From d13495871e768bdad605a4edb83646ac6389b26e Mon Sep 17 00:00:00 2001 From: ai-modelscope Date: Mon, 30 Dec 2024 14:00:46 +0800 Subject: [PATCH] Update AWQ models --- .gitattributes | 35 +- README.md | 610 ++++++++++- added_tokens.json | 22 + config.json | 246 +++++ configuration.json | 1 + configuration_intern_vit.py | 120 +++ configuration_internvl_chat.py | 96 ++ configuration_phi3.py | 211 ++++ conversation.py | 383 +++++++ examples/image1.jpg | Bin 0 -> 78073 bytes examples/image2.jpg | Bin 0 -> 125656 bytes examples/red-panda.mp4 | 3 + generation_config.json | 9 + model-00001-of-00002.safetensors | 3 + model-00002-of-00002.safetensors | 3 + model.safetensors.index.json | 548 ++++++++++ modeling_intern_vit.py | 430 ++++++++ modeling_internvl_chat.py | 349 +++++++ modeling_phi3.py | 1610 ++++++++++++++++++++++++++++++ preprocessor_config.json | 19 + special_tokens_map.json | 41 + tokenizer.model | 3 + tokenizer_config.json | 214 ++++ 23 files changed, 4893 insertions(+), 63 deletions(-) create mode 100644 added_tokens.json create mode 100644 config.json create mode 100644 configuration.json create mode 100644 configuration_intern_vit.py create mode 100644 configuration_internvl_chat.py create mode 100644 configuration_phi3.py create mode 100644 conversation.py create mode 100644 examples/image1.jpg create mode 100644 examples/image2.jpg create mode 100644 examples/red-panda.mp4 create mode 100644 generation_config.json create mode 100644 model-00001-of-00002.safetensors create mode 100644 model-00002-of-00002.safetensors create mode 100644 model.safetensors.index.json create mode 100644 modeling_intern_vit.py create mode 100644 modeling_internvl_chat.py create mode 100644 modeling_phi3.py create mode 100644 preprocessor_config.json create mode 100644 special_tokens_map.json create mode 100644 tokenizer.model create mode 100644 tokenizer_config.json diff --git a/.gitattributes b/.gitattributes index 53d7257..b125d9a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1,47 +1,38 @@ *.7z filter=lfs diff=lfs merge=lfs -text *.arrow filter=lfs diff=lfs merge=lfs -text *.bin filter=lfs diff=lfs merge=lfs -text -*.bin.* filter=lfs diff=lfs merge=lfs -text *.bz2 filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text *.ftz filter=lfs diff=lfs merge=lfs -text *.gz filter=lfs diff=lfs merge=lfs -text *.h5 filter=lfs diff=lfs merge=lfs -text *.joblib filter=lfs diff=lfs merge=lfs -text *.lfs.* filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text *.model filter=lfs diff=lfs merge=lfs -text *.msgpack filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text *.onnx filter=lfs diff=lfs merge=lfs -text *.ot filter=lfs diff=lfs merge=lfs -text *.parquet filter=lfs diff=lfs merge=lfs -text *.pb filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text *.pt filter=lfs diff=lfs merge=lfs -text *.pth filter=lfs diff=lfs merge=lfs -text *.rar filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text saved_model/**/* filter=lfs diff=lfs merge=lfs -text *.tar.* filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text *.tflite filter=lfs diff=lfs merge=lfs -text *.tgz filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text *.xz filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text -*.zstandard filter=lfs diff=lfs merge=lfs -text -*.tfevents* filter=lfs diff=lfs merge=lfs -text -*.db* filter=lfs diff=lfs merge=lfs -text -*.ark* filter=lfs diff=lfs merge=lfs -text -**/*ckpt*data* filter=lfs diff=lfs merge=lfs -text -**/*ckpt*.meta filter=lfs diff=lfs merge=lfs -text -**/*ckpt*.index filter=lfs diff=lfs merge=lfs -text -*.safetensors filter=lfs diff=lfs merge=lfs -text -*.ckpt filter=lfs diff=lfs merge=lfs -text -*.gguf* filter=lfs diff=lfs merge=lfs -text -*.ggml filter=lfs diff=lfs merge=lfs -text -*.llamafile* filter=lfs diff=lfs merge=lfs -text -*.pt2 filter=lfs diff=lfs merge=lfs -text -*.mlmodel filter=lfs diff=lfs merge=lfs -text -*.npy filter=lfs diff=lfs merge=lfs -text -*.npz filter=lfs diff=lfs merge=lfs -text -*.pickle filter=lfs diff=lfs merge=lfs -text -*.pkl filter=lfs diff=lfs merge=lfs -text -*.tar filter=lfs diff=lfs merge=lfs -text -*.wasm filter=lfs diff=lfs merge=lfs -text *.zst filter=lfs diff=lfs merge=lfs -text -*tfevents* filter=lfs diff=lfs merge=lfs -text \ No newline at end of file +*tfevents* filter=lfs diff=lfs merge=lfs -text +examples/red-panda.mp4 filter=lfs diff=lfs merge=lfs -text +model-00001-of-00002.safetensors filter=lfs diff=lfs merge=lfs -text +model-00002-of-00002.safetensors filter=lfs diff=lfs merge=lfs -text diff --git a/README.md b/README.md index dce551e..cdd5c6f 100644 --- a/README.md +++ b/README.md @@ -1,47 +1,575 @@ --- -license: Apache License 2.0 - -#model-type: -##如 gpt、phi、llama、chatglm、baichuan 等 -#- gpt - -#domain: -##如 nlp、cv、audio、multi-modal -#- nlp - -#language: -##语言代码列表 https://help.aliyun.com/document_detail/215387.html?spm=a2c4g.11186623.0.0.9f8d7467kni6Aa -#- cn - -#metrics: -##如 CIDEr、Blue、ROUGE 等 -#- CIDEr - -#tags: -##各种自定义，包括 pretrained、fine-tuned、instruction-tuned、RL-tuned 等训练方法和其他 -#- pretrained - -#tools: -##如 vllm、fastchat、llamacpp、AdaSeq 等 -#- vllm +license: mit +pipeline_tag: image-text-to-text +library_name: transformers +base_model: + - OpenGVLab/InternViT-300M-448px + - microsoft/Phi-3-mini-128k-instruct +new_version: OpenGVLab/InternVL2_5-4B +base_model_relation: merge +language: + - multilingual +tags: + - internvl + - custom_code --- -### 当前模型的贡献者未提供更加详细的模型介绍。模型文件和权重，可浏览“模型文件”页面获取。 -#### 您可以通过如下git clone命令，或者ModelScope SDK来下载模型 -SDK下载 -```bash -#安装ModelScope -pip install modelscope -``` +# Mini-InternVL-Chat-4B-V1-5 + +[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[📜 InternVL 1.0\]](https://huggingface.co/papers/2312.14238) [\[📜 InternVL 1.5\]](https://huggingface.co/papers/2404.16821) [\[📜 Mini-InternVL\]](https://arxiv.org/abs/2410.16261) [\[📜 InternVL 2.5\]](https://huggingface.co/papers/2412.05271) + +[\[🆕 Blog\]](https://internvl.github.io/blog/) [\[🗨️ Chat Demo\]](https://internvl.opengvlab.com/) [\[🤗 HF Demo\]](https://huggingface.co/spaces/OpenGVLab/InternVL) [\[🚀 Quick Start\]](#quick-start) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/) + +## Introduction + +

+ image/png +

+ +You can run multimodal large models using a 1080Ti now. + +We are delighted to introduce the Mini-InternVL-Chat series. In the era of large language models, many researchers have started to focus on smaller language models, such as Gemma-2B, Qwen-1.8B, and InternLM2-1.8B. Inspired by their efforts, we have distilled our vision foundation model [InternViT-6B-448px-V1-5](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-5) down to 300M and used [InternLM2-Chat-1.8B](https://huggingface.co/internlm/internlm2-chat-1_8b) or [Phi-3-mini-128k-instruct](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct) as our language model. This resulted in a small multimodal model with excellent performance. + +As shown in the figure below, we adopted the same model architecture as InternVL 1.5. We simply replaced the original InternViT-6B with InternViT-300M and InternLM2-Chat-20B with InternLM2-Chat-1.8B / Phi-3-mini-128k-instruct. For training, we used the same data as InternVL 1.5 to train this smaller model. Additionally, due to the lower training costs of smaller models, we used a context length of 8K during training. + +![image/png](https://cdn-uploads.huggingface.co/production/uploads/64006c09330a45b03605bba3/rDyoe66Sqev44T0wsP5Z7.png) + +## Model Details + +- **Model Type:** multimodal large language model (MLLM) + +- **Model Stats:** + + - Architecture: [InternViT-300M-448px](https://huggingface.co/OpenGVLab/InternViT-300M-448px) + MLP + [Phi-3-mini-128k-instruct](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct) + - Image size: dynamic resolution, max to 40 tiles of 448 x 448 (4K resolution). + - Params: 4.2B + +- **Training Strategy:** + + - Learnable component in the pre-training stage: MLP + - Learnable component in the fine-tuning stage: ViT + MLP + LLM + - For more details on training hyperparameters, please see our [blog](https://internvl.github.io/blog/2024-05-25-Mini-InternVL-1.5/). + +## Performance + +![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/ngl8oZvNrjItWtLUQqB2V.png) + +- We simultaneously use [InternVL](https://github.com/OpenGVLab/InternVL) and [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) repositories for model evaluation. Specifically, the results reported for DocVQA, ChartQA, InfoVQA, TextVQA, MME, AI2D, MMBench, CCBench, MMVet, and SEED-Image were tested using the InternVL repository. OCRBench, RealWorldQA, HallBench, and MathVista were evaluated using the VLMEvalKit. + +Limitations: Although we have made efforts to ensure the safety of the model during the training process and to encourage the model to generate text that complies with ethical and legal requirements, the model may still produce unexpected outputs due to its size and probabilistic generation paradigm. For example, the generated responses may contain biases, discrimination, or other harmful content. Please do not propagate such content. We are not responsible for any consequences resulting from the dissemination of harmful information. + +## Quick Start + +We provide an example code to run Mini-InternVL-Chat-4B-V1-5 using `transformers`. + +> Please use transformers>=4.37.2 to ensure the model works normally. + +### Model Loading + +#### 16-bit (bf16 / fp16) + ```python -#SDK模型下载 -from modelscope import snapshot_download -model_dir = snapshot_download('Shanghai_AI_Laboratory/Mini-InternVL-Chat-4B-V1-5') -``` -Git下载 -``` -#Git模型下载 -git clone https://www.modelscope.cn/Shanghai_AI_Laboratory/Mini-InternVL-Chat-4B-V1-5.git +import torch +from transformers import AutoTokenizer, AutoModel +path = "OpenGVLab/Mini-InternVL-Chat-4B-V1-5" +model = AutoModel.from_pretrained( + path, + torch_dtype=torch.bfloat16, + low_cpu_mem_usage=True, + use_flash_attn=True, + trust_remote_code=True).eval().cuda() ``` -

如果您是本模型的贡献者，我们邀请您根据模型贡献文档，及时完善模型卡片内容。

\ No newline at end of file +#### BNB 8-bit Quantization + +```python +import torch +from transformers import AutoTokenizer, AutoModel +path = "OpenGVLab/Mini-InternVL-Chat-4B-V1-5" +model = AutoModel.from_pretrained( + path, + torch_dtype=torch.bfloat16, + load_in_8bit=True, + low_cpu_mem_usage=True, + use_flash_attn=True, + trust_remote_code=True).eval() +``` + +#### BNB 4-bit Quantization + +```python +import torch +from transformers import AutoTokenizer, AutoModel +path = "OpenGVLab/Mini-InternVL-Chat-4B-V1-5" +model = AutoModel.from_pretrained( + path, + torch_dtype=torch.bfloat16, + load_in_4bit=True, + low_cpu_mem_usage=True, + use_flash_attn=True, + trust_remote_code=True).eval() +``` + +#### Multiple GPUs + +The reason for writing the code this way is to avoid errors that occur during multi-GPU inference due to tensors not being on the same device. By ensuring that the first and last layers of the large language model (LLM) are on the same device, we prevent such errors. + +```python +import math +import torch +from transformers import AutoTokenizer, AutoModel + +def split_model(model_name): + device_map = {} + world_size = torch.cuda.device_count() + num_layers = {'Mini-InternVL-2B-V1-5': 24, 'Mini-InternVL-4B-V1-5': 32, 'InternVL-Chat-V1-5': 48}[model_name] + # Since the first GPU will be used for ViT, treat it as half a GPU. + num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5)) + num_layers_per_gpu = [num_layers_per_gpu] * world_size + num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5) + layer_cnt = 0 + for i, num_layer in enumerate(num_layers_per_gpu): + for j in range(num_layer): + device_map[f'language_model.model.layers.{layer_cnt}'] = i + layer_cnt += 1 + device_map['vision_model'] = 0 + device_map['mlp1'] = 0 + device_map['language_model.model.tok_embeddings'] = 0 + device_map['language_model.model.embed_tokens'] = 0 + device_map['language_model.output'] = 0 + device_map['language_model.model.norm'] = 0 + device_map['language_model.lm_head'] = 0 + device_map[f'language_model.model.layers.