nexaai 1.0.19rc19__cp310-cp310-macosx_14_0_universal2.whl → 1.0.21rc1__cp310-cp310-macosx_14_0_universal2.whl

nexaai/__init__.py

@@ -24,6 +24,13 @@ from .common import ModelConfig, GenerationConfig, ChatMessage, SamplerConfig, P
 # Import logging functionality
 from .log import set_logger, get_error_message
 
+# Import runtime errors
+from .runtime_error import (
+    NexaRuntimeError,
+    ContextLengthExceededError,
+    GenerationError
+)
+
 # Create alias for PluginID to be accessible as plugin_id
 plugin_id = PluginID
 
@@ -52,6 +59,11 @@ __all__ = [
     # Logging functionality
     "set_logger",
     "get_error_message",
+    
+    # Runtime errors
+    "NexaRuntimeError",
+    "ContextLengthExceededError",
+    "GenerationError",
 
     "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__ = "1.0.19-rc19" 
+__version__ = "1.0.21-rc1" 

nexaai/mlx_backend/vlm/interface.py

@@ -482,8 +482,12 @@ class VLM(ProfilingMixin):
 
     def apply_chat_template(self, messages: Sequence[ChatMessage], tools: Optional[str] = None, enable_thinking: bool = True) -> str:
         """Apply chat template to messages with optional tools support."""
+        if self.model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking", "qwen3vl-8b", "qwen3vl-8b-thinking"]:
+            return apply_chat_template_qwen3_vl(messages, num_images=0, num_audios=0, tools=tools, enable_thinking=enable_thinking)
+        if self.model_name == "qwen3vl-moe":
+            return apply_chat_template_qwen3_vl_moe(messages, num_images=0, num_audios=0, tools=tools, enable_thinking=enable_thinking)
+        
         if hasattr(self.processor, "apply_chat_template"):
-            # Convert ChatMessage objects to dictionaries for the processor
             messages_dict = [{"role": msg.role, "content": msg.content} for msg in messages]
             
             parsed_tools = None
@@ -492,7 +496,6 @@ class VLM(ProfilingMixin):
             
             result = apply_chat_template(self.processor, self.model.config, messages_dict, add_generation_prompt=True, enable_thinking=enable_thinking, tools=parsed_tools)
             return result
-        # Fallback: join messages
         return "\n".join([f"{m.role}: {m.content}" for m in messages])
 
     def apply_chat_template_with_media(self, messages: Sequence[ChatMessage], num_images: int = 0, num_audios: int = 0, tools: Optional[str] = None, enable_thinking: bool = True) -> str:

nexaai/runtime_error.py

@@ -0,0 +1,24 @@
+"""Runtime errors for Nexa SDK operations."""
+
+
+class NexaRuntimeError(Exception):
+    """Base class for Nexa runtime errors."""
+    
+    def __init__(self, message: str, error_code: int = None):
+        self.error_code = error_code
+        super().__init__(message)
+
+
+class ContextLengthExceededError(NexaRuntimeError):
+    """Raised when the input context length exceeds the model's maximum."""
+    
+    def __init__(self, message: str = "Input context length exceeded model's maximum", error_code: int = None):
+        super().__init__(message, error_code)
+
+
+class GenerationError(NexaRuntimeError):
+    """Raised when generation fails."""
+    
+    def __init__(self, message: str = "Generation failed", error_code: int = None):
+        super().__init__(message, error_code)
+

nexaai/vlm.py

@@ -99,7 +99,8 @@ class VLM(BaseModel):
     def apply_chat_template(
         self,
         messages: List[MultiModalMessage],
-        tools: Optional[List[Dict[str, Any]]] = None
+        tools: Optional[List[Dict[str, Any]]] = None,
+        enable_thinking: bool = True
     ) -> str:
         """Apply the chat template to multimodal messages."""
         pass

