nexaai 1.0.4rc14__cp310-cp310-win_amd64.whl → 1.0.4rc15__cp310-cp310-win_amd64.whl

nexaai/_version.py

@@ -1,4 +1,4 @@
 # This file is generated by CMake from _version.py.in
 # Do not modify this file manually - it will be overwritten
 
-__version__ = "1.0.4-rc14" 
+__version__ = "1.0.4-rc15" 

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/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-1.0.4rc14.dist-info → nexaai-1.0.4rc15.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.4rc14
+Version: 1.0.4rc15
 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-1.0.4rc14.dist-info → nexaai-1.0.4rc15.dist-info}/RECORD

@@ -1,10 +1,10 @@
 nexaai/__init__.py,sha256=d1bC_PUNduXYYPLrzKLyS0RapvcrKzLQGJREsoBZvXM,1977
-nexaai/_stub.cp310-win_amd64.pyd,sha256=AqYjxocL8y9D6PVMpNfrK8bjWlkh5QKtCE2nIG1ALJw,10752
-nexaai/_version.py,sha256=uRRQiVCCiR6jdFwtfQT2hqN_HwM1guhUpuxR4tuqrL8,147
+nexaai/_stub.cp310-win_amd64.pyd,sha256=C8_5OyVPadKYiSnRiPWYNr6Y7TJCu76TAZV3lBhcqjI,10752
+nexaai/_version.py,sha256=65huyjVsErx2svghMx8d4OaeLV72OhoKTE2Wt_RlEgk,147
 nexaai/asr.py,sha256=1XnwbrSoweBfIVAH6LbILv0DMStTQe_Uq5U_f-EyArY,1873
 nexaai/base.py,sha256=qQBCiQVNzgpkQjZX9aiFDEdbAAe56TROKC3WnWra2Zg,1021
 nexaai/common.py,sha256=00cP8uT9NdldBI3dRNHrQFx-uhdgtOGGxRAx4p96nw4,1586
-nexaai/cv.py,sha256=qGDVK0pKAEx0DrSi2rpSVLV4Kf53UAVIst9GdCMAGN0,3021
+nexaai/cv.py,sha256=90lrW6o77E6uNMk5MYWsLp2f-fhLacjWTT1ENDhVYEg,3120
 nexaai/embedder.py,sha256=FtJtMKrniejTCi8_-ePLOymfkH8j1VzUqteOqGy5cO4,2279
 nexaai/image_gen.py,sha256=oliLxFN7Bd_3wzP4F6frMJ7GPvRn-1kn_8kAtdcy_pY,4258
 nexaai/llm.py,sha256=7V60E1cI1tt6CZ1ti2-tPqkYS56TcJE_kIhvyRyIBeQ,3194
@@ -16,18 +16,18 @@ nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/asr_impl/mlx_asr_impl.py,sha256=UpGOtl4JZmcmDRa56z5OyIozFVjfsYpKw_vYU_7HoWk,3282
 nexaai/asr_impl/pybind_asr_impl.py,sha256=C0Qnx-WDNmyC72dxuZVfUwuAoUSMvpo8IfOOkEbqsFA,1493
 nexaai/binds/__init__.py,sha256=tYvy0pFhoY29GstDT5r-oRiPRarPLECvJAkcamJItOg,83
-nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=dkXWLzyqnwbjD8X-R2EgXVisNQ0m8dp4vPDUk9lTFtk,201216
-nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=rIpVBezf-CnRLOXp-UUm1QzMyo6jfVv45gmYcjazGbw,182784
-nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=ZYJ3LPPoCJ71I_-fzWBIP1CAH2074Gp6nOofnAcOYdc,160256
-nexaai/binds/nexa_bridge.dll,sha256=5s3i1ipX6gy0PsTN6NwpdI3sHb1GWThWuDR6V_IpHZY,176640
-nexaai/binds/nexa_llama_cpp/ggml-base.dll,sha256=gbrUPK8zgkj9bBR5xUDGD0D7nvNGUq92AN3WQn8MQko,514560
-nexaai/binds/nexa_llama_cpp/ggml-cpu.dll,sha256=35M4StfLC6dabERz-lsIMN-0cqZ-5TuzBiMxNJfavDk,663552
-nexaai/binds/nexa_llama_cpp/ggml-cuda.dll,sha256=ArpiDq4u5H69gMvTNLI4n6KVvEAW9kX0W47bkoGte18,315100160
-nexaai/binds/nexa_llama_cpp/ggml-vulkan.dll,sha256=SKtwhTRMZ5NjAd-tfOwRSh5YClFPOs4rW8RVuMTHaPs,26204160
-nexaai/binds/nexa_llama_cpp/ggml.dll,sha256=pvoyxBM3WRq_RkivDQlYi5V5zLKAtZ_onLoZBf68OmA,66560