{num_layers - 1}'] = 0 + + return device_map + +path = "OpenGVLab/Mini-InternVL-Chat-4B-V1-5" +device_map = split_model('Mini-InternVL-Chat-4B-V1-5') +model = AutoModel.from_pretrained( + path, + torch_dtype=torch.bfloat16, + low_cpu_mem_usage=True, + use_flash_attn=True, + trust_remote_code=True, + device_map=device_map).eval() +``` + +### Inference with Transformers + +```python +import numpy as np +import torch +import torchvision.transforms as T +from decord import VideoReader, cpu +from PIL import Image +from torchvision.transforms.functional import InterpolationMode +from transformers import AutoModel, AutoTokenizer + +IMAGENET_MEAN = (0.485, 0.456, 0.406) +IMAGENET_STD = (0.229, 0.224, 0.225) + +def build_transform(input_size): + MEAN, STD = IMAGENET_MEAN, IMAGENET_STD + transform = T.Compose([ + T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img), + T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC), + T.ToTensor(), + T.Normalize(mean=MEAN, std=STD) + ]) + return transform + +def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size): + best_ratio_diff = float('inf') + best_ratio = (1, 1) + area = width * height + for ratio in target_ratios: + target_aspect_ratio = ratio[0] / ratio[1] + ratio_diff = abs(aspect_ratio - target_aspect_ratio) + if ratio_diff < best_ratio_diff: + best_ratio_diff = ratio_diff + best_ratio = ratio + elif ratio_diff == best_ratio_diff: + if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]: + best_ratio = ratio + return best_ratio + +def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False): + orig_width, orig_height = image.size + aspect_ratio = orig_width / orig_height + + # calculate the existing image aspect ratio + target_ratios = set( + (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if + i * j <= max_num and i * j >= min_num) + target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) + + # find the closest aspect ratio to the target + target_aspect_ratio = find_closest_aspect_ratio( + aspect_ratio, target_ratios, orig_width, orig_height, image_size) + + # calculate the target width and height + target_width = image_size * target_aspect_ratio[0] + target_height = image_size * target_aspect_ratio[1] + blocks = target_aspect_ratio[0] * target_aspect_ratio[1] + + # resize the image + resized_img = image.resize((target_width, target_height)) + processed_images = [] + for i in range(blocks): + box = ( + (i % (target_width // image_size)) * image_size, + (i // (target_width // image_size)) * image_size, + ((i % (target_width // image_size)) + 1) * image_size, + ((i // (target_width // image_size)) + 1) * image_size + ) + # split the image + split_img = resized_img.crop(box) + processed_images.append(split_img) + assert len(processed_images) == blocks + if use_thumbnail and len(processed_images) != 1: + thumbnail_img = image.resize((image_size, image_size)) + processed_images.append(thumbnail_img) + return processed_images + +def load_image(image_file, input_size=448, max_num=12): + image = Image.open(image_file).convert('RGB') + transform = build_transform(input_size=input_size) + images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num) + pixel_values = [transform(image) for image in images] + pixel_values = torch.stack(pixel_values) + return pixel_values + +# If you want to load a model using multiple GPUs, please refer to the `Multiple GPUs` section. +path = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5' +model = AutoModel.from_pretrained( + path, + torch_dtype=torch.bfloat16, + low_cpu_mem_usage=True, + use_flash_attn=True, + trust_remote_code=True).eval().cuda() +tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False) + +# set the max number of tiles in `max_num` +pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda() +generation_config = dict(max_new_tokens=1024, do_sample=True) + +# pure-text conversation (纯文本对话) +question = 'Hello, who are you?' +response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +question = 'Can you tell me a story?' +response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +# single-image single-round conversation (单图单轮对话) +question = '\nPlease describe the image shortly.' +response = model.chat(tokenizer, pixel_values, question, generation_config) +print(f'User: {question}\nAssistant: {response}') + +# single-image multi-round conversation (单图多轮对话) +question = '\nPlease describe the image in detail.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +question = 'Please write a poem according to the image.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像) +pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda() +pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda() +pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0) + +question = '\nDescribe the two images in detail.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, + history=None, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +question = 'What are the similarities and differences between these two images.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, + history=history, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +# multi-image multi-round conversation, separate images (多图多轮对话，独立图像) +pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda() +pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda() +pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0) +num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)] + +question = 'Image-1: \nImage-2: \nDescribe the two images in detail.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, + num_patches_list=num_patches_list, + history=None, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +question = 'What are the similarities and differences between these two images.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, + num_patches_list=num_patches_list, + history=history, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +# batch inference, single image per sample (单图批处理) +pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda() +pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda() +num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)] +pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0) + +questions = ['\nDescribe the image in detail.'] * len(num_patches_list) +responses = model.batch_chat(tokenizer, pixel_values, + num_patches_list=num_patches_list, + questions=questions, + generation_config=generation_config) +for question, response in zip(questions, responses): + print(f'User: {question}\nAssistant: {response}') + +# video multi-round conversation (视频多轮对话) +def get_index(bound, fps, max_frame, first_idx=0, num_segments=32): + if bound: + start, end = bound[0], bound[1] + else: + start, end = -100000, 100000 + start_idx = max(first_idx, round(start * fps)) + end_idx = min(round(end * fps), max_frame) + seg_size = float(end_idx - start_idx) / num_segments + frame_indices = np.array([ + int(start_idx + (seg_size / 2) + np.round(seg_size * idx)) + for idx in range(num_segments) + ]) + return frame_indices + +def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32): + vr = VideoReader(video_path, ctx=cpu(0), num_threads=1) + max_frame = len(vr) - 1 + fps = float(vr.get_avg_fps()) + + pixel_values_list, num_patches_list = [], [] + transform = build_transform(input_size=input_size) + frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments) + for frame_index in frame_indices: + img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB') + img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num) + pixel_values = [transform(tile) for tile in img] + pixel_values = torch.stack(pixel_values) + num_patches_list.append(pixel_values.shape[0]) + pixel_values_list.append(pixel_values) + pixel_values = torch.cat(pixel_values_list) + return pixel_values, num_patches_list + +video_path = './examples/red-panda.mp4' +pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1) +pixel_values = pixel_values.to(torch.bfloat16).cuda() +video_prefix = ''.join([f'Frame{i+1}: \n' for i in range(len(num_patches_list))]) +question = video_prefix + 'What is the red panda doing?' +# Frame1: \nFrame2: \n...\nFrame8: \n{question} +response, history = model.chat(tokenizer, pixel_values, question, generation_config, + num_patches_list=num_patches_list, history=None, return_history=True) +print(f'User: {question}\nAssistant: {response}') + +question = 'Describe this video in detail.' +response, history = model.chat(tokenizer, pixel_values, question, generation_config, + num_patches_list=num_patches_list, history=history, return_history=True) +print(f'User: {question}\nAssistant: {response}') +``` + +#### Streaming Output + +Besides this method, you can also use the following code to get streamed output. + +```python +from transformers import TextIteratorStreamer +from threading import Thread + +# Initialize the streamer +streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10) +# Define the generation configuration +generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer) +# Start the model chat in a separate thread +thread = Thread(target=model.chat, kwargs=dict( + tokenizer=tokenizer, pixel_values=pixel_values, question=question, + history=None, return_history=False, generation_config=generation_config, +)) +thread.start() + +# Initialize an empty string to store the generated text +generated_text = '' +# Loop through the streamer to get the new text as it is generated +for new_text in streamer: + if new_text == model.conv_template.sep: + break + generated_text += new_text + print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line +``` + +## Finetune + +Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTurner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning. + +## Deployment + +### LMDeploy + +LMDeploy is a toolkit for compressing, deploying, and serving LLMs & VLMs. + +```sh +pip install lmdeploy>=0.5.3 +``` + +LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline. + +#### A 'Hello, world' Example + +```python +from lmdeploy import pipeline, TurbomindEngineConfig +from lmdeploy.vl import load_image + +model = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5' +image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg') +pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192)) +response = pipe(('describe this image', image)) +print(response.text) +``` + +If `ImportError` occurs while executing this case, please install the required dependency packages as prompted. + +#### Multi-images Inference + +When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased. + +> Warning: Due to the scarcity of multi-image conversation data, the performance on multi-image tasks may be unstable, and it may require multiple attempts to achieve satisfactory results. + +```python +from lmdeploy import pipeline, TurbomindEngineConfig +from lmdeploy.vl import load_image +from lmdeploy.vl.constants import IMAGE_TOKEN + +model = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5' +pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192)) + +image_urls=[ + 'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg', + 'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg' +] + +images = [load_image(img_url) for img_url in image_urls] +# Numbering images improves multi-image conversations +response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images)) +print(response.text) +``` + +#### Batch Prompts Inference + +Conducting inference with batch prompts is quite straightforward; just place them within a list structure: + +```python +from lmdeploy import pipeline, TurbomindEngineConfig +from lmdeploy.vl import load_image + +model = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5' +pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192)) + +image_urls=[ + "https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg", + "https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg" +] +prompts = [('describe this image', load_image(img_url)) for img_url in image_urls] +response = pipe(prompts) +print(response) +``` + +#### Multi-turn Conversation + +There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface. + +```python +from lmdeploy import pipeline, TurbomindEngineConfig, GenerationConfig +from lmdeploy.vl import load_image + +model = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5' +pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192)) + +image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg') +gen_config = GenerationConfig(top_k=40, top_p=0.8, temperature=0.8) +sess = pipe.chat(('describe this image', image), gen_config=gen_config) +print(sess.response.text) +sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config) +print(sess.response.text) +``` + +#### Service + +LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup: + +```shell +lmdeploy serve api_server OpenGVLab/Mini-InternVL-Chat-4B-V1-5 --server-port 23333 +``` + +To use the OpenAI-style interface, you need to install OpenAI: + +```shell +pip install