nexaai/vlm_impl/mlx_vlm_impl.py

@@ -72,7 +72,8 @@ class MlxVlmImpl(VLM):
     def apply_chat_template(
         self,
         messages: List[MultiModalMessage],
-        tools: Optional[List[Dict[str, Any]]] = None
+        tools: Optional[List[Dict[str, Any]]] = None,
+        enable_thinking: bool = True
     ) -> str:
         """Apply the chat template to multimodal messages."""
         if not self._mlx_vlm:
@@ -116,7 +117,7 @@ class MlxVlmImpl(VLM):
                     num_images=total_images,
                     num_audios=total_audios,
                     tools=tools,
-                    enable_thinking=False  # Default to False, could be made configurable
+                    enable_thinking=enable_thinking
                 )
             else:
                 # Use regular apply_chat_template for text-only messages

nexaai/vlm_impl/pybind_vlm_impl.py

@@ -8,6 +8,11 @@ from nexaai.binds import vlm_bind, common_bind
 from nexaai.runtime import _ensure_runtime
 from nexaai.vlm import VLM
 from nexaai.base import ProfilingData
+from nexaai.runtime_error import ContextLengthExceededError, GenerationError
+
+# Error codes from ml.h
+ML_SUCCESS = 0
+ML_ERROR_LLM_TOKENIZATION_CONTEXT_LENGTH = -200004
 
 
 class PyBindVLMImpl(VLM):
@@ -91,7 +96,8 @@ class PyBindVLMImpl(VLM):
     def apply_chat_template(
         self,
         messages: List[MultiModalMessage],
-        tools: Optional[List[Dict[str, Any]]] = None
+        tools: Optional[List[Dict[str, Any]]] = None,
+        enable_thinking: bool = True
     ) -> str:
         """Apply the chat template to multimodal messages."""
         payload = []
@@ -111,7 +117,7 @@ class PyBindVLMImpl(VLM):
 
             payload.append({"role": role, "content": blocks})
 
-        result = vlm_bind.ml_vlm_apply_chat_template(self._handle, payload, tools)
+        result = vlm_bind.ml_vlm_apply_chat_template(self._handle, payload, tools, enable_thinking)
         return result
 
     def generate_stream(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> Generator[str, None, None]:
@@ -143,6 +149,18 @@ class PyBindVLMImpl(VLM):
                     on_token=on_token,
                     user_data=None
                 )
+                
+                # Check for errors in result
+                error_code = result.get("error_code", ML_SUCCESS)
+                if error_code != ML_SUCCESS:
+                    error_message = result.get("error_message", "Unknown error")
+                    if error_code == ML_ERROR_LLM_TOKENIZATION_CONTEXT_LENGTH:
+                        exception_container[0] = ContextLengthExceededError(error_message, error_code)
+                    else:
+                        exception_container[0] = GenerationError(error_message, error_code)
+                    token_queue.put(('end', None))
+                    return
+                
                 self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
             except Exception as e:
                 exception_container[0] = e
@@ -186,6 +204,15 @@ class PyBindVLMImpl(VLM):
             user_data=None
         )
 
+        # Check for errors in result
+        error_code = result.get("error_code", ML_SUCCESS)
+        if error_code != ML_SUCCESS:
+            error_message = result.get("error_message", "Unknown error")
+            if error_code == ML_ERROR_LLM_TOKENIZATION_CONTEXT_LENGTH:
+                raise ContextLengthExceededError(error_message, error_code)
+            else:
+                raise GenerationError(error_message, error_code)
+
         self._profiling_data = ProfilingData.from_dict(result.get("profile_data", {}))
         return result.get("text", "")
 

{nexaai-1.0.19rc19.dist-info → nexaai-1.0.21rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.19rc19
+Version: 1.0.21rc1
 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
@@ -19,7 +19,7 @@ Requires-Dist: numpy
 Requires-Dist: httpx
 Provides-Extra: mlx
 Requires-Dist: mlx; extra == "mlx"
-Requires-Dist: mlx-lm; extra == "mlx"
+Requires-Dist: mlx-lm==0.27.0; extra == "mlx"
 Requires-Dist: mlx-vlm; extra == "mlx"
 Requires-Dist: mlx-embeddings; extra == "mlx"
 Requires-Dist: tokenizers; extra == "mlx"

