nexaai 1.0.12rc1__cp310-cp310-win_amd64.whl → 1.0.13rc1__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__ = "1.0.12-rc1" 
+__version__ = "1.0.13-rc1" 

nexaai/binds/__init__.py

@@ -1,3 +1,4 @@
 from .common_bind import *
 from .llm_bind import *
 from .embedder_bind import *
+from .vlm_bind import *

nexaai/utils/model_manager.py

@@ -1124,15 +1124,64 @@ def download_from_huggingface(
 ##########################################################################
 
 
+def _download_model_if_needed(
+    model_path: str,
+    param_name: str,
+    progress_callback: Optional[Callable[[Dict[str, Any]], None]] = None,
+    token: Union[bool, str, None] = None
+) -> str:
+    """
+    Helper function to download a model from HuggingFace if it doesn't exist locally.
+    
+    Args:
+        model_path: The model path that may be local or remote
+        param_name: Name of the parameter (for error messages)
+        progress_callback: Callback function for download progress updates
+        token: HuggingFace authentication token for private repositories
+        
+    Returns:
+        str: Local path to the model (either existing or downloaded)
+        
+    Raises:
+        RuntimeError: If download fails
+    """
+    # Check if model_path exists locally (file or directory)
+    if os.path.exists(model_path):
+        # Local path exists, return as-is
+        return model_path
+    
+    # Model path doesn't exist locally, try to download from HuggingFace
+    try:
+        # Parse model_path to extract repo_id and filename
+        repo_id, file_name = _parse_model_path(model_path)
+        
+        # Download the model
+        downloaded_path = download_from_huggingface(
+            repo_id=repo_id,
+            file_name=file_name,
+            local_dir=None,  # Use default cache directory
+            enable_transfer=True,
+            progress_callback=progress_callback,
+            show_progress=True,
+            token=token
+        )
+        
+        return downloaded_path
+        
+    except Exception as e:
+        # Only handle download-related errors
+        raise RuntimeError(f"Could not load model from '{param_name}={model_path}': {e}")
+
+
 def auto_download_model(func: Callable) -> Callable:
     """
     Decorator that automatically downloads models from HuggingFace if they don't exist locally.
     
-    This decorator should be applied to __init__ methods that take a name_or_path parameter.
-    If name_or_path doesn't exist as a local file/directory, it will attempt to download
-    it from HuggingFace Hub using the download_from_huggingface function.
+    This decorator should be applied to __init__ methods that take a name_or_path parameter
+    and optionally an mmproj_path parameter. If these paths don't exist as local files/directories,
+    it will attempt to download them from HuggingFace Hub using the download_from_huggingface function.
     
-    The name_or_path can be in formats like:
+    The name_or_path and mmproj_path can be in formats like:
     - "microsoft/DialoGPT-small" (downloads entire repo)  
     - "microsoft/DialoGPT-small/pytorch_model.bin" (downloads specific file)
     - "Qwen/Qwen3-4B-GGUF/Qwen3-4B-Q4_K_M.gguf" (downloads specific file)
@@ -1149,21 +1198,6 @@ def auto_download_model(func: Callable) -> Callable:
     """
     @functools.wraps(func)
     def wrapper(*args, **kwargs):
-        # Find name_or_path in arguments
-        # Assuming name_or_path is the first argument after self
-        if len(args) >= 2:
-            name_or_path = args[1]
-            args_list = list(args)
-            path_index = 1
-            is_positional = True
-        elif 'name_or_path' in kwargs:
-            name_or_path = kwargs['name_or_path']
-            path_index = None
-            is_positional = False
-        else:
-            # No name_or_path found, call original function
-            return func(*args, **kwargs)
-        
         # Extract progress_callback and token from arguments
         progress_callback = None
         if 'progress_callback' in kwargs:
@@ -1173,39 +1207,63 @@ def auto_download_model(func: Callable) -> Callable:
         if 'token' in kwargs:
             token = kwargs.pop('token')  # Remove from kwargs to avoid passing to original func
         