openai +``` + +Then, use the code below to make the API call: + +```python +from openai import OpenAI + +client = OpenAI(api_key='YOUR_API_KEY', base_url='http://0.0.0.0:23333/v1') +model_name = client.models.list().data[0].id +response = client.chat.completions.create( + model=model_name, + messages=[{ + 'role': + 'user', + 'content': [{ + 'type': 'text', + 'text': 'describe this image', + }, { + 'type': 'image_url', + 'image_url': { + 'url': + 'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg', + }, + }], + }], + temperature=0.8, + top_p=0.8) +print(response) +``` + +## License + +This project is released under the MIT License. This project uses the pre-trained Phi-3-mini-128k-instruct as a component, which is also licensed under the MIT License. + +## Citation + +If you find this project useful in your research, please consider citing: + +```BibTeX +@article{chen2024expanding, + title={Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling}, + author={Chen, Zhe and Wang, Weiyun and Cao, Yue and Liu, Yangzhou and Gao, Zhangwei and Cui, Erfei and Zhu, Jinguo and Ye, Shenglong and Tian, Hao and Liu, Zhaoyang and others}, + journal={arXiv preprint arXiv:2412.05271}, + year={2024} +} +@article{gao2024mini, + title={Mini-internvl: A flexible-transfer pocket multimodal model with 5\% parameters and 90\% performance}, + author={Gao, Zhangwei and Chen, Zhe and Cui, Erfei and Ren, Yiming and Wang, Weiyun and Zhu, Jinguo and Tian, Hao and Ye, Shenglong and He, Junjun and Zhu, Xizhou and others}, + journal={arXiv preprint arXiv:2410.16261}, + year={2024} +} +@article{chen2024far, + title={How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites}, + author={Chen, Zhe and Wang, Weiyun and Tian, Hao and Ye, Shenglong and Gao, Zhangwei and Cui, Erfei and Tong, Wenwen and Hu, Kongzhi and Luo, Jiapeng and Ma, Zheng and others}, + journal={arXiv preprint arXiv:2404.16821}, + year={2024} +} +@inproceedings{chen2024internvl, + title={Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks}, + author={Chen, Zhe and Wu, Jiannan and Wang, Wenhai and Su, Weijie and Chen, Guo and Xing, Sen and Zhong, Muyan and Zhang, Qinglong and Zhu, Xizhou and Lu, Lewei and others}, + booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, + pages={24185--24198}, + year={2024} +} +``` diff --git a/added_tokens.json b/added_tokens.json new file mode 100644 index 0000000..f1a42ff --- /dev/null +++ b/added_tokens.json @@ -0,0 +1,22 @@ +{ + "": 32019, + "": 32012, + "": 32015, + "": 32017, + "": 32013, + "": 32018, + "": 32011, + "": 32014, + "": 32016, + "<|assistant|>": 32001, + "<|endoftext|>": 32000, + "<|end|>": 32007, + "<|placeholder1|>": 32002, + "<|placeholder2|>": 32003, + "<|placeholder3|>": 32004, + "<|placeholder4|>": 32005, + "<|placeholder5|>": 32008, + "<|placeholder6|>": 32009, + "<|system|>": 32006, + "<|user|>": 32010 +} diff --git a/config.json b/config.json new file mode 100644 index 0000000..d5ca75c --- /dev/null +++ b/config.json @@ -0,0 +1,246 @@ +{ + "_commit_hash": null, + "architectures": [ + "InternVLChatModel" + ], + "auto_map": { + "AutoConfig": "configuration_internvl_chat.InternVLChatConfig", + "AutoModel": "modeling_internvl_chat.InternVLChatModel", + "AutoModelForCausalLM": "modeling_internvl_chat.InternVLChatModel" + }, + "system_message": "You are an AI assistant whose name is Phi-3.", + "downsample_ratio": 0.5, + "dynamic_image_size": true, + "force_image_size": 448, + "llm_config": { + "_name_or_path": "microsoft/Phi-3-mini-128k-instruct", + "add_cross_attention": false, + "architectures": [ + "Phi3ForCausalLM" + ], + "attn_implementation": "flash_attention_2", + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_phi3.Phi3Config", + "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM" + }, + "bad_words_ids": null, + "begin_suppress_tokens": null, + "bos_token_id": 1, + "chunk_size_feed_forward": 0, + "cross_attention_hidden_size": null, + "decoder_start_token_id": null, + "diversity_penalty": 0.0, + "do_sample": false, + "early_stopping": false, + "embd_pdrop": 0.0, + "encoder_no_repeat_ngram_size": 0, + "eos_token_id": 32000, + "exponential_decay_length_penalty": null, + "finetuning_task": null, + "forced_bos_token_id": null, + "forced_eos_token_id": null, + "hidden_act": "silu", + "hidden_size": 3072, + "id2label": { + "0": "LABEL_0", + "1": "LABEL_1" + }, + "initializer_range": 0.02, + "intermediate_size": 8192, + "is_decoder": false, + "is_encoder_decoder": false, + "label2id": { + "LABEL_0": 0, + "LABEL_1": 1 + }, + "length_penalty": 1.0, + "max_length": 20, + "max_position_embeddings": 131072, + "min_length": 0, + "model_type": "phi3", + "no_repeat_ngram_size": 0, + "num_attention_heads": 32, + "num_beam_groups": 1, + "num_beams": 1, + "num_hidden_layers": 32, + "num_key_value_heads": 32, + "num_return_sequences": 1, + "original_max_position_embeddings": 4096, + "output_attentions": false, + "output_hidden_states": false, + "output_scores": false, + "pad_token_id": 32000, + "prefix": null, + "problem_type": null, + "pruned_heads": {}, + "remove_invalid_values": false, + "repetition_penalty": 1.0, + "resid_pdrop": 0.0, + "return_dict": true, + "return_dict_in_generate": false, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "long_factor": [ + 1.0299999713897705, + 1.0499999523162842, + 1.0499999523162842, + 1.0799999237060547, + 1.2299998998641968, + 1.2299998998641968, + 1.2999999523162842, + 1.4499999284744263, + 1.5999999046325684, + 1.6499998569488525, + 1.8999998569488525, + 2.859999895095825, + 3.68999981880188, + 5.419999599456787, + 5.489999771118164, + 5.489999771118164, + 9.09000015258789, + 11.579999923706055, + 15.65999984741211, + 15.769999504089355, + 15.789999961853027, + 18.360000610351562, + 21.989999771118164, + 23.079999923706055, + 30.009998321533203, + 32.35000228881836, + 32.590003967285156, + 35.56000518798828, + 39.95000457763672, + 53.840003967285156, + 56.20000457763672, + 57.95000457763672, + 59.29000473022461, + 59.77000427246094, + 59.920005798339844, + 61.190006256103516, + 61.96000671386719, + 62.50000762939453, + 63.3700065612793, + 63.48000717163086, + 63.48000717163086, + 63.66000747680664, + 63.850006103515625, + 64.08000946044922, + 64.760009765625, + 64.80001068115234, + 64.81001281738281, + 64.81001281738281 + ], + "short_factor": [ + 1.05, + 1.05, + 1.05, + 1.1, + 1.1, + 1.1500000000000001, + 1.2000000000000002, + 1.2500000000000002, + 1.3000000000000003, + 1.3500000000000003, + 1.5000000000000004, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.000000000000001, + 2.0500000000000007, + 2.0500000000000007, + 2.0500000000000007, + 2.1000000000000005, + 2.1000000000000005, + 2.1000000000000005, + 2.1500000000000004, + 2.1500000000000004, + 2.3499999999999996, + 2.549999999999999, + 2.5999999999999988, + 2.5999999999999988, + 2.7499999999999982, + 2.849999999999998, + 2.849999999999998, + 2.9499999999999975 + ], + "type": "su" + }, + "rope_theta": 10000.0, + "sep_token_id": null, + "sliding_window": 262144, + "suppress_tokens": null, + "task_specific_params": null, + "temperature": 1.0, + "tf_legacy_loss": false, + "tie_encoder_decoder": false, + "tie_word_embeddings": false, + "tokenizer_class": null, + "top_k": 50, + "top_p": 1.0, + "torch_dtype": "bfloat16", + "torchscript": false, + "transformers_version": "4.37.2", + "typical_p": 1.0, + "use_bfloat16": true, + "use_cache": true, + "vocab_size": 32020 + }, + "max_dynamic_patch": 12, + "min_dynamic_patch": 1, + "model_type": "internvl_chat", + "ps_version": "v2", + "select_layer": -1, + "template": "phi3-chat", + "torch_dtype": "bfloat16", + "use_backbone_lora": 0, + "use_llm_lora": 0, + "use_thumbnail": true, + "vision_config": { + "architectures": [ + "InternVisionModel" + ], + "attention_dropout": 0.0, + "drop_path_rate": 0.0, + "dropout": 0.0, + "hidden_act": "gelu", + "hidden_size": 1024, + "image_size": 448, + "initializer_factor": 1.0, + "initializer_range": 0.02, + "intermediate_size": 4096, + "layer_norm_eps": 1e-06, + "model_type": "intern_vit_6b", + "norm_type": "layer_norm", + "num_attention_heads": 16, + "num_channels": 3, + "num_hidden_layers": 24, + "output_attentions": false, + "output_hidden_states": false, + "patch_size": 14, + "qk_normalization": false, + "qkv_bias": true, + "return_dict": true, + "torch_dtype": "bfloat16", + "transformers_version": "4.37.2", + "use_bfloat16": true, + "use_flash_attn": true + } +} diff --git a/configuration.json b/configuration.json new file mode 100644 index 0000000..4aef15d --- /dev/null +++ b/configuration.json @@ -0,0 +1 @@ +{"framework": "pytorch", "task": "image-text-to-text", "allow_remote": true} \ No newline at end of file diff --git a/configuration_intern_vit.py b/configuration_intern_vit.py new file mode 100644 index 0000000..7e630c4 --- /dev/null +++ b/configuration_intern_vit.py @@ -0,0 +1,120 @@ +# -------------------------------------------------------- +# InternVL +# Copyright (c) 2024 OpenGVLab +# Licensed under The MIT License [see LICENSE for details] +# -------------------------------------------------------- + +import os +from typing import Union + +from transformers.configuration_utils import PretrainedConfig +from transformers.utils import logging + +logger = logging.get_logger(__name__) + + +class InternVisionConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`InternVisionModel`]. It is used to + instantiate a vision encoder according to the specified arguments, defining the model architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + Args: + num_channels (`int`, *optional*, defaults to 3): + Number of color channels in the input images (e.g., 3 for RGB). + patch_size (`int`, *optional*, defaults to 14): + The size (resolution) of each patch. + image_size (`int`, *optional*, defaults to 224): + The size (resolution) of each image. + qkv_bias (`bool`, *optional*, defaults to `False`): + Whether to add a bias to the queries and values in the self-attention layers. + hidden_size (`int`, *optional*, defaults to 3200): + Dimensionality of the encoder layers and the pooler layer. + num_attention_heads (`int`, *optional*, defaults to 25): + Number of attention heads for each attention layer in the Transformer encoder. + intermediate_size (`int`, *optional*, defaults to 12800): + Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder. + qk_normalization (`bool`, *optional*, defaults to `True`): + Whether to normalize the queries and keys in the self-attention layers. + num_hidden_layers (`int`, *optional*, defaults to 48): + Number of hidden layers in the Transformer encoder. + use_flash_attn (`bool`, *optional*, defaults to `True`): + Whether to use flash attention mechanism. + hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`): + The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, + `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported. + layer_norm_eps (`float`, *optional*, defaults to 1e-6): + The epsilon used by the layer normalization layers. + dropout (`float`, *optional*, defaults to 0.0): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + drop_path_rate (`float`, *optional*, defaults to 0.0): + Dropout rate for stochastic depth. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the attention probabilities. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + initializer_factor (`float`, *optional*, defaults to 0.1): + A factor for layer scale. + """ + + model_type = 'intern_vit_6b' + + def __init__( + self, + num_channels=3, + patch_size=14, + image_size=224, + qkv_bias=False, + hidden_size=3200, + num_attention_heads=25, + intermediate_size=12800, + qk_normalization=True, + num_hidden_layers=48, + use_flash_attn=True, + hidden_act='gelu', + norm_type='rms_norm', + layer_norm_eps=1e-6, + dropout=0.0, + drop_path_rate=0.0, + attention_dropout=0.0, + initializer_range=0.02, + initializer_factor=0.1, + **kwargs, + ): + super().