{nexaai-1.0.19rc19.dist-info → nexaai-1.0.21rc1.dist-info}/RECORD

@@ -1,6 +1,6 @@
-nexaai/__init__.py,sha256=L8oB7GFZZMGnUpCg0PecDbI_ycKuQak-ZEJ4Y12_QIw,2184
-nexaai/_stub.cpython-310-darwin.so,sha256=fhGvLrWdd7Lb4h9i_s3x3cud95dAqxuHrSteZIR-UVc,66768
-nexaai/_version.py,sha256=HOqTOpd7Lb8gdF93CQYr-x-Z_1O0X0vCrAAHaWQlI4I,144
+nexaai/__init__.py,sha256=gOd7sNsqEESopw_24xgnOSkIRENrk4Fa-RMtmVv62eA,2421
+nexaai/_stub.cpython-310-darwin.so,sha256=DMnGNcPFvhY3a3UlfwjsnBARJK6Ms-g6tgwmrdUEpx4,66768
+nexaai/_version.py,sha256=sVSvlzCHN_LFLOFxO9UEdPfKi2BBoPi4MtOW-fdz1i8,143
 nexaai/asr.py,sha256=NljMXDErwPNMOPaRkJZMEDka9Nk8xyur7L8i924TStY,2054
 nexaai/base.py,sha256=N8PRgDFA-XPku2vWnQIofQ7ipz3pPlO6f8YZGnuhquE,982
 nexaai/common.py,sha256=Y0NJNLTi4Nq4x1WL6PQsSvGUto0eGmWhjpsC6jcekfA,3444
@@ -11,17 +11,18 @@ nexaai/llm.py,sha256=S1o_k2VQoF5w2wO25f142OO1R75TP89Ii69VZv8pIGo,3567
 nexaai/log.py,sha256=Kwo2CIfWN6iP4M4F5EUIV8KIO5hAsvz6HZAaOwJ27Og,2628
 nexaai/rerank.py,sha256=vWaBucoQ1wz-2iYnZqyFIcEjm-4Xcs1KDbFN5X8zzDQ,1872
 nexaai/runtime.py,sha256=JvllhlNPgYGLbgGyX2yNvmGzT0lZ5XbvTvEo8sZG_Ho,2067
+nexaai/runtime_error.py,sha256=sO87LyCA0qzm0hVqBrmG2FDzGQH865EMbTMop2OfZto,779
 nexaai/tts.py,sha256=ZnBpWUxIfHhh7KfEjddtH7hHOTa91zg7ogGLakMIALo,2167
-nexaai/vlm.py,sha256=OCxwML-Z5uVGp3fjzJVtbCxfTLpgxkhQ8Wo6MVysoiw,4733
+nexaai/vlm.py,sha256=MreJ_S5-C0KH3haFuJwHqVtL099MrrmBQ23vK1PINCc,4771
 nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/asr_impl/mlx_asr_impl.py,sha256=eosd8-TIWAOwV0HltmoFrLwzXHcU4jyxtncvuZE9pgA,3257
 nexaai/asr_impl/pybind_asr_impl.py,sha256=pE9Hb_hMi5yAc4MF83bLVOb8zDtreCkB3_u7XED9YpA,1516
 nexaai/binds/__init__.py,sha256=eYuay_8DDXeOUWz2_R9HFSabohxs6hvZn391t2L0Po0,104
 nexaai/binds/common_bind.cpython-310-darwin.so,sha256=zxJuD0nSV--VZKxBfWZUavU7_bHj_JTi0FhkjvG4VJw,235264
 nexaai/binds/embedder_bind.cpython-310-darwin.so,sha256=tPa0c0Dv_GiW66fgmAGWGCHXRGNApznqoQS0eQx9GFM,202064
-nexaai/binds/libnexa_bridge.dylib,sha256=8oYPIjtIf6VBG30Id9x6K9D9irfYZBFHxvW-8iZ5rbQ,272648
+nexaai/binds/libnexa_bridge.dylib,sha256=l4Psl0JU6ki2K4oEA4EPmi01Fah55D8fc3eiEtfhIk4,272648
 nexaai/binds/llm_bind.cpython-310-darwin.so,sha256=TAWfa1Hzq00TjtC1xVsiAeLp6hv2LrL5afDz4omUghc,182784