-        # Check if name_or_path exists locally (file or directory)
-        if os.path.exists(name_or_path):
-            # Local path exists, use as-is without downloading
+        # Handle name_or_path parameter
+        name_or_path = None
+        name_path_index = None
+        is_name_positional = False
+        
+        # Find name_or_path in arguments
+        # Assuming name_or_path is the first argument after self
+        if len(args) >= 2:
+            name_or_path = args[1]
+            args_list = list(args)
+            name_path_index = 1
+            is_name_positional = True
+        elif 'name_or_path' in kwargs:
+            name_or_path = kwargs['name_or_path']
+            is_name_positional = False
+        
+        # Handle mmproj_path parameter
+        mmproj_path = None
+        if 'mmproj_path' in kwargs:
+            mmproj_path = kwargs['mmproj_path']
+        
+        # If neither parameter is found, call original function
+        if name_or_path is None and mmproj_path is None:
             return func(*args, **kwargs)
         
-        # Model path doesn't exist locally, try to download from HuggingFace
-        try:
-            # Parse name_or_path to extract repo_id and filename
-            repo_id, file_name = _parse_model_path(name_or_path)
-            
-            # Download the model
-            downloaded_path = download_from_huggingface(
-                repo_id=repo_id,
-                file_name=file_name,
-                local_dir=None,  # Use default cache directory
-                enable_transfer=True,
-                progress_callback=progress_callback,  # Use the extracted callback
-                show_progress=True,
-                token=token  # Use the extracted token
-            )
-            
-            # Replace name_or_path with downloaded path
-            if is_positional:
-                args_list[path_index] = downloaded_path
-                args = tuple(args_list)
-            else:
-                kwargs['name_or_path'] = downloaded_path
-            
-        except Exception as e:
-            # Only handle download-related errors
-            raise RuntimeError(f"Could not load model from '{name_or_path}': {e}")
+        # Download name_or_path if needed
+        if name_or_path is not None:
+            try:
+                downloaded_name_path = _download_model_if_needed(
+                    name_or_path, 'name_or_path', progress_callback, token
+                )
+                
+                # Replace name_or_path with downloaded path
+                if is_name_positional:
+                    if name_path_index is not None:
+                        args_list[name_path_index] = downloaded_name_path
+                        args = tuple(args_list)
+                else:
+                    kwargs['name_or_path'] = downloaded_name_path
+                    
+            except Exception as e:
+                raise e  # Re-raise the error from _download_model_if_needed
+        
+        # Download mmproj_path if needed
+        if mmproj_path is not None:
+            try:
+                downloaded_mmproj_path = _download_model_if_needed(
+                    mmproj_path, 'mmproj_path', progress_callback, token
+                )
+                
+                # Replace mmproj_path with downloaded path
+                kwargs['mmproj_path'] = downloaded_mmproj_path
+                
+            except Exception as e:
+                raise e  # Re-raise the error from _download_model_if_needed
         
-        # Call original function with updated path (outside try-catch to let model creation errors bubble up)
+        # Call original function with updated paths (outside try-catch to let model creation errors bubble up)
         return func(*args, **kwargs)
     
     return wrapper

nexaai/vlm_impl/pybind_vlm_impl.py

@@ -1,7 +1,6 @@
 from typing import Generator, Optional, List, Dict, Any, Union
 import queue
 import threading
-import base64
 from pathlib import Path
 
