nexaai 1.0.18rc1__cp310-cp310-macosx_13_0_x86_64.whl → 1.0.19__cp310-cp310-macosx_13_0_x86_64.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__ = "1.0.18-rc1" 
+__version__ = "1.0.19" 

nexaai/asr.py

@@ -35,7 +35,8 @@ class ASR(BaseModel):
                    tokenizer_path: Optional[str] = None,
                    language: Optional[str] = None,
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'ASR':
         """Load ASR model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/cv.py

@@ -73,7 +73,8 @@ class CVModel(BaseModel):
                    _: str,  # TODO: remove this argument, this is a hack to make api design happy
                    config: CVModelConfig,
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'CVModel':
         """Load CV model from configuration, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/embedder.py

@@ -22,7 +22,7 @@ class Embedder(BaseModel):
         pass
 
     @classmethod
-    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP):
+    def _load_from(cls, model_path: str, tokenizer_file: str = "tokenizer.json", plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP, **kwargs):
         """
         Load an embedder from model files, routing to appropriate implementation.
 

nexaai/image_gen.py

@@ -71,7 +71,8 @@ class ImageGen(BaseModel):
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
                    device_id: Optional[str] = None,
                    float16: bool = True,
-                   quantize: bool = False
+                   quantize: bool = False,
+                   **kwargs
         ) -> 'ImageGen':
         """Load image generation model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string

nexaai/llm.py

@@ -15,10 +15,12 @@ class LLM(BaseModel):
     @classmethod
     def _load_from(cls,
                    local_path: str,
+                   model_name: Optional[str] = None,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
-                   device_id: Optional[str] = None
+                   device_id: Optional[str] = None,
+                   **kwargs
         ) -> 'LLM':
         """Load model from local path, routing to appropriate implementation."""
         # Check plugin_id value for routing - handle both enum and string
@@ -26,10 +28,10 @@ class LLM(BaseModel):
         
         if plugin_value == "mlx":
             from nexaai.llm_impl.mlx_llm_impl import MLXLLMImpl
-            return MLXLLMImpl._load_from(local_path, tokenizer_path, m_cfg, plugin_id, device_id)
+            return MLXLLMImpl._load_from(local_path, model_name, tokenizer_path, m_cfg, plugin_id, device_id)
         else:
             from nexaai.llm_impl.pybind_llm_impl import PyBindLLMImpl
-            return PyBindLLMImpl._load_from(local_path, tokenizer_path, m_cfg, plugin_id, device_id)
+            return PyBindLLMImpl._load_from(local_path, model_name, tokenizer_path, m_cfg, plugin_id, device_id)
 
     def cancel_generation(self):
         """Signal to cancel any ongoing stream generation."""

nexaai/llm_impl/mlx_llm_impl.py

@@ -16,6 +16,7 @@ class MLXLLMImpl(LLM):
     @classmethod
     def _load_from(cls,
                    local_path: str,
+                   model_name: Optional[str] = None,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.MLX,
@@ -40,6 +41,7 @@ class MLXLLMImpl(LLM):
             instance = cls(m_cfg)
             instance._mlx_llm = MLXLLMInterface(
                 model_path=local_path,
+                # model_name=model_name, # FIXME: For MLX LLM, model_name is not used
                 tokenizer_path=tokenizer_path or local_path,
                 config=mlx_config,
                 device=device_id

nexaai/llm_impl/pybind_llm_impl.py

@@ -19,6 +19,7 @@ class PyBindLLMImpl(LLM):
     @classmethod
     def _load_from(cls,
                    local_path: str,
+                   model_name: Optional[str] = None,
                    tokenizer_path: Optional[str] = None,
                    m_cfg: ModelConfig = ModelConfig(),
                    plugin_id: Union[PluginID, str] = PluginID.LLAMA_CPP,
@@ -55,6 +56,7 @@ class PyBindLLMImpl(LLM):
         plugin_id_str = plugin_id.value if isinstance(plugin_id, PluginID) else plugin_id
         handle = llm_bind.ml_llm_create(
             model_path=local_path,
+            model_name=model_name,
             tokenizer_path=tokenizer_path,
             model_config=config,
             plugin_id=plugin_id_str,

