nexaai 0.0.0__cp310-cp310-macosx_14_0_universal2.whl → 1.0.4__cp310-cp310-macosx_14_0_universal2.whl

nexaai/_version.py

@@ -1,4 +1,4 @@
 # This file is generated by CMake from _version.py.in
 # Do not modify this file manually - it will be overwritten
 
-__version__ = "0.0.0" 
+__version__ = "1.0.4" 

nexaai/cv.py

@@ -69,6 +69,7 @@ class CVModel(BaseModel):
 
     @classmethod
     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",
                    device_id: Optional[str] = None

nexaai/utils/progress_tracker.py

@@ -62,8 +62,8 @@ class DownloadProgressTracker:
         self.last_display_length = 0
         
         # Speed tracking
-        self.last_downloaded = 0
-        self.last_time = time.time()
+        self.last_downloaded = None  # Use None to indicate no previous measurement
+        self.last_time = None  # Use None to indicate no previous time measurement
         self.speed_history = []
         self.max_speed_history = 10
         
@@ -101,18 +101,27 @@ class DownloadProgressTracker:
     def calculate_speed(self, current_downloaded: int) -> float:
         """Calculate download speed in bytes per second."""
         current_time = time.time()
-        time_diff = current_time - self.last_time
         
-        if time_diff > 0 and self.last_downloaded > 0:
-            bytes_diff = current_downloaded - self.last_downloaded
-            speed = bytes_diff / time_diff
-            
-            # Add to speed history for smoothing
-            self.speed_history.append(speed)
-            if len(self.speed_history) > self.max_speed_history:
-                self.speed_history.pop(0)
+        # Check if we have a previous measurement to compare against
+        if self.last_time is not None and self.last_downloaded is not None:
+            time_diff = current_time - self.last_time
             
-            # Return smoothed speed
+            # Only calculate if we have a meaningful time difference (avoid division by very small numbers)
+            if time_diff > 0.5:  # At least 500ms between measurements
+                bytes_diff = current_downloaded - self.last_downloaded
+                
+                # Only calculate speed if bytes actually changed
+                if bytes_diff >= 0:  # Allow 0 for periods with no progress
+                    speed = bytes_diff / time_diff
+                    
+                    # Add to speed history for smoothing
+                    self.speed_history.append(speed)
+                    if len(self.speed_history) > self.max_speed_history:
+                        self.speed_history.pop(0)
+        
+        # Return the average of historical speeds if we have any
+        # This ensures we show the last known speed even when skipping updates
+        if self.speed_history:
             return sum(self.speed_history) / len(self.speed_history)
         
         return 0.0

nexaai/vlm_impl/mlx_vlm_impl.py

@@ -76,8 +76,10 @@ class MlxVlmImpl(VLM):
             raise RuntimeError("MLX VLM not loaded")
         
         try:
-            # Convert MultiModalMessage to MLX format
             mlx_messages = []
+            total_images = 0
+            total_audios = 0
+            
             for msg in messages:
                 # Create a simple object with role and content attributes
                 class MLXChatMessage:
@@ -85,19 +87,38 @@ class MlxVlmImpl(VLM):
                         self.role = role
                         self.content = content
                 
-                # For MLX VLM, we need to extract text content from multimodal messages
-                # This is a simplified approach - the actual implementation may need
-                # more sophisticated handling of different content types
+                # Extract text content and count media files
                 text_content = ""
+                first_content = True
+                
                 for content_item in msg["content"]:
-                    if content_item["type"] == "text":
+                    content_type = content_item.get("type", "")
+                    
+                    if content_type == "text":
+                        if not first_content:
+                            text_content += " "
                         text_content += content_item.get("text", "")
-                    # Note: image/audio/video content is typically handled separately
-                    # in the generation phase, not in the chat template
+                        first_content = False
+                    elif content_type == "image":
+                        total_images += 1
+                    elif content_type == "audio":
+                        total_audios += 1
                 
                 mlx_messages.append(MLXChatMessage(msg["role"], text_content))
             
-            return self._mlx_vlm.apply_chat_template(mlx_messages)
+            if total_images > 0 or total_audios > 0:
+                # Use apply_chat_template_with_media when media is present
+                return self._mlx_vlm.apply_chat_template_with_media(
+                    mlx_messages,
+                    num_images=total_images,
+                    num_audios=total_audios,
+                    tools=tools,
+                    enable_thinking=False  # Default to False, could be made configurable
+                )
+            else:
+                # Use regular apply_chat_template for text-only messages
+                return self._mlx_vlm.apply_chat_template(mlx_messages)
+                
         except Exception as e:
             raise RuntimeError(f"Failed to apply chat template: {str(e)}")
 
