Create app.py

app.py

@@ -0,0 +1,1163 @@
+import asyncio
+import base64
+import io
+import json
+import logging
+import time
+import uuid
+import threading
+from contextlib import asynccontextmanager
+from typing import AsyncGenerator, Dict, List, Optional, Union
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+from fastapi import FastAPI, HTTPException, Request, UploadFile, File
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import JSONResponse
+from huggingface_hub import hf_hub_download, list_repo_files
+from pydantic import BaseModel, Field
+from sse_starlette.sse import EventSourceResponse
+from transformers import AutoImageProcessor, PreTrainedTokenizerFast
+from PIL import Image
+import aiohttp
+
+from config import settings
+
+# Configure logging
+logging.basicConfig(
+    level=getattr(logging, settings.log_level.upper()),
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s"
+)
+logger = logging.getLogger(__name__)
+
+
+# ==============================================================================
+# Pydantic Models for OpenAI-compatible API
+# ==============================================================================
+
+class ImageContent(BaseModel):
+    type: str = "image"
+    image_url: Optional[str] = None  # data:image/jpeg;base64,... or URL
+
+
+class TextContent(BaseModel):
+    type: str = "text"
+    text: str
+
+
+class VisionMessage(BaseModel):
+    role: str = Field(..., description="Role: 'system', 'user', or 'assistant'")
+    content: Union[str, List[Union[ImageContent, TextContent, dict]]] = Field(..., description="Message content")
+
+
+class VisionCompletionRequest(BaseModel):
+    model: str = Field(default="lfm-vision", description="Model identifier")
+    messages: List[VisionMessage] = Field(..., description="Conversation messages")
+    temperature: Optional[float] = Field(default=None, ge=0.0, le=2.0)
+    top_p: Optional[float] = Field(default=None, ge=0.0, le=1.0)
+    top_k: Optional[int] = Field(default=None, ge=0)
+    max_tokens: Optional[int] = Field(default=None, ge=1)
+    stream: bool = Field(default=False, description="Enable streaming response")
+    stop: Optional[Union[str, List[str]]] = Field(default=None)
+
+
+class ChatMessage(BaseModel):
+    role: str = Field(..., description="Role: 'system', 'user', or 'assistant'")
+    content: str = Field(..., description="Message content")
+
+
+class ChatCompletionRequest(BaseModel):
+    model: str = Field(default="lfm-vision", description="Model identifier")
+    messages: List[ChatMessage] = Field(..., description="Conversation messages")
+    temperature: Optional[float] = Field(default=None, ge=0.0, le=2.0)
+    top_p: Optional[float] = Field(default=None, ge=0.0, le=1.0)
+    top_k: Optional[int] = Field(default=None, ge=0)
+    max_tokens: Optional[int] = Field(default=None, ge=1)
+    stream: bool = Field(default=False, description="Enable streaming response")
+
+
+class ChatCompletionChoice(BaseModel):
+    index: int
+    message: ChatMessage
+    finish_reason: Optional[str] = None
+
+
+class ChatCompletionResponse(BaseModel):
+    id: str
+    object: str = "chat.completion"
+    created: int
+    model: str
+    choices: List[ChatCompletionChoice]
+    usage: Dict[str, int]
+
+
+class ModelInfo(BaseModel):
+    id: str
+    object: str = "model"
+    created: int
+    owned_by: str = "liquid-ai"
+
+
+class ModelListResponse(BaseModel):
+    object: str = "list"
+    data: List[ModelInfo]
+
+
+# ==============================================================================
+# ONNX Vision Model Manager
+# ==============================================================================
+
+# ONNX dtype mapping
+ONNX_DTYPE = {
+    "tensor(float)": np.float32,
+    "tensor(float16)": np.float16,
+    "tensor(int64)": np.int64
+}
+
+
+class Lfm2VlProcessorWrapper:
+    """
+    Custom processor wrapper that combines ImageProcessor + Tokenizer.
+    This bypasses the AutoProcessor tokenizer auto-detection bug in LFM models.
+    """
+    
+    def __init__(self, image_processor, tokenizer):
+        self.image_processor = image_processor
+        self.tokenizer = tokenizer
+    
+    def apply_chat_template(self, messages, add_generation_prompt=True, tokenize=False, **kwargs):
+        """
+        Apply chat template for vision-language model.
+        Converts vision message format [{"type": "image"}, {"type": "text", "text": "..."}]
+        to text with <image> placeholders as expected by the tokenizer.