nexaai/mlx_backend/vlm/generate_qwen3_vl.py

@@ -1,6 +1,5 @@
 import argparse
 import json
-import sys
 import os
 import mlx.core as mx
 import mlx.nn as nn
@@ -10,38 +9,21 @@ import requests
 import numpy as np
 from pathlib import Path
 from huggingface_hub import snapshot_download
-
-# Add current directory to path for imports
-curr_dir = os.path.dirname(os.path.abspath(__file__))
-sys.path.append(curr_dir)
-sys.path.append(os.path.dirname(curr_dir))
-
-# Add the qwen3vl model directory to path
-qwen3vl_dir = os.path.join(curr_dir, "modeling", "models", "qwen3_vl")
-sys.path.append(qwen3vl_dir)
+from dataclasses import dataclass
+from typing import Any, Generator, List, Optional, Sequence, Tuple, Union
 
 # Import required modules for quantized loading
 from transformers import AutoTokenizer
 
-# Try relative imports first, fallback to sys.path approach for Nuitka compatibility
-try:
-    from .modeling.models.qwen3_vl.llm_common.generate import nexa_generate_step
-    from .modeling.models.qwen3_vl.llm_common.cache import make_prompt_cache
-    from .modeling.models.qwen3_vl.qwen3vl import (
-        VEGModel, LLMModel, ModelArgs, VisionConfig, TextConfig, handle_multimodal_embeds
-    )
-    from .modeling.models.qwen3_vl.processor import Qwen3VLProcessor
-except ImportError:
-    # Fallback for Nuitka compiled environment - use sys.path approach
-    from llm_common.generate import nexa_generate_step
-    from llm_common.cache import make_prompt_cache
-    from qwen3vl import VEGModel, LLMModel, ModelArgs, VisionConfig, TextConfig, handle_multimodal_embeds
-    from processor import Qwen3VLProcessor
-
-from ml import ChatMessage
-from dataclasses import dataclass
-from typing import Any, Generator, List, Optional, Sequence, Tuple, Union
+# Import from the nested modeling structure
+from .modeling.models.qwen3_vl.llm_common.generate import nexa_generate_step
+from .modeling.models.qwen3_vl.llm_common.cache import make_prompt_cache
+from .modeling.models.qwen3_vl.qwen3vl import (
+    VEGModel, LLMModel, ModelArgs, VisionConfig, TextConfig, handle_multimodal_embeds
+)
+from .modeling.models.qwen3_vl.processor import Qwen3VLProcessor
 from .generate import GenerationResult
+from ml import ChatMessage
 
 # Custom exception for context length exceeded
 class ContextLengthExceededError(Exception):
@@ -61,17 +43,156 @@ def _ensure_list(x: Union[str, List[str], None]) -> Optional[List[str]]:
     return x if isinstance(x, list) else [x]
 
