
+        except Exception as e:
+            raise RuntimeError(f"Failed to load MLX LLM: {str(e)}")
+
+    def eject(self):
+        """Release the model from memory."""
+        if self._mlx_llm:
+            self._mlx_llm.destroy()
+            self._mlx_llm = None
+
+    def apply_chat_template(self, messages: list[ChatMessage]) -> str:
+        """Apply the chat template to messages."""
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        
+        try:
+            # Convert to MLX ChatMessage format
+            mlx_messages = []
+            for msg in messages:
+                # Create a simple object with role and content attributes
+                class MLXChatMessage:
+                    def __init__(self, role, content):
+                        self.role = role
+                        self.content = content
+                mlx_messages.append(MLXChatMessage(msg["role"], msg["content"]))
+            
+            return self._mlx_llm.apply_chat_template(mlx_messages)
+        except Exception as e:
+            raise RuntimeError(f"Failed to apply chat template: {str(e)}")
+
+    def generate_stream(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> Generator[str, None, None]:
+        """Generate text with streaming."""
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        
+        try:
+            import queue
+            import threading
+            
+            # Convert GenerationConfig to MLX format
+            
+            mlx_gen_config = MLXGenerationConfig()
+            mlx_gen_config.max_tokens = g_cfg.max_tokens
+            mlx_gen_config.stop = g_cfg.stop_words
+            mlx_gen_config.image_paths = g_cfg.image_paths
+            mlx_gen_config.audio_paths = g_cfg.audio_paths
+            
+            if g_cfg.sampler_config:
+                mlx_sampler_config = MLXSamplerConfig()
+                mlx_sampler_config.temperature = g_cfg.sampler_config.temperature
+                mlx_sampler_config.top_p = g_cfg.sampler_config.top_p
+                mlx_sampler_config.top_k = g_cfg.sampler_config.top_k
+                mlx_sampler_config.repetition_penalty = g_cfg.sampler_config.repetition_penalty
+                mlx_sampler_config.presence_penalty = g_cfg.sampler_config.presence_penalty
+                mlx_sampler_config.frequency_penalty = g_cfg.sampler_config.frequency_penalty
+                mlx_sampler_config.seed = g_cfg.sampler_config.seed
+                mlx_sampler_config.grammar_path = g_cfg.sampler_config.grammar_path
+                mlx_sampler_config.grammar_string = g_cfg.sampler_config.grammar_string
+                mlx_gen_config.sampler_config = mlx_sampler_config
+            
+            # Create a queue for streaming tokens
+            token_queue = queue.Queue()
+            exception_container = [None]
+            self.reset_cancel()  # Reset cancel flag before generation
+            
+            def token_callback(token: str, user_data: Any = None) -> bool:
+                if self._cancel_event.is_set():
+                    token_queue.put(('end', None))
+                    return False
+                try:
+                    token_queue.put(('token', token))
+                    return True
+                except Exception as e:
+                    exception_container[0] = e
+                    return False
+            
+            # Run generation in a separate thread
+            def generate():
+                try:
+                    self._mlx_llm.generate_stream(prompt, mlx_gen_config, token_callback)
+                except Exception as e:
+                    exception_container[0] = e
+                finally:
+                    token_queue.put(('end', None))
+            
+            thread = threading.Thread(target=generate)
+            thread.start()
+            
+            # Yield tokens as they come from the queue
+            while True:
+                if exception_container[0]:
+                    raise exception_container[0]
+                    
+                try:
+                    msg_type, token = token_queue.get(timeout=0.1)