+        """
+        # Transform vision messages to text format
+        text_messages = []
+        for msg in messages:
+            role = msg.get("role", "user") if isinstance(msg, dict) else getattr(msg, "role", "user")
+            content = msg.get("content", "") if isinstance(msg, dict) else getattr(msg, "content", "")
+            
+            if isinstance(content, list):
+                # Vision message format: [{"type": "image"}, {"type": "text", "text": "..."}]
+                text_parts = []
+                for item in content:
+                    if isinstance(item, dict):
+                        item_type = item.get("type", "")
+                        if item_type == "image":
+                            text_parts.append("<image>")
+                        elif item_type == "text":
+                            text_parts.append(item.get("text", ""))
+                    else:
+                        text_parts.append(str(item))
+                content = "".join(text_parts)
+            
+            text_messages.append({"role": role, "content": content})
+        
+        return self.tokenizer.apply_chat_template(
+            text_messages,
+            add_generation_prompt=add_generation_prompt,
+            tokenize=tokenize,
+            **kwargs
+        )
+    
+    def __call__(self, images=None, text=None, **kwargs):
+        """
+        Process images and text for the vision-language model.
+        
+        CRITICAL: The vision encoder produces N image embeddings (e.g., 256 for a 512x512 image).
+        Each embedding needs its own <image> token position in input_ids.
+        
+        This method:
+        1. Processes images FIRST to determine N (number of image tokens)
+        2. Expands single <image> in text to N consecutive <image> tokens
+        3. Tokenizes the expanded text
+        
+        Returns a dict with pixel_values, input_ids, attention_mask, etc.
+        """
+        result = {}
+        return_tensors = kwargs.pop('return_tensors', None)
+        num_image_tokens = 0
+        
+        # Step 1: Process images FIRST to get the number of image tokens
+        if images is not None:
+            image_outputs = self.image_processor(images=images, return_tensors=return_tensors)
+            result.update(image_outputs)
+            
+            # Calculate number of image tokens from pixel_values shape
+            # pixel_values shape: [batch, num_patches, hidden_dim]
+            # The MLP projector in LFM2.5-VL reduces patches by factor of 4
+            # Reference: https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B
+            if 'pixel_values' in image_outputs:
+                pv = image_outputs['pixel_values']
+                num_patches = pv.shape[1] if hasattr(pv, 'shape') else pv.size(1)
+                # MLP projector reduces by factor of 4: 1024 patches → 256 tokens
+                num_image_tokens = num_patches // 4
+                logger.debug(f"Image processing: {num_patches} patches → {num_image_tokens} image tokens")
+        
+        # Step 2: Expand <image> placeholder(s) to match token count
+        if text is not None:
+            # Ensure text is a string
+            if isinstance(text, list):
+                text = text[0] if len(text) == 1 else " ".join(text)
+            
+            # Expand each <image> placeholder to N <image> tokens
+            if num_image_tokens > 0 and "<image>" in text:
+                # Count existing <image> placeholders
+                image_count = text.count("<image>")
+                # Each placeholder represents one image, expand to num_image_tokens
+                tokens_per_image = num_image_tokens // image_count if image_count > 0 else num_image_tokens
+                expanded_image = "<image>" * tokens_per_image
+                text = text.replace("<image>", expanded_image)
+                logger.debug(f"Expanded {image_count} <image> placeholder(s) to {tokens_per_image} tokens each")
+            
+            text_outputs = self.tokenizer(
+                text, 
+                return_tensors=return_tensors,
+                padding=kwargs.get('padding', False),
+                truncation=kwargs.get('truncation', False),
+                max_length=kwargs.get('max_length', None)
+            )
+            result.update(text_outputs)
+        
+        return result
+
+
+class ONNXVisionModelManager:
+    """Manages ONNX Vision-Language model with 3 sessions: embed_tokens, embed_images, decoder."""
+    
+    def __init__(self):
+        self._embed_tokens = None
+        self._embed_images = None
+        self._decoder = None
+        self._processor = None
+        self._cache_template = None
+        self._lock = threading.Lock()
+    
+    @property
+    def is_loaded(self) -> bool:
+        return all([self._embed_tokens, self._embed_images, self._decoder])
+    
+    def download_models(self) -> Dict[str, str]:
+        """Download ONNX model files from HuggingFace."""
+        model_id = settings.model_id
+        encoder_var = settings.encoder_variant
+        decoder_var = settings.decoder_variant
+        
+        logger.info(f"Downloading model: {model_id}")
+        logger.info(f"  Encoder variant: {encoder_var}")
+        logger.info(f"  Decoder variant: {decoder_var}")
+        
+        paths = {}
+        
+        # Download embed_tokens (use same variant as encoder or fp16)
+        embed_suffix = f"_fp16" if encoder_var in ["fp16", "q8", "q4"] else ""
+        paths["embed_tokens"] = hf_hub_download(model_id, f"onnx/embed_tokens{embed_suffix}.onnx")
+        
+        # Download embed_images (vision encoder)
+        img_suffix = f"_{encoder_var}" if encoder_var != "fp32" else ""
+        paths["embed_images"] = hf_hub_download(model_id, f"onnx/embed_images{img_suffix}.onnx")
+        
+        # Download decoder
+        dec_suffix = f"_{decoder_var}" if decoder_var != "fp32" else ""
+        paths["decoder"] = hf_hub_download(model_id, f"onnx/decoder{dec_suffix}.onnx")
+        
+        # Download all data files - use exact prefix matching to avoid downloading wrong variants
+        # Expected files for selected variants only (e.g., decoder_q8.onnx_data, not decoder.onnx_data)
+        expected_prefixes = [
+            f"onnx/embed_tokens{embed_suffix}.onnx_data",
+            f"onnx/embed_images{img_suffix}.onnx_data",
+            f"onnx/decoder{dec_suffix}.onnx_data"
+        ]
+        
+        for f in list_repo_files(model_id):
+            if f.startswith("onnx/") and ".onnx_data" in f:
+                # Check if this file STARTS WITH one of our expected prefixes
+                # This handles split files like decoder_q8.onnx_data, decoder_q8.onnx_data_1, etc.