 from nexaai.common import ModelConfig, GenerationConfig, MultiModalMessage, PluginID
@@ -102,26 +101,16 @@ class PyBindVLMImpl(VLM):
                 t = c["type"]
                 if t == "text":
                     blocks.append({"type": "text", "text": c.get("text","") or ""})
+                elif t == "image":
+                    # Pass through the original structure - let vlm-bind.cpp handle field extraction
+                    blocks.append(c)
                 else:
-                    # image/audio/video
-                    src = c.get("url") or c.get("path")
-                    if not src:
-                        raise ValueError(f"No url/path for {t}")
-                    # read local file or strip data URI
-                    if Path(src).exists():
-                        raw = Path(src).read_bytes()
-                        b64 = base64.b64encode(raw).decode("ascii")
-                        blocks.append({"type": t, "text": b64})
-                    elif src.startswith("data:"):
-                        b64 = src.split(",",1)[1]
-                        blocks.append({"type": t, "text": b64})
-                    else:
-                        # remote URL
-                        blocks.append({"type": t, "text": src})
+                    raise ValueError(f"Unsupported content type: {t}. Use 'text' or 'image' to match the golden reference in vlm.cpp")
 
             payload.append({"role": role, "content": blocks})
 
-        return vlm_bind.ml_vlm_apply_chat_template(self._handle, payload, tools)
+        result = vlm_bind.ml_vlm_apply_chat_template(self._handle, payload, tools)
+        return result
 
     def generate_stream(self, prompt: str, g_cfg: GenerationConfig = GenerationConfig()) -> Generator[str, None, None]:
         """Generate text with streaming."""

{nexaai-1.0.12rc1.dist-info → nexaai-1.0.13rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nexaai
-Version: 1.0.12rc1
+Version: 1.0.13rc1
 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-1.0.12rc1.dist-info → nexaai-1.0.13rc1.dist-info}/RECORD

