nexaai 1.0.19rc19__cp310-cp310-win_amd64.whl → 1.0.21rc1__cp310-cp310-win_amd64.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/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=mbzzeXrEHHI_E3BQ0_OukD9wNajKJJVk0ykxT0rz8uM,2267
-nexaai/_stub.cp310-win_amd64.pyd,sha256=oBsyAojVBnmWvTxsQwqjvKwHuXXVwViAcaJ0UOU9gl4,10752
-nexaai/_version.py,sha256=ly1OAJ4Z8m8oUF6yqhQJ-skQN0goQcrzdf_W4s3nFMc,148
+nexaai/__init__.py,sha256=Yt2YPVcRfwdA6DQoUR4ZI34CHBUYoY2cVsrG7Zp6IrI,2516
+nexaai/_stub.cp310-win_amd64.pyd,sha256=wlEBCN5E8QuK2OoupinLzjTYqPdSdtTOrOmgkwrd9T0,10752
+nexaai/_version.py,sha256=RJD43EAWJ7pU1bQt-f7S--4-GPeOQxjr-rdudgziXbA,147
 nexaai/asr.py,sha256=_fsGaxpiU137bUtO5ujtFSYCI1RLsyeEm3Gf4GhHVRk,2118
 nexaai/base.py,sha256=qQBCiQVNzgpkQjZX9aiFDEdbAAe56TROKC3WnWra2Zg,1021
 nexaai/common.py,sha256=muQqFY-WllwL5IO83tImexbuUcoEQsKg73u4gWL2lNs,3548
@@ -11,37 +11,38 @@ nexaai/llm.py,sha256=Qwm1q_NStLfD-JYZQIvxniWnAmwNl1V6LUON3Me7w_I,3663
 nexaai/log.py,sha256=F_Qe169kLbnFV25WJmS_ZtmBfOdcGic8BYFIsYVoD_o,2720
 nexaai/rerank.py,sha256=_zGWmX6eDigY2kViMKCtNssp4JMEeVycZZfJH9eAZOY,1927
 nexaai/runtime.py,sha256=_BoAtTUv5ZR7wtOlJL5TldR3AZTP0OnMWjB9p71k-8E,2135
+nexaai/runtime_error.py,sha256=7TvjMJO7MpvVRQCYEDwcFPOgZsIILQtN-ZJ74JlARrE,803
 nexaai/tts.py,sha256=afs6sx0w0Tvs_aJlyZRPm62qQpTrs-fW_jDHrMkc4AA,2241
-nexaai/vlm.py,sha256=wUMbRURluxvULbS7vgyHsWdEPgLGrrC_S79bhdEN7Vg,4860
+nexaai/vlm.py,sha256=e-IMTiyCkP1vvrVHn98quyJoISGQzVf6UETSQ3lHudo,4899
 nexaai/asr_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/asr_impl/mlx_asr_impl.py,sha256=XwMX3LYMeulp8cDS0TCCYcjvttFHAyDWQ_oMvABwQmI,3349
 nexaai/asr_impl/pybind_asr_impl.py,sha256=20o5SOPzhF9x41ra8L_qIM7YxCkYeLb5csSrNde-dds,1560
 nexaai/binds/__init__.py,sha256=ENl-uoIF9-3XGIXitVgZ2QmJ6p7Yet4h1-X7nUDZ0Hk,108
-nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=wwXQwxZCPVidbGyC28hcwbp8XilXFjco4X3FkCSEKfg,205824
-nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=URVof-RRuK7OHISj2Qcrnr5Beo_J1CiMYqZNTY3ZT7I,182784
+nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=A7AuP8TpdE1O0jSRYSX0PmV0s8zXmrr_KcTXY93oa1I,205824
+nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=bpk2lKSKLZPCNn97GdS72UIR7x5V0aWS_Ft6jxU5cWY,182784
 nexaai/binds/libcrypto-3-x64.dll,sha256=X7hDuQPMn6eRe58dTGJxTmRPKFWgd5i6Gjoy9QPLmPg,7315968
 nexaai/binds/libssl-3-x64.dll,sha256=GRw0cOaGUaR2QHRvBFuU_9S9iNgkyVgm0NIpZMnReAw,1313792
-nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=IYdme70wPZFFHHIGUsbi6hiBHJZz0H478C1pBfkWXDg,162816
-nexaai/binds/nexa_bridge.dll,sha256=TdDzK71xKW_rE0Qvxnr1M05QsbXL7IuCCuu4WkS9WQU,187904
-nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=eemVZdUWNdjN42msYJDsPCq3IBvdmpg5vxiQUx854HA,168960
-nexaai/binds/cpu_gpu/ggml-base.dll,sha256=er-uYRvlG9B8b5ZZ3RBKf7mS97U5LSN51VZy-oNKS8o,532480
-nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=vzcRC3K4i4WnfStrM3Rt0-_DQ3nl3P3OvNyW7DTEz2U,672768
-nexaai/binds/cpu_gpu/ggml-cuda.dll,sha256=R4KEMvK_m47c5WO1SiYUO_FuRDZ5b5SAjxoiP_5n8jo,313528832
-nexaai/binds/cpu_gpu/ggml-vulkan.dll,sha256=3IP-XdGWuojTGCucHQ_erpWmFuJwadbU6DJsPPHZ_KY,36627968
-nexaai/binds/cpu_gpu/ggml.dll,sha256=3h0Vykj3K7IYLEAXE9cKzJShhmw4psKZmNTLNxXbItI,66560
-nexaai/binds/cpu_gpu/mtmd.dll,sha256=F5Y9ZdVaxKO7GnMMK5YrvhxVgVDNb85za14PxU3vvcE,561152
-nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=DOc_ROdv-CMly-Zw9DN-JltYBjYinkXF44kETdWzjp0,1611776
-nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=bcvp3FmhT8RXbbYuwDJ3UaNeg95oMF9KTFo1152sbIY,1411072
-nexaai/binds/nexaml/ggml-base.dll,sha256=er-uYRvlG9B8b5ZZ3RBKf7mS97U5LSN51VZy-oNKS8o,532480
-nexaai/binds/nexaml/ggml-cpu.dll,sha256=vzcRC3K4i4WnfStrM3Rt0-_DQ3nl3P3OvNyW7DTEz2U,672768
-nexaai/binds/nexaml/ggml-cuda.dll,sha256=R4KEMvK_m47c5WO1SiYUO_FuRDZ5b5SAjxoiP_5n8jo,313528832
-nexaai/binds/nexaml/ggml-vulkan.dll,sha256=3IP-XdGWuojTGCucHQ_erpWmFuJwadbU6DJsPPHZ_KY,36627968
-nexaai/binds/nexaml/ggml.dll,sha256=3h0Vykj3K7IYLEAXE9cKzJShhmw4psKZmNTLNxXbItI,66560
+nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=7isvwd_kZYtGAmc9O6u67KPr0RQbTpreASOcTjcVqo8,162816
+nexaai/binds/nexa_bridge.dll,sha256=YrnvRvUw4aDoPSEYrOEUMia-FKbGTz88C-CN_jQNwWg,187904
+nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=WcCrXej3pFt7JDr0VbeW_c8_Y5dPAt8n4THD2qf6JSo,170496
+nexaai/binds/cpu_gpu/ggml-base.dll,sha256=h8m2MYmzg2syDaGLGvfogWbi88ZlA7e_ZnL79hDFwgE,532480
+nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=fMHSkx2QYWUNC63l0tgO6F94Bb4kcQE3sH70ptnH9DQ,672768
+nexaai/binds/cpu_gpu/ggml-cuda.dll,sha256=q4-dtifsuifcuRZn4St-5gJpiEA1hKE2ld1VEqkkqUk,313528832
+nexaai/binds/cpu_gpu/ggml-vulkan.dll,sha256=_z41BQzz1Ud1w5UEoMiJmEGNE-ybxJKYKN1ssagdKGg,36627968
+nexaai/binds/cpu_gpu/ggml.dll,sha256=_3ejjzy1owBjJAS8dGMmQBWBLAkhs1BRyo0qfT-jGkg,66560
+nexaai/binds/cpu_gpu/mtmd.dll,sha256=XQKiNXM34-Cj27vrLTuH5I3KBdCN1gAgp6eTz7eYF-g,561152
+nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=Gf9C8LiOUuM9QYNKd1jgCcgsfyT0ymFJ4ebHPib8Q24,1611776
+nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=pBNvjvjomppglDJIQ3rnzzr2gJT3b-TqdEBe2sr3Az8,1411072
+nexaai/binds/nexaml/ggml-base.dll,sha256=h8m2MYmzg2syDaGLGvfogWbi88ZlA7e_ZnL79hDFwgE,532480