+                if any(f.startswith(prefix) for prefix in expected_prefixes):
+                    logger.info(f"Downloading: {f}")
+                    hf_hub_download(model_id, f)
+        
+        return paths
+    
+    def load_model(self) -> None:
+        """Load the ONNX models and processor."""
+        with self._lock:
+            if self.is_loaded:
+                return
+            
+            logger.info("=" * 60)
+            logger.info("Loading LFM2.5-VL-1.6B Vision-Language ONNX model...")
+            logger.info(f"Model: {settings.model_id}")
+            logger.info(f"Encoder: {settings.encoder_variant} (Q8 = ~95% accuracy)")
+            logger.info(f"Decoder: {settings.decoder_variant}")
+            logger.info("=" * 60)
+            
+            start_time = time.time()
+            
+            # Download models
+            paths = self.download_models()
+            
+            # Configure ONNX Runtime for CPU
+            sess_options = ort.SessionOptions()
+            sess_options.intra_op_num_threads = settings.num_threads
+            sess_options.inter_op_num_threads = settings.num_threads
+            sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+            
+            # Load ONNX sessions
+            self._embed_tokens = ort.InferenceSession(
+                paths["embed_tokens"],
+                sess_options=sess_options,
+                providers=['CPUExecutionProvider']
+            )
+            
+            self._embed_images = ort.InferenceSession(
+                paths["embed_images"],
+                sess_options=sess_options,
+                providers=['CPUExecutionProvider']
+            )
+            
+            self._decoder = ort.InferenceSession(
+                paths["decoder"],
+                sess_options=sess_options,
+                providers=['CPUExecutionProvider']
+            )
+            
+            # Load processor components separately to bypass TokenizersBackend bug
+            # LFM models incorrectly specify TokenizersBackend as tokenizer_class
+            logger.info("Loading image processor...")
+            image_processor = AutoImageProcessor.from_pretrained(
+                settings.model_id,
+                trust_remote_code=True
+            )
+            
+            logger.info("Loading tokenizer with PreTrainedTokenizerFast...")
+            tokenizer = PreTrainedTokenizerFast.from_pretrained(
+                settings.model_id,
+                trust_remote_code=True
+            )
+            
+            # Create our custom processor wrapper
+            self._processor = Lfm2VlProcessorWrapper(
+                image_processor=image_processor,
+                tokenizer=tokenizer
+            )
+            logger.info(f"✓ Processor created: {type(self._processor).__name__}")
+            
+            # Initialize cache template for decoder
+            self._init_cache_template()
+            
+            load_time = time.time() - start_time
+            logger.info("=" * 60)
+            logger.info(f"✓ Model loaded in {load_time:.2f}s")
+            logger.info(f"  Threads: {settings.num_threads}")
+            logger.info(f"  Provider: CPU")
+            logger.info("=" * 60)
+    
+    def _init_cache_template(self) -> None:
+        """Initialize KV cache template for decoder."""
+        self._cache_template = {}
+        for inp in self._decoder.get_inputs():
+            if inp.name in {"inputs_embeds", "attention_mask", "position_ids"}:
+                continue
+            
+            shape = [d if isinstance(d, int) else 1 for d in inp.shape]
+            for i, d in enumerate(inp.shape):
+                if isinstance(d, str) and "sequence" in d.lower():
+                    shape[i] = 0
+            
+            dtype = ONNX_DTYPE.get(inp.type, np.float32)
+            self._cache_template[inp.name] = (shape, dtype)
+    
+    def _create_empty_cache(self) -> Dict[str, np.ndarray]:
+        """Create a new empty KV cache."""
+        return {
+            name: np.zeros(shape, dtype=dtype)
+            for name, (shape, dtype) in self._cache_template.items()
+        }
+    
+    @property
+    def processor(self):
+        if self._processor is None:
+            raise RuntimeError("Processor not loaded")
+        return self._processor
+    
+    def process_image(self, image: Image.Image) -> Dict[str, np.ndarray]:
+        """Process image to embeddings."""
+        # Ensure RGB
+        if image.mode != "RGB":
+            image = image.convert("RGB")
+        
+        return image
+    
+    def generate(
+        self,
+        images: List[Image.Image],
+        messages: List[dict],
+        max_tokens: int = 512,
+        temperature: float = 0.1,
+        top_k: int = 50,
+        top_p: float = 0.1,
+        stop_tokens: Optional[List[int]] = None
+    ) -> List[int]:
+        """Generate tokens using ONNX Vision model."""