-nexaai/binds/vlm_bind.cpython-310-darwin.so,sha256=nd6eG_m2EiPthzkSZ97hlXWUOZQir4cQfFJZ4p6eR2U,182704
+nexaai/binds/vlm_bind.cpython-310-darwin.so,sha256=VpTl0U8YmiYTyp5vNlmLRxiwGD8-4am8nBvdWYY1d-I,182704
 nexaai/binds/cpu_gpu/libggml-base.dylib,sha256=9urbatWTB4QJ6uariOYlB_ZN1FFci5GaD8OfGZ9oPSQ,650736
 nexaai/binds/cpu_gpu/libggml-cpu.so,sha256=qiYxbTe4Nt7n36zJVvq3zovgSZEmrN2is6gzTern7UI,677728
 nexaai/binds/cpu_gpu/libggml-metal.so,sha256=zfaX7rIBYQazH2lf-vza007BMhPTK1ASd2T0HLLIA4E,673104
@@ -190,12 +191,12 @@ nexaai/binds/nexaml/libggml-metal.so,sha256=zfaX7rIBYQazH2lf-vza007BMhPTK1ASd2T0
 nexaai/binds/nexaml/libggml.dylib,sha256=aOTj_6RrAMkfDO0ZI28_3nfcC-l4Y3dRCiS3C0d0_eI,58592
 nexaai/binds/nexaml/libmp3lame.0.dylib,sha256=G-21u3MAZ2hiv1fbPEpOUGdToZfLtz2XI6BwW9xIqIA,305136
 nexaai/binds/nexaml/libmpg123.0.dylib,sha256=L4AgLcdcjGOQkEovUGDCNlH959500i0GkKBhqiHxBY4,306464
-nexaai/binds/nexaml/libnexa-mm-process.dylib,sha256=-tNsX7gmHbgGDVOo9WEMyppWb1u4ldRCyTl3VIRkeHc,10222440
-nexaai/binds/nexaml/libnexa-sampling.dylib,sha256=cnOFmvb5urwkB1S1aruZA7s6jaa6e_q-sv-2NFW785U,7957848
-nexaai/binds/nexaml/libnexa_plugin.dylib,sha256=7aQX5UQWVipWmq0KcyPtbqmpZMsAtyHMFJ4zusFGGDw,249512
-nexaai/binds/nexaml/libnexaproc.dylib,sha256=aWhBGz9c8Dd0JIigOeYey0CIraTZ_PYOyRjaS9lho0M,968320
+nexaai/binds/nexaml/libnexa-mm-process.dylib,sha256=WCEgYK13YBfg-DiaGXXJ00-QYo0ucEP-QUS1jVEUR_k,10222392
+nexaai/binds/nexaml/libnexa-sampling.dylib,sha256=mWJ3jsF4bp3RwWebgjd4PalZqtLiXGK4Y5J9IkYqgvk,7957800
+nexaai/binds/nexaml/libnexa_plugin.dylib,sha256=uJKzGOiWjrJyZqwpUrhum5tRTYKpcWWpdiF9UK4i2mI,249512
+nexaai/binds/nexaml/libnexaproc.dylib,sha256=KNAyTwv2Tk94VNbUNFldDbnhQmJwv2ykysfvW_267DY,897248
 nexaai/binds/nexaml/libomp.dylib,sha256=RcvfaCz0XuqTYtVPFYyUm2V6S9YgUQ0x5pHATr9FnDk,754912
-nexaai/binds/nexaml/libqwen3-vl.dylib,sha256=wRQUVHtJJZRi8QCoHHURmpzpd7LeirBbMvIsCmWNJsc,770352
+nexaai/binds/nexaml/libqwen3-vl.dylib,sha256=mFGkFKAqc3h_FOEiKg7A3ProF8qzefa6FAOwDSqk1HM,770592
 nexaai/binds/nexaml/libqwen3vl-vision.dylib,sha256=t0QfNZG5gpvHPBqX-Tdm1qjaXHCVROTtzL5bW78v4-I,490344