@@ -1,6 +1,6 @@
 nexaai/__init__.py,sha256=Lt8NU57eTMtWrDYzpFeYR9XtGAPXqizynP83TPU0UW0,2105
-nexaai/_stub.cp310-win_amd64.pyd,sha256=l_sOvn0sGZaAzxUURKum8xng89EtFH_aE1lcWv_ZpKA,10752
-nexaai/_version.py,sha256=56eZbYiAoFM_PEDgdfRtf4Lkj9WNLYjfs6jyjl979Rs,147
+nexaai/_stub.cp310-win_amd64.pyd,sha256=2kqB_I7bt9IOBZdXuVSl6GeZyklNg_9BsL_GASPecWU,10752
+nexaai/_version.py,sha256=LWPXus8DDROkMfTf423Cv01gb3gMsNaD34NhoyTjtG4,147
 nexaai/asr.py,sha256=_fsGaxpiU137bUtO5ujtFSYCI1RLsyeEm3Gf4GhHVRk,2118
 nexaai/base.py,sha256=qQBCiQVNzgpkQjZX9aiFDEdbAAe56TROKC3WnWra2Zg,1021
 nexaai/common.py,sha256=6keIpdX5XS5us4z79EMoa6RSkVze9SbbXax13IJ9yvs,3525
@@ -15,21 +15,22 @@ nexaai/vlm.py,sha256=STjXCw67ABrHrEll8A2NGiwmfo7MotfYgBh1k1aNxkk,4775
 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=tYvy0pFhoY29GstDT5r-oRiPRarPLECvJAkcamJItOg,83
-nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=rWuxRVTVPuakItIeyiXeouD25e4Ekl-niQix_GOriIs,201216
-nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=Bz5tKsR7W_ohrsshGq314-ZkfG39H5fs6dH9kJh584A,182784
-nexaai/binds/libcrypto-3-x64.dll,sha256=-Lau6pL5DpDXzpg9MED63gCeL8oRrSLI_e2LeaxIHqk,7314432
-nexaai/binds/libssl-3-x64.dll,sha256=Tzzyu5jRpUugFxr_65hbFlAtFpjxIDpOYMU1E0ijkJw,1313792
-nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=Wpm1mEohC6DWlwYuK8BU_PJQoW5tjsarcjXhLFhdSgY,162816
-nexaai/binds/nexa_bridge.dll,sha256=UbbvGrUkwFk_tMJeALO5HM7O9rMfJ7xt8YXpB5LheEQ,168448
-nexaai/binds/nexa_llama_cpp/ggml-base.dll,sha256=qpJM5qmcne4UqEbsYEBeoDHOe0KyWjqPk-DrKd8TD_k,532480
-nexaai/binds/nexa_llama_cpp/ggml-cpu.dll,sha256=lbNh62Q7bW0hB_Dbps8HqUkLMNwh_VOQhT3gOTrpOt4,672768
-nexaai/binds/nexa_llama_cpp/ggml-cuda.dll,sha256=Hs_WgTlv9d5MkS2S-hENjwMVX953bhUZtM5959Q7jFs,313528832
-nexaai/binds/nexa_llama_cpp/ggml-vulkan.dll,sha256=siZ1qWTIg4LWcOBlwVsdf7mryOIUrhhiMfdITMdH-TM,36627456
-nexaai/binds/nexa_llama_cpp/ggml.dll,sha256=sASNHyXx2Kor4O5Ao5X2AhmLZwMg3hdJe2jMSxdJmi0,66560
-nexaai/binds/nexa_llama_cpp/llama.dll,sha256=S_SaHxVswdt8LlLazOsLhA72pOoD4d2iba9JKjWGTlg,1611776
-nexaai/binds/nexa_llama_cpp/mtmd.dll,sha256=5bAv5F32eaZdlJ_lbwO53Ph3b3c62t14tmZiflkVL8U,561152
-nexaai/binds/nexa_llama_cpp/nexa_plugin.dll,sha256=1AruxKWsmmuvod95SwAQaffwbrtm38EZSJT9M2-ngM0,1405440
+nexaai/binds/__init__.py,sha256=ENl-uoIF9-3XGIXitVgZ2QmJ6p7Yet4h1-X7nUDZ0Hk,108
+nexaai/binds/common_bind.cp310-win_amd64.pyd,sha256=GDLsSU0D_XeYgTEtn8Popmk9xRj3NiLfZgAVTr9J53o,201216
+nexaai/binds/embedder_bind.cp310-win_amd64.pyd,sha256=f4DGCuQjCJ8W686QILwqX_o_GD0f0bI0bUepqxkZpHs,182784
+nexaai/binds/libcrypto-3-x64.dll,sha256=PYwJJtM4xpo2IQfxm-of6VR6rtVcHCyodwcvBHmfzig,7315968
+nexaai/binds/libssl-3-x64.dll,sha256=mp_RoRjYcCeAuqMy3QDYvxQ-cjAMUNCuwA320oXVVpg,1313792
+nexaai/binds/llm_bind.cp310-win_amd64.pyd,sha256=syZ7WqdxBZ2kgSWV7Xz4bdCFTjfUneZAxJQp_RK34to,162816
+nexaai/binds/nexa_bridge.dll,sha256=5_ek6-h0kMlHh4P1YtSJxuAx7U51W7bkjd1Lmu5dx9c,168448