__init__(**kwargs) + + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.dropout = dropout + self.drop_path_rate = drop_path_rate + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.num_channels = num_channels + self.patch_size = patch_size + self.image_size = image_size + self.initializer_range = initializer_range + self.initializer_factor = initializer_factor + self.attention_dropout = attention_dropout + self.layer_norm_eps = layer_norm_eps + self.hidden_act = hidden_act + self.norm_type = norm_type + self.qkv_bias = qkv_bias + self.qk_normalization = qk_normalization + self.use_flash_attn = use_flash_attn + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> 'PretrainedConfig': + config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs) + + if 'vision_config' in config_dict: + config_dict = config_dict['vision_config'] + + if 'model_type' in config_dict and hasattr(cls, 'model_type') and config_dict['model_type'] != cls.model_type: + logger.warning( + f"You are using a model of type {config_dict['model_type']} to instantiate a model of type " + f'{cls.model_type}. This is not supported for all configurations of models and can yield errors.' + ) + + return cls.from_dict(config_dict, **kwargs) diff --git a/configuration_internvl_chat.py b/configuration_internvl_chat.py new file mode 100644 index 0000000..9360429 --- /dev/null +++ b/configuration_internvl_chat.py @@ -0,0 +1,96 @@ +# -------------------------------------------------------- +# InternVL +# Copyright (c) 2024 OpenGVLab +# Licensed under The MIT License [see LICENSE for details] +# -------------------------------------------------------- + +import copy + +from transformers import AutoConfig, LlamaConfig +from transformers.configuration_utils import PretrainedConfig +from transformers.utils import logging + +from .configuration_intern_vit import InternVisionConfig +from .configuration_phi3 import Phi3Config + +logger = logging.get_logger(__name__) + + +class InternVLChatConfig(PretrainedConfig): + model_type = 'internvl_chat' + is_composition = True + + def __init__( + self, + vision_config=None, + llm_config=None, + use_backbone_lora=0, + use_llm_lora=0, + select_layer=-1, + force_image_size=None, + downsample_ratio=0.5, + template=None, + dynamic_image_size=False, + use_thumbnail=False, + ps_version='v1', + min_dynamic_patch=1, + max_dynamic_patch=6, + **kwargs): + super().__init__(**kwargs) + + if vision_config is None: + vision_config = {'architectures': ['InternVisionModel']} + logger.info('vision_config is None. Initializing the InternVisionConfig with default values.') + + if llm_config is None: + llm_config = {'architectures': ['Phi3ForCausalLM']} + logger.info('llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).') + + self.vision_config = InternVisionConfig(**vision_config) + if llm_config.get('architectures')[0] == 'LlamaForCausalLM': + self.llm_config = LlamaConfig(**llm_config) + elif llm_config.get('architectures')[0] == 'Phi3ForCausalLM': + self.llm_config = Phi3Config(**llm_config) + else: + raise ValueError('Unsupported architecture: {}'.format(llm_config.get('architectures')[0])) + self.use_backbone_lora = use_backbone_lora + self.use_llm_lora = use_llm_lora + self.select_layer = select_layer + self.force_image_size = force_image_size + self.downsample_ratio = downsample_ratio + self.template = template + self.dynamic_image_size = dynamic_image_size + self.use_thumbnail = use_thumbnail + self.ps_version = ps_version # pixel shuffle version + self.min_dynamic_patch = min_dynamic_patch + self.max_dynamic_patch = max_dynamic_patch + + logger.info(f'vision_select_layer: {self.select_layer}') + logger.info(f'ps_version: {self.ps_version}') + logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}') + logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}') + + def to_dict(self): + """ + Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`]. + + Returns: + `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance, + """ + output = copy.deepcopy(self.__dict__) + output['vision_config'] = self.vision_config.to_dict() + output['llm_config'] = self.llm_config.to_dict() + output['model_type'] = self.__class__.model_type + output['use_backbone_lora'] = self.use_backbone_lora + output['use_llm_lora'] = self.use_llm_lora + output['select_layer'] = self.select_layer + output['force_image_size'] = self.force_image_size + output['downsample_ratio'] = self.downsample_ratio + output['template'] = self.template + output['dynamic_image_size'] = self.dynamic_image_size + output['use_thumbnail'] = self.use_thumbnail + output['ps_version'] = self.ps_version + output['min_dynamic_patch'] = self.min_dynamic_patch + output['max_dynamic_patch'] = self.max_dynamic_patch + + return output diff --git a/configuration_phi3.py b/configuration_phi3.py new file mode 100644 index 0000000..c657051 --- /dev/null +++ b/configuration_phi3.py @@ -0,0 +1,211 @@ +# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License atd +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" Phi-3 model configuration""" + + +from transformers.configuration_utils import PretrainedConfig +from transformers.utils import logging + +logger = logging.get_logger(__name__) + +PHI3_PRETRAINED_CONFIG_ARCHIVE_MAP = { + 'microsoft/Phi-3-mini-4k-instruct': 'https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/config.json', + 'microsoft/Phi-3-mini-128k-instruct': 'https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/config.json', +} + + +class Phi3Config(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3 + model according to the specified arguments, defining the model architecture. Instantiating a configuration with the + defaults will yield a similar configuration to that of the + [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct). + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + Args: + vocab_size (`int`, *optional*, defaults to 32064): + Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`Phi3Model`]. + hidden_size (`int`, *optional*, defaults to 3072): + Dimension of the hidden representations. + intermediate_size (`int`, *optional*, defaults to 8192): + Dimension of the MLP representations. + num_hidden_layers (`int`, *optional*, defaults to 32): + Number of hidden layers in the Transformer decoder. + num_attention_heads (`int`, *optional*, defaults to 32): + Number of attention heads for each attention layer in the Transformer decoder. + num_key_value_heads (`int`, *optional*): + This is the number of key_value heads that should be used to implement Grouped Query Attention. If + `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if + `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When + converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed + by meanpooling all the original heads within that group. For more details checkout [this + paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to + `num_attention_heads`. + resid_pdrop (`float`, *optional*, defaults to 0.0): + Dropout probability for mlp outputs. + embd_pdrop (`int`, *optional*, defaults to 0.0): + The dropout ratio for the embeddings. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio after computing the attention scores. + hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): + The non-linear activation function (function or string) in the decoder. + max_position_embeddings (`int`, *optional*, defaults to 4096): + The maximum sequence length that this model might ever be used with. + original_max_position_embeddings (`int`, *optional*, defaults to 4096): + The maximum sequence length that this model was trained with. This is used to determine the size of the + original RoPE embeddings when using long scaling. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + rms_norm_eps (`float`, *optional*, defaults to 1e-05): + The epsilon value used for the RMSNorm. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). Only + relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not. + tie_word_embeddings (`bool`, *optional*, defaults to `False`): + Whether to tie weight embeddings + rope_theta (`float`, *optional*, defaults to 10000.0): + The base period of the RoPE embeddings. + rope_scaling (`dict`, *optional*): + The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must + contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and + the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size + divided by the number of attention heads divided by 2. + bos_token_id (`int`, *optional*, defaults to 1): + The id of the "beginning-of-sequence" token. + eos_token_id (`int`, *optional*, defaults to 32000): + The id of the "end-of-sequence" token. + pad_token_id (`int`, *optional*, defaults to 32000): + The id of the padding token. + sliding_window (`int`, *optional*): + Sliding window attention window size. If `None`, no sliding window is applied. + + Example: + + ```python + >>> from transformers import Phi3Model, Phi3Config + + >>> # Initializing a Phi-3 style configuration + >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct") + + >>> # Initializing a model from the configuration + >>> model = Phi3Model(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = 'phi3' + keys_to_ignore_at_inference = ['past_key_values'] + + def __init__( + self, + vocab_size=32064, + hidden_size=3072, + intermediate_size=8192, + num_hidden_layers=32, + num_attention_heads=32, + num_key_value_heads=None, + resid_pdrop=0.0, + embd_pdrop=0.0, + attention_dropout=0.0, + hidden_act='silu', + max_position_embeddings=4096, + original_max_position_embeddings=4096, + initializer_range=0.02, + rms_norm_eps=1e-5, + use_cache=True, + tie_word_embeddings=False, + rope_theta=10000.0, + rope_scaling=None, + bos_token_id=1, + eos_token_id=32000, + pad_token_id=32000, + sliding_window=None, + **kwargs, + ): + self.vocab_size = vocab_size + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + + if num_key_value_heads is None: + num_key_value_heads = num_attention_heads + + self.num_key_value_heads = num_key_value_heads + self.resid_pdrop = resid_pdrop + self.embd_pdrop = embd_pdrop + self.attention_dropout = attention_dropout + self.hidden_act = hidden_act + self.max_position_embeddings = max_position_embeddings + self.original_max_position_embeddings = original_max_position_embeddings + self.initializer_range = initializer_range + self.rms_norm_eps = rms_norm_eps + self.use_cache = use_cache + self.rope_theta = rope_theta + self.rope_scaling = rope_scaling + self._rope_scaling_validation() + self.sliding_window = sliding_window + + super().__init__( + bos_token_id=bos_token_id, + eos_token_id=eos_token_id, + pad_token_id=pad_token_id, + tie_word_embeddings=tie_word_embeddings, + **kwargs, + ) + + def _rope_scaling_validation(self): + """ + Validate the `rope_scaling` configuration. + """ + if self.rope_scaling is None: + return + + if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3: + raise ValueError( + '`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, ' + f'got {self.rope_scaling}' + ) + rope_scaling_type = self.rope_scaling.get('type', None) + rope_scaling_short_factor = self.rope_scaling.get('short_factor', None) + rope_scaling_long_factor = self.rope_scaling.get('long_factor', None) + if rope_scaling_type is None or rope_scaling_type not in ['su', 'yarn']: + raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}") + if not ( + isinstance(rope_scaling_short_factor, list) + and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor) + ): + raise ValueError( + f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}" + ) + if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2: + raise ValueError( + f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}" + ) + if not ( + isinstance(rope_scaling_long_factor, list) + and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor) + ): + raise ValueError( + f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}" + ) + if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2: + raise ValueError( + f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}" + ) diff --git a/conversation.py b/conversation.py new file mode 100644 index 0000000..24b3bd7 --- /dev/null +++ b/conversation.py @@ -0,0 +1,383 @@ +""" +Conversation prompt templates. + +We kindly request that you import fastchat instead of copying this file if you wish to use it. +If you have changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates. +""" + +import dataclasses +from enum import IntEnum, auto +from typing import Any, Dict, List, Tuple, Union + + +class SeparatorStyle(IntEnum): + """Separator styles.""" + + ADD_COLON_SINGLE = auto() + ADD_COLON_TWO = auto() + ADD_COLON_SPACE_SINGLE = auto() + NO_COLON_SINGLE = auto() + NO_COLON_TWO = auto() + ADD_NEW_LINE_SINGLE = auto() + LLAMA2 = auto() + CHATGLM = auto() + CHATML = auto() + CHATINTERN = auto() + DOLLY = auto() + RWKV = auto() + PHOENIX = auto() + ROBIN = auto() + FALCON_CHAT = auto() + CHATGLM3 = auto() + INTERNVL_ZH = auto() + MPT = auto() + + +@dataclasses.dataclass +class Conversation: + """A class that manages prompt templates and keeps all conversation history.""" + + # The name of this template + name: str + # The template of the system