+                    if msg_type == 'end':
+                        break
+                    elif msg_type == 'token':
+                        yield token
+                except queue.Empty:
+                    if not thread.is_alive():
+                        break
+                    continue
+                    
+            thread.join()
+            
+            if exception_container[0]:
+                raise exception_container[0]
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate streaming text: {str(e)}")
+
+    def generate(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> str:
+        """
+        Generate text without streaming.
+
+        Args:
+            prompt (str): The prompt to generate text from.
+            g_cfg (GenerationConfig): Generation configuration.
+
+        Returns:
+            str: The generated text.
+        """
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        
+        try:
+            # Convert GenerationConfig to MLX format
+            
+            mlx_gen_config = MLXGenerationConfig()
+            mlx_gen_config.max_tokens = g_cfg.max_tokens
+            mlx_gen_config.stop = g_cfg.stop_words
+            mlx_gen_config.image_paths = g_cfg.image_paths
+            mlx_gen_config.audio_paths = g_cfg.audio_paths
+            
+            if g_cfg.sampler_config:
+                mlx_sampler_config = MLXSamplerConfig()
+                mlx_sampler_config.temperature = g_cfg.sampler_config.temperature
+                mlx_sampler_config.top_p = g_cfg.sampler_config.top_p
+                mlx_sampler_config.top_k = g_cfg.sampler_config.top_k
+                mlx_sampler_config.repetition_penalty = g_cfg.sampler_config.repetition_penalty
+                mlx_sampler_config.presence_penalty = g_cfg.sampler_config.presence_penalty
+                mlx_sampler_config.frequency_penalty = g_cfg.sampler_config.frequency_penalty
+                mlx_sampler_config.seed = g_cfg.sampler_config.seed
+                mlx_sampler_config.grammar_path = g_cfg.sampler_config.grammar_path
+                mlx_sampler_config.grammar_string = g_cfg.sampler_config.grammar_string
+                mlx_gen_config.sampler_config = mlx_sampler_config
+            
+            # Simple token callback that just continues
+            def token_callback(token: str, user_data: Any = None) -> bool:
+                return not self._cancel_event.is_set()
+            
+            # Use MLX streaming generation and return the full result
+            return self._mlx_llm.generate_stream(prompt, mlx_gen_config, token_callback)
+            
+        except Exception as e:
+            raise RuntimeError(f"Failed to generate text: {str(e)}")
+
+    def save_kv_cache(self, path: str):
+        """
+        Save the key-value cache to the file.
+
+        Args:
+            path (str): The path to the file.
+        """
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        
+        try:
+            success = self._mlx_llm.save_kv_cache(path)
+            if not success:
+                raise RuntimeError("Failed to save KV cache")
+        except Exception as e:
+            raise RuntimeError(f"Failed to save KV cache: {str(e)}")
+
+    def load_kv_cache(self, path: str):
+        """
+        Load the key-value cache from the file.
+
+        Args:
+            path (str): The path to the file.
+        """
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        
+        try:
+            success = self._mlx_llm.load_kv_cache(path)
+            if not success:
+                raise RuntimeError("Failed to load KV cache")
+        except Exception as e:
+            raise RuntimeError(f"Failed to load KV cache: {str(e)}")
+
+    def reset(self):
+        """
+        Reset the LLM model context and KV cache.
+        """
+        if not self._mlx_llm:
+            raise RuntimeError("MLX LLM not loaded")
+        
+        try:
+            self._mlx_llm.reset()
+        except Exception as e:
+            raise RuntimeError(f"Failed to reset MLX LLM: {str(e)}")