+        tokenizer = self._processor.tokenizer
+        
+        if stop_tokens is None:
+            stop_tokens = [tokenizer.eos_token_id]
+        
+        # Process inputs through processor
+        prompt = self._processor.apply_chat_template(messages, add_generation_prompt=True)
+        inputs = self._processor(
+            images=images if images else None,
+            text=prompt,
+            return_tensors="pt"
+        )
+        
+        # Convert to numpy with correct dtypes
+        input_ids = inputs["input_ids"].numpy().astype(np.int64)
+        
+        # Get token embeddings
+        token_outputs = self._embed_tokens.run(None, {"input_ids": input_ids})
+        token_embeds = token_outputs[0]
+        
+        # Process images if present
+        if images and "pixel_values" in inputs:
+            pixel_values = inputs["pixel_values"].numpy().astype(np.float32)
+            pixel_attention_mask = inputs.get("pixel_attention_mask", None)
+            spatial_shapes = inputs.get("spatial_shapes", None)
+            
+            image_feed = {"pixel_values": pixel_values}
+            if pixel_attention_mask is not None:
+                image_feed["pixel_attention_mask"] = pixel_attention_mask.numpy().astype(np.int64)
+            if spatial_shapes is not None:
+                image_feed["spatial_shapes"] = spatial_shapes.numpy().astype(np.int64)
+            
+            image_outputs = self._embed_images.run(None, image_feed)
+            image_embeds = image_outputs[0]
+            
+            # Replace <image> tokens with image embeddings
+            image_token_id = tokenizer.convert_tokens_to_ids("<image>")
+            image_positions = np.where(input_ids[0] == image_token_id)[0]
+            for i, pos in enumerate(image_positions):
+                if i < len(image_embeds):
+                    token_embeds[0, pos] = image_embeds[i]
+        
+        # Initialize KV cache
+        cache = self._create_empty_cache()
+        seq_len = token_embeds.shape[1]
+        generated_tokens = []
+        
+        for step in range(max_tokens):
+            if step == 0:
+                embeds = token_embeds.astype(np.float32)
+            else:
+                last_token = np.array([[generated_tokens[-1]]], dtype=np.int64)
+                embeds = self._embed_tokens.run(None, {"input_ids": last_token})[0].astype(np.float32)
+            
+            attn_mask = np.ones((1, seq_len + len(generated_tokens)), dtype=np.int64)
+            
+            feed = {"inputs_embeds": embeds, "attention_mask": attn_mask, **cache}
+            outputs = self._decoder.run(None, feed)
+            
+            # Get logits and apply temperature
+            logits = outputs[0][0, -1]
+            
+            if temperature > 0:
+                logits = logits / temperature
+                
+                # Apply top-k
+                if top_k > 0:
+                    indices_to_remove = np.argsort(logits)[:-top_k]
+                    logits[indices_to_remove] = -np.inf
+                
+                # Apply top-p (nucleus sampling)
+                if top_p < 1.0:
+                    sorted_indices = np.argsort(logits)[::-1]
+                    sorted_logits = logits[sorted_indices]
+                    probs = np.exp(sorted_logits - np.max(sorted_logits))
+                    probs = probs / probs.sum()
+                    cumulative_probs = np.cumsum(probs)
+                    sorted_indices_to_remove = cumulative_probs > top_p
+                    sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                    sorted_indices_to_remove[0] = False
+                    indices_to_remove = sorted_indices[sorted_indices_to_remove]
+                    logits[indices_to_remove] = -np.inf
+                
+                # Sample
+                probs = np.exp(logits - np.max(logits))
+                probs = probs / probs.sum()
+                next_token = int(np.random.choice(len(probs), p=probs))
+            else:
+                next_token = int(np.argmax(logits))
+            
+            generated_tokens.append(next_token)
+            
+            # Update cache
+            for i, out in enumerate(self._decoder.get_outputs()[1:], 1):
+                name = out.name.replace("present_conv", "past_conv").replace("present.", "past_key_values.")
+                if name in cache:
+                    cache[name] = outputs[i]
+            
+            if next_token in stop_tokens:
+                break
+        
+        return generated_tokens
+    
+    def generate_stream(
+        self,
+        images: List[Image.Image],
+        messages: List[dict],
+        max_tokens: int = 2000,
+        temperature: float = 0.1,
+        top_k: int = 50,
+        top_p: float = 0.1,
+        stop_tokens: Optional[List[int]] = None
+    ):
+        """Streaming generation for Vision model."""