prompt + system_template: str = '{system_message}' + # The system message + system_message: str = '' + # The names of two roles + roles: Tuple[str] = ('USER', 'ASSISTANT') + # All messages. Each item is (role, message). + messages: List[List[str]] = () + # The number of few shot examples + offset: int = 0 + # The separator style and configurations + sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE + sep: str = '\n' + sep2: str = None + # Stop criteria (the default one is EOS token) + stop_str: Union[str, List[str]] = None + # Stops generation if meeting any token in this list + stop_token_ids: List[int] = None + + def get_prompt(self) -> str: + """Get the prompt for generation.""" + system_prompt = self.system_template.format(system_message=self.system_message) + if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE: + ret = system_prompt + self.sep + for role, message in self.messages: + if message: + ret += role + ': ' + message + self.sep + else: + ret += role + ':' + return ret + elif self.sep_style == SeparatorStyle.ADD_COLON_TWO: + seps = [self.sep, self.sep2] + ret = system_prompt + seps[0] + for i, (role, message) in enumerate(self.messages): + if message: + ret += role + ': ' + message + seps[i % 2] + else: + ret += role + ':' + return ret + elif self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE: + ret = system_prompt + self.sep + for role, message in self.messages: + if message: + ret += role + ': ' + message + self.sep + else: + ret += role + ': ' # must be end with a space + return ret + elif self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE: + ret = '' if system_prompt == '' else system_prompt + self.sep + for role, message in self.messages: + if message: + ret += role + '\n' + message + self.sep + else: + ret += role + '\n' + return ret + elif self.sep_style == SeparatorStyle.NO_COLON_SINGLE: + ret = system_prompt + for role, message in self.messages: + if message: + ret += role + message + self.sep + else: + ret += role + return ret + elif self.sep_style == SeparatorStyle.NO_COLON_TWO: + seps = [self.sep, self.sep2] + ret = system_prompt + for i, (role, message) in enumerate(self.messages): + if message: + ret += role + message + seps[i % 2] + else: + ret += role + return ret + elif self.sep_style == SeparatorStyle.RWKV: + ret = system_prompt + for i, (role, message) in enumerate(self.messages): + if message: + ret += ( + role + + ': ' + + message.replace('\r\n', '\n').replace('\n\n', '\n') + ) + ret += '\n\n' + else: + ret += role + ':' + return ret + elif self.sep_style == SeparatorStyle.LLAMA2: + seps = [self.sep, self.sep2] + if self.system_message: + ret = system_prompt + else: + ret = '[INST] ' + for i, (role, message) in enumerate(self.messages): + tag = self.roles[i % 2] + if message: + if i == 0: + ret += message + ' ' + else: + ret += tag + ' ' + message + seps[i % 2] + else: + ret += tag + return ret + elif self.sep_style == SeparatorStyle.CHATGLM: + # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308 + # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926 + round_add_n = 1 if self.name == 'chatglm2' else 0 + if system_prompt: + ret = system_prompt + self.sep + else: + ret = '' + + for i, (role, message) in enumerate(self.messages): + if i % 2 == 0: + ret += f'[Round {i//2 + round_add_n}]{self.sep}' + + if message: + ret += f'{role}：{message}{self.sep}' + else: + ret += f'{role}：' + return ret + elif self.sep_style == SeparatorStyle.CHATML: + ret = '' if system_prompt == '' else system_prompt + self.sep + '\n' + for role, message in self.messages: + if message: + ret += role + '\n' + message + self.sep + '\n' + else: + ret += role + '\n' + return ret + elif self.sep_style == SeparatorStyle.CHATGLM3: + ret = '' + if self.system_message: + ret += system_prompt + for role, message in self.messages: + if message: + ret += role + '\n' + ' ' + message + else: + ret += role + return ret + elif self.sep_style == SeparatorStyle.CHATINTERN: + # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771 + seps = [self.sep, self.sep2] + ret = system_prompt + for i, (role, message) in enumerate(self.messages): + # if i % 2 == 0: + # ret += "" + if message: + ret += role + ':' + message + seps[i % 2] + '\n' + else: + ret += role + ':' + return ret + elif self.sep_style == SeparatorStyle.DOLLY: + seps = [self.sep, self.sep2] + ret = system_prompt + for i, (role, message) in enumerate(self.messages): + if message: + ret += role + ':\n' + message + seps[i % 2] + if i % 2 == 1: + ret += '\n\n' + else: + ret += role + ':\n' + return ret + elif self.sep_style == SeparatorStyle.PHOENIX: + ret = system_prompt + for role, message in self.messages: + if message: + ret += role + ': ' + '' + message + '' + else: + ret += role + ': ' + '' + return ret + elif self.sep_style == SeparatorStyle.ROBIN: + ret = system_prompt + self.sep + for role, message in self.messages: + if message: + ret += role + ':\n' + message + self.sep + else: + ret += role + ':\n' + return ret + elif self.sep_style == SeparatorStyle.FALCON_CHAT: + ret = '' + if self.system_message: + ret += system_prompt + self.sep + for role, message in self.messages: + if message: + ret += role + ': ' + message + self.sep + else: + ret += role + ':' + + return ret + elif self.sep_style == SeparatorStyle.INTERNVL_ZH: + seps = [self.sep, self.sep2] + ret = self.system_message + seps[0] + for i, (role, message) in enumerate(self.messages): + if message: + ret += role + ': ' + message + seps[i % 2] + else: + ret += role + ':' + return ret + elif self.sep_style == SeparatorStyle.MPT: + ret = system_prompt + self.sep + for role, message in self.messages: + if message: + if type(message) is tuple: + message, _, _ = message + ret += role + message + self.sep + else: + ret += role + return ret + else: + raise ValueError(f'Invalid style: {self.sep_style}') + + def set_system_message(self, system_message: str): + """Set the system message.""" + self.system_message = system_message + + def append_message(self, role: str, message: str): + """Append a new message.""" + self.messages.append([role, message]) + + def update_last_message(self, message: str): + """Update the last output. + + The last message is typically set to be None when constructing the prompt, + so we need to update it in-place after getting the response from a model. + """ + self.messages[-1][1] = message + + def to_gradio_chatbot(self): + """Convert the conversation to gradio chatbot format.""" + ret = [] + for i, (role, msg) in enumerate(self.messages[self.offset :]): + if i % 2 == 0: + ret.append([msg, None]) + else: + ret[-1][-1] = msg + return ret + + def to_openai_api_messages(self): + """Convert the conversation to OpenAI chat completion format.""" + ret = [{'role': 'system', 'content': self.system_message}] + + for i, (_, msg) in enumerate(self.messages[self.offset :]): + if i % 2 == 0: + ret.append({'role': 'user', 'content': msg}) + else: + if msg is not None: + ret.append({'role': 'assistant', 'content': msg}) + return ret + + def copy(self): + return Conversation( + name=self.name, + system_template=self.system_template, + system_message=self.system_message, + roles=self.roles, + messages=[[x, y] for x, y in self.messages], + offset=self.offset, + sep_style=self.sep_style, + sep=self.sep, + sep2=self.sep2, + stop_str=self.stop_str, + stop_token_ids=self.stop_token_ids, + ) + + def dict(self): + return { + 'template_name': self.name, + 'system_message': self.system_message, + 'roles': self.roles, + 'messages': self.messages, + 'offset': self.offset, + } + + +# A global registry for all conversation templates +conv_templates: Dict[str, Conversation] = {} + + +def register_conv_template(template: Conversation, override: bool = False): + """Register a new conversation template.""" + if not override: + assert ( + template.name not in conv_templates + ), f'{template.name} has been registered.' + + conv_templates[template.name] = template + + +def get_conv_template(name: str) -> Conversation: + """Get a conversation template.""" + return conv_templates[name].copy() + + +register_conv_template( + Conversation( + name='Hermes-2', + system_template='<|im_start|>system\n{system_message}', + system_message='Answer the questions.', + roles=('<|im_start|>user\n', '<|im_start|>assistant\n'), + sep_style=SeparatorStyle.MPT, + sep='<|im_end|>', + stop_token_ids=[ + 2, + 6, + 7, + 8, + ], + stop_str='<|endoftext|>', + ) +) + + +register_conv_template( + Conversation( + name='internlm2-chat', + system_template='<|im_start|>system\n{system_message}', + system_message='You are an AI assistant whose name is InternLM (书生·浦语).', + roles=('<|im_start|>user\n', '<|im_start|>assistant\n'), + sep_style=SeparatorStyle.MPT, + sep='<|im_end|>', + stop_token_ids=[ + 2, + 92543, + 92542 + ] + ) +) + + +register_conv_template( + Conversation( + name='phi3-chat', + system_template='<|system|>\n{system_message}', + system_message='You are an AI assistant whose name is Phi-3.', + roles=('<|user|>\n', '<|assistant|>\n'), + sep_style=SeparatorStyle.MPT, + sep='<|end|>', + stop_token_ids=[ + 2, + 32000, + 32007 + ] + ) +) diff --git a/examples/image1.jpg b/examples/image1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fd9891ef7e00774157a9dcd726b2ea9fa0c5ecff GIT binary patch literal 78073 zcma&NcUTio)HX~n5^Crgsz#(HgpTwkRp}k6p@Vb?MSAZwp#+fLyA%=WHPlc91(jYD zDFFddUw+T?Uhf~@U*EZ|OlEgy&YYb&bKhroXa0TscZ|oWVDD+`fM*|$hlhtxh(|+6 zKuB_z{eSZB6J9hPfQ*ccoD4ut4xpvDM?uR#4FFIxFav=MKp-#ls^Y z#J`jOXDuNSJ^?Wv2`SlsXYb_yUAVhB0zx8U8ax650(_#ogz$f=-krrGq$Q%G;RF(M zK)FaD^a{e<@(i$tq#}&8cl|*?L`ZxmMECHC|Iu7(twWy^$|NH$8M&=yAMRvZ=wI ztwDVtu~7;0#)0?B$L}gFe|SIc?TzgRn++6eif-C1=`EsMdV@-x3rPyN*HtbG0fs*% zP3-$_2I%xFFC4|D5ATbd8(U;C@Te!XwMD#Ni5a!GNR>?W3$#Tn9ymOFbGQG9X~$|XbE<~h|}u(;4|q_{<92g{StzhWUJ zK6kVB&NnqH)CS7@v!y1E^A7oG6oMvcreI`~S9VvAYm7!dkws!TUU27T0KitRsO&rL zQf!0LZ;8JmJ_dOl->^wEr$whmmtL#b#%m!EWxZ{jfzkk~0^6&qdwK?2+1|9cywI2> zERC;7Z*a_SfAp+@cphnAeOjTQuczRYN=OiQ^sKp^wR+5qOJP}~#cr~ON`Y@9F&2L2 zT~=Pw_MSiDqCl-qhSPkKs`BjGS9CV@!58@liRulnWlGI<^#cY5?zwR8Y@>4DH{sU? zJF$|#S4|bWtsNOV`#~7;UX6KjPS9-=Ap!_}DSb5hsT;9jE?D^ma{l(QF=J0x7}CjT zqe`qX+VJs<>2lWF&0CL?y(uwN-%6YQ>`&{6eB9eZWmYAKI>Ykml8{LG+Xk1;d|#$G z&UmWkTRvrsp}Ou8Ip6ZAV9IC-@X6s5a_$yd&U-3>ZZ%)H4g9EeN8-{?g87zZ>?zc$ zkEk+K`R%*s?WTA1-foG^U5#o*#oJkWoyS& zf7x1NcqHGP+)z7Qn+xkxE(JxMcs!dhNm8lU&o``mD`Ydru{+ldr#doWak36EAsuPe z`-A^PJe$y7*Yr_h?LOWbjufHG+N#o0y5M~@JTFMFKd~b7YF5Jv5MvpQ8klC6z4ZU; z*HE)dv#t$|-!GZnK~MV*2kUJ<<`KD1BN^+RRvOq~_0-Q>k8fs{?TZxKyqqe@!{8#F zmz)kcR12T_ynaeMLZ+roaDvNfhU)AeN7jD2Rpeaf2j8C)DA;1rlSPX%3XE(J`F|Qz z{YdL6{{`*fzhWD5-JG|7Q#`S(d~SF+t#wh6+TH-3v)X#7W!0_roT6fhULhLjQPT*W#|D1KRX%p8nx_U~yxR(Jri?zJM8wQA-l36cB)-pK{2F(U@;^MHRK# z&A6!TL%W^2q7p4f-rupKFxI)fVvb|h(?qf8xoDczj?&$W_xJ4`9`I+JUp>>Qb$_+- zd+Bn|arm9VhZKpmoDGBJvALxMA1;I_ZIb5sk&_3aW!D^+aNXWeXn}o2_1jzau}{cA zGxYl<3#CF+uDLc~o{qB)=lgzf*ORkVSVfZ$KFeV?{mrGtuwuH?#xP?(q%CJr=8eYQS?1QCh7;T}WLgwZ|&yHyh z9M#+)YQ518^HMOZRV?$!t~K*Lfck!}d}F<$XS#Bpzc^=sz)1bz&JAg*Dir&{wzFMe zTDYUPFZaomcrkTtV)#Nr*d|RUa0`0IE?IG4wf5Z|_wzY~(%LJrPp0&jyBw>Eul_*= z2(!A!4a9!+Q@Bl?ZMy&S;Xgd{*eMEI=_0Gmqtmx;tAxL_J@mVJS@IlLzq|R6XJj3n zCWrY2wOBot{w7RRzczl~j7p+`G{fXEp~?IlhqTYj8dnOdoDNS~z0_fx<7lbZ<=iXX zNYfmcT~x)^HjyYyP|rc~2Uf?hO4jt(eVIQ}#0BSjHp0W%Pd^12vs+5f`#%swz*Rd6 zM&DuVGXGs_-DDGomb9E%nWkHd@#Q$ByQR4-Nw zLKZY8AYa04%+sFMU58wY^Nd0!m-vUrV-Wsf`FW-r#gjL)f?^z_*TJ4b`V+sUf_~b4 zm67|9)-b+C`~k8gSSKY|`cg(%;Zf4S@(!z~;HSna4gZ-=HRB|g@{3^ja*|AC75d`) z?d;^A4|{pC;QJ(@^!TLtT(RR}73W~v;Q(#m^7V6tuJJVK{^-xO+akFPl`l-%6ATN$ zsvY;@;=h{y+}X|3XV)JY{dn^?;7wkZ-@(8_rC@N=C##57J)b#L#YAfF+=_SjW|6pa zOvw`6^?a$IbGKfult_=IMy2bMqRKQY*CLjKDZSP0w!n|7-JZ$?X4?a~`ctr;uvnd=3T&G@-bPZ<4om4iq)4jkWnDZXw^_H9P&mReAP@L; znG=CAUT;rY!Ef2dvKYsg^@D*Ib1Mjg_Pw=dy?btop-*-=r{#uJMXtG+TGPc!6hyv8 zk?ADrlYMr4#;k`DrnYmkyWH+13hjD3W7OEZQc;PSKN8LLvX?4lO;X2>CsWA6z;y&R7H|@Jz_vNH7eifGq@8tJB zcdUFq`?uoT?~@aaeMh?vzQDac9?yJRsvTQgONyJ%k!*C2sWK9bx%d!Q~UfkGq6 zQuw@?HtBtR=|UJW6OVKKdjT?vtu~u`-JgtAujlVid&@spVT_xIoy=J(^&XmRFL5zW zPwd6h6~-^2)7^+PGM2cY!K_XH7!SjeF5|wHOy=z|N0>a4CpQJjH)Kg%_atvi!fWOo zJ7K$RaVy1W!)Vd5d3M1+!fnhQJE4tvX`ZGF!`0`FsQ1gU9=ZPKy8zpu?`$z`3d!qX zUFrg}(!L#(I>P?2i@@F0{J>|ciq}tP;ao|HH=zgg$rc6kLF-9r)@M4O8J#75qGyZ0 z9PC}srr(H17Qx;O?XB5z>4T)(r`34iMsxQ0$!@-b!N2@Rv_^7Rg0mjEU%wQ|Yo^^I zJ!H&3lb;_PEi|N`Mh%JuyZdW~Alm#qXKkFkcze_#G4MA^Da|OiDS6dDJ5>IygXKL3 zi@7TT1rXv;qKa<}Ir0w7k5G=fMfFtydkexb_FtR%8FqMW9Ox{KZZA{)+IL+ust=`P zx^*dIq)$&b0v4|G*k61>FC7(P{xk<2o2F^))aFj6Ki4Vf=`c2N$g$jFf7A{umUV}C zIF{AN)OD_{ckHgy$-4EXaLg~9n}M5b{q2s@E&nV_G)}a`K6YPz4~p_VKXw#|Ea9N(&sZ*D!vbOu!