@@ -107,9 +128,6 @@ class MlxVlmImpl(VLM):
             raise RuntimeError("MLX VLM not loaded")
         
         try:
-            # Get MLX config classes
-            _, MLXSamplerConfig, MLXGenerationConfig, _ = get_mlx_configs()
-            
             # Convert GenerationConfig to MLX format
             mlx_gen_config = MLXGenerationConfig()
             mlx_gen_config.max_tokens = g_cfg.max_tokens
@@ -130,25 +148,57 @@ class MlxVlmImpl(VLM):
                 mlx_sampler_config.grammar_string = g_cfg.sampler_config.grammar_string
                 mlx_gen_config.sampler_config = mlx_sampler_config
             
-            # Create a token callback for streaming
-            def token_callback(token: str) -> bool:
-                # Check if generation should be cancelled
-                return not self._cancel_event.is_set()
+            import queue
+            import threading
+            
+            # Create a queue for streaming tokens
+            token_queue = queue.Queue()
+            exception_container = [None]
+            self.reset_cancel()  # Reset cancel flag before generation
             
-            # Use MLX VLM streaming generation
-            result = self._mlx_vlm.generate_stream(prompt, mlx_gen_config, token_callback)
+            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
             
-            # MLX VLM interface returns a GenerationResult, extract the text
-            if hasattr(result, 'text') and result.text:
-                # Split the result into words and yield them
-                words = result.text.split()
-                for i, word in enumerate(words):
-                    if self._cancel_event.is_set():
+            # Run generation in a separate thread
+            def generate():
+                try:
+                    self._mlx_vlm.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
-                    if i == 0:
-                        yield word
-                    else:
-                        yield " " + word
+                    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)}")
@@ -168,9 +218,6 @@ class MlxVlmImpl(VLM):
             raise RuntimeError("MLX VLM not loaded")
         
         try:
-            # Get MLX config classes
-            _, MLXSamplerConfig, MLXGenerationConfig, _ = get_mlx_configs()
-            
             # Convert GenerationConfig to MLX format
             mlx_gen_config = MLXGenerationConfig()
             mlx_gen_config.max_tokens = g_cfg.max_tokens
@@ -191,15 +238,12 @@ class MlxVlmImpl(VLM):
                 mlx_sampler_config.grammar_string = g_cfg.sampler_config.grammar_string
                 mlx_gen_config.sampler_config = mlx_sampler_config
             
-            # Use MLX VLM generation
-            result = self._mlx_vlm.generate(prompt, mlx_gen_config)
+            # Simple token callback that just continues
+            def token_callback(token: str, user_data: Any = None) -> bool:
+                return not self._cancel_event.is_set()
             
-            # MLX VLM interface returns a GenerationResult, extract the text
-            if hasattr(result, 'text'):
-                return result.text
-            else:
-                # Fallback if result is just a string
-                return str(result)
+            # Use MLX streaming generation and return the full result
+            return self._mlx_vlm.generate_stream(prompt, mlx_gen_config, token_callback)
             
         except Exception as e:
             raise RuntimeError(f"Failed to generate text: {str(e)}")

{nexaai-0.0.0.dist-info → nexaai-1.0.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 0.0.0
+Version: 1.0.4
 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
@@ -17,10 +17,11 @@ Requires-Dist: tqdm
 Requires-Dist: hf_xet
 Requires-Dist: numpy
 Requires-Dist: httpx
-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: mlx
+Requires-Dist: mlx-lm
+Requires-Dist: scipy
+Requires-Dist: soundfile
+Requires-Dist: Pillow
+Requires-Dist: opencv-python
+Requires-Dist: shapely
+Requires-Dist: pyclipper