+        tokenizer = self._processor.tokenizer
+        
+        if stop_tokens is None:
+            stop_tokens = [tokenizer.eos_token_id]
+        
+        # Process inputs through processor
+        prompt = self._processor.apply_chat_template(messages, add_generation_prompt=True)
+        inputs = self._processor(
+            images=images if images else None,
+            text=prompt,
+            return_tensors="pt"
+        )
+        
+        # Convert to numpy with correct dtypes
+        input_ids = inputs["input_ids"].numpy().astype(np.int64)
+        
+        # Get token embeddings
+        token_outputs = self._embed_tokens.run(None, {"input_ids": input_ids})
+        token_embeds = token_outputs[0]
+        
+        # Process images if present
+        if images and "pixel_values" in inputs:
+            pixel_values = inputs["pixel_values"].numpy().astype(np.float32)
+            pixel_attention_mask = inputs.get("pixel_attention_mask", None)
+            spatial_shapes = inputs.get("spatial_shapes", None)
+            
+            image_feed = {"pixel_values": pixel_values}
+            if pixel_attention_mask is not None:
+                image_feed["pixel_attention_mask"] = pixel_attention_mask.numpy().astype(np.int64)
+            if spatial_shapes is not None:
+                image_feed["spatial_shapes"] = spatial_shapes.numpy().astype(np.int64)
+            
+            image_outputs = self._embed_images.run(None, image_feed)
+            image_embeds = image_outputs[0]
+            
+            # Replace <image> tokens with image embeddings
+            image_token_id = tokenizer.convert_tokens_to_ids("<image>")
+            image_positions = np.where(input_ids[0] == image_token_id)[0]
+            for i, pos in enumerate(image_positions):
+                if i < len(image_embeds):
+                    token_embeds[0, pos] = image_embeds[i]
+        
+        # Initialize KV cache
+        cache = self._create_empty_cache()
+        seq_len = token_embeds.shape[1]
+        generated_tokens = []
+        
+        # Pre-allocate attention mask
+        max_possible_len = seq_len + max_tokens
+        attn_mask = np.ones((1, max_possible_len), dtype=np.int64)
+        
+        # Pre-compute flags
+        use_temp = temperature > 0
+        use_top_k = top_k > 0
+        use_top_p = top_p < 1.0
+        
+        feed = {}
+        
+        for step in range(max_tokens):
+            current_len = seq_len + step
+            
+            if step == 0:
+                embeds = token_embeds.astype(np.float32)
+            else:
+                last_token = np.array([[generated_tokens[-1]]], dtype=np.int64)
+                embeds = self._embed_tokens.run(None, {"input_ids": last_token})[0].astype(np.float32)
+            
+            # Update Feed Dict
+            feed.clear()
+            feed["inputs_embeds"] = embeds
+            feed["attention_mask"] = attn_mask[:, :current_len]
+            feed.update(cache)
+            
+            # Inference
+            outputs = self._decoder.run(None, feed)
+            logits = outputs[0][0, -1]
+            
+            # Sampling
+            if use_temp:
+                logits /= temperature
+                
+                if use_top_k and top_k < len(logits):
+                    top_k_idx = np.argpartition(logits, -top_k)[-top_k:]
+                    mask = np.ones(logits.shape, dtype=bool)
+                    mask[top_k_idx] = False
+                    logits[mask] = -np.inf
+                
+                if use_top_p:
+                    valid_mask = logits > -np.inf
+                    if valid_mask.any():
+                        valid_logits = logits[valid_mask]
+                        valid_indices = np.where(valid_mask)[0]
+                        
+                        sorted_indices = np.argsort(valid_logits)[::-1]
+                        sorted_logits = valid_logits[sorted_indices]
+                        
+                        exp_logits = np.exp(sorted_logits - np.max(sorted_logits))
+                        probs = exp_logits / exp_logits.sum()
+                        
+                        cumulative = np.cumsum(probs)
+                        cutoff = np.searchsorted(cumulative, top_p)
+                        cutoff = min(cutoff + 1, len(sorted_logits))
+                        
+                        accepted_indices = sorted_indices[:cutoff]
+                        accepted_probs = probs[:cutoff]
+                        accepted_probs /= accepted_probs.sum()
+                        
+                        sample_idx = np.searchsorted(np.cumsum(accepted_probs), np.random.rand())
+                        next_token = int(valid_indices[accepted_indices[sample_idx]])
+                    else:
+                        next_token = int(np.argmax(logits))
+                else:
+                    valid_mask = logits > -np.inf
+                    valid_logits = logits[valid_mask]
+                    valid_indices = np.where(valid_mask)[0]
+                    exp_logits = np.exp(valid_logits - np.max(valid_logits))
+                    probs = exp_logits / exp_logits.sum()
+                    sample_idx = np.searchsorted(np.cumsum(probs), np.random.rand())
+                    next_token = int(valid_indices[sample_idx])
+            else:
+                next_token = int(np.argmax(logits))
+            
+            generated_tokens.append(next_token)
+            yield next_token
+            
+            if next_token in stop_tokens:
+                break
+            
+            # Update Cache
+            for i, out in enumerate(self._decoder.get_outputs()[1:], 1):
+                name = out.name.replace("present_conv", "past_conv").replace("present.", "past_key_values.")
+                if name in cache:
+                    cache[name] = outputs[i]
+    
+    def unload(self) -> None:
+        """Unload models from memory."""