nexaai/llm_impl/pybind_llm_impl.py

@@ -0,0 +1,207 @@
+from typing import Generator, Optional
+import queue
+import threading
+
+from nexaai.common import ModelConfig, GenerationConfig, ChatMessage
+from nexaai.binds import llm_bind, common_bind
+from nexaai.runtime import _ensure_runtime
+from nexaai.llm import LLM
+
+
+class PyBindLLMImpl(LLM):
+    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
+
+    @classmethod
+    def _load_from(cls,
+                   local_path: str,
+                   tokenizer_path: Optional[str] = None,
+                   m_cfg: ModelConfig = ModelConfig(),
+                   plugin_id: str = "llama_cpp",
+                   device_id: Optional[str] = None
+        ) -> 'PyBindLLMImpl':
+        """Load model from local path."""
+        _ensure_runtime()
+
+        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
+        if m_cfg.n_gpu_layers is not None:
+            config.n_gpu_layers = m_cfg.n_gpu_layers
+
+        # handle chat template strings
+        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
+
+        # Create handle : returns py::capsule with automatic cleanup
+        handle = llm_bind.ml_llm_create(
+            model_path=local_path,
+            tokenizer_path=tokenizer_path,
+            model_config=config,
+            plugin_id=plugin_id,
+            device_id=device_id
+        )
+        return cls(handle, m_cfg)
+
+    def eject(self):
+        """Release the model from memory."""
+        # py::capsule handles cleanup automatically
+        del self._handle
+        self._handle = None
+
+    def apply_chat_template(self, messages: list[ChatMessage]) -> 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)
+
+    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]
+        self.reset_cancel()  # Reset cancel flag before generation
+
+        def on_token(token: str, user_data) -> bool:
+            if self._cancel_event.is_set():
+                token_queue.put(('end', None))
+                return False  # Stop generation
+            try:
+                token_queue.put(('token', token))
+                return True  # Continue generation
+            except Exception as e:
+                exception_container[0] = e
+                return False  # Stop generation
+
+        config = self._convert_generation_config(g_cfg)
+
+        # Run generation in thread
+        def generate():
+            try:
+                llm_bind.ml_llm_generate(
+                    handle=self._handle,
+                    prompt=prompt,
+                    config=config,
+                    on_token=on_token,
+                    user_data=None
+                )
+            except Exception as e:
+                exception_container[0] = e
+            finally:
+                token_queue.put(('end', None))
+
+        thread = threading.Thread(target=generate)
+        thread.start()
+
+        # Yield tokens as they come
+        try:
+            while True:
+                msg_type, token = token_queue.get()
+                if msg_type == 'token':
+                    yield token
+                elif msg_type in ('error', 'end'):
+                    break
+        finally:
+            thread.join()
+
+        if exception_container[0]:
+            raise exception_container[0]
+
+    def generate(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> str:
+        """
+        Generate text without streaming.
+
+        Args:
+            prompt (str): The prompt to generate text from. For chat models, this is the chat messages after chat template is applied.
+            g_cfg (GenerationConfig): Generation configuration.
+
+        Returns:
+            str: The generated text.
+        """
+        config = self._convert_generation_config(g_cfg)
+        result = llm_bind.ml_llm_generate(
+            handle=self._handle,
+            prompt=prompt,
+            config=config,
+            on_token=None,  # No callback for non-streaming
+            user_data=None
+        )
+        return result.get("text", "")
+
+    def save_kv_cache(self, path: str):
+        """
+        Save the key-value cache to the file.
+
+        Args:
+            path (str): The path to the file.
+        """
+        llm_bind.ml_llm_save_kv_cache(self._handle, path)
+
+    def load_kv_cache(self, path: str):
+        """
+        Load the key-value cache from the file.
+
+        Args:
+            path (str): The path to the file.
+        """
+        llm_bind.ml_llm_load_kv_cache(self._handle, path)
+
+    def reset(self):
+        """
+        Reset the LLM model context and KV cache. If not reset, the model will skip the number of evaluated tokens and treat tokens after those as the new incremental tokens.
+        If your past chat history changed, or you are starting a new chat, you should always reset the model before running generate.
+        """
+        llm_bind.ml_llm_reset(self._handle)
+
+    def _convert_generation_config(self, g_cfg: GenerationConfig):
+        """Convert GenerationConfig to binding format."""
+        config = common_bind.GenerationConfig()
+
+        # Set basic generation parameters
+        config.max_tokens = g_cfg.max_tokens
+
+        if g_cfg.stop_words:
+            config.stop = g_cfg.stop_words
+
+        if g_cfg.image_paths:
+            config.image_paths = g_cfg.image_paths
+
+        if g_cfg.audio_paths:
+            config.audio_paths = g_cfg.audio_paths
+
+        if g_cfg.sampler_config:
+            sampler = common_bind.SamplerConfig()
+            sampler.temperature = g_cfg.sampler_config.temperature
+            sampler.top_p = g_cfg.sampler_config.top_p
+            sampler.top_k = g_cfg.sampler_config.top_k
+            sampler.repetition_penalty = g_cfg.sampler_config.repetition_penalty
+            sampler.presence_penalty = g_cfg.sampler_config.presence_penalty
+            sampler.frequency_penalty = g_cfg.sampler_config.frequency_penalty
+            sampler.seed = g_cfg.sampler_config.seed
+
+            if g_cfg.sampler_config.grammar_path:
+                sampler.grammar_path = g_cfg.sampler_config.grammar_path
+
+            if g_cfg.sampler_config.grammar_string:
+                sampler.grammar_string = g_cfg.sampler_config.grammar_string
+
+            config.sampler_config = sampler
+
+        return config

nexaai/mlx_backend/asr/__init__.py

@@ -0,0 +1,12 @@
+# patching the _resume method in phonemizer because logger.setLevel(logging.ERROR) doesn't work - the logger instance is created and stored in the package.
+try:
+    from phonemizer.backend.espeak.words_mismatch import BaseWordsMismatch
+    
+    def silent_resume(self, nmismatch, nlines):
+        """Silent version of _resume that suppresses warnings"""
+        pass
+    
+    BaseWordsMismatch._resume = silent_resume
+    
+except ImportError:
+    pass