-nexaai/binds/nexa_llama_cpp/llama.dll,sha256=zOiAWXJivXJCPRMPLEXC9Ct1jWXNl8NyknEEy8whfDA,1587712
-nexaai/binds/nexa_llama_cpp/mtmd.dll,sha256=fiSHNh1iDND5aqBwhE0Bki-up2olfoLDIKTsWmjQip0,560128
-nexaai/binds/nexa_llama_cpp/nexa_plugin.dll,sha256=G6LAMb8NNRUTiVIyTBF-6CZ3OzdeHJ5ZliHJDDPkUWk,1086464
+nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=tsCwzARi5fu-66q24KKLnWnBhNOt4UEbXbssgv4MPTw,201216
+nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=-JroOPxqlTEIU1BirfvjE0vqF__WCvsJaiuE3toWoyI,182784
+nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=e7BWEt6TzJocCe_Qb3dB1c-73n6ETL9uPOMuTEsT0Ew,160256
+nexaai/binds/nexa_bridge.dll,sha256=HPGaSjp3uuNGWXKhxvNbyiwxuyLFWr6TzIBxTlaweqQ,176640
+nexaai/binds/nexa_llama_cpp/ggml-base.dll,sha256=WiLohLZ6F5t9Vc0o65CjOMijPjuSH6-G586FOwOJLFE,514560
+nexaai/binds/nexa_llama_cpp/ggml-cpu.dll,sha256=vUHtE3B6dAuJu-WCmXCD7GKayM5HBcoYHw7DVXMvG-4,663552
+nexaai/binds/nexa_llama_cpp/ggml-cuda.dll,sha256=2O7SF6N6VUYPq2yf3l-KFvEpee2X6Yh_7vSbfgWVi2o,315100160
+nexaai/binds/nexa_llama_cpp/ggml-vulkan.dll,sha256=vhNQXlF9_0f5WQkzckQ4lGzmI0ccTPWDODhWlej1ebs,26204160
+nexaai/binds/nexa_llama_cpp/ggml.dll,sha256=VYw6ZxxvlZf4Gi2jkTaOeWSB_pZCIuex4hBd6cWwSWI,66560
+nexaai/binds/nexa_llama_cpp/llama.dll,sha256=Zp6IS6RSJ72vDulABng_9Q59tKvi3znbm13pJwazzNY,1587712
+nexaai/binds/nexa_llama_cpp/mtmd.dll,sha256=9350YqtbNICY1eV1gLB7flBu6NpRUoMf3ESIyH3uoiI,560128
+nexaai/binds/nexa_llama_cpp/nexa_plugin.dll,sha256=DeKJPc4fgbgC14iBPghfjGID9vuT3FEkvx9FT6BDk1s,1086464
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/cv_impl/mlx_cv_impl.py,sha256=vYN8ASbDr-VlQcia0ydpE3tUfnRcRIoRVQDAOhlZB_4,3250
 nexaai/cv_impl/pybind_cv_impl.py,sha256=oXT7Hcurg2YH_qgvwpGtgeQcIFxt6uzT9xN-cLvRHcU,1029
@@ -51,9 +51,9 @@ nexaai/utils/decode.py,sha256=0Z9jDH4ICzw4YXj8nD4L-sMouDaev-TISGRQ4KzidWE,421
 nexaai/utils/model_manager.py,sha256=Ksl-tKq-a3miTUxEn6-SSOC_KVdn6RPjcUdkWmDDwCk,49767
 nexaai/utils/progress_tracker.py,sha256=FmJBoOlzfQdc-TmccEav0cBR_iSNrrcskG3Fm1OrEJA,15482
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/vlm_impl/mlx_vlm_impl.py,sha256=DobOLnbf1v2uQrcB0TByTzJDp2Vhrjyjk3DvW4feahQ,9152
+nexaai/vlm_impl/mlx_vlm_impl.py,sha256=O0NA0zkJ4mhv5lT4EBSVineGe_F0Rsz4ITQvaX_VgIk,10651
 nexaai/vlm_impl/pybind_vlm_impl.py,sha256=WGC21LaQnr4LEiytXX8wVBMLJDG75GIweW3kEMxJGXE,8561
-nexaai-1.0.4rc14.dist-info/METADATA,sha256=AHPVkJq7Csbteh1-yBdVtZFIp0UKq6QWyVPTMJmQbCc,978
-nexaai-1.0.4rc14.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
-nexaai-1.0.4rc14.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.4rc14.dist-info/RECORD,,
+nexaai-1.0.4rc15.dist-info/METADATA,sha256=eHwHTGzl4PN_GfTApdOqE5dQTesECMWRo6r6CYZe7rY,910
+nexaai-1.0.4rc15.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
+nexaai-1.0.4rc15.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.4rc15.dist-info/RECORD,,