+        with self._lock:
+            if self._embed_tokens is not None:
+                del self._embed_tokens
+                del self._embed_images
+                del self._decoder
+                del self._processor
+                self._embed_tokens = None
+                self._embed_images = None
+                self._decoder = None
+                self._processor = None
+                logger.info("Models unloaded")
+
+
+# Global model manager
+model_manager = ONNXVisionModelManager()
+
+
+# ==============================================================================
+# Image Processing Utilities
+# ==============================================================================
+
+def resize_image_for_model(image: Image.Image, max_dim: int = 512) -> Image.Image:
+    """
+    Resize image to max dimension while preserving aspect ratio.
+    Uses LANCZOS (highest quality) resampling for best visual fidelity.
+    
+    This optimization ensures:
+    - Consistent processing time (~3-4s) regardless of input size
+    - Single-patch processing (256 tokens) instead of tiling
+    - Reduced memory usage
+    
+    Args:
+        image: PIL Image to resize
+        max_dim: Maximum dimension (width or height), default 512
+        
+    Returns:
+        Resized PIL Image (or original if already small enough)
+    """
+    width, height = image.size
+    
+    # Skip if already small enough
+    if width <= max_dim and height <= max_dim:
+        logger.debug(f"Image {width}x{height} already within {max_dim}px limit")
+        return image
+    
+    # Calculate new dimensions (preserve aspect ratio)
+    ratio = min(max_dim / width, max_dim / height)
+    new_width = int(width * ratio)
+    new_height = int(height * ratio)
+    
+    logger.info(f"Resizing image: {width}x{height} → {new_width}x{new_height} (LANCZOS)")
+    
+    # Resize with high-quality LANCZOS filter
+    return image.resize((new_width, new_height), Image.Resampling.LANCZOS)
+
+async def load_image_from_url(url: str) -> Image.Image:
+    """Load image from URL, convert to RGB, and resize for optimal processing."""
+    async with aiohttp.ClientSession() as session:
+        async with session.get(url) as response:
+            if response.status != 200:
+                raise HTTPException(status_code=400, detail=f"Failed to fetch image from URL: {url}")
+            data = await response.read()
+            image = Image.open(io.BytesIO(data))
+            # Convert to RGB to ensure consistent channel format
+            if image.mode != 'RGB':
+                image = image.convert('RGB')
+            # Resize for optimal model processing (max 512x512)
+            image = resize_image_for_model(image)
+            return image
+
+
+def load_image_from_base64(data_url: str) -> Image.Image:
+    """Load image from base64 data URL, convert to RGB, and resize for optimal processing."""
+    # Format: data:image/jpeg;base64,/9j/4AAQ...
+    if "," in data_url:
+        header, encoded = data_url.split(",", 1)
+    else:
+        encoded = data_url
+    
+    image_data = base64.b64decode(encoded)
+    image = Image.open(io.BytesIO(image_data))
+    # Convert to RGB to ensure consistent channel format
+    if image.mode != 'RGB':
+        image = image.convert('RGB')
+    # Resize for optimal model processing (max 512x512)
+    image = resize_image_for_model(image)
+    return image
+
+
+async def process_image_content(content: Union[ImageContent, dict]) -> Optional[Image.Image]:
+    """Process image content from request."""
+    if isinstance(content, dict):
+        content = ImageContent(**content)
+    
+    if content.type != "image":
+        return None
+    
+    if not content.image_url:
+        return None
+    
+    url = content.image_url
+    
+    # Check if it's a base64 data URL
+    if url.startswith("data:"):
+        return load_image_from_base64(url)
+    else:
+        # It's a regular URL
+        return await load_image_from_url(url)
+
+
+# ==============================================================================
+# Application Lifecycle
+# ==============================================================================
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """Application lifespan handler."""
+    logger.info("Starting LFM2.5-VL Vision API Server (ONNX Runtime)...")
+    
+    loop = asyncio.get_event_loop()
+    await loop.run_in_executor(None, model_manager.load_model)
+    
+    yield
+    
+    logger.info("Shutting down...")
+    model_manager.unload()
+
+
+# ==============================================================================
+# FastAPI Application
+# ==============================================================================
+
+app = FastAPI(
+    title=settings.app_name,
+    description="Fast CPU inference for LiquidAI LFM2.5-VL-1.6B Vision-Language model using ONNX Runtime",
+    version=settings.app_version,
+    lifespan=lifespan,
+    docs_url="/docs",
+    redoc_url="/redoc",
+)
+
+origins = [
+    "http://127.0.0.1:5500",
+    "http://127.0.0.1:5501",
+    "http://localhost:5500",
+    "http://localhost:5173",
+    "https://toolboxesai.com"
+]
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=origins,
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+
+@app.middleware("http")
+async def add_cors_for_null_origin(request: Request, call_next):
+    """Handle CORS for null origin (when HTML is opened from file://)."""