{nexaai-0.0.0.dist-info → nexaai-1.0.4.dist-info}/RECORD

@@ -1,10 +1,10 @@
 nexaai/__init__.py,sha256=JTjJWdiBXHZyc_91Oe-GNOcODFp9gbUQM43bzNY7S8Q,1906
-nexaai/_stub.cpython-310-darwin.so,sha256=f1y4Gg4RGKC4h9GeU7gnRfm9a1QNFwogco7bkH1KnDE,66768
-nexaai/_version.py,sha256=8uyKXwDbAQT8F8B_GXLSK40bIOCYzI_DlSjnfsNoc3s,138
+nexaai/_stub.cpython-310-darwin.so,sha256=zHzIOZ9uQ-tqMPBbqJ8yGaEu-Rx-r45Es1FYTp741yI,66768
+nexaai/_version.py,sha256=lhnMmDQ6cHv0o4YRcKJ1pQ1UjlE_R6m7l1igB1Oe6tM,138
 nexaai/asr.py,sha256=Yg8Yml_nklzJYl3C_lwvEApTdNjY2czAurDaoEjkiIU,1813
 nexaai/base.py,sha256=N8PRgDFA-XPku2vWnQIofQ7ipz3pPlO6f8YZGnuhquE,982
 nexaai/common.py,sha256=VPM7NaUNaLTT7quW-u4D2uOeNrQqPjvfcgJlYGS3Qy8,1525
-nexaai/cv.py,sha256=CYfLSDU0_QJkcaIDIJ-a-JjA9FMvIUrSfG71_7-79hI,2934
+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
@@ -18,15 +18,15 @@ nexaai/asr_impl/pybind_asr_impl.py,sha256=ybvthYgtVbH_JgpSsl0nxjZYvXyk8KGRSKdsJ-
 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=zgnLSGXKm0e89H08A85b9pKYgcNE5W6Uv0Upc2FgiCg,251256
+nexaai/binds/libnexa_bridge.dylib,sha256=jBHp9IthZkkQu-RIrzHYPOS4PehnQ6TFhMWbVGUTZCk,251256
 nexaai/binds/llm_bind.cpython-310-darwin.so,sha256=Bv08rn9OBAHy01eAQeANiJSrCxskn1xSx4Gl1Vcrhm0,166064
-nexaai/binds/nexa_llama_cpp/libggml-base.dylib,sha256=i2aH2Gt0WRVFkuEfgWsE2hnRE2uZRXH83WQfrBrT-oI,631840
-nexaai/binds/nexa_llama_cpp/libggml-cpu.so,sha256=1qFcy_h2kozVyVaTF5oyQA7xrutzkw_LUn_jIw-9mTs,659216
-nexaai/binds/nexa_llama_cpp/libggml-metal.so,sha256=fCbK0DD8Uospc653JD0iaf7EJxRiXajIm9OEGE2nd9Q,673104
-nexaai/binds/nexa_llama_cpp/libggml.dylib,sha256=aOxGtGHFL5sn7ihZTOrcFYqn941O6RMpK3CwWYniN0g,58592
-nexaai/binds/nexa_llama_cpp/libllama.dylib,sha256=hmKPTc8OUHL985DAiVlh8tT4gVV5Fm6Vk7GGBXBPZds,1746592
-nexaai/binds/nexa_llama_cpp/libmtmd.dylib,sha256=TuDI-WK7LSNISNCIhTCzvCwgZSFRLC5Xsk5aGh1Ja_Y,586784
-nexaai/binds/nexa_llama_cpp/libnexa_plugin.dylib,sha256=sIseaikcmDXra4Zed7nt0KNN2u2bDTwRvPfUJx0B6dk,1750280
+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=iS6dkM6ZualLlzo67SNAqsrAS51WmuUWbylR-GH8s18,1806696
 nexaai/binds/nexa_mlx/libnexa_plugin.dylib,sha256=aw8if8RwXjb02CehbqGPHOeEKRUcTpzeJZLOkjTBm8A,596328
 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
@@ -510,11 +510,11 @@ nexaai/tts_impl/pybind_tts_impl.py,sha256=Be5QiXzDz6h1LTIQzUBd0ZyBs7rUpNA-pULCXF
 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=YO24mqqxiUseb1hd9ehLS2dD1t_4lYej_JWi76H7VnI,14374
+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=4lFZ8ZQnYZ-Uoh9j2Fh2UzpdpMAy_v1Jz-lrqX33XcI,8947
+nexaai/vlm_impl/mlx_vlm_impl.py,sha256=7gm_tFNox3LC78DQEtlMQ-eBK55zDY0xWlJghUAOP5Y,10402
 nexaai/vlm_impl/pybind_vlm_impl.py,sha256=C-3fa0AIypI33OAGuGfVxo1V7zN0wjQMgruKlDIlW4Q,8333
-nexaai-0.0.0.dist-info/METADATA,sha256=zfeUTCcL6g-tj984P5MIcA8SG4sfC2gp9BMGS9LLzB4,948
-nexaai-0.0.0.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
-nexaai-0.0.0.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-0.0.0.dist-info/RECORD,,
+nexaai-1.0.4.dist-info/METADATA,sha256=oFRx34_x3HEkQOp3cRCjUDgyZt13SWdvGiOVeGe6o7g,879
+nexaai-1.0.4.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
+nexaai-1.0.4.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.4.dist-info/RECORD,,