|TM>u;E(E#zc*ifxPWF7hT}?99tac;yws3>z z6qS-1EZ?5>C6UtG6StKafw94!Oa~V-Z5kjnQPWWQqSEN<_OP_FOA{0Ce3P=o>&)uS zhiu5d#;e!U!iR_FjfJV75mJ}VXX$iNgEbL&+MoxLR>I;bk`=^cYq@yZB~AQ@H@TLP z;dZs(gy+rnRl3(P8t+1t19|c~8Hjp9gTyDTjMrv;@+`I8EOpG*=f@Y1CVJmL{a~Co zz3Wz$D%mC@d3a!R$_Kf8r*?>W!1xm_$G@T$h|a~{TMYLM|Mo*LUeagFbq(CBcnzb7bUAKhE3c@K}N7JChovThV3=5=tKH>RbX!nuI~7JcREKY zV`m;Q^>L85Ls^cGc69~XZc&*MD;4$j&434xntORy6{-W$WW?;tns*YB(Eq>zgh%aS zkZ;?F99~Nv31>}{1~iV;KI1U_W(2Mnn4(wynk`nb6});S1a}YL+-o;*XBU4(aM=FV z+!;5x*mZN>9+{t(C%B5`D7#(-nE4#G#A@}NDU>h-K8ie$|2)~f0>?Xa8*%He@E;KH zs`1inCsR`~ax-#Mssq4N-8C%24@{aV9rj{ViTwp00A(!-J@s=|STG9+o1%J8g<13b zRdXseB^{XuPZ^ospNq%RBU$#2WJqH_?W8=F;^w;vbJcL1yVo&UPAKDzGU4f-sh6}E zYxA^Q?dbZpp;*WJjJ|yUZj+@Z(i`V$_J&19c0r=xz$V=xo?D$qI;DCzxj{Me1&VHNzFZOa{*@T`OU%~Ql;6S8tAafmsxX}T1{xwxbu@wEQ7wFz#1j22@Bcx zwHv_zwna5%lhtvDQyoG}B~nk*Y3v7Z5c45f2f^bJcG(z0mx06vx6l|CRzlap9h~&Z zG+U4P=u4x%moIkbMLNJ4aD#^Y?#GQ=oO5-&!USS(hLew;h}Z3J%ls5?fgN3jth_1g z5YR_R>;}CI%{#|#+_|LVNe7(J9dA!`(?(w5%T>PYhdtIoO+DP1B2Ld zN336maKwE@&G_0x;$du)bj}(GtI$FCX|=Qz{U?Peym^dPZ{PWewNz3Qe?!J4p365H z7d1w%MlkJk_N+du==rSaL0`a6#a$Q9-5=?ViBI1-)^AEDRg-t=?eTqyE38Zv&sy}t1G>-0aQq+vDqoK^tP}f& z_C0ECZ}IU-ObUB$6Mdo2|Io}Og|p1N?5o(>qHbUVC*wL@xl4x`!uF!m|6AR`-_Uo3 ze<;65clK%5yLvLSJ>O^62?N~kKewZM=Qrv-*83JEFb7h8Kp84K7QePrtl*FqFFHx7 zIIp{pulFlNo@zPE1r@bO6~~hY>?o94{99aNn>aoWzeyG@oiDf287AtPcKX8k_2AuL z>ZwcbBb)6=Q=0%fMT%|{T;FT!&-f!#l`c|2ePP%8UAgfoG16zdOWv$Pl4I3wpZ82d z`UlubvVXDc@K0xO?Y=j}Ff=X*c=oqy!t3}@{DUBl%l)HG;;OTB{GYyxFVYQ1 zr^YkC@@Hm#7FNmjEqy^tAPRL?q(a*-JHiEO-0dW{>IB8ao6k_+W&Kn{-O{w2cu0c1 z$e*~TG#Q>RSp11eEBJjsC5Hd8NSiNqYl`2@`T7iELe}R$_^!*R^i!=Xp)v5;FdG+V zg-VVuYZ=6|BS$KG-Y22RttEpc_0rRo;zaw?hAkk%gn~8Qi-nT0!qyS-AZ9F+|J^9Y z9TfA!_WEX3U$;Ig$jRZLoab?j;_?J;a%<3s2?q0&p8Y~0Wd5kUR{nCp!5FKl4^vB1 zAnxDO5LF(zm7oNY9q9o?wn-_ogTd%FWLal{?e6)4rPtcG)HOUVO={mt|3NpD$`XQalE-WI|9d>YH!n_Wc9Ll65{ zKfAni-|QxRqb+gX-oT$Z>T%uZx0yp&`agZ)2)%C{XWB&USWlQ zwTp1nk5}@+wae_0qkc{tR|t{!0TG{`!#yQy`lQ_795*kTtyy0Sz255%IH^6ftC!zu z;xq5r-qCJQ1?RQiNIDL0xH6zEDDlcS*YYeC71k~?i4z1cZK*aQP1%pE#|-;E8`Uov zbBxzp31FI0O^WLtC0Z{#n`5d-oWXt^tFt5>#^s*odXqUph|L~u=lUMgT|9A^DZ4wJ zX6a%9<3*n**0js@r6^!pYX;=i{Ojposh#*JoR46vf2_#X_00L|=|-gzIOkV%vbz;> zYU)BZ?4`8&eY+a9K81(=C(@sH$7(g_;k(BEr}Nvs{l2pghMlbn2H)>*z38n(*k8{o<)BeR96hJ+hs0Q& z*V1(LKf=f@I??Bd+qAfdG-;rcj*9}HQP5-iuq{G<9{fr5Zm1_`Y6(@;TP@fGA{8DF zBx+)6G^nCz>UbyQ3#rgZs@=(-$Z&828|5(6*S>9o(j59*FaaDN0Qi!MoIc9BcDyv* z`;%JPRq6d4X2+doK`pZn0iSJl98Gr$Pp2loO=8~2()?~qVJa5v&5_FaTBa^WJjW`L zpCtPv`eJfpzDd1LqcF?siIgx0vp-_xDgG~+>!-16N7L>}>b(qV@_gGNV}0jOZ>~iC zJk_>kpmCLEMf-j{IP&iutrRNpky8rU!yvTYg@vWPIrnkZ;|@{x0_u%qxIntM)yi2U zc5NJ%94QMe!FsO6d<_2Qdzs0o2_s=6?f%iShYdc)C1n+fi!F9*-zya{ZO(RnR7`HQ zn8@GH6s#2WE=yTYb=AL?)X;NE==w*vNYX5*L~01?Yrvi;ImM!s<_@YpBrdn@C%Fs; znW3I`rtO5vrS&}T;fKbq$hLH6eMoVWh`QMk72qeailtu*tWf-zF>{@cTbyb5p8KWc z-D1I(3A&ZmXD>fnLD#v(kbZIf)A@ZV^v~A=!LkSs0py;b?{ z>kgF{zFA8<3nP*p69y=K<=fr5xCA>Tp(-$Mu~D%j8&WGNlP-zxv%bJ2xn;mB)6usq z6Q1Y?9O@D|;&fhl=|Qo0Dl3xy@=F^k<#YQ$pab)WU?lYo%O8)<+3|_cHttr22r(Lx z7o~~{Vp1JepsDD=S~k2?&XW9^=|m3xCAYsu^Q8y3D>ciOHYD`~s^^l+0eP>*T#C$w zhztvTv7d)=9Z z`W!2W@v2t3qFU3x2( zI@~@21o>Q^7jFBX*1t)0vBMhZqODCttsfhK1?KhEO(PkC`2-y4Tz%9qowHW)&P+LI zC)AD!$6k>ryJq11QrDnBVhwmbSnD9MZo%o*yD}=<;Nbeecl;nDzHB#QCxu+Y^21RI ze0_J}iFerrE+ZtX<7>I``}6wmn*cX2MZ0VEiSuXP$ePA@)pxLK8z*dc4~O_IU$SYw zW_q33F=FT!azsjj6_N_#K(PIonr&+3VJ8(d!F##w?}y z=KF^}hgQL#kA&o#qK8Lx{ppk4DA~zH%aIJOrWnQ+EaD)K^(*=PjIGw#Q!L}M=c1_V zW{$|(I7@s5!W4;&@1>O)DJgsFX%%Jiv`ISE!gKNcPWx5BYAe8}SBz{gk5xCMMz{3~ zFUN|KzI21LIwl|p*^>LEI~UPO;-C~2zMR14W$01l!fXks^wWH0#?Mqf`z|u-8EbCt zS7rC>&I?hIr+-uIl8l8!>3ws1RPS?f{7#y7Kem>c3QP3)s0ja%l2Q@VK)DI(zuY3X zy4fA;INQjRT8E^;uwt!zh&k_zObd3&qot?T?{ye6_8rwl-a?AoOuD|WmPszN&is^p zyn`Idx2F4E4kt|ck_k7f{e2n4^*ieGRn>b-Zc@H=#JVh*T*>Wk(G^pz#yim0s11#eg*Tq9$l3;IX-#DN4G(;2gM5^JBuejrL${?l5rld9Bm!%Hqz! zBX6b(T|i|BD6RE9I@D6s&!O~&-lCDsaQim1V_{K~=9*qZY}W<+`ROwEtd6^mw6ak% zjn6!A4!Csv(x&8np2fJXP{lopWH=6G3_hFqR6d)PmJnjFl1GW)7<#>OBz>4b=^CD*FRre zr`WE08B4JWhy+Fd@hxG;LBE^k?9E*ka?fi~-E*<@A{^Z~hn{ zWsz*upK0>AscDY1wH2^7G^-wkt#dYwskoSMWuxHgdZEQh7YwyjMLQ*QmmvXDVX4>e zLpJUSgCzw-@5a>W(rU;iRd=yj&y(C-x#5^4=e-+P@&X62>E6Q~((0WYE>;C%cdmnj|rx9HSk9o-WQ5A!;S3ddZJ^~ zis#cD-<^L@Dngt&fY`%NRH6%)XsMv3*;%!4Jf}^fWu~4H_b^@^Vx^?8Nbqy}cvQj23{TtSZ5d8bC4^0^mA^_$fuHS--@}KA4sfOp)*{;Z zr@j~_-m~Q>)8zAmj#c*)oUDVuP+cM0$Y|~M&1++rKG zvb3lvPq}p#{uRvq_Gjr}Tsc2&knNMC&J2p3efw0pn4)6^h6K||*IcIps&-WFwTQAnMW0cs`ww z7rPW-pIvIoW+pq}gQmHWdTM5yO9r-oV$MZ8qHF~{(1tpQ6;NMfpOEd-l#|3BuuGQ? zkUqqjr0xrGr)OE2D{sUF)`Zm$)DyH%Rg2CxX7RLa<2nb9L!a&($*Jv%fr+qwN6jCd zeNEt$3d6eEfDyiV@%*KHUDbhw6LMT)?amCsT_&%D4eOR++0HUEft@ZAp`dhpv)8_1 zOlVV>4jjj=4cZ52t`}ZA0}arr0!#bvc-^&0QH3tC?Kz(43A2`9c`j9riJLr;p0alj zbJQ1-7vrxPz|<^l>RS!!by_DH7n!^sOtBpC?GY`!##I%G7{IkK95s} z8=CT_=JKnI7`k)BvV5PZs}yvlb$pX6YdFPN!_*{I8ROP=ji7Cpj-*a&%7AM~OR^j_ zU-|^8z*o0imOY{TUoU$9CLO)uaCe~eRBV~kQloO;_Bh*rzm00M^O{aPOIL+kWFBQ& zw7Sr_@~+`$y9@Dbhwli4(Rc6mTLm6JGMdNQ2l#5&h&>MwKs+HjH2Gkpky)ti{z!zD zvGTV>>c#G#fl&JvY|YtA?i<+Na>(dCiQ|JhWs@T#ijQJoyNdU*io@(0Q%+EHexdA` z-opjUkVX|Qvw{lEMbUn&wV%^r>k=R1hbaaVXJDBY?>9+A5DB+Z#WYS2YiS??J$oCP zRj{4p+-b5k0HPVs-Ch z*Uks~jV|)@Yd2O%FfSWD5fHxjmMiL^;naxWNIdy`J0Z;S;o0?XhU(9nswPu`EZzIA zaBV2@Zh?)h4!@CF$X8HgLmM(gO-Wh#6kpM>^r9s7`2H_B=|a~dwG%6B zoIk-B{RfQzt9uvSwnF5~O^l#zE9MfewCY>o5xyXkBjuci1GkHZ?PJOd%cLGrl}plZ zp0RLE`kuNPIV74@Mg^wgM%U-$EJQ_r*uUnx`Cz=~aU(aA`5pwCAI7M=E=-AS3NhwQ z%js9n!nC~?Oj-yZYxnNLQ*xaBQ5>ek;MEj)7>zoSFH4OB*gw?(mG|U{d;PSj*JPGy zEyre?8#8I{f?#2)O4+tHVb(@7>>UN^Mnp zuoES&Ssi()=dUOatGGk7AB1g$0wo$QcYhgoqR-H^{8wCGT&|Vj%BR=|X7i7l6h&lY}y)00}%@HJi11V2W>j}q5<`0TvCS5efSdnA7Br4%0k!@y=0&2i$ii$9Bu zc;UPC$|DXjp6DTG5)X>OCU{d(#8=!W-V5+)yr!>XqFb~sLR5hFG&`w8RO4Te2kG&)(7#hLLtq&a9K!HK(qz)RK3Knp4h&Xep97wpZ_v(-*LJ?%U0$Ss{why%xc=)H5K(I*^nTe<5sLW(|dA(1U)ei zA#JK{%ZXSOS7SSwx(C9$`m5fbp=G&CJh{gM#WOVGcG?G6N zB`vLhPcTe>e>bVEBe_ttG!q+?+0md8J~ES6W+zOGi`)oVbb5p>60&NJbXO|tLLW<) zCCtz4+`wnYWIWmWmW_pEW4ey>;f_9;by{<8TmVqhYn{gYjZSK0aLF-}yDCxT zhELCclBIT7E9!%{@3PTQJHva9MOmsit6)NlKafWol!0lKmzamw_IlhCY~-{Tg; zHSCb{b7ZaL3L);Cpv9{G*+G!UV`^TX=Ik27W_iBg7kUe9RmJA>WmvSOd65H;d#J*6 zNs^DPC0&4TKKkdb_-NmREYqs@e%StW`$}zaNqjr)`utKSErsu+*|n zW5%6&yxD=2F3T&6BW~bfi^^$Vm{{r>6M4n5;eXun&Zk~PUzZ^uVVv9iI?I1w#CB};+iAs?QcXki83BNK?o}*D{~a&d-<8m&Jr}HrPYqAaY_}i%H-l`U*>YQGR+|t zOi33_Hge#S;+a}9PhpsRx*w&SJx@6kn#zaAxQQuo z6{>EZch`CGZQ}`aEqA}La(-zYC|K2Vod9cU_%s|WC_@brzQYL+-KV(4XXPO1#Ebk# z_eMzkoOs3r^IU}%-nL>DND@tuhxN-mFXLWxfRwyHL!e9A6pJ2o-F8nZdfjAp#1TF# zt?q_DO}kHJj&HuClPwkwEIU`AXnsg&8%kBYvDnWkLf$WY_t4l5y|U}cBls*=D)8W# zYu>Ihu)+P2fRmJ~U6c94Y3q&}`SWMwv}4NSQuiy&s0m5};U>71HmTGG+YWwLzUn`& zw4sZg$>S28`Iw8MJAm~H);l||@xr;k&OlfN|H7;2_D#sw93yv;uFVpyT`N+wR?cm4 zQRz3_-!|G0vdFes;5y-Xqn9Hz$eLeg$Wq&4M`&SZXq5f=&C*QuK~dMYQNd6d=`elO zy=7?T7c2Jwzhd5om99E2&|IE_gmZh53s%J~ESX^EXs;Tbdo(8V!NF{6Z5^5t;5&j? zt~IlfeivU8SU#MwoSEBh;EO9VpIj)9j82fUr5N_26$NKFfHY~|m@Nu3Et%~JvcJ@I zVQWt`8`$s~t4uxV#kq(>z6^q7z4Skfo&bt7*9DjmxaMJLxqTFocC93g}6auswbB41vR^(lbNBNl&QA> z+vh6}ztJ?fADX9V1)0vhgp(b@3m$x z63@7UjCE~9S!c`&D1J(|+PFd#EFNlQFot`%tWG~!_`Uu*7aMg;=hvh&Li=7ric^_yD>zmO0@5_L9)MSwX3MAezSJk2H#@=;5oDs< z`czhhKEje?7kr0{u*2F8@WpP%M1>kgnA?Jnb}_ZiC_)kp&itij3S+zj-es4LCM$x* z@b8o6Huqsh*q6HZ7)7%g=JW%NlPq;O#skdo2Y9lR256n`U&!+J-UR2@Nv!U-lkOa; zF%u>0dVTC4lscRo%zP4;e&7G(CrRpnm=juzdFtZ%B^XJl$oW}x z-j6(MW_~sl5UueWxTuBsPi|pEJCZEp|C3KB9E8K+HjwAtpK2f%x|wy6>tC5sk+yDY z2HyN@+fS&-xWT(ytS6)eJ;E7gf|h|YiEnvcP8fV8qt_CW-(vIKTY<@tUprQYv7o{+ zn0Nq@;+;Nt*h4LaPg_~!!=HcB^$o;}N?2FDoge2>XJ8&zsBS0B{P4+0+1Q4+Gkfda zu9xtzTj0xF1bf0kaV&xmHg_>=@t5oRbcR&0M`PAHzUuIL4NzSnp2+`&o{Jx(x|s z-o1+Cy(?2YawS+X3ow2bb4yRBV(SG20+Y8OoV3L~nVSAA5y7-j38XSn!kMxZ4}bI5>4Ee~y5X=yM%K3X7?@B8D5esV?vurufi*?qxxgxwlE4 zS!mskr>$mnb8}E_9aLJ|kTWxf*FPjPk+%+-qzd1)l{n6!lP7>XmW5_O6D(25CDU!u z^6HhXCeCHWe-H#yEE*U(31k_qIxNj=E~{{t7D{#$|HknO(DtX6w&c6!U~uja)Fj-y zu{dKO%l^aDJKR(SMjUecRu`I(JlFE2#J_Im%Zk^+$Ux6P(;+ zfrSkbNpRx0F4(cP0(8CrkPNUvl8|tMl);h*NUd#0Zb@e5{yC&g)S+GJ0lIkaR?Byo zgaa9^7HbqNgbmGsW+!A#$#m@!4|~5NFH}}JRxiTPl@kSyan6mgAiq;7&nik0lBY;2 zB5ZB#YGis|N5O940SRzCYa8g8GH7fO^VY5eMWaf{L-d`KBmR+?t>dOLJ9U?umKqK2 zWiKJwFqd5^a^LYYVI_i@RQF7Y2ohgx%Yyu~_dv^cOcDG_bFYyDCeuf_8Ac_|-j%Kc zspr6Zu8ReNsQbysDiND@6`tY2v57ExMJbbh+RwB^CfHkABp-|>;jb`Ujy?REtzOy$ zN~FX{h9QE7Z?c6C(c8CXP!G}N{}v7hjLCD*q<+M+>j|LZQbA()KrnhQBuQY29{%qi zq#HiaFUi9az+1LJ5Y}m`CCK7z`!!gqn7N+EtmroPg_7p^FZGE_dh(U_lv$XzHT$Bz&S-}Re48R zJR~wk$KZ1IWUF*vN*o)1w|nEA?=S^wtc6D2(IXljI864K7P{19>lGb9LWtaIMM_bn zkQH0PD750SGGoh9n_hqKpidQqR0e&ER=6YrJ3`3g0R;F^yh0nA3x~+iv8eCNUo+noGjtKcj&E9ye&hjL>6vWrnG@41*F$fV%mo0y^zXS!VEklQ#SmW><78v zL#j%_fmj#~t|nGk@-=5vrBwS^pfyAhK?b&l!63}R5i7t$5xF1t zD92y39CF0fq!2ktGuX&}J3>eL?`x5k#FRRt<0@@0*oWc|$q=Ui40Y6w#Fb;o6-#P& zT*>|vz#|c0j6xnO?1SMpk_Ux~qmFX@RecEikhT+OI*Wrmos|8+TnH)V7C({}!u@bl z`DQ<=85tdNdV<;J^rApz(SNP8Cj)>!q@*)n;FBS*qi(}5hPfg2$vlLq@{5-RZKozILmG>lBzEpzcBFqTWsdd)X z`f*%Rm*2ir%Mi(N&qYMd;YfEJHxNKc#r_%9P>hzA)@wGd7FzRW$ZMbu=pSA>Zq^RV zB7q7YIBD@A=h&X@LxyX;=D3QqZ701tJpri=hlEFKDXbBmB8F7;J`5D^VbsD0uq<(# zuaQG@=sBoM|BJv&iHM5Ctl4^M7ojOKw# z5bDlk1XU!QIzoneUdD$!`W6Obz!0%#;EguX+(Q51MN2D(U~oVX;2$3MCbdPvxQ?w) z%pJpEA(%l&&bxjgq-8kAzZQ;7mMH!ksdu4NjME*R3N87^T-Z6Vm#xdwF!OC}pC+W*Y0NuW#j+oVt{|392*?ax%ClO23?yaNqrz+<) z|46o?ZQg8gs)ad?cZ*x{BH= zzWq=089QaT!ZzG?^R5?2Y(Ug@=>97FUtf=R%fCWGQCPQm@bTX*weF_UvA z;#7uOo)C~@y-IUSt#Atnu3Dw*Ia7IM1mOdYJ_gJX5JiB&95nSd07$&rt2Q7_ys{0m zB1iZjRCq45XWMv+Jzer33Pyft{ab$=VxM4ba{|F(Fta0h;Bt)FKw>)o78D|7Yxk9t zcGzaF|86um?^HOCz2(4w2F4FADHu31z2=Z%B;roVQnC!$+kRZMP^^sdFc<#%3%ptm zc7Dt)sUnPDA>lu{YaHN29=FfTxeo~YrhJ2n`-+hGo+XnrAo?JlTA%!mFXKtI2%$8W zCQ`uLXnO|aAqO&*llChF9X@~yJ10kl_Y0F9rN**-D*$%mY5c~s-nZE&+LD-oCG$b& zb?z1K&Z^4W33Afl2y7)%*2LvGX)ePD5nCk>wf-B~(gWLHFRP>4NgPYVRsM5Hs6@OG zZ$HP)H9Hz^!;&gJz*Ld(4T<`$1(QHU^>UY9g%*k-x+b`3Nf68&`_@TN=-dF7hD?@l zH;jvUibY2r;`_R)sFZ;1ea~LYRi?rj09`4Z8y}$zV`qT)%9-AUsCSI`_`|RkPL3Z3 z8JI5{0!&PRfiM_-6P>pg^mMF0b=>+dF%-_NBq_g1GI6!*xPw9?ui~d{qx$L(B9id{MGUKsA3DUHHT{XkkN{zG%m61XFtbfG z26G{Ye{n8DG3xp$?Gln$FlX&801?T1&q@i4jS!b zjoc5@u!C@?oyS+TJC%LVn$b^3C96Xo$j>SED>75)K&NN!WpOKhDOGkR&P`ARl118KRY= z$=0d$U~H{mu#F7<74!RVy3kv@n(z@wr`yQB4mVp8IVu*WV|m9Ni*#JwaWZtYWJzn{ z^Kvvc18K%^=IbncE^pM|2Lq=xEa(LC;fLr%?gv@}4>oB}#uOHr%ZDye4>QPFh}2#q zVbD4fnTI5bV2*t_nB#Zc-QE){6kzvp0563I1G#&i9Ukk~LXsysUP(j`-LgcK45XMq z_fw1>16j#YX?t??H!}3{{i>33FVM)poJ6{SyWMzfwtZX>*eC!&6l!Z%1&oxkW#8d0 zLGA;FhYd0Pv~eL9GXME784?+;Ktuv(Nk~ZB#HbFNTe}T*=e_{EjU=zfFu#ruQPWaj z+K+E*_mh;fWl2a&{4cc%jK&uXVIGyDP6dNu z5(#O@Xsy@G;lpjT{EsWa((z$x)mLdq3!IcMlI&Pzh9je8Fp zZJUHIgKiLxAhWq*55EGgN=e>9J&Q0a03!!{eF)z<C>O7Vn^n4V zZW>>o&wp?vB;k8s8$7{i4ACLc zz*}F8oF#St0MRf_yO@1W1oMC{dZK(U_Ls@bEo2SjrBXX}8TB ze_!tEWAYo?DC*Cqe>r4e)C!mMTGX*pn?$haJ1Pug-dE+1E)lQMWM)(bj2~ zk&f*ktx?-XsZHX-$1KD0cS^7BT>v!|4Hr`-}Qa z|7!ViF+|&KQgQXNU$LN2F)d8pzhjiStM^G2r39kozy8DfwrR@}FCQI}bIK6$_vu|^ zeP<=sRE41GxDP{fch^R=N&^uhbRnCqq~pph8k5Id*`5cu?xojG?FT=^eTBo z;8uxd>xx<_-sZZ-&G)DD%z3Zm(HcN!{XVQyXAZ~i^FP4~{(r%W_3j#Bg5d~(X9m;9 zwbl>r&6Q0b%^8B%{RuxgA>Sb|>^`Qd$6w)Vucqal6m&yBSFW+$mohh!k?pj3#JbGy z@_mePc4s7a+^C<{JxJSV>r9W3Ify(9Ci$S-j$kl)_J*G^OOlyo$R@rmU2u;+*Huu* zaHGk58>s+UoI5A)wt?wCoF8w*)5>+D6=;w4maiyW%QYePe-K1Q?d33$^Ms>kh3b;* zlIl^g|7WFFOdis_?<=2#G;N34@ED%ncl!E0i#|*6wzMXChzv7DGH&D^&Uhr5^p;<# zrn41mu%@`$-|%(p(Mkep>p|v9%a9zy2QTG?FMB+%Cd(jCMsmCh@>PZmLm6G@j6v}e zqRg&bZmzMG>jBlv+fC`@=5&7aQu>Xx^hAEWNb|N*;r@@!?e9>onq(xbZ|-$8a+`ds z+u(hDNkQ6d^6p5kNBG^g!uM4b^f1)2+~-e>$x0?=xq=JQy^IH!i6;9?qHid_W%UU0 zgQTxIvO$+$>WE_5yy#U29A7uNJ-h;zI831%L-&6uIul1G+&_-bnIm$aC6se+jZ{jm z%9U$ln=xa^MvkoTRR}qgVmT`}o6UyJ6}jas#}0FptAnYpioTKh`TY^k^SnQw_xt^N zOmwmw-iW-%+=iVx%ouEfE8<$N9mMc{^=Z9+;Cx4gkSEW4dunehioPcnFZ4sU;X1Sj zQMaK~b5v^l4pzU}?xhav7RCLoSD?(CAng%m8@@D{YF38tOwJ4|@Mh$4>z{d^PEp-0 zFgoJ!MGLmab*vNfKC8W$@%4m!&WobZpKB1)$0E^jq3=PjrsmtIDGj4dCnt^s zKl2Dscjv%i&IBi|&L!n2^1$mxm$HvM`Chflj(9SjW_0B_!(E>l;owWT7ibq&C)T!Y zYSr-MNyZ0%rD3f}TXW;IO&<93h&ilXb zNHEGH8b9fC&hr=CnbM9X_QT%AcP5{-}i}mOb39 zsem3@4l}6Mzf?`o+KE|aR@u1bT*JSe=6GMZ?M|n=&zkf@t*UaBC$Xy0l0%s06k=S@ z-c#adnUPaG>;|}?QlCFkqu>bM)j0Iu0*b!t8sGBa-cc9D>un?+^mP?+*Ajd)+TPN}h~Tpi(iCF$BZd-q9K0@DB zJgE=+vle<&=j6Ozm&WQy$3~@=IYSf=Dk}E_-<(5oUmOX%&Q&loJ{g!!Zw3rW9bKu~ z_Kr|Uv^sks+ozC6T2l(C|LmUt&edRlxfLywskaJ3-U?2AFg^>#$*A$OMt~PfmqjMe z)@52K@A%|?UT7pEJOh&>FDzIu)dZ`_tQmpN8~GK|)5E=g5_)OSXMx5q_4O)h3~L&0 zGGol3C=#4MzugH$%*tnK-^E3!hBk`oaCCIpSG+nX(6^Isg>J0Ll`4Y@m$}IO+(1C2Y@s0R9MA)d=3MRD@s|3oi zpn~vyY+Ex0JqtM@xK>_jXIkb=sTq&lU}-@m>WsEgyq)FQ^s5r>um5@Yue8GZ##pc{ zH57RI29x(rBD+c|6QgY-?XECeMW69RJfM;xG32LS$K7ddg-j_?N7x7$pT28)rO`k2 zS`_{GL$i-^0E6E}`Q$;XvBC5ov?4~hN2DaCVi~-%gq1b6Zk_ETwhgyQel*=bEyJOGdG1NdFGsawzrjJR-RL_ggO!%K;87tw@8?#4qYpmw1tJB1eWyzWtG~Opf zX&c3`?FLR;v*j+Ww$t#8mT#Cu`#D9#Q=P7-#DLfMkWF^4Hu8vu_Al2NA@q-OWv3Qc zr&T|t%6&OU$f>QNQ-i4J=7RJwwGgXp`ZhD#g}-@v@`$NL5MMgnNO3f*g(*W9#fH{PLoa=IsVBXTn$1DU&~1g6vPa8G2P~kDS~0G>UZ>c1Aw)w%`wP$BFB+`)_bggnNP( z8{sF;co#6;7h`$_oLem;y9$r~ned5s6r&pY%S{K7U(0#Kq9crv3gq_<`W}Xt9$6#= zsrU5~II7-C(uG&0OOIKMf8en|MoE_jTj&KFsf8$v7b?Xb{jQ*=DVVZrmgsojLikOm zc@$6jRi2cOhhqWMqGzh%HTKDppoN&kke(`KGWL^#b&#ghQCm{)!?C;!sHDNFKKu0j zNS+cvhtk(OtqNus-maWzl|K8KG+&YFgMS?B&`Cxyx0wZu-O-tS-4U04eA7GiE!A4o zOYc@s2!xwm@-y|%YTm);O?qnU-ZRerQQ6yMxq0LW>H4{QGnkIlfxm#`*%)!X&h*Zk zQe}?