+    origin = request.headers.get("origin", "")
+    response = await call_next(request)
+    
+    if origin == "null" or not origin:
+        response.headers["Access-Control-Allow-Origin"] = "*"
+        response.headers["Access-Control-Allow-Methods"] = "GET, POST, PUT, DELETE, OPTIONS"
+        response.headers["Access-Control-Allow-Headers"] = "*"
+        response.headers["Access-Control-Expose-Headers"] = "*"
+    
+    return response
+
+
+# ==============================================================================
+# Helper Functions
+# ==============================================================================
+
+def generate_id() -> str:
+    return f"chatcmpl-{uuid.uuid4().hex[:12]}"
+
+
+async def extract_images_and_text(messages: List[VisionMessage]) -> tuple[List[Image.Image], List[dict]]:
+    """Extract images and convert messages to processor format."""
+    images = []
+    processed_messages = []
+    
+    for msg in messages:
+        if isinstance(msg.content, str):
+            # Simple text message
+            processed_messages.append({
+                "role": msg.role,
+                "content": msg.content
+            })
+        else:
+            # Mixed content (images + text)
+            content_parts = []
+            for item in msg.content:
+                if isinstance(item, dict):
+                    item_type = item.get("type", "")
+                else:
+                    item_type = item.type
+                
+                if item_type == "image":
+                    image = await process_image_content(item)
+                    if image:
+                        images.append(image)
+                        content_parts.append({"type": "image"})
+                elif item_type == "text":
+                    text = item.get("text", "") if isinstance(item, dict) else item.text
+                    content_parts.append({"type": "text", "text": text})
+            
+            processed_messages.append({
+                "role": msg.role,
+                "content": content_parts
+            })
+    
+    return images, processed_messages
+
+
+async def stream_vision_completion(request: VisionCompletionRequest) -> AsyncGenerator[str, None]:
+    """Streaming vision completion."""
+    request_id = generate_id()
+    created = int(time.time())
+    
+    loop = asyncio.get_running_loop()
+    async_queue = asyncio.Queue()
+    
+    # Extract images and process messages
+    images, processed_messages = await extract_images_and_text(request.messages)
+    
+    tokenizer = model_manager.processor.tokenizer
+    
+    # Config
+    max_tokens = request.max_tokens or settings.max_tokens
+    temperature = request.temperature if request.temperature is not None else settings.temperature
+    top_k = request.top_k if request.top_k is not None else settings.top_k
+    top_p = request.top_p if request.top_p is not None else settings.top_p
+    
+    # Prepare stop tokens
+    stop_tokens = [tokenizer.eos_token_id]
+    if request.stop:
+        if isinstance(request.stop, str):
+            encoded = tokenizer.encode(request.stop, add_special_tokens=False)
+            if encoded:
+                stop_tokens.append(encoded[0])
+        elif isinstance(request.stop, list):
+            for stop_str in request.stop:
+                encoded = tokenizer.encode(stop_str, add_special_tokens=False)
+                if encoded:
+                    stop_tokens.append(encoded[0])
+
+    def generate_tokens():
+        try:
+            for token in model_manager.generate_stream(
+                images,
+                processed_messages,
+                max_tokens=max_tokens,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p,
+                stop_tokens=stop_tokens
+            ):
+                loop.call_soon_threadsafe(async_queue.put_nowait, ("token", token))
+        except Exception as e:
+            logger.error(f"Stream generation error: {e}")
+            loop.call_soon_threadsafe(async_queue.put_nowait, ("error", str(e)))
+        finally:
+            loop.call_soon_threadsafe(async_queue.put_nowait, ("done", None))
+
+    threading.Thread(target=generate_tokens, daemon=True).start()
+    
+    try:
+        while True:
+            msg_type, data = await async_queue.get()
+
+            if msg_type == "token":
+                text = tokenizer.decode([data], skip_special_tokens=True)
+                if text:
+                    chunk = {
+                        "id": request_id,
+                        "object": "chat.completion.chunk",
+                        "created": created,
+                        "model": request.model,
+                        "choices": [{
+                            "index": 0,
+                            "delta": {"content": text},
+                            "finish_reason": None
+                        }]
+                    }
+                    yield {"data": json.dumps(chunk)}
+            
+            elif msg_type == "done":
+                final = {
+                    "id": request_id,
+                    "object": "chat.completion.chunk",
+                    "created": created,
+                    "model": request.model,
+                    "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]
+                }
+                yield {"data": json.dumps(final)}
+                yield {"data": "[DONE]"}
+                break
+                
+            elif msg_type == "error":
+                logger.error(f"Stream error: {data}")
+                yield {"data": json.dumps({"error": {"message": data}})}
+                break
+                
+    except asyncio.CancelledError:
+        logger.info(f"Stream cancelled for request {request_id[:8]}")
+        raise
+    except Exception as e:
+        logger.error(f"Streaming error: {e}")
+        yield {"data": json.dumps({"error": {"message": str(e)}})}
+
+
+# ==============================================================================
+# API Endpoints
+# ==============================================================================
+
+
+
+
+@app.get("/health")
+async def health():
+    if not model_manager.is_loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded")
+    return {"status": "healthy"}
+
+
+@app.get("/v1/models", response_model=ModelListResponse)
+async def list_models():
+    return ModelListResponse(
+        data=[
+            ModelInfo(id="lfm-vision", created=int(time.time())),
+            ModelInfo(id="lfm-2.5-vl-1.6b-onnx", created=int(time.time()))
+        ]
+    )
+
+
+@app.post("/v1/vision/completions")
+async def vision_completions(request: VisionCompletionRequest):
+    """Vision-language completion with image support."""