+nexaai/binds/vlm_bind.cp310-win_amd64.pyd,sha256=zb0hJz9eKBSpWDeV0wtgF9QaLnz0Q09Glcp7b1LWihc,168960
+nexaai/binds/nexa_llama_cpp/ggml-base.dll,sha256=kKTLxMmbNEAyyIgk_l6W--Apa5a0nGgV1qdICra3C8Q,532480
+nexaai/binds/nexa_llama_cpp/ggml-cpu.dll,sha256=nrgEyUrV6lzyX5XLgCxfSsj9hz_eU0GCRhU906dHx0Q,672768
+nexaai/binds/nexa_llama_cpp/ggml-cuda.dll,sha256=Zw5tuw8RFRSNGvqFavwbxCAUptn9OYNjHn6zLW0Od5s,313528832
+nexaai/binds/nexa_llama_cpp/ggml-vulkan.dll,sha256=fr-akwWSxnZwNYErSB502bWgx-Rgz8ab0gAIwl1_ASw,36627456
+nexaai/binds/nexa_llama_cpp/ggml.dll,sha256=6moSgelf0ORjwM2Uepzs1iF83eEk6RWrTnPLYd6S97g,66560
+nexaai/binds/nexa_llama_cpp/llama.dll,sha256=hHfbJB4MHAG10VN-sAJb3tjnF0ZDsqOVV7iAZU-rq1s,1611776
+nexaai/binds/nexa_llama_cpp/mtmd.dll,sha256=A95Q7TGpEAXrLLQPqBFEAa9NV4n6ww6JOp30G0zb40s,561152
+nexaai/binds/nexa_llama_cpp/nexa_plugin.dll,sha256=GF4O0KEbL2Pv0A23YMgQJ1JG1hsy5aEdZI5voz33oxA,1405440
 nexaai/cv_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/cv_impl/mlx_cv_impl.py,sha256=QLd_8w90gtxH8kmssaDYatCTRvQNIJuUGKZNnYrmx6E,3317
 nexaai/cv_impl/pybind_cv_impl.py,sha256=aSOCAxmHrwJbEkSN6VX3Cykqlj_9RIpVrZXILul04GA,1096
@@ -51,14 +52,14 @@ nexaai/tts_impl/pybind_tts_impl.py,sha256=n3z4zmPQayQJgAwcvETw0IBUCp8IYROuYFSg0t
 nexaai/utils/avatar_fetcher.py,sha256=D01f8je-37Nd68zGw8MYK2m7y3fvGlC6h0KR-aN9kdU,3925
 nexaai/utils/decode.py,sha256=0Z9jDH4ICzw4YXj8nD4L-sMouDaev-TISGRQ4KzidWE,421
 nexaai/utils/manifest_utils.py,sha256=zMgQpf5dAgF2RjGhk73zBggxRDGMRKDGxh2a8m8kmYg,10045
-nexaai/utils/model_manager.py,sha256=TmaT1fFculHgfAdutpNXP4d07HIhEMPVTfPvFKE-zR0,51480
+nexaai/utils/model_manager.py,sha256=Gr8PfCHIUVKB1xUQL3WDbi-7hQ5dYey3_w9r42hMWyY,53569
 nexaai/utils/model_types.py,sha256=arIyb9q-1uG0nyUGdWZaxxDJAxv0cfnJEpjCzyELL5Q,1416
 nexaai/utils/progress_tracker.py,sha256=BztrFqtjwNUmeREwZ5m7H6ZcrVzQEbpZfsxndWh4z0A,15778
 nexaai/utils/quantization_utils.py,sha256=jjQaz7K4qH6TdP8Tnv5Ktb2viz8BaVBSOrb_jm3ns28,7889
 nexaai/vlm_impl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nexaai/vlm_impl/mlx_vlm_impl.py,sha256=oY_qb9z_iF0zArBuY5CCYIvZcA3R0i_NKXrr_r-QSgg,10989
-nexaai/vlm_impl/pybind_vlm_impl.py,sha256=Hu8g8OXyPn8OzLQOpRSE5lfGmhjChiKj7fMRB8mC_cI,9147
-nexaai-1.0.12rc1.dist-info/METADATA,sha256=6sAd6d9nS11AE5aIHCxvqT5MeuRxoYeezwFsQGjsGkg,1233
-nexaai-1.0.12rc1.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
-nexaai-1.0.12rc1.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
-nexaai-1.0.12rc1.dist-info/RECORD,,
+nexaai/vlm_impl/pybind_vlm_impl.py,sha256=7Bo0kpSrmOdr--bWSpQBvcaexkPPRt3x1yt9e_jIyDs,8686
+nexaai-1.0.13rc1.dist-info/METADATA,sha256=60xKZLiss9_ltNHe6S6o7X-jc2Cbjg5vALDQja6CrXA,1233
+nexaai-1.0.13rc1.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
+nexaai-1.0.13rc1.dist-info/top_level.txt,sha256=LRE2YERlrZk2vfuygnSzsEeqSknnZbz3Z1MHyNmBU4w,7
+nexaai-1.0.13rc1.dist-info/RECORD,,