nexaai/mlx_backend/asr/interface.py

@@ -0,0 +1,122 @@
+from typing import Any, List, Optional, Sequence
+import argparse
+import sys
+import os
+
+import mlx.core as mx
+import numpy as np
+
+from ml import ASR, ASRConfig, ASRResult, Path as MLPath
+from mlx_audio.stt.utils import load_model
+from mlx_audio.stt.models.whisper.tokenizer import LANGUAGES
+from mlx_audio.stt.models.whisper.whisper import Model
+import soundfile as sf
+import scipy.signal
+
+from profiling import ProfilingMixin, StopReason
+
+
+class MlxAsr(ASR, ProfilingMixin):
+    """MLX Audio implementation of ASR interface."""
+    
+    def __init__(
+        self,
+        model_path: MLPath,
+        tokenizer_path: Optional[MLPath],
+        language: Optional[str],
+        device: Optional[str] = None,
+    ) -> None:
+        # Initialize profiling mixin
+        ProfilingMixin.__init__(self)
+        
+        if os.path.isfile(model_path):
+            model_path = os.path.dirname(model_path)
+        
+        super().__init__(model_path, tokenizer_path, language, device)
+        
+        # Load model immediately in constructor
+        self.model: Model = load_model(model_path)
+        self.model_path = model_path
+        
+    def destroy(self) -> None:
+        """Destroy the model and free resources."""
+        if self.model is not None:
+            del self.model
+            self.model = None
+            mx.clear_cache()
+            
+    def close(self) -> None:
+        """Close the model."""
+        self.destroy()
+        
+    def transcribe(
+        self,
+        audio_path: MLPath,
+        language: Optional[str] = None,
+        config: Optional[ASRConfig] = None,
+        clear_cache: bool = True,
+    ) -> ASRResult:
+        """Transcribe audio file to text."""
+        if self.model is None:
+            raise RuntimeError("Model not loaded")
+        
+        # Start profiling
+        self._start_profiling()
+        self._decode_start()
+        
+        try:
+            result = self.model.generate(audio_path)
+            
+            if clear_cache:
+                mx.clear_cache()
+            
+            self._decode_end()
+            self._set_stop_reason(StopReason.ML_STOP_REASON_COMPLETED)
+            self._end_profiling()
+        except Exception as e:
+            self._end_profiling()
+            raise RuntimeError(f"Failed to transcribe audio file {audio_path}: {e}")
+        
+        # Extract confidence scores and timestamps
+        confidence_scores = []
+        timestamps = []
+        
+        # Handle different result types: Whisper (STTOutput) vs Parakeet (AlignedResult)
+        if hasattr(result, 'segments') and result.segments:
+            # Whisper STTOutput format
+            for segment in result.segments:
+                if 'avg_logprob' in segment:
+                    # Convert log probability to confidence score (0-1)
+                    confidence = max(0.0, min(1.0, np.exp(segment['avg_logprob'])))
+                    confidence_scores.append(confidence)
+                else:
+                    confidence_scores.append(0.5)  # Default confidence
+                    
+                start_time = segment.get('start', 0.0)
+                end_time = segment.get('end', 0.0)
+                timestamps.append((start_time, end_time))
+        elif hasattr(result, 'sentences') and result.sentences:
+            # Parakeet AlignedResult format
+            for sentence in result.sentences:
+                confidence_scores.append(0.5)  # Default confidence for Parakeet
+                timestamps.append((sentence.start, sentence.end))
+        else:
+            # Single segment case or empty result
+            confidence_scores.append(0.5)
+            timestamps.append((0.0, 0.0))  # Default timestamps
+            
+        return ASRResult(
+            transcript=result.text,
+            confidence_scores=confidence_scores,
+            timestamps=timestamps,
+            duration_us=self._get_audio_duration_us(audio_path)
+        )
+        
+    def list_supported_languages(self) -> List[str]:
+        """List supported languages."""
+        return list(LANGUAGES.keys())
+
+    def _get_audio_duration_us(self, audio_path: MLPath) -> int:
+        with sf.SoundFile(audio_path) as f:
+            duration_us = f.frames / f.samplerate * 1e6
+        return int(duration_us)

nexaai/mlx_backend/common/utils.py

@@ -0,0 +1,25 @@
+import atexit
+
+# Store the original atexit.register function
+_original_atexit_register = atexit.register
+
+def _filtered_atexit_register(func, *args, **kwargs):
+    """
+    Clean atexit interceptor that skips nanobind handlers to prevent segfaults due to MLX atexit cleanups.
+    This should be registered early during Python runtime initialization.
+    """
+    # Skip nanobind handlers silently
+    func_type_str = str(type(func))
+    if 'nanobind' in func_type_str or func_type_str.startswith("<class 'nb_"):
+        return lambda: None
+
+    # Allow all other handlers to register normally
+    return _original_atexit_register(func, *args, **kwargs)
+
+def install_atexit_filter():
+    """Install the atexit filter to prevent problematic nanobind registrations."""
+    atexit.register = _filtered_atexit_register
+
+def uninstall_atexit_filter():
+    """Restore the original atexit.register function."""
+    atexit.register = _original_atexit_register