nexaai 1.0.20__cp310-cp310-macosx_13_0_x86_64.whl → 1.0.21rc2__cp310-cp310-macosx_13_0_x86_64.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__ = "1.0.20" 
+__version__ = "1.0.21-rc2" 

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/utils/manifest_utils.py

@@ -157,12 +157,16 @@ def create_gguf_manifest(repo_id: str, files: List[str], directory_path: str, ol
                 # Use the new enum-based quantization extraction
                 quantization_type = extract_quantization_from_filename(current_file_name)
                 quant_level = quantization_type.value if quantization_type else "UNKNOWN"
-                
-                model_files[quant_level] = {
-                    "Name": current_file_name,
-                    "Downloaded": True,
-                    "Size": file_size
-                }
+
+                # FIXME: hardcode to handle the multiple mmproj files problem
+                if quant_level == "UNKNOWN" and "mmproj" in current_file_name.lower():
+                    pass
+                else:
+                    model_files[quant_level] = {
+                        "Name": current_file_name,
+                        "Downloaded": True,
+                        "Size": file_size
+                    }
     
     # Determine PluginId with priority: kwargs > downloaded_manifest > model_file_type > default
     plugin_id = kwargs.get('plugin_id')

nexaai/utils/model_manager.py

@@ -410,6 +410,20 @@ def _remove_specific_file(target_model: DownloadedModel, file_name: str, local_d
     except OSError:
         file_size = 0
     
+    # Check if we should remove entire folder instead (for .gguf files)
+    # If removing a .gguf file and no other non-mmproj .gguf files remain, remove entire folder
+    if file_name.endswith('.gguf'):
+        updated_files = [f for f in target_model.files if f != file_name]
+        # Find remaining .gguf files that don't contain "mmproj" in filename
+        remaining_non_mmproj_gguf = [
+            f for f in updated_files 
+            if f.endswith('.gguf') and 'mmproj' not in f.lower()
+        ]
+        
+        # If no non-mmproj .gguf files remain, remove entire repository
+        if len(remaining_non_mmproj_gguf) == 0:
+            return _remove_entire_repository(target_model, local_dir)
+    
     # Remove the file
     try:
         os.remove(file_path)
@@ -846,6 +860,41 @@ class HuggingFaceDownloader:
                 pass
         return {}
     
+    def _download_manifest_if_needed(self, repo_id: str, local_dir: str) -> bool:
+        """
+        Download nexa.manifest from the repository if it doesn't exist locally.
+        
+        Args:
+            repo_id: Repository ID
+            local_dir: Local directory where the manifest should be saved
+            
+        Returns:
+            bool: True if manifest was downloaded or already exists, False if not found in repo
+        """
+        manifest_path = os.path.join(local_dir, 'nexa.manifest')
+        
+        # Check if manifest already exists locally
+        if os.path.exists(manifest_path):
+            return True
+        
+        # Try to download nexa.manifest from the repository
+        try:
+            print(f"[INFO] Attempting to download nexa.manifest from {repo_id}...")
+            self.api.hf_hub_download(
+                repo_id=repo_id,
+                filename='nexa.manifest',
+                local_dir=local_dir,
+                local_dir_use_symlinks=False,
+                token=self.token,
+                force_download=False
+            )
+            print(f"[OK] Successfully downloaded nexa.manifest from {repo_id}")
+            return True
+        except Exception as e:
+            # Manifest doesn't exist in repo or other error - this is fine, we'll create it
+            print(f"[INFO] nexa.manifest not found in {repo_id}, will create locally")
+            return False
+    
     def _fetch_and_save_metadata(self, repo_id: str, local_dir: str, is_mmproj: bool = False, file_name: Optional[Union[str, List[str]]] = None, **kwargs) -> None:
         """Fetch model info and save metadata after successful download."""
         # Initialize metadata with defaults to ensure manifest is always created
@@ -946,6 +995,9 @@ class HuggingFaceDownloader:
             if progress_tracker:
                 progress_tracker.stop_tracking()
             
+            # Download nexa.manifest from repo if it doesn't exist locally
+            self._download_manifest_if_needed(repo_id, file_local_dir)
+            
             # Save metadata after successful download
             self._fetch_and_save_metadata(repo_id, file_local_dir, self._current_is_mmproj, self._current_file_name, **kwargs)
             
@@ -1055,6 +1107,9 @@ class HuggingFaceDownloader:
             if progress_tracker:
                 progress_tracker.stop_tracking()
             
+            # Download nexa.manifest from repo if it doesn't exist locally
+            self._download_manifest_if_needed(repo_id, repo_local_dir)
+            
             # Save metadata after successful download
             self._fetch_and_save_metadata(repo_id, repo_local_dir, self._current_is_mmproj, self._current_file_name, **kwargs)
             

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-1.0.20.dist-info → nexaai-1.0.21rc2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.20
+Version: 1.0.21rc2
 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