 
+def get_model_configs(model_name: str):
+    """Get model configurations based on model name"""
+    
+    # 4B model configs (default)
+    if model_name in ["qwen3vl", "qwen3vl-4b", "qwen3vl-4b-thinking"]:
+        vision_config = VisionConfig(
+            hidden_size=1024,
+            intermediate_size=4096,
+            num_heads=16,
+            num_hidden_layers=24,
+            patch_size=16,
+            temporal_patch_size=2,
+            in_channels=3,
+            hidden_act="gelu",
+            spatial_merge_size=2,
+            out_hidden_size=2560,
+            num_position_embeddings=2304,
+            deepstack_visual_indexes=[5, 11, 17],
+        )
+
+        text_config = TextConfig(
+            model_type="qwen3vl",
+            hidden_size=2560,
+            num_hidden_layers=36,
+            intermediate_size=9728,
+            num_attention_heads=32,
+            num_key_value_heads=8,
+            rms_norm_eps=1e-6,
+            vocab_size=151936,
+            max_position_embeddings=32768,
+            rope_theta=5000000.0,
+            head_dim=128,
+            tie_word_embeddings=True,
+            attention_bias=False,
+            attention_dropout=0.0,
+            rope_scaling={"mrope_section": [24, 20, 20],
+                          "rope_type": "default", "type": "default"},
+        )
+        
+    # 8B model configs
+    elif model_name in ["qwen3vl-8b", "qwen3vl-8b-thinking"]:
+        vision_config = VisionConfig(
+            hidden_size=1152,
+            intermediate_size=4304,
+            num_heads=16,
+            num_hidden_layers=27,
+            patch_size=16,
+            temporal_patch_size=2,
+            in_channels=3,
+            hidden_act="gelu",
+            spatial_merge_size=2,
+            out_hidden_size=4096,
+            num_position_embeddings=2304,
+            deepstack_visual_indexes=[8, 16, 24],
+        )
+
+        text_config = TextConfig(
+            model_type="qwen3vl",
+            hidden_size=4096,
+            num_hidden_layers=36,
+            intermediate_size=12288,
+            num_attention_heads=32,
+            num_key_value_heads=8,
+            rms_norm_eps=1e-6,
+            vocab_size=151936,
+            max_position_embeddings=262144,
+            rope_theta=5000000,
+            head_dim=128,
+            tie_word_embeddings=False,
+            attention_bias=False,
+            attention_dropout=0.0,
+            rope_scaling={"mrope_section": [24, 20, 20], "rope_type": "default", "mrope_interleaved": True},
+        )
+    else:
+        # Fallback to 4B config
+        return get_model_configs("qwen3vl-4b")
+        
+    return vision_config, text_config
+
+def get_weight_filenames(model_name: str, model_path: Path):
+    """Get appropriate weight filenames based on model name and available files"""
+    
+    # Determine model size and type based on the actual file structure
+    if "4b" in model_name:
+        size_prefix = "4b"
+    elif "8b" in model_name:
+        size_prefix = "8b"
+    else:
+        size_prefix = "4b"
+    
+    # Determine model type
+    if "thinking" in model_name:
+        model_type = f"{size_prefix}_thinking"
+    else:
+        model_type = f"{size_prefix}_instruct"
+    
+    # Try different weight file patterns matching the actual file structure
+    llm_patterns = [
+        # New naming convention matching actual files
+        f"qwen3vl-llm-{model_type}-q4_0.safetensors",
+        f"qwen3vl-llm-{model_type}-q8_0.safetensors", 
+        f"qwen3vl-llm-{model_type}-f16.safetensors",
+        # Legacy naming convention
+        f"qwen3vl-llm-{size_prefix.upper()}-q4_0.safetensors",
+        f"qwen3vl-llm-{size_prefix.upper()}-q8_0.safetensors",
+        f"qwen3vl-llm-{size_prefix.upper()}-f16.safetensors",
+        f"qwen3vl-llm-{size_prefix.upper()}-f32.safetensors",
+    ]
+    
+    vision_patterns = [
+        f"qwen3vl-vision-{model_type}-f16.safetensors",
+        f"qwen3vl-vision-{size_prefix.upper()}-f16.safetensors",
+    ]
+    
+    # Find LLM weights
+    llm_weights_path = None
+    quantization_bits = None
+    
+    for pattern in llm_patterns:
+        candidate_path = model_path / pattern
+        if candidate_path.exists():
+            llm_weights_path = candidate_path
+            if "q4_0" in pattern:
+                quantization_bits = 4
+            elif "q8_0" in pattern:
+                quantization_bits = 8
+            else:
+                quantization_bits = 16
+            break
+    
+    # Find vision weights
+    vision_weights_path = None
+    for pattern in vision_patterns:
+        candidate_path = model_path / pattern
+        if candidate_path.exists():
+            vision_weights_path = candidate_path
+            break
+    
+    return llm_weights_path, vision_weights_path, quantization_bits
+
+# Update the load_qwen3_vl function signature and implementation:
 def load_qwen3_vl(
     path_or_repo: str,
     adapter_path: Optional[str] = None,
     lazy: bool = False,
     revision: Optional[str] = None,
+    model_name: Optional[str] = None,
     **kwargs,
 ) -> Tuple[Qwen3VLBundledModel, Qwen3VLProcessor]:
-    """Load Qwen3-VL quantized models and processor.
-
-    Parameters are aligned with .generate.load for compatibility.
-    """
+    """Load Qwen3-VL quantized models and processor with support for different model sizes."""
     