+    if not model_manager.is_loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded")
+    
+    if request.stream:
+        return EventSourceResponse(
+            stream_vision_completion(request),
+            media_type="text/event-stream",
+            ping=30000,
+            ping_message_factory=lambda: '{"type": "ping"}'
+        )
+    
+    try:
+        # Extract images and process messages
+        images, processed_messages = await extract_images_and_text(request.messages)
+        
+        tokenizer = model_manager.processor.tokenizer
+        
+        max_tokens = request.max_tokens or settings.max_tokens
+        temperature = request.temperature if request.temperature is not None else settings.temperature
+        top_k = request.top_k if request.top_k is not None else settings.top_k
+        top_p = request.top_p if request.top_p is not None else settings.top_p
+        
+        start_time = time.time()
+        
+        loop = asyncio.get_event_loop()
+        tokens = await loop.run_in_executor(
+            None,
+            lambda: model_manager.generate(
+                images,
+                processed_messages,
+                max_tokens=max_tokens,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p
+            )
+        )
+        
+        response_text = tokenizer.decode(tokens, skip_special_tokens=True)
+        gen_time = time.time() - start_time
+        
+        logger.debug(f"Generated {len(tokens)} tokens in {gen_time:.2f}s")
+        
+        return ChatCompletionResponse(
+            id=generate_id(),
+            created=int(time.time()),
+            model=request.model,
+            choices=[
+                ChatCompletionChoice(
+                    index=0,
+                    message=ChatMessage(role="assistant", content=response_text),
+                    finish_reason="stop"
+                )
+            ],
+            usage={
+                "prompt_tokens": 0,  # Would need to track input tokens
+                "completion_tokens": len(tokens),
+                "total_tokens": len(tokens)
+            }
+        )
+        
+    except Exception as e:
+        logger.error(f"Vision completion error: {e}")
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@app.post("/v1/chat/completions")
+async def chat_completions(request: ChatCompletionRequest):
+    """Text-only chat completion (for compatibility)."""
+    if not model_manager.is_loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded")
+    
+    # Convert to vision request format (no images)
+    vision_messages = [
+        VisionMessage(role=m.role, content=m.content)
+        for m in request.messages
+    ]
+    
+    vision_request = VisionCompletionRequest(
+        model=request.model,
+        messages=vision_messages,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        max_tokens=request.max_tokens,
+        stream=request.stream
+    )
+    
+    return await vision_completions(vision_request)
+
+
+@app.post("/v1/vision/upload")
+async def upload_image(
+    file: UploadFile = File(...),
+    prompt: str = "What is in this image?"
+):
+    """Direct image upload endpoint."""
+    if not model_manager.is_loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded")
+    
+    # Validate file type
+    content_type = file.content_type or ""
+    file_ext = Path(file.filename or "").suffix.lower().lstrip(".")
+    
+    if file_ext not in settings.supported_formats and not any(fmt in content_type for fmt in settings.supported_formats):
+        raise HTTPException(
+            status_code=400,
+            detail=f"Unsupported image format. Supported: {settings.supported_formats}"
+        )
+    
+    # Read and process image
+    contents = await file.read()
+    if len(contents) > settings.max_image_size_mb * 1024 * 1024:
+        raise HTTPException(
+            status_code=400,
+            detail=f"Image too large. Max size: {settings.max_image_size_mb}MB"
+        )
+    
+    try:
+        image = Image.open(io.BytesIO(contents))
+    except Exception as e:
+        raise HTTPException(status_code=400, detail=f"Invalid image: {e}")
+    
+    # Create request
+    messages = [{
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": prompt}
+        ]
+    }]
+    
+    tokenizer = model_manager.processor.tokenizer
+    
+    tokens = model_manager.generate(
+        [image],
+        messages,
+        max_tokens=settings.max_tokens,
+        temperature=settings.temperature,
+        top_k=settings.top_k,
+        top_p=settings.top_p
+    )
+    
+    response_text = tokenizer.decode(tokens, skip_special_tokens=True)
+    
+    return {
+        "id": generate_id(),
+        "model": "lfm-vision",
+        "response": response_text
+    }
+
+
+# ==============================================================================
+# Run Server
+# ==============================================================================
+
+if __name__ == "__main__":
+    import uvicorn
+    
+    logger.info(f"Starting server on {settings.host}:{settings.port}")
+    
+    uvicorn.run(
+        "app:app",
+        host=settings.host,
+        port=settings.port,
+        reload=False,
+        log_level=settings.log_level
+    )