nexaai 1.0.20__cp310-cp310-win_amd64.whl → 1.0.21rc2__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__ = "1.0.20" 
+__version__ = "1.0.21-rc2" 

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=mbzzeXrEHHI_E3BQ0_OukD9wNajKJJVk0ykxT0rz8uM,2267
-nexaai/_stub.cp310-win_amd64.pyd,sha256=Efpmsr8o9_ipQ75vdg7mSVW9cDblaJR8WTv-Vbw0z5g,10752
-nexaai/_version.py,sha256=YQby4GMX_eKc_N1oaIy8c7_k8XB5jYqAA-LTQGOHYpk,143
+nexaai/__init__.py,sha256=Yt2YPVcRfwdA6DQoUR4ZI34CHBUYoY2cVsrG7Zp6IrI,2516
+nexaai/_stub.cp310-win_amd64.pyd,sha256=s9lKUQQYMxt6aJWxxMYb5LHq80Em0sCz5pQGrQjSIb8,10752
+nexaai/_version.py,sha256=RntcW2nBfo_nNGJb6146ZZwHwjDCavL1Tj1XtV9VmJI,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=oiNbKJtiPnP4aIOkKMxR4Vl3O-6gNWUOtiAqx3aG8AA,205824
-nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=TpTTLaLmFVFa8BiTE5zX-6FZ9IZKoq9MUBqyimJpA7A,182784
+nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=hI07QkGT4tWMoWA0zHZNEZh0ToslONCJxj5HMPqqbnc,205824
+nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=dM2dndm4Ch5F391i90d2ZF8U9tADmLCChzwpaj18oqk,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=Kh4F7bqWvVQA9eZkEwTSTpQRuznOhVYeELhOsGMQyzo,162816
-nexaai/binds/nexa_bridge.dll,sha256=F71KwL9PeUz9e1eDt5rSLP1nee5hc8_ekfpMYh2sJ0k,187904
-nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=FyyI4GrYnVCMUIXfzdkxNERlgMMGjA4qSSlmFhJYDBA,168960
-nexaai/binds/cpu_gpu/ggml-base.dll,sha256=Ws7NWfpdEn_QTL8gnh_sUz_J0K179Yl241uM4P0JGVw,532480
-nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=RanyT0m4REZJYz04ixpbhd87lemnUrzhCNXtdsDqbLY,672768
-nexaai/binds/cpu_gpu/ggml-cuda.dll,sha256=ex7fxyZe2KPcxkgfxSBrjf2oIgf7jbE1Rt74ywKyAls,313528832
-nexaai/binds/cpu_gpu/ggml-vulkan.dll,sha256=wGkKnEdCLKw0vWLH51GRuuzR8Lum9ta6Xh1Lge1zeFg,36627968
-nexaai/binds/cpu_gpu/ggml.dll,sha256=XBQajP9P05YqqzJ-0sADsb0luzbwWomxdI5nwpsnzqw,66560
-nexaai/binds/cpu_gpu/mtmd.dll,sha256=Z5r0VXuuFYZqwQBAutN2P378mF9zFFayxYRTa8BjeQk,561152
-nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=36jvlF5lKsOiQZuAY-QdXRNm1vQwZY5xFjX5MHUt7eI,1611776
-nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=Bv65TVCcZZ_CVjjMbbnwV4Jb5wZ0ILONDn0w95YABwQ,1411072
-nexaai/binds/nexaml/ggml-base.dll,sha256=Ws7NWfpdEn_QTL8gnh_sUz_J0K179Yl241uM4P0JGVw,532480
-nexaai/binds/nexaml/ggml-cpu.dll,sha256=RanyT0m4REZJYz04ixpbhd87lemnUrzhCNXtdsDqbLY,672768
-nexaai/binds/nexaml/ggml-cuda.dll,sha256=ex7fxyZe2KPcxkgfxSBrjf2oIgf7jbE1Rt74ywKyAls,313528832
-nexaai/binds/nexaml/ggml-vulkan.dll,sha256=wGkKnEdCLKw0vWLH51GRuuzR8Lum9ta6Xh1Lge1zeFg,36627968
-nexaai/binds/nexaml/ggml.dll,sha256=XBQajP9P05YqqzJ-0sADsb0luzbwWomxdI5nwpsnzqw,66560
+nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=DJvEg2tgTcpOvS1NJE0kSkd8hIvaQpSFlmUoTpS2xsU,162816
+nexaai/binds/nexa_bridge.dll,sha256=JMlu4matSkWSfqgnLaaJ4dB4uu0THI8aAdTATlD5VhI,187904
+nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=qBOnkTT7F2FJKOaD8utj2HmCbtDWoI3HFvTeQlAg2aw,170496