     model_path = Path(path_or_repo)
     if not model_path.exists():
@@ -79,75 +200,28 @@ def load_qwen3_vl(
             model_path = Path(snapshot_download(
                 repo_id=path_or_repo, repo_type="model", revision=revision))
         else:
-            # Fallback to local modelfiles directory
-            model_path = Path(qwen3vl_dir) / "modelfiles"
+            # Fallback to local modelfiles directory relative to this file
+            curr_dir = Path(__file__).parent
+            model_path = curr_dir / "modeling" / "models" / "qwen3_vl" / "modelfiles"
             if not model_path.exists():
-                model_path = Path(curr_dir) / "modelfiles"
-
-    # Model configs (kept identical to main)
-    vision_config = VisionConfig(
-        hidden_size=1024,
-        intermediate_size=4096,
-        num_heads=16,
-        num_hidden_layers=24,
-        patch_size=16,
-        temporal_patch_size=2,
-        in_channels=3,
-        hidden_act="gelu",
-        spatial_merge_size=2,
-        out_hidden_size=2560,
-        num_position_embeddings=2304,
-        deepstack_visual_indexes=[5, 11, 17],
-    )
+                model_path = curr_dir / "modelfiles"
 
-    text_config = TextConfig(
-        model_type="qwen3vl",
-        hidden_size=2560,
-        num_hidden_layers=36,
-        intermediate_size=9728,
-        num_attention_heads=32,
-        num_key_value_heads=8,
-        rms_norm_eps=1e-6,
-        vocab_size=151936,
-        max_position_embeddings=32768,
-        rope_theta=5000000.0,
-        head_dim=128,
-        tie_word_embeddings=True,
-        attention_bias=False,
-        attention_dropout=0.0,
-        rope_scaling={"mrope_section": [24, 20, 20],
-                      "rope_type": "default", "type": "default"},
-    )
+    # Get model configurations based on model name
+    if model_name:
+        vision_config, text_config = get_model_configs(model_name)
+    else:
+        # Default to 4B config
+        vision_config, text_config = get_model_configs("qwen3vl-4b")
 
     vision_model = VEGModel(vision_config)
     llm_model = LLMModel(text_config)
 
-    # Try to load LLM model from available files in order of preference
-    preferred_order = [
-        ("qwen3vl-llm-4B-q4_0.safetensors", 4),
-        ("qwen3vl-llm-4B-q8_0.safetensors", 8),
-        ("qwen3vl-llm-4B-f32.safetensors", 32)
-    ]
-
-    llm_weights_path = None
-    quantization_bits = None
-    
-    # Try loading in order of preference
-    for filename, bits in preferred_order:
-        candidate_path = model_path / filename
-        if candidate_path.exists():
-            llm_weights_path = candidate_path
-            quantization_bits = bits
-            break
-    
-    if llm_weights_path is None:
-        # Fallback to original hardcoded path for backward compatibility
-        llm_weights_path = model_path / "qwen3vl-llm-4B-q4_0.safetensors"
-        quantization_bits = 4
-
-    vision_weights_path = model_path / "qwen3vl-vision-4B-f16.safetensors"
+    # Get appropriate weight filenames
+    llm_weights_path, vision_weights_path, quantization_bits = get_weight_filenames(
+        model_name or "qwen3vl-4b", model_path
+    )
 
-    if not vision_weights_path.exists() or not llm_weights_path.exists():
+    if not vision_weights_path or not llm_weights_path:
         raise FileNotFoundError(
             f"Missing safetensors. Vision: {vision_weights_path}, LLM: {llm_weights_path}"
         )
@@ -163,8 +237,14 @@ def load_qwen3_vl(
     
     llm_model.load_weights(str(llm_weights_path), strict=True)
 
-    # Tokenizer and processor
-    tokenizer = AutoTokenizer.from_pretrained(path_or_repo)
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(str(model_path))
+    except Exception:
+        try:
+            tokenizer = AutoTokenizer.from_pretrained(path_or_repo)
+        except Exception:
+            raise Exception("Failed to load tokenizer from the same path where model weights are loaded and original path_or_repo.")
+    
     processor = Qwen3VLProcessor(tokenizer=tokenizer)
 
     return Qwen3VLBundledModel(vision_model=vision_model, llm_model=llm_model), processor