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/cv_impl/mlx_cv_impl.py,sha256=gKECQOv8iaWwG3bl7xeqVy2NN_9K7tYerIFzfn4eLo4,3228
@@ -416,7 +417,7 @@ nexaai/mlx_backend/vlm/__init__.py,sha256=_25kvMEviX16Hg3bro8Ws70V0eeIEqYKV8ZDXq
 nexaai/mlx_backend/vlm/generate.py,sha256=DqHFEAuqk-nko8ho6U9GAXTDAWz4d8GTe_hCt-XFyCw,19071
 nexaai/mlx_backend/vlm/generate_qwen3_vl.py,sha256=srN8-RFv8eOeH2rdyygCJ7Yt7kW7MQzS3i50UHBVfIM,13151
 nexaai/mlx_backend/vlm/generate_qwen3_vl_moe.py,sha256=ZSbM8JjTlkxUaVO9UNZM6YSbd60am3Z4ztJJEBsnJHg,9015
-nexaai/mlx_backend/vlm/interface.py,sha256=_rnqaIkvy3OUsH2b08l623oKjoe_la0G2W9iusD5qwI,22741
+nexaai/mlx_backend/vlm/interface.py,sha256=D6TCUWbiGLkgmAk_b9yMb36Y4TLGT9gFPxnTaDSaCSM,23070
 nexaai/mlx_backend/vlm/main.py,sha256=8bmSTtyebp8eyL2jL36DZbNHapOpFXNmjM2NyzCFqGs,12919
 nexaai/mlx_backend/vlm/modeling/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/mlx_backend/vlm/modeling/convert.py,sha256=ia5i9cgTufFGmKyhkYUaW0nfNqT_bMo8i-Hg_zy5JC4,1863
@@ -564,9 +565,9 @@ nexaai/utils/model_types.py,sha256=ONWjjo8CFPdhxki6qo7MXnSZaEzjBcxa_Kkf_y5NXus,1
 nexaai/utils/progress_tracker.py,sha256=jdUqtmPqyhwC9uSKvQcJEYETwSt-OhP4oitdJ94614o,15394
 nexaai/utils/quantization_utils.py,sha256=FYcNSAKGlBqFDUTx3jSKOr2lnq4nyiyC0ZG8oSxFwiU,7825
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/vlm_impl/mlx_vlm_impl.py,sha256=pLtWm_ckz8a0U-AtAOMVseFDO4OVPvHyYO2KlfBaGYk,10833
-nexaai/vlm_impl/pybind_vlm_impl.py,sha256=FAbhpRJzHgI78r0mUvKybO97R1szvNhH0aTn_I52oT4,8597
-nexaai-1.0.19rc19.dist-info/METADATA,sha256=V_rgXCiklp4A99jP4b_rAOsNBpscrNaLGd4Pp7aLGDo,1202
-nexaai-1.0.19rc19.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
-nexaai-1.0.19rc19.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.19rc19.dist-info/RECORD,,
+nexaai/vlm_impl/mlx_vlm_impl.py,sha256=sgHqnX5OCSGLccCnTuRiktIbqThNn3AAIvYE2_Dy4TI,10833
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=MDbreWSqugakXU_PqH6mPoCxjKEEbYfQIco_NDck8_s,9905
+nexaai-1.0.21rc1.dist-info/METADATA,sha256=I2YizqGmn9LBQh1tfAPNxPZYE_limQe4ELxBlzYqtKM,1209
+nexaai-1.0.21rc1.dist-info/WHEEL,sha256=T2p57lol9__xkoU6aJTyN1Pm43ZpRU3q6km7mIbrAMs,114
+nexaai-1.0.21rc1.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.21rc1.dist-info/RECORD,,