@&o-R}cu%he>+6^3;=>z{9n=L+FN$fmw3y>McKG~LvtBx5pAt&|=#FGzgUai9 z369B?@KL`YFVs&j)`DiNSaw$lLP=HSQqWl+Y|r;XtY++s*w=QJ{d%We)DxVB1Z78b z0XC0(;u- z8Us-b$}fkOfCOzC@3?s!Qd>;fp|qOXGZR_AD+eRu{4!6m0?exIZ-zkpPxaLnK{%7j$4=HPE2Np3<7RE7xC z6dV)K>t1xIph}fE=0RhrjTs1kwCh4$a?iId`v8~fjlEkP`72N08w~*qWWBqqM=beJ zf?xEXLD!66Gowxr2}(+m{7Vsjqrzs10|n_<0;Mu8mVHIp$G(*%SpvoCpu^2_L7Ae0)Zl=gae7Dd?9nO@V?eQrd)oGx$-z=D9}+ zcUjppTp9=7OhSFR(|)loV(-5kUyYa#8fI1BxmArp!^)sFGMEt>6X|J{^r0ZMw<|RR z@w4^Z$$S!zpNd9k#Hw+Iz;ix}67iX+g$F34r?exwIcy=Bbsv!=VyqDojOg;R+4M7x zY1eOziPVb@s`FBLb4|Z7sHS2A8xu!=r4l-mIfN=Fqd;5zn~>~R)1<+3y&AeqhbxL! zuYcAHr>0CIOQX*!xK-3AS|z?wNI~6B4ANH6qN_Y5r7oZw0*71=@NT8C3sVeG=KHX)4n1{Y=|KI$Kakxq>dzmbDZtZDgEH z7-tyDG)TXfWTlK_A|=)Yea|0BUvm<>--tn%nox6o9aA*))&YMVAlW513s&2f_FTF%fPh6I(e=V*wzCC9xmiS1X&wF_)T8~Tj;k)78pmV%(maa`44$9d? z$AgJli4LAjpI=8_?RjNJs-tAt3!PE_k-ip7?*>T)}h>?VicPc;?b#%?bZLGbUQUE76jn(-V zl=3;=z~12)se29vs|HJ<;~s9DDII3$2wb_ux~pPR1%chk-d+Jg?|Toi)J2@O-y08L zgfa&P*{ab&qW+RT3STlWP<@1cN+_fBGyR1Ed@$+I{xywA`bUaX$sL*^ca+*6^jeT)uIoW{$TtKnH zkHe|2_e%WEH5Qz_tqMH&$)K741yE(vVx|gq;lO1D23Qn6=p!K4eZuVrn76D)Y*N9`P_!Nqc3ad?}TpatdwXE zbEfHxtsMY|@k5Pqa6K~w^Yf(lC@IfBQUNV^##HN`w>5oV%1CIZYq$&m< z?MZRa*9HG_v3UM5kW$KB%f5sEqM+bW3M;rit6Df~YR6K|T+%mF(C!dvF!KBt(+(uC z|CLbc9;(b0y)Lemt+*n=5J5#)TtPpJ(uj|FbhAFAlxp$#ve89Vs*_v;CWS9wDMX`V zL)b}nK#01j$W=S5g0;s5Q0WZHViRPtZ{R3H`(q6cq_H_sX$sJauP!Rd{s#;M^)*3A zAnH?Az0Rgp!Di&KhO6hMyUYNarC6pLNenMm7Nj~ADUG*j40~jY^EWe76sA2B7$;>Hcp)D= zHpvw5XU;JQ@mM23GkGnt>IZGdPfFA+a|Q~*k?`UYZ07GjgG1(HMWgT zvQ4R0cM6LKmzlli{;jB&t_P9V2`6uMjhY#Gd<&JZAHW?o}9v;P8M{5>*3UJaA> z=^3>}Xn^Jp(sZUUG$g+bbV~TOkP&(|OK{Q7Wzg<5OO>QhtW-sR){1&z*RUO7tAl4s z2e?5E?``_|vQhN3;#G}2Y~wvQu@3KRm=AQ1*JQK|fAOatq{6uHi`%ZIFkXE-tvN&G zqyB8+p-4^HUQ@{^)s}&XAbW5zkW%_x>LoXM|q)tRkS`kTH&Y7f+SDkqz6?z zu}?_nM-MVdZ{H`XmH)*mWDH+Wr%#r(-abtVj6Z)CTF0HTh4tGrUByI4Np_UEP1)3L z`fMsxiLS4deASwoIN_j>dHE8{coS%N{Y2<*=6V2d)~qY;Vt0uPUZv~$aIl-_HdTLi zDg4a5$6&cHv;q>Q3cYF8?50pHtfO#&UA%K5IYeT1LqC+UMuQ@X?}?RZeglbAF^$1i zz0g5|j8{adsafdJ`7<&%B}$!kI)-yVkqz0jbAJKfml|duWy_pIEmZe+FA#(4a1zm` zpC6nWdI#%Xoh7om)qL_IIy^hFC}?Z;kBc@AtjV=OE*!`bYsK?O*(xNJ|$B&k7% zcm&a5i3(Q1dUn2H8!fz~0Rkf+UReW&|s^6SfJ+PT+w*H=pTPb8M=o*QpkwDyVn zIq0d>Q{7vUJUiF;Ys;PPeBCFO{+RhQ^nTDAvg6Kk$AK@2Vx@Xmz{9*BbF2559WzXLhO?4kRf-Q#uH z;xB^ly!i}|YIbQ3*Bf>J%nbb;?+9#VV6zCW4*SRj+?j^f0LJD!kI?q*oMgel8rbAz zR6|&R4h@CXQ5LVpSd00ANB1WYFLu1*1r~n9!&rVb?v_qkLX`sDMO4`&0I$do(V_)A zBRVg8e5uin8zI%wG1xD!p(pZBVox+cQ}%w#{sYQe`0Kmv}5rDU?Ss^r=-0FX0Sw8SvPzhL6eqiKwohsoy zNfm$>@S9I~v0e8pd13PgMVJdBay(y5;gW$?niK0lIpb5`@Kfx=Ad%fmKJ;2adsZl2 zKzopAE&Mw9ZZ&z;gMFBX=Olcs3=5vCN`7rCDy~7CQ&@)un@Yw@FNqKQv2iJ1XLu5x z1{>7}B;AMVFC&4Z=H7@^i4WRFNjaIOz{=F(FV;FEkqWqX6x}_>Z|hzEr0$;PdT-h` z+QS^Eu32Y7I^A$%Q^qzH_004UN=j z!TE=Vk%UoV|4wXQSgsT|%$ub#CpbRXQjTbX%sX0=G9c4M*%LYLGpr>HA1?F@@VWo1 zq31VhF=x36L93X#&z3pL`8h78OJVvdA02Odg2A6=;zKLMr=^J@Qu_s|!>uM-k3nlR zjrI+;QG`+>-3>5s;K+`MrPaGI?ls?S+k-4#wswa;f}d6yjOO8*ASJBZ(`yTy2Y(of zi=+dkhZo&!R{^Qn3dRIUMv9fcYj9Fjy@bHSgTl3Vo21PWp5sNagYn#dW`R)GI(#jd zBa##fa=kG}xa?G&b5$#|IabzsF72*H0aY*d1J7Vm&cg=Q0-r`zK^t|h-RT3kcv9kB zV(S_r_yE`{lJc!Z(#!{lxcQ+|5xn4Tm#sv6EgVVhx>yHnkSqTn2ukVs$oslcAkk_Y z7UW(2rpe}Mb)%LmgYV(hSmm1zeH|3`Uw}i_FOVTb!2iV5tMyF^soD7{^i}5P1u#6u zQTZPH`iXgE`43n%G|986TC6q?k3cLmiFVzz;q1_??sy>_blLN??z6@0 zZ`<+u@AcYD@yG`leyl|JOOvH#U7{Y-moMBGc94+BJoW0Qgkb%(wUEsi~t+kd7e`0j`l&Y@Y8H$XLODDnyZVjjDl=Dh1ol zYg_SWb_k6NTbYmG=%<()?T#&+J9p&WH+{*z&wYPqtST$aP)spkM9^PA$PeuP$*8AU zX3mYIVC6G6-9VOUWjp(?(B#huRR0!+OlX{o$q9Iz?RF5~sHE@uKs>xkR#vxIX1F!$ zml0P8ZZ{F$ZtC_Q_QoYkJN~oTCYE0}36}YAD#4Rom~u-HsBqNY0x6Q0M#?PH zugiJOm#D(Whl{u3WoSVCB@R9qviBrYI(f`XrK`uzt>q;I=88i@^xb&!cOUlW;F9zM z`r#NkS{oi-^g+Df-DIcYB~7({PXKB>%L>xFy}$rdi?0ZESbxkBWEKQujNyZTn}0H zF_EmbGYnxFs0{V1kBEVVbiV70nEL{@mmdI?tGz+Zg z4-}jpT@VKMjB z71^>$w6~^kzJRu1^T8ac(7PTx6Hithc|_(m<1;sYUTj6yd@~AomSPntJLwt3qmu*#Rw)$k!3oAq~8Wp2o;NtvTb|um@@wYiu2%Qu57s zwfmqb`heWtG%@-FlM&P_Ff37Wh~ETpF|?M?7IsZ}H!wA%uIJw#jssb4NHn?;o7^Fg z3M<>m*Qa5dXxRc__NyO_ogO!p%{_X^yiI%%3J|rGyh$a)-d!Om9R$L9)KFB_qxfd2 zv19^AIO>J>&I|YoT<(LG@3=@fH`jnZ&QdwEjjG|L$f@OV{-0p zPlf%IGnsEjR~7EoBYgG7niEXtIbpf3C~Y-NIY+6?p8S?Px>G!F*(U7rTD@a~N2yi% zx;o}GN4GLSB>}xI)qTomTbW;G&;4Vu#}nx91i3N2`o+OufJ({F2b@k=r}jbZbPk-< zAQ*kW&JQbXuUw?2E&(D4^dD3}Q8 zj|{_COh`is)uG&lsVY!hxjvg{tgG+xtp3WYg1v@PT_*T4$xCSms{*Lj*C_OG#?QRB zE5@p9y@aDX7U<^6v=aqBg)N{;#>{FcBKo!QQG&%?GoR!rjp9kIk)oJ^QA9IhT*C!@ zJtYyb_2NbJKeE+4SS9MCrETU2_e0o8(iDdQ`;(MtaG`vB^;6 z_ybc&Dw5M%GgMa4-Su-btFddIeI9h~HS_vsAI9vu2D}}SQ!&IEdE3qtR?M#HyPvAt za~bx{;kyhnE@-fWTrIQfaiWh*X6<&(V6~sOyxM&pmv)7E_0tunfdC6DG?FqqL+`EL zvTalOBaqha8@{zZYU4Qm zJH5GcVVH6HPq&~kg_6@Jt_0|`tYir*4Vx6W*Mu3ae%Yb9&iA|`acCm*VWhUC43-a+ zs8dI*@m%5&6X7$g&;Iv9Eo55TS~JhpdRAkna+u{_L($x5Jn0U|NN}{I@s>S)&-+2H zV7~WLH`S~wvs2SiPGE*>A$j*8?4{vMJYR+qadFP)!MQ^$1^OSLu)cIP6(cmz7G9dqLpA55k`39O?%b00|BG`lMrdP)=ya^IAk~X z^OtGcK+@LOuMpe8{$ZeJAd90}+b$nOoXB7JHN$9@!GZ!gj+jba(JK!ZlZtyHL(|N> ztw>vvg;^xQ)%cVKeA9fPPx$r!&Jtel`cF|C5%{o+{4)o1w zOrB`^S3OPBw$l7bE9=k6qy98}Y)WHbfq9~IfW(r!Sf1ik&BQ;oP35r8dKvSds_m%* zx#DA4%<7SP!aS2UYl#VZ)(gwh)xfgJ6ODvdD-SK)tcg64)LVLnHIZ-F4_q0QfCjyI zKtwb+-&CxGp+DWGS-U3hW1h~eD`A}%A>eKu7LGRs{2o_$U#<-K)&NRgv{xTxKdsMK zj}9=Br;=X3)*MO4UEkZIWvH_e+4}m>6?|Z?LNpZ2pdK6V&>3XM)O8i8HcQq%7ikL0;H@t2wcgfu4{j zq5DdXlZ$=QAWSdqmqzo)hQ~Y4ap=-^v?J{1*L;5g@0g*Sf2>(y)l}NxL;0J2LiM+G znqlv3+IbJ&5R*3&;)xL`e%v@m0Z~>8YW-#= z3%)seL)$EF(UTs&7q@8W`4Q}Y{rKZHSD}I@Wr%9>Ugo?X*I%Ig-QKejSI_5^a1EvT ziyX?Et#9uZPGg#VJ|%hGyM{&N;Htm#4aQnmnpf4aLBI~`SaS;{o9fiwD878(>=aBj z<|xbP8GOv^p1K}TU?ujuSaxs`wK>ETH@)o@tj29oVs2nYKwL@Zcv+bJV8<@a(Us9U zp5N?cenXAc0AZR*jG#~-40O9Nuo1%ZBN-U_8mHyWHg%F(koNh)ec8I|E{ z4AKyH3idr;r1o7eeDJVJp~z~alhD@W%Wi2G0nP{F*kY(7sBwj59&OveYM@omx-V%F zAITGbeI2EzI&CKW6Grk~>ym2rI`^1ShamSg<(sF5k@V`n%(cw|ju*y@AV3XMw1*RV zyc^Ib&X}`2d=fEmEiZ*~k8?XZ1?uVjq8*!~i_Oo!Lay(AKF#%zMpf+n6YU(!2QXHJ zN8DhK=JJw{#t(_6-51qHYLpEot^<9lFV{@?xURd51UWDM^F}_VN&+}wCm5GA9~0KN zu=!tD9W@51eHP!XD1Q00!QOd#{hSgJaU-Z~#Gd{a&~2wOc5ig}Tezvs^{&9WV~33l zwq|d0EoUH5wRh>pd+29usG`1j;~{u%>=-6&#gNX$Lc?7avo7tnHmGn`fxCF)+FLF1 zY?OuZJN0?67+t5xMUInAn?vzMajG{#A}mxzP3=@WFuqQLsX*dB>?4XXd~13-g*R`6 zHinz+>t1N#oQ1p8F30S0-9^)2;twH`T%x}P-FvMBL55(>cLTJ z!*1sW@U2^j1vr4de%zeuDNk(ZKrI$XsMzY1^_WMBQQlV%=rrduBA;)^_*8g$YAwFY zqd#(HBzAJ<{6153pze~WX$^-_4`gzQ)f}}ag;E!9W7Z}Jdm*Yj{(JCtb=~cR<`V%; z?=4^E=w4$o^=?>Y4hZ2qPwIs9Z?|rSm-ba(HEC6Qg?Og{a>gDY#L$V3rtsjCLwK1_g0wg@@fIA zw`GB+;=MDjZ`P4lS!FU;Bff)UN7tDqo1SL=ZkQ>Lbaq@_PomO4K~{!+=^Flyb!}xjQIu8uh{8vV1c8b-k;B)j9anPh-&EE*G3AIKVCeA?dzX`dV756 zBbsPB3{6VMDpKl83}pOns`sDPHuCaWzM5W~Ry0v9w@~3W0ON;HSzYyO-|a zBoO^tH-oTu2&dRYJ93zH1!@jdbaV_3FYLAMMP41A_sIH0iBRBOQx>jusP4QmlJ8IU zX*NAkmlpzk2P+tX0-lSVsShuGU;lXcct9)V-D-Qr|u|88)qJkpo_z` z==BdjAo#5Iz5HtI2OeIb>VsVk@&vpi6lMeM+ido+!wh!-0~UZObDpmi<(?36gL=uYwC6Ms4>s7g%}A=6i`n{o9La>I>l~#TW>v8-dMm z_{VnTZ;3$Xukxb39(R{{zUa;4BZ~=EC+2|A$5-z9tx)yOLSn8$oy3yXj5hHqvJzw= zpCNY*+BZ3jv@$e>UMo49d_2Rs?O7}O@mBFc`f6p3**bcxeM+fBqw6D+=fieB11!XY_&f0S;^3~J>2Eu5C^#5M zIgkAS{~0pg-MvMg>v&w9FEq-(`oy=}rTYbS>N0F9nrg!Vbzw>b%XHx9*ce|?5A7`X z2drvnDyoXi6$KTA|7mL9F2Qih@j0$^&}V-tFMB|9u%^G>JKR8p@N>ru*un@<=ek~V zjt=__BP_IC0lLL2mI$u2Y{!qVSkaKCK(-Cg&+$Z4r~W|-Za$dhcX4vJuwJR)DWXMU z3-YFEmTQChek)qcsc{p=RC$RCZAx;yt7}X3o+u{Z-t9Deyv;Cf-$6i0v;bq2W~AJo z5qj^?X5>36>715)+$yI;MlI$k+tExIKIortzoiu$fSGGMh0bt(Yu9;S-q$;}OsrF| z_aLJ8$G@W~y+Hy{^@=;^@*8faSAtf&6U$!!k6YQY7i{l`In}I(2nGoC@Y(tIABX(6 z9y9mgs2K=6xR*^L)Y0P-x}0xS-Y{?+vLF4~hJp`^3*$NbP2*$71LtI2VrCvOiHUNZ zDlsz#IDg*>o-~eywxV0-jqTzM$WQ+g&g0c^ z8C7$ot$Eqz^I@`5FP3Gq5TsD9W>%DkpTqH$LuXj>AFfZC{)&y-x?JLYH*QKJ5Nfovg@K)Z-s5$=V|gNW^OxVVkYj9#B%h zOG_h#!xATI6y+o?tUKS1#wYl21!RBO!P*?_c%lBQ-chfG5K+cdBxv?`CQ9`=YB_z2 zv#Fpz%l=U2XlhUKhI)SYP{^Kj(51$P`g8vR)=;)$Pr@9&O@%??a<}}#a~#2t<+Q%+ z1!=idypdaoUN{^QDHy=p@r@BBSmWds7N|d~-Hi4{&SpB}K)st$Ox;TbmnbWs7Xqs> zW*<6s$THV`i$6?VJu*BLADS<^cpU3>`^SzM^h@Kx*J#KGAbo(w3Yy4*5+~p)G=VJF zm8T00=B1g>k#elN(ZeI*M{5Io-?Fd~_K+LomBc-AaXu{s(3FRJ*6+M^-bgv<3FIer zJ&4BAXbES5zCu*4c(unLv1?{-Z3=JR){S;dOG;NfYQR+MYTW$6rL!95>=CAVQ<7(7 zq-(1DYBW07JPkEa%D5?ruBhzt<`gEqpH_BX{LLNmi9z)-(Q`F+^&GmcO(5vLoub3UAUm3cHa?*%O5w{2Lm0CNFWL#qZo4v(lIMufMS zUfQQ@jSGs4G^L%RiNINK=a`y7T-K~~R8^ayGBv9~{gUQI;}avrFVAY=AcB-V{0^R?#hIR`4-(M9z6>Bpc?%C<*A6Ci6 zh%CQTv3!9a^)CJ;y9=tr=l^+Auj&R)$GoH7iO)p|$jl9BHNl?v%@YR}YnHiTzWk?>PBZ38tg7Bt4Rh-&HXQuEnT_29`USwkj=JHqu&)Ab+r!No&)sVp*E{W;ko<(X zU;OFZCCcaK4X(zhUwSzI7k6}er_i2|%0ZNZy9q7Hupwx}mQx9+QEeEl*GotU8}p^} z49n6cecI*gwKDkiE-PG7wG5{^d_>IJ4;gf7Rt?R4(WyT;;$ zzPd>t@*+dY>;~0SCl{uOK=SvEL1W<*KCe5NQtIvcHZJ6H9nwOVQ=lA_Mv5WD%V_`8 z!sfQp1=e53Epmnk72_Ygp!xq^bAI9Q4_)djRKrYsDQA!S;}ziLX2lqgB1azGB5Ec= zS`;c>-^g{)`bw*2yG^u6anPf4DgmsTq9ehX6omyu(~f#y}@P;S1!jt z!D#8@!TQvUgNM!m_YGXX@0CrG1wawcx*LikH0J^OQGTO+u)$SaX$42la^wN>YO|za zuXBO=A`>i+zm%kg;Q)=xbupiK5>GzV04 zRf(DD{df`74j`4%95lR6Zz?bUvKbjP zU9iwzyok9|vA|%1AzOjpL?4|J4+Hx+*`P!X&dlFyZ{4GF=qzfq>9pk5U%=bAGYh-L zN3Mj|+$!ExmN{SlV<)Wz@uAU+^`Jz3%nR=Anp218B*bI94(~8T$0BB{nZ>bW(iqrU z@`TqcdWm-_TrynXid^cV!tq?`jE#gG&h?FiF1X_wd~oPI)0sT3N5gl*4?)aJcjrbN z*lBg|A8J{@RjT0+TO%bVqf=6G+=*nSl<{0|p{d|0qla@M-0E1O&hJnP>ea4Wk#`RO zPvqrM!ZI}&_;Wdo`|-Y-1Hbp6bsAd4Wv`_7!O44u+*h7=e|8mBW@cpui-48Ir*inrD|fah}_)Ooe27j3X8JupTJaK92%1Ian778D7B?zX~S(eY5l$Rj-Y=*i$yrH z+D%*>;=k%zgSv`Gh4lqEdTY!koHi;D|MX(@*nFjPjQ-PSXFPdu;pZ54OEktRzE0{FSh4O(R~$Q`znEE5!(M1O%zuZ3HOe~|05Mr;8gePs zq15&tnU`;jSEHYR$hw@{PYl|Mf`o7;W2Pq_DjA}*j=z!fXDQs$z0)ERiwnMV^}OeK zx5XqkQ{etlqoJ^~vOl$O&wr(;g`U)J%+wiQYm({GW5CO z5785^;Bk(>Q@2EF{bB&;&nt^vONiHFueH>p>GUIvckdHU?AgUEkfi<3R%cJ0R8Ofh z80uIkxkyiKoWpFe#4-?1o7k>Q>Vv4N!4FN9$qU{Oks#YqgwDD7l8Z20)2<(6fFc|E z=Zjw7Z)8Hj)|(7cyHrt@z1V3&)}ey2Z+jYXP4&H|4Fr&JALu5A%(}612PN~HsDL`p zEb>%?q-2^NLp_jyCVyP5%KO_lrB%8h zZCqBdP@`0OU;GfjlXKtfWZwA4OY9l9Ja)Od3G_F2^jq%`J`5X$bT{M*m5a$S0? zzW}nw5yHoyR}C_~?k{HfE10q+2q&dg-bYlTV7#HvEoZ`Y-SKqwoUq zb=iwvCS7UJ`cCA<&ft7T(-7=>X!th==EL~R0=})|$8Lff5C3u{#?-azNctS5^Cbzq zx9xOai~Cn#SEth7ZG+<}dv2SSY21d6nj*foOt&b=9ZkyfLVt_QjmyBK>6N}9cxDQ` zkPTQY1;o=_Td_TT=cABS&Gmw*<9`9CN7G+ZbYpXivn11O7w1-r$KLrWxI(|RiciwFgU(|{Me>#s+`94&q+dTEF&9#wm9Z=|UKwj6#N zw@hjTL*2*#@a6M%ZXgCgHkntb>P_2-$Om{Ocr4ox=XXWvvA);3 zis-=NPKT}AM}nxtTU}$Dv~DxCncuUKCPJ!G5^F%`fK)j{<;H>d?4#ohW@I9uQ)(;& zev+RWQX$}~GVU<1S~o1Xc1gN9SgU5qI1KXX7{*KyGJ)sw3hx&ybj}i0daHk%VzMGz z6tC8wGX4slLb7queh^n3M+L=0n?$`Fjp7=9gXgN#!#k%v(`mXw{qDm$#$Gkp?$FBG06ZTI)nhh%7dW#8K`YDV%asGOg@P5+y-wY-9K7)wuvBGf=IFS&X6;DXZM6<)Cxi6aSBfpGh#QNtMP0$a~DVc zE{&305_)yn_R8sYBOSP26kW7gXA&Me?LkZZ)*uf`;W*TPNqmecz@@g4$il&r2(lWWxfl?j(Lw+h-l znH`WZKFS$R7XneCwKmYOxb;QQbTw-xl!Uk`tua)daBI$R*K79A&_TR&(FbT@!mznc z1D3+rDN)v%zT0YF;pNV(MoJ|6lfZ|4&FY1pU^%~=sM+Tftj69w3OV5(t%d>N=P@>( z-kuN5H{}#Vn)fCi&L0bIP%eST<1??yOsCgVcFhSjn?x4Q>;~<#0O2jz@8k`S*@Qe? z7$-&hij`IR*lRCnu1ygci(Ys3!cGGtfwWCMo{YG9`Q}WKjGv^}bS%-~X+}*&y?XRxEnF45&iA|9G*v8(1ZJqYHFRAbh>WuRF7-IOoplJZI#f**g zArSbS&zzarVa!>cB-Fa;3JI!7OFYt|?aG8qaW4fV)Hv@Rwahwmv8Mz?Kk<%{ zSH3{FeB$U_a7q5$KXZ+KEU(T3=dO~^KYYQp@Nn#(0&dK7?hKApytng-f49?%Tr*cg zebs~_uQwy%;Fx`2ow)~YBmBQI-?w1f(6!{ReLq+z=oQgde z0*ueRP!MJm1h6xX??K`GmrPcqc?O!^qAi)bU*11rJky4{O5vaHO+z1X=hy3-#KIn3 zi3PVkBUjBq?_0WuBj%6!2BoS=qf2e&TS4^Jfvp zlBy};ld@%Bct9=3hgsy_T)S!jVxtBhMK zS3aoY*3U)rgHCw;)|!}J<{fEYx$9~3;#+oW z{B8f{cW2YtXWlJs%qdN+WjFM^m$hs*K^lkMa2c^r>Mx-e$RdoEK&KJ z!I<)fx0l6D0JogK>1$ctn+3)Dt^g(6GLR|#FcxB1B&EH}T1#qI z?mc;-V)A%uRx;zcO4)yiOg_;l*?NJ{%Ox^GeVA%k9y}m?fZe?JC)a7JwmG3dU1Ee{ zJ130roxfRd`<3SFq5<+x38}4e2kqINLqxQ1>k#;ay!q+oun~+NP(ch!O4|k!i%4J zEtq)inN##@e+myoRd@2C{aI|+Wy{IrcElAgKZjjJ@W#?s@}F*yqMdTa#~%!K0!<^P zE2sC%uA$AByMteRzHWuhMTMX9{;sfD$ExIjLWbW22 z$=8`{N>d80!Hpdy#uDEPJZ)>b?cZOXqtvm+JNkO%yM}m?CJ)_zKOkxxPmKuzol71s zvSm>4<*WZPVgEdZdc#6kK)WBDOC+nNF`V8ueGK`hpIbWD1{;h4w2Wju_DogM+gKtH zH(+Uf+RY;tn+@&wpoy$eRhM*Z(~n@hS`HT~7Gy+YSxQ=>tLudprb?^Cnu zR>}YjwcUo;?PDCD^0>=6Cq1AUpHMN1PJDz)Eh{S6cWO46%J$m%MLuQzy<#84HcYf} z2DfKgaW$|YwQ;d;BB!`$*Y{iUb?*G_AuOqh9wVAljHqPR{wJ~7+VE^Bg%!?993`6r zkYK`<3_^J2=aqviSYX#wM+;9Dp%%imfEuu7)~T_nMWtz|+q0=-S!vsMZ6{>ETWIV* zu=a~o+`h&;3|XHUGZLjdm2%#?gGu#kq>8oHMVJF)nKp1m-r9B>OtSKm$iRnolyO$| zx=(ZI`h{aokBw^;sKsV}Fo~Q`cdXlb8(QyGu8wC0F+b6M%9mB64XdrEZ2d3QE{2h5 zm|<4K3jnk~A7`txu)9C3`%4DT>fWY?e_Q&sGNbw}X06ZF8m$KOHCeT-==E58B%2=d z*lUgd0N8!OH*IX(LW2N-zY+*O#E)z0=-b-Uxog$;BI?~)+!zcQGX@^2a9Ci7Anz2A zsKgY%0;`(3_4aAvFBl4>nv}~m!L{{M0-0TG2?q%BYeh8Lo^wxl7`0R5UzyjwEb6IfByhg>S;QD=i6)*_CDRJ(qR*ci`@7akx&5E zBdG;|pR)eCHX7|q>MEHb2r8$LFH39dTYcZG{=G-8YBm7@^8;oU?aEk8w%`ZsI~!Hh z-t-f*`vAVW{{ZzHVSQbtY1K+iD_E}QPZC;|oy}dXG{4$i!JL>5Ap}G@h{6@MYsL?C zH*Rd~+5U0R35kw^el19xpfEHnL{@B6>|TpNYBp;~n}Evp3k7h*V1r6aP#OJvHC1HT zfG*S40mdl7YFe_+xl-I(q%2ak!Xko}@HSY^>f9s#PT0o9%qR9 z29@jb29^x^O5<9~HJIDXDJC?jFps&Lg}w>>jtlQ>Ydvf08iD+l@Si2H#P>Xdp6r^wyn;ZUX*Ot{{Xi1-nd!4>~{~V z^!x8`uh82?s+c=NZveLzoS4E4uzown#+F?