{nexaai-1.0.20.dist-info → nexaai-1.0.21rc2.dist-info}/RECORD

@@ -1,6 +1,6 @@
-nexaai/__init__.py,sha256=L8oB7GFZZMGnUpCg0PecDbI_ycKuQak-ZEJ4Y12_QIw,2184
-nexaai/_stub.cpython-310-darwin.so,sha256=gdqah1VJ1qyRogQe5O82w0KyOozQj89SVlY2A9fDp8k,49832
-nexaai/_version.py,sha256=v412uZvaY1JQ8H4RwZLyc9qL0riThUDmOS3PorCTMUk,139
+nexaai/__init__.py,sha256=gOd7sNsqEESopw_24xgnOSkIRENrk4Fa-RMtmVv62eA,2421
+nexaai/_stub.cpython-310-darwin.so,sha256=2oHxHG6XxyeEiIYzsYcGtwqcujIDFwMkL7aekNiK8EA,49832
+nexaai/_version.py,sha256=rUlsUioGYFQcgW8x3C0Lhx4EJxLXKPkq0Tn9gN5475g,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=BoXByRlNGDaNS1YyZyCF-s7h0vXP9NLPlJMQQ5pqusU,235488
 nexaai/binds/embedder_bind.cpython-310-darwin.so,sha256=b2NoXFAJvPLi_P1X7lXLKmAUU0v2HJI3Zwa10gfqHdw,202032
-nexaai/binds/libnexa_bridge.dylib,sha256=AEVIbaSMqE70UymaH9QT1t8CvgEuZ6RCL-w9wPfrjd0,271952
+nexaai/binds/libnexa_bridge.dylib,sha256=NtyCUjyZi3veVq3G4nfeGexuoJ70yLXQTV5_EOg3SpY,271952
 nexaai/binds/llm_bind.cpython-310-darwin.so,sha256=p1ZTGMolEkWywkmwzOUjTr3RpSEH21BHZAggVzo89Ks,183088
-nexaai/binds/vlm_bind.cpython-310-darwin.so,sha256=LGd-tykePnQFfGca25HnPIBfXsfrMzbwyx6d5Ld3xps,183000
+nexaai/binds/vlm_bind.cpython-310-darwin.so,sha256=CQdy70sGqhM9SvoDN3xmsOj79IWEYeOVhwou_kgkjHI,199392
 nexaai/binds/cpu_gpu/libggml-base.dylib,sha256=YDclLDlP7XlDpXiKfTOTt6mW7jgXlmwSoT_VuRrGrmM,629528
 nexaai/binds/cpu_gpu/libggml-cpu.so,sha256=cnLUQ7WdX-5iiDaH8v45u1kX1NUmK8DanpzSMGCpXPE,1039800
 nexaai/binds/cpu_gpu/libggml-metal.so,sha256=Xhhl_tLg1xmCIQVrKjqPFaLHAlx_2wUFiwDyUk0wJ-E,713680
@@ -248,7 +249,7 @@ 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
@@ -390,15 +391,15 @@ nexaai/tts_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/tts_impl/mlx_tts_impl.py,sha256=i_uNPdvlXYtL3e01oKjDlP9jgkWCRt1bBHsExaaiJi8,3101
 nexaai/tts_impl/pybind_tts_impl.py,sha256=mpn44r6pfYLIl-NrEy2dXHjGtWtNCmM7HRyxiANxUI4,1444
 nexaai/utils/decode.py,sha256=61n4Zf6c5QLyqGoctEitlI9BX3tPlP2a5aaKNHbw3T4,404
-nexaai/utils/manifest_utils.py,sha256=SCcFN09xNI0DiTA1U7DZwWiQsRH0CInWSny_9q0BwNM,21273
-nexaai/utils/model_manager.py,sha256=NnbPv1iuwo6T523gLsWjnff-gGvPGUjez-rFg8-ffpE,59568
+nexaai/utils/manifest_utils.py,sha256=OOp_BmFWH1ZHMYkS2VGAby5Rpm4f4GLCRBJEBYm-kys,21489
+nexaai/utils/model_manager.py,sha256=CsRflxY329DgDeZxl_PGFKis7MQW7XROkrvZUCPEpvo,62022
 nexaai/utils/model_types.py,sha256=ONWjjo8CFPdhxki6qo7MXnSZaEzjBcxa_Kkf_y5NXus,1483
 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-1.0.20.dist-info/METADATA,sha256=nrAx2CD06ULRSTnz5K7wGVAv8CELbeE_aZHn3R9iLdo,1206
-nexaai-1.0.20.dist-info/WHEEL,sha256=0KYp5feZ1CMUhsfFXKpSQTbSmQbXy4mv6yPPVBXg2EM,110
-nexaai-1.0.20.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.20.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-1.0.21rc2.dist-info/METADATA,sha256=hkj3IE2vIhQaOsChkXdv2vG7c3kKvn6bg74LkPLq4yY,1209
+nexaai-1.0.21rc2.dist-info/WHEEL,sha256=0KYp5feZ1CMUhsfFXKpSQTbSmQbXy4mv6yPPVBXg2EM,110
+nexaai-1.0.21rc2.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.21rc2.dist-info/RECORD,,