+nexaai/binds/cpu_gpu/ggml-base.dll,sha256=-0MMrq_hqJXMk3FSymyq4AoiMk_1r97HT-c_bR55A5w,532480
+nexaai/binds/cpu_gpu/ggml-cpu.dll,sha256=B81v_5sWbOumvKzHRBnK1UWvDsdxmCmOJSn5r-7qy84,672768
+nexaai/binds/cpu_gpu/ggml-cuda.dll,sha256=NH-HpsN_7UYfUy4Q9X2JYH7d-TNt7iRYxfAGpX4WpZY,313528832
+nexaai/binds/cpu_gpu/ggml-vulkan.dll,sha256=iJPrRl--pkNcDc0ZtdW2pw9piHDPd0D252VqgIOhMXo,36627968
+nexaai/binds/cpu_gpu/ggml.dll,sha256=gwZrKugz4jQrltFUeBjRJzvnOEOd6JgviggGEdCYBL8,66560
+nexaai/binds/cpu_gpu/mtmd.dll,sha256=v8UkuqbMFoU3Mot8C844ne8s8eagvtgh0idxk0flMXs,561152
+nexaai/binds/cpu_gpu/nexa_cpu_gpu.dll,sha256=uoKWytNu82wk2_Tjm3xZI3lAKJhn8uTS7Kklsc4Mx30,1611776
+nexaai/binds/cpu_gpu/nexa_plugin.dll,sha256=xCc9UoKj7kyxA2esE8teCo5yARoI5769cME9f_O4CeY,1411072
+nexaai/binds/nexaml/ggml-base.dll,sha256=-0MMrq_hqJXMk3FSymyq4AoiMk_1r97HT-c_bR55A5w,532480
+nexaai/binds/nexaml/ggml-cpu.dll,sha256=B81v_5sWbOumvKzHRBnK1UWvDsdxmCmOJSn5r-7qy84,672768
+nexaai/binds/nexaml/ggml-cuda.dll,sha256=NH-HpsN_7UYfUy4Q9X2JYH7d-TNt7iRYxfAGpX4WpZY,313528832
+nexaai/binds/nexaml/ggml-vulkan.dll,sha256=iJPrRl--pkNcDc0ZtdW2pw9piHDPd0D252VqgIOhMXo,36627968
+nexaai/binds/nexaml/ggml.dll,sha256=gwZrKugz4jQrltFUeBjRJzvnOEOd6JgviggGEdCYBL8,66560
 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=6nvov9eZ7Y5eJVFzCzrmaGesweplmSOv_2L6s6wsXEU,601600
-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=UMKMlN-KbhGPQAFPBs3aL9KqyjeBW-R31bQEOPDK93g,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
@@ -62,15 +63,15 @@ nexaai/tts_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/tts_impl/mlx_tts_impl.py,sha256=LcH9bVdIl3Q6lOzSUB_X2s-_nWFmlCl1yL7XSUK0fYI,3195
 nexaai/tts_impl/pybind_tts_impl.py,sha256=n3z4zmPQayQJgAwcvETw0IBUCp8IYROuYFSg0tAy_8Y,1487
 nexaai/utils/decode.py,sha256=0Z9jDH4ICzw4YXj8nD4L-sMouDaev-TISGRQ4KzidWE,421
-nexaai/utils/manifest_utils.py,sha256=E07t_bNq-EDrqVYZW7uyX8zsKsFklvqqwSvic0in0tQ,21800
-nexaai/utils/model_manager.py,sha256=qzlGP4sZLYsO72vg-EUCVA0RUknVKtLYMmDDfPmG45g,60999
+nexaai/utils/manifest_utils.py,sha256=BoKqnI6N5CkEPTDZCRWSx7wga6aH8h6cr4nbff7gVwg,22020
+nexaai/utils/model_manager.py,sha256=yxLaxOOhfMZ9i_2UnQ0qQSFLE-Rx9x9f_bGG5ut-nrk,63508
 nexaai/utils/model_types.py,sha256=q2m7diYLOpLvRl1ixL2eMq5_kdTj8KqPBGWX4p6Ob08,1532
 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-1.0.20.dist-info/METADATA,sha256=K2V5GVx_4VMBYf-ki7CpTOKuWrpxhc4lSGYpY5zl_rs,1238
-nexaai-1.0.20.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
-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=gI09QJcok68QG-36csXr3c2yPCZXTE0dlq1MUBEo_Ms,11091
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=4qPt5C5m4y4kWD9EFfVbcAkMq2Dgd5EP9tYo6yZIxgo,10162
+nexaai-1.0.21rc2.dist-info/METADATA,sha256=MzoQDZSvw3pwbFRRx2-3F1i9GcqVgYFZ-QR_PX8Tw-Y,1241
+nexaai-1.0.21rc2.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
+nexaai-1.0.21rc2.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.21rc2.dist-info/RECORD,,