+nexaai/binds/nexaml/ggml-cpu.dll,sha256=fMHSkx2QYWUNC63l0tgO6F94Bb4kcQE3sH70ptnH9DQ,672768
+nexaai/binds/nexaml/ggml-cuda.dll,sha256=q4-dtifsuifcuRZn4St-5gJpiEA1hKE2ld1VEqkkqUk,313528832
+nexaai/binds/nexaml/ggml-vulkan.dll,sha256=_z41BQzz1Ud1w5UEoMiJmEGNE-ybxJKYKN1ssagdKGg,36627968
+nexaai/binds/nexaml/ggml.dll,sha256=_3ejjzy1owBjJAS8dGMmQBWBLAkhs1BRyo0qfT-jGkg,66560
 nexaai/binds/nexaml/nexa-mm-process.dll,sha256=DDD5qPbelZAIrRWGggOd3Pg9U_fh_ZEYIplwqiyD5LY,4643328
 nexaai/binds/nexaml/nexa-sampling.dll,sha256=Notkz287laSUG2_ED3oFXWVLJM3t7x_USkvX6wKVOLA,4265984
-nexaai/binds/nexaml/nexa_plugin.dll,sha256=6FUKlvQwcyEuFLIzL5KsVDvquDKjICFhIz4bAwOHLiU,600576
-nexaai/binds/nexaml/nexaproc.dll,sha256=cVU4ngSdsQ0MttqOv5IUZhmUE8lZLq-OjC8ueaAsx1s,2668544
-nexaai/binds/nexaml/qwen3-vl.dll,sha256=say_RLwLsIq_svPlgqPl12gfrOCQtcCiwa_q91Ql7YM,5873152
+nexaai/binds/nexaml/nexa_plugin.dll,sha256=8XyOTP11koqdTaKUEgxFm2Fwl9v_O4GDgg040yw5F5g,601600
+nexaai/binds/nexaml/nexaproc.dll,sha256=rRcdSfhkxnRwktXW4lp7Np8sNY_uefNXYwAmPp34AHg,2670080
+nexaai/binds/nexaml/qwen3-vl.dll,sha256=KxwQ0PHvzM1zwLpH7vlNOlfaVzgEUHTfTV76e61KeyA,5873664
 nexaai/binds/nexaml/qwen3vl-vision.dll,sha256=5AX4MtAOWTcXSAkS2qqvl2pFti4HR55dtkCaOO03Evc,1063424
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/cv_impl/mlx_cv_impl.py,sha256=QLd_8w90gtxH8kmssaDYatCTRvQNIJuUGKZNnYrmx6E,3317
@@ -68,9 +69,9 @@ nexaai/utils/model_types.py,sha256=q2m7diYLOpLvRl1ixL2eMq5_kdTj8KqPBGWX4p6Ob08,1
 nexaai/utils/progress_tracker.py,sha256=BztrFqtjwNUmeREwZ5m7H6ZcrVzQEbpZfsxndWh4z0A,15778
 nexaai/utils/quantization_utils.py,sha256=FxnZ6-uAE_bl_vQ5jsRXlpU0NBn-U4Y8iN9_O6aCdPA,8070
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-nexaai/vlm_impl/mlx_vlm_impl.py,sha256=MgqJO7OzuPd79gOZZKhSXXMNSP2eBuhhrdCX8XHn6aQ,11090
-nexaai/vlm_impl/pybind_vlm_impl.py,sha256=NuQ_Ep1TnjmGAkjJuUS0Lb6z7iPu3wroLVOx7kiAwlE,8827
-nexaai-1.0.19rc19.dist-info/METADATA,sha256=2-hG8alVSHrKOSLulm6sVIUyyegHYxvYcQ-6YkZhDO8,1234
-nexaai-1.0.19rc19.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
-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=gI09QJcok68QG-36csXr3c2yPCZXTE0dlq1MUBEo_Ms,11091
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=4qPt5C5m4y4kWD9EFfVbcAkMq2Dgd5EP9tYo6yZIxgo,10162
+nexaai-1.0.21rc1.dist-info/METADATA,sha256=ZWD3yi3ygRwzzLE5gkFx3V1hXnejEJH9Fm5zkX6Yo4I,1241
+nexaai-1.0.21rc1.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
+nexaai-1.0.21rc1.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.21rc1.dist-info/RECORD,,