-Ri>BEEzVEEns51O^&78b_2V;CF&K(EnEj={ zm6vOdcU4w5HT0gQO*-$^ze`@aZ`UnNwADPzVpr@`pxwfa#b9ohZD4yJZuZ&k1r|$v zX~h?9>RU_d+co=Bv$J9?Zq_cpt-j{3^qq&YztnwQYptVBs;z2_E#Wy4H810_rEOaY z`;Y$soHY#S=+I+5k51#DFeqTUrwp7A!!wsoL`tr~=lLTQ6m8g&t(Gw_{Hn-;=#Cz7 z$iGaneSAyxU8NePsJB=e+NGtlM&+8XeW_)1SS3!*c8Y*mjI=)uH6pvIV-0MstUzOe z_JFHnJ9h@QVEmeEh7p$_s^h`j4%hxD&*X4>qm`Pjd1(F9$jCygh5)7h0Qe`n40VYpiW-Y_Q#XuP%g|y>{KrU#`$sI(5}s`#)c* zp0nQS2h#Rb+XI3G%*chn8yg{O{>@6eJEuOwZr8J7_u<5usD{C-MayQfxk{ayENqpb z??(R1iLt%6QT87DYNpxRt8|^+jc6{TVXU~*DbstBk59dE{{YkLnMRhaQHfy8V0AFJ zW;u$WL!4D#CiOO2P+5ulSLhmk5EGmlc=T z&Jch$Z5(xoWpz_6sB^RMWCqEzTdL69Un7zL4Ne)<#08M1XN&*<2ixC-L?KlPgmDRY zL7}p&c0sXr>)>>#Yf|mQS~Uh(GFM&QfNI{9u)>tkbp4IZRo7l@7C0xs>wQU<_57u^ z>gzt%{RwWzwcW7V_FUC~vs&we`pY&dPhanK%Tur{Un1bxCg$iucCNbr02`tmLQR@0 z&v#bIT`UtqrW+J#krP}A;Ov#wCkz|FQL|$3_DC+4QM%Jt(|cQ`9*0)7{-(c;E#jfn zA0|TDR7NxWFL=v|k$`e&$!!Hb6Ok&g&)Gj#^`Efo{>A#v=b^BP#Nt(8u4F;ug+?{` z7*;RUVp&LeOq5<${W&1!AlO#H@CH;2h8MQxtz+eRk_|4I7$TkD2W~N3Qo8zqWpfz35niq>h=y z%*@H3Vzs2sHS=#DD(?hXS0rW-L<6=aAX5Qh1%OIebJx8Dt6^CFAo&=?HK-eTTCT{g z$SbjzYDX7L3KWxwrvc7lb~c5wQo$mX<-;-QIA>%xT3BNxP=UbbMKHC~`EkGt({jKO zYf~LD4pcAX5}s1Pt45Wy*1IQ}MiR@i`uS;<1aNz{wi?!AaLlX{#HF`4EN$&wyF*#J zj;8fyEa9Knpt7@LrSx_hCB?dku6(Uw_ASp_nksuPcIoUh%`Wf(6)^yWcJbuwQ_Mc9 z1${)XW~N>?+<|w!#Z*ntb+c@sUc!~*7l|gd1X>hZYZ;QnGZ0$9%0mS}(^|7(d{j;+ zjF9&5Ad>Yea%WuY+gF=a_@#G<){LnH3n5CbtM?!O05)oaCJ7LZpw7ksNd)8(k%13* z7c9tBs^SjNf+kGvP6(B-SzVc$?!r*dnyg41@$%P-ov}^b(4JJ;#BI_8Ru3bXxg@C& zL0jd!OD$y@HwRssj~hlMouUECjc+2Y1#GVstppieLDWzsl>{EwXrGn)V*YUZh~ z7F9+j`|(X+R%}=ZQaG+z(YT=efvn=?B0exbCbHby?TnkxS zO100j<<%E*`YNXVhG*9yYrAByr+ij4y_NPqtM($rqV;&B7_p83IgC^bPi>whTYOTx ze&W2g!`#&2VH`oYz;XZ}XE0VEK^Kucb1}psq^zn#aIpUXn=K&)d4tKo(X!{j=zpw% zw0BK0EDkkxkChy83&<-KmA`M?Km7Y&+YQ|foh^bP?K@0}s<l}K!6w03mW3*@y>uu>SSa;m(&)hgFz zKpmDO=v{-aO53Nq_*r2DRZ|z!R|3lAbTchglr;7ofY`GY?w=;ko~r@IRJe`V)k3D* z^SKY>vuc4^%2Qb`qb!PZVIduVD=~uIs~{bni);biN!)u4TRfXdZMENam4R%v+S=($ zHxW|Zk2Q?RTeYRha>W&x>e#7Xw(YNWpSr1T*Y-8IqiZM12=Xkya3fpn(;FV%kazck ztwGq+QzHcAaCn7r7bXAE;S(_<4}uuArC#442vo!+kc zdxo;UH#Ob_s?meqcM)wl{gd@yUGMe(0Ac-a*!nHD>`o>oNF=~!MCCa^(2!toMno?G z?r=wW^#PQ$uEK#fRJD5}DSep$%9H~|3t}h$DuFnI6o5tv|0dqW#GlY(rfO%1`l zX7c!b$SZXPIxf_=tM*rUAKsc*U*9i@{{UgAO|Np+KE>M>*~QqJgx0SSYhx0ePYyz^ z7@$z9s}mCo3aOT4MP(6FXEo~<16H7TAcV0hj0QQv3?PVDnS(hP_41VNTS21gE;ZSR zmi7&pz{;$^_I|TR+iHH-`X=w8v6BRjhvAblxCR40<_yCaG9WO)6_rPnnNp0a)~Tw> zV5Y42iytFaXmG(!B^Kt*y2k`x!%~BD?Ey&_2W^W3z37CsC1#DGEcQC9v#q09Sy}=t z{8pY^p7%OQ`sA9>;ii}E4RNY=!3pkqG z)ar7uqJoM;S6B)b0KkC800d>`kCi;jDpRsry&^5AQ(Z}^KpmnkCJ8)q3mWH;5M5XX z3<>}mvL;3_4ElvTD-=50VTE=XUAc(agkY;MEGfFb-2VXl;i?T=D;2_+CNMe)nU28s zRs_27CC?Zq(k_Yr0{I%`vJn*)lRbl7L#E9n5@*mO9*4Gj%`o?_BZfa&nMR%@G)BcdkT0lB~%W6 z#IK)`HfxT~zl}AOH4ga1x3sxKJ@&gRE=__ZPh(k)SWirznVEpf^@}P{ss&A&^}&Tz zWmT&(X##vqrn=A{0;}K33R-yMC<3~37K{O2BO$_;C&t*XO6C-4t*~m>TPp(`Fe(fO z5RhV+rEtKp84@!LE2y#awfLqOL8i>kgwbc^%CF9_WuCTH2u>1 z4IRA=f+cCm+zgy&u5t!5{#}%yGd=kUlQ{?&M6d}?!2ns4O<9#HDcE27YYbU!MsWsE z{#Cisu~m$a0gfa^!Fvl9td(8MJeGOhx2DFsS6a+UHb`Z;S*%RD)y8sQ+}j@CYZG5{ z4X=1yMy3kE*;7 zc*kakkR*dS6994%SkD-oJ7I(oe!GHHFa!ur4ki_+)FTjxL7$n7jWzFjr5g&us|FR0 z0|N<_+Eo6-`q!)di(mF1)%}m5*~BpX6c_;w4seKEkVFs&6dA@Sz)T3XsFkyTj>WhA zXia(t{R-x?2P#{IyNFuUrRB$nu#k*6%7->(z%AGl9XRo{7mF;wM$60{5z?X6j!K1T47_E3ifmc47NTJ#M60M8E1^-CJG=Wm6jmnAzbwCBK81-4{7+3g0tx4IOp4rAyI&%Dyte` z7`9nMQM$H@uLOq%EBAqjJGwd^%UAag(6;?0hcNXJJ50{;a8(Zf|}{<@HPGh&8QhwK;gtXcx2b8Vi>ZG^QdLpcn!Q-^E`s^9X!GR*Oz9qct3 zU0t~1Zk$(t29=_s&~kMctf7l0`TTudnWx%XR?C$W)$4Y8bIX)PG6CE%h`wfyngt|`V1I+Wp1gyw2Gc!Hs z5rhCS9UF1Zy7y?!T2p@cv)AtYm066j+X|kusKwy}kQ;;AwAY%g_wOQkhbJ-QJ zQJz#xsfKRi`iu2h7FOp1^etAI%^W}!cYV{_fh-6BTM%VSepT^ATX>#qWh{X?W-rr! zuduxMscT)>Wi@LlH4rhXu$AsUT8*=$mcTfjrxj`R22>H4 ze<-66eHH>{7VOEHM&YfIjw!ube}H-=tS9Y%qG)^6_Fl7h)$JG{Ga??jfu>^pXSfh{ zBGZH-S7+tBa_iErK&ksN%T&u3g57HzYbjK?^K3NYO8Cf?a^ksb21MXUsZ35wDLEiU z2OB-kLOu8K8fVE~io)tDy1>U}@U~pYr3p=XyXn>#c!wnbMmUanNAD_9OVr_Mqm7#v zDpQwXY7j^{VF-kLJumD}>&C05&bl)o&tRqz+?CJzg<$G}#dcm`SBTZNRZ6h$Gl`B5IYa324`b44rTy@6SqQu zimtm*)3H_SP3$6&U~V(+0#KagKn@<{6HF3EE}>Gl`h0GLn(` zY_qlCel?s6or9qy0PATPYySY}P*T#m#CpIba|JENmHz;vuj5eU60ZI()gcZ9ZsPiR z4Y9&pZn3>$vZk^OY`+S>AvQS`Vo6o}5MpS5sr$wyou!5;%T&NHu}rS9WxE>nUt{Yc zE7tSIab`gF)u}!< z5H2QMO2{Axz>;j@F7n_5jyACu8!EwN#A`4RVgLxPtTkzw##Didpu^S#xN%=$T9ww* zR~6KWC||>Qz(=nOdH|I!Y~pYi3!rc;ax4rbuEbe6VhaeBUQ8AeZy6Tr4jUD!t7Xl| zc@YSNd{v6Krv19wn^$EnuZS?JfGZ*out9>bClxaoouM-T%whx69hqSaAjp;24Eu#a zm?CE*JYfNWQXxQQSWYGbF);`N42Tr4Wb7Ciayk*2gO~))BV>$X5j8hM*vi#4I|ZNs zjF?R13jX8z1#E1s>wk!tiw7KqQ3M7nxNEEBHMRh2d)Jr_N*v8*0j^SEMzT{&AShm2 zOiEp>6e$To%GHTj#ASe$mDtgaM%{&D)NFc_#1xHh#>_1@491QZYh zfyhA*PV+Ey0uQkWjIfD`lLLV=W=WZtn1oKo55#3!41<$7LT8i%H>-e!pWM0nivC#qR)>kIPxH zW6{RSn+a_PCcers4&%xK2*%4iyZixOQxVT|Ss!7~Tq3?_D;M7YWC7)tN5m!okliD6imC4lysgin9^ z_Mf@e_kM?a(`;j)f_A|Bi5H29D9LKi7qg31U5kXYWUf&d@!}1FvEp}n+KW7@Nm;K) znW`&X@fQh&mhmX5fsW;3j+^V1*A3413c8>4x7a@oeMZf%eh%T}v?t;h^KfE<01?!v z7e=$ShXj!nnMpkAy-=q5BW-&Iupy$je#+k*1E#k7C7RUp@WVb zcFKqeAU!F6HxzkR%DHsNR*V6j)79&5YyQRhw$G}w zsDQy148#UfSTS-LShY*{z(p*Jt3_^h5G9oy)D;{Gl_3PpN{S^^<*gnh@)JKMB!yat zU>0x@lM?j|gIEc$g{QAjv)ycQt4(k%W68fGOiGJEQ#G$rg)Bob4vaXlSHYsoBIs5S zYSfc}i&r&cS{LA7XGqOWl-Hwu)Y`|o{pY+L%MSkl9K?Goe@Xk5h4nY7zMg78%7D;5 zO0byB$3h?`cAel8h!Yr@nCwIb35oSH2on#Gs3U0_rUvVm%V$e?Bt%Q@l) zQ#k?|u^5@2gCcP-f*@=PcUriXrHfX58@Twjn6(yBGnShQxh-&-z z@-}D^h2XyK+{!|PcI4nqaW%*jUY{kFqM}@?ZbcQ#1y+!(4O|tnc_n26x-jk?-E3`& zt-4iEuDi8S#rk=c>$*>?Z1Oc{b_ z8Hh)CMEX53FhmSydS(+dKISL+E1Aa=1{Dx5e0OAqvI|(?*BB50Ak4tWEOH4T{{WhH z9;TyeyOz5d3cA>@Uv?n2&bW2`i}f1zrp~SUw?Jhzh60K)#{~$@eVX*iAgR}=TW0y@ z%YxLClyIa1&;y<@9@?xiL-A#_!^DuNKvg3h!j3tK+&|Dbm!Tols(WtM(tiSsSt+rA z#WmNqLGsuIfLUc=TVl0U)@TUU-PzTdxL2EO1%Mgy;JH>6h2m@PiQ`tA8cIgifd%Sh zn1JB!Fl7D|cCGy%?qEH8HQb?FDhPngz-D@6!}JO1pMZuQfajs>nfE@Y+#rZ&)FK>- z{?UM7CJZ<|3rygXh+8u!E#7iuZEe+PtkSE)$E{eLtg@AL_rI=d`yG3ICgP|OUjc0} zzD;OlA%k3Ua;$Y6QWm1XsW+5xw0Q}o%F8z%t0>;gTs?Zf?M;eUm;so1R!BGyJ7mho zkZU@w-w7wqZ^_ghlSNu@=WyGD0ejH#`-l{Jb&*jX9Zlvghjy46E*RydDl0ODgP zcGstgdIPg1sRsv*RL>anhMO8n^8Gff@HykyshBYJ110O@@Wk}K`)}0Q>OW=D_G*r< zq;&cq%=9qK^)tK%7BCh7X9PrJg8{_tGCH2n9P~d~?oMIEEeRqFiI7hqNi&uYd@~p< z%*Zi-ou;nu2&0f&Szu7X={3&GEYnf@C+)wZ?dWXh5e`W06AWu-&EM?V<>g{&ZzXN2 zt8uI-ZtVX6{5~%rO3;dv1xZO=uD0uKc?$&)5f`+w#aJ$@7W+yg`->|d`%T7et*x!M z7r0gdoUO6a!w$P}JcJgZU~CGt7A&C&v7QQ)xpY~`G7;F?ywkgx(R@qbfPu)#(Uv9h9 zj)SIp zxMed9HMb2Wi|V;iiV*1H+OWJ!dkwW)Z2h0~TK0vsU9E6o5De9T_STp+&J8KKt6VFF zT!MmDIYh66@KtoBm~7X^twl$$)+(0!8ftd-cbT)hy>l9ZhgN!4j6p*yJ0MWtOaP!b z4xFds8M%UApA{~FKx;B7CkF{*R23mFlJt!3KF9)e>~ z$O?y#tAZJpGKLO}2@5UP?QJZw(_kfYG2a=OIaRjLRyq%caXnwX-u0GxkJz=nqOJ@F zp>aVPkDsWE15XE6bg-}HY zz5zBN%`3#?^J-fZ-qo)#^=n}+vej*J?ABHth^ZZXJeg47xIoO$7$}N{csQOai;!Vh z0~X_27QEooai)z8wl#Xz0dy);T@P5QDr7Pt;Ku_zLXt^unh(sR{C3jg(o`?5LH5b=?aFv~mz?q$8{(~!(0LJTAl3}G=3OfZ-XIOZ~u4{?J5 zF`ckl5&#TA)!!J}+z!5|GZBLcDHjbleQu|)qxL`2w*6BhjO?7Upd!}5S0qxnSGL;9 z-^ejOB*TMXO>)C6U^@Eg{e(kuQBJi2TDDl#)QY5lZZ-uCACnsEz&2A*vfx!;0<^Z+ zH7)>p+wi`omXzsh)CFidVqyU}&NBo;1_UEBxQ8Y(&M=PULkFaY$gK@*k&#+mNd$5e z6}r7RjaXrJu1Howr3`eeuF9wcd_?-`iRu3UV&C;m>i+;>(e`Q`Ou?8WcO6e|os9aK z1{H%KpUO3CMn)Oa7E+8r5F;zsa4B7=A#9V_izj0UB48QxIZj=P&rU-mh)fPd{g!SBk^|Z?Lbye^2{`h}zq9?5E@>ex$&|9J4ZV$GXU`Be6o3<{TJXu{ed{vooGXNQLtmjyaJzm~8BBg3>{}s2dHRC~5e!QWH~O zSK3=!>h4!+rGy&qMuY-Y@K^1JcUa0w)lVkiLD#bfYOf)r*ppje0p$FK%VSc;vhE0w z`2c!vu%?JG=<*eM`l};am4Os=#2rU@m?4A_GY=sZx(E!;CT2wJV4Sft4q+1*hbB1g z%ZbOel_RU`u#B~|YD$K{jEd0mSW2{dL6>XGwTi}42@@R$#0iem)%$(FS8t~Mhfmn2 zHRR2J_7e@yTPtP@m&kDpz%x8xaA5-saSj-i%sEolR3YOTMZ2>OBn2$(Yl35kdVRm3aX7?^-{ z;|D>|Mh6H^L7cmTGl+0zCLsb85UxQ^z%Z6;qsu6CjJ3G7mRXVy6=2K89xN=^25f5y zWel^4giJ&5Q#0G3c1HC70J&RC*WRmUzQmq5uGbL=NE{evd6;2fBal~}rUzjSggVto zV`8kkXo!{o4WF5Ej2pK&&LMLNCIcC#IPO96kXpiFwi^`3e9td$DG{tV7a8mItlhuc zZ&$9Ay_Xy4r8RZGHEp)ful6m-{9-t8z-GmsCT6!4(&dm_V?59$O>V7i-P2y> zv09`dP#3DYgScNJHHfO<#YwWzT1tCN!s_g*VHwsURWo=zQEN&NMBa@U}hvzi)T{{XXRji~{Qw%0*N zu(SI#YztP(ou;L)#aVtpcE?^=frKC!atQA@oZ}D-!mZQ}#k4ZQAeiV31|PnL7?xRGp0oDbYfXFg z_9;?Bc^JXmc8Us&!t+)ZV-X7x>fN2dDz)7jmx`lXCCeMPhsqh70Cm3sNuCG{I$)LpPN z-u5l$3d{Ks>(RRdM5f~TRB zj^u1IN_NGL8>pEAW^xlMXJ}3ce9yR61_0wQf+s9ML4bAb2nJ)P5N8r0=)jm%vTToXmG$-qo?%Ue>a{ghoS4<6Pfh^oUsEbgloXtZj*8OC*h?w|TTUSu=%)~3Vqy^+>K=;Bjq5*X zU^VyZx}#_@Z4sSBGHiwiU}5hWYert)AnbN+j03Pt zV(k`(Y5gWJX(U}*L0W)|Rz*6+&2`o5YUO42S5X8mBef_PW&@c`qO^bhF{buW>u;i) z@7fDInP51+)uH1UaW-lywI){UTD@v&F4lz&c50{8mlkZHH`PmV55O1X$x1zTdhuC7 ziv4cqk*o4mX0AgzBTFgbNrycNImwxsnd!!4!9JjTL*8NrA_o%147Sx^u{SNQQ@KUG zGcz*@5E=CWJp{xe7(j(gbXJyX=3T3Z_3~EJ?De%|oSQ2cuwq!V6A@U!(EUdEW_tes zYqOyG9_EeIHW+SUKeVk4;6t$4^X5I>+Tjh*-wozH%4N0`BW^a(hhUkDTqdoIM!E=9 zK~*MJAgq|Q_F7cLs7mR`*RG#mtkzw(ZE6uh_a*TDP!KcD439}bs>D}DzQdna`q#45 zp2fIcmZ&?}#dwDh`R~i4>00;pB0tP<-tg?e(%N6d9x>HD9*%S!g6jD)Z z11VDRPW0nKaCdXH8l^nf-kwN#Cb%9n+EGQREhrRzYueaf&jv94wQ2`Yg@p%X85Tm% zUNNOY37(}wp6DIaAeg#~J<;PrC|vG|P4>H#>=)$T>Au(MzGW6baqxmISwWtXza~MM z8SOO91MQtCQo6EHWF0*PKNsqO3Nx}O)ZrK&=dx+?QG~t^#Ttbyzb5!0wDuGmX_O-B z6egbguTevI`@i1lzA$v0F70NaK^8*<=P#gCzFE3ZuUJ0TZCSQ1H8-*ySkx{PxOXl_mQ?2ju$0{Pd@o{%? z;1BN@b<*miL~cro`E4`k#`)4XM$GWus&@|uD7Rm~fIWQF?03LT^2HYGU}^Q5QM=dC z4R$q-*XZK%Q;g@0hD}d=0?1Kb^-sl4vGT_gydR|+kb`9#^Q0R*ZwFDHG^KYq=TE&y zr~ldj2mu2E20s9#F6hRc<`|~p-{45)jwaqhK zE7Ce5qNfzHp0zr!RE=RstFxhn-_c7W_RD!_x|P=KF`!=*aSAYmbH^Pk?st=*GyXC+&YPDAT;&(#rwmYR^|v=grs?CON^hPbFhZQt zp-(00V8ZG+_*0%}8M*1FN!^s7%6Yk~#@ujXolp-7%8wHhqTG4X9BA5uY*sm%yJ#9nrzemod=XIy!y=ZmSwJ|8N0F0sa(G-mXryG$WKCpQNNH;_RQ36Nai<+$L%XX)kB=b&B3Q<*E66)#B3g?;~SItnr8YdJg zu3~{a8BtKUMQRQ+$vdl~Qh+9@$xb_ySyd7^IVio5sk$ec1S%&9iC!h*UM1pPGAWFT zxm7BXMT)xWR;sXFOqvY&6z7AREIrE1saC-f+-|5-QDCs+g2RnUy3I1HGGGHLvJL`a z5b4MPbbvbwV3;ZVNFz;3N9H$4h3e{_iaIYO% zCkm?S>b0t~aO_mz z3bzU#H%#j3--#rqbxY=}UMwi7P^ug!kY!LASSkRls0V7jcsn^vB3)H7wRk$O9Yfrx zSm1=E;x`HS^WjwCF{fh90y9sNX2XkJQ-cLin*`8LAcF>E0KRIxZmPOhG#?eqfmK$z z)IVZ(uMC5t#WKxS=($8)7smu!G&o&5x@+*QRRXnmRZ{8oD)6eZZWL7ZM+OS&rCkE) zxZm4g2?^Cgq9hh8QE)=73e8nitMGgkY&c!GyZ3Od6+@>1yAz5ofoxL|P`au66;q;w zL1UO#??j%e(FjPEP7`9Tufpo3UaGZJ&dHA(swWDm?ufxtg}77LFN$CsVZqQ3MI!ec zNeD#Hrc_rT1Pp9}j=Fi-4VT|4`4bxt+=3-JLtm&Gu)Y14GB zlSP`Us8pg1s+bl-uF1)B>~s`07Aoq8;<}9!x}i==Yp>i_-u9*s;3N&zm4&hX~ zZYA+SQ-wn8K{qrJMz#Aq{konPS4?W?0qvJp_A0Kfs(X`z@k|$sgOtpf>Wh_DSu3X) zp(V*aCd#}hS4pC&O)IAy7?(`l>g$rIwUrU75gA=Ps!$;^*8oV#X&{0YlJ&BoR5esp zG*l|+IjRFR89E`AZksi9+1uKD6N+Y-8d=p!yPT5)RKYV{2*eh4eAG#PY4Ab97;vr8 zG-O5w%IKF3tx89~4+(npZ@lOn=c!@O~WlSB-dGQvSfAY*kON8->-^HB$T^ zB-kd+JnDlE11xw4T9Z1by0kX&-Vx&%Jy^(B@e`25MH1^#v9@V{z@uG67sl$@J z7fPvBx_C>qIY7D(CPm~5Iu}(DV7Vt}3EHR&5paMjR7)G78GF-a6oC{q1yab>xu8LH zNV@6_a!gcVnNjRSlT=kWMMv0YDyOwDEcQ%Gr@AE)ZdZwEmvVE0U>FP)17L%?VuOHc zx@iZk_BAgiFZL5^huQ2zw>X5tq3zHA20p)`TXCaw>z}6vS0H z3n>IbV^f+SWQie@f|$`-iij3vc12yvG76&TY7JE`Wj_bV{e{)|SM51)a6a0cY?~DR z)fISGfZ<*&jVhQ}l>8W=CTJDFO{g;HE@ZVmZ^z)eJ;{hoF7>I|@l1NVQ-Bc--nnpd znj3MZYf(4N1Dda}#46c1O@d%mQ^xB_!7y5`b4)|J=}uOl8IwrVr-~v#o@g5fT20iT zLz87F94wtPvZ+oJghCM|-3liGF;ALh+=0k@EUCahu}&4>+-R%zR8>_@E{RoFS65d} z_BQ(g#+)3haH6*cehu7-6;60o*s5W|m>_wk#Gnw6EP!`Sf;^LA5Dk7c;7%yEDW$-s z;7T0$b5%}=l5iTRN9GC0u4;=?twu}DfbK>zqj2h?yM;wW;Q=a1nORcEhVMZK#S?Vg zC_m9H?6hcu)oB3}DKn;lbM99PhlXi_s-mKMDywjIRrV=$RZnHz0Q@V!HSyy_#;YXg zwX#6O-50uGdPW#0#Xbs*$P`Hk05MI1A(oe*K~V5@OQtw3YpC5YuLng{YGmT+ig9$# z!2zjsp7o@hQ*d<}v`Pb25PKj8sPApX#vxGDAyuYDHU&91RSDuPh(lDJI7hdD zB@&DKbXQe0_G+r11w={VWnK-3vqgRvR8*%HZY~02Zb^KiDz1u}nBru3RduUO22(il8ajuL(`1m6NilR25cMRrWzw zR!tzuqR+;hE5^y1>CMU-M+Q|2oF9YP7~YD-2@uYl-B-CYs?`w8hLD3%4OB&n?^;cl zg;_qDA{kkNvU3`tL~~TSs7^3LjW`ul!3t9p7*RPKSu;A+907tEju$}`^BpF|3Nku^ zI%qgGKovpcY=8>4On^jSQDH<}H`!fLF<3gOWklR8s2yq`hd@f}HG-=D0OQ!{;o0m_ zH^nmMx$vM(!HD5Q#JWDivSJgeRXcHTq8*n|XQ<|?jP!$wCryhXwL>#m0xXQKS1_6T zP$L!1I84r*2IQEIOpKcsK#+!>47CKpG+q?sP$FV0VCkKX#ltm$N(%Tv`97uOK)-YOCe%DVx=KcN$MD*s-&Vos5Mf7kRZg5)Gf`1F zHB>>>FT+swoDXb2umHttn89?4;Hsb4aI5Uw*lZExoLOm2lGLt?Ba-9=6i_GKz)Atg zxXn~E;F#i=3n!LrjA9+plQ9L8JEj=~U4ol2n>p-8R}u~a*G)h&)JKdX8g+wvP$8HM z{IXCcI)*E3?Z9ouaStJPP?#+!&rFF8ZxYf2#DS#uBbX?^daaF+UA9k(&1)mI3^-Yp z0@b)G3Je7-pV(rlV2rFVDVb4Ui(NYAh$>IYAS_xE;N*%}Z}{Y{yAw%?dKDm3^n-PA-@&QWOYA z%oPn&raswE6J#9KcJU~bO@d*HafS$@f^kBOUdovQqbypgIGrRZLEA{bGy`E&X<~$j zhDt3GMkNt>2^MUPkP;0c9FQb9k#X)on@O9wK(!v6)a8*BQY~4AQPp;Da-1-CkGXd( zvMxiQ0wn@4qA|{@QTj#Gv6GQO%@A(oZHS7Cp{nZ2nIL{;%}`l)I*SZh12J3`J+@?a zhlpG%-;EVV4JU((iXaW0LxmMrk+xjOkp0V&A||mWiZUvDd4)Zxxts|El;u%Ugr_9d z>WfjK;QSmQ6RkKpZa!p;bQ=cheq~BCfY)&$LXSz)nl0h>AM6!uh=sN z$bCchE^I>b*{0My!gG9kQIHj8iiAzty{k@08BpK?lnpgE!4r{DAko&6Zlx!%B?ys) zPy;0}J?IQBd8T$bl_;8(9m@!~kp1dJ=HszN&BAu~rx>cJxj~ugE}A&06X1tOdIL>D zuhYX~A|wJ{t2D{)$vwFuwJj&!p9Cixj1Y;VW`PA#RTm3PXk=N@+?P|> zrZ_E7yVPbDfOPHwP>$n*R^(?B5OKh|2US+3M8L5r=4_fZ;Z;SX9`xp@WQ%aG1L@`9 zm9htc+ueweA*zn%(ABhYR2|$XnK!jW zn<>V{f~QnV0-?R7rVUU4Su?OzFPTKd*9KZem0XgZ{2V4BiUo@w zccu6kZop0_FzQ>rg65}gU&@Rp9WoRng;VivHB{gOr%m?dP`y#8bm+Z8GRpvf5jAG! zIUy$Lp$p9VOMr|)Z!qeK$eyI+F$Iru0QVKBeqoVXWNDh5ewh^-hLEpnAmChPnb0NJ zshH8N+MzsX-RKD2o-d0nAXULCF>5-~FB+A(3^xmRV5ZhGaoZAvrAL)X=Cw%C-Lqt4 z6XTJWa4JC2?i6A%{mL+YqFs|wiR$DtTBKdA*u29>EC5H6a%#Dfghx@0HDW@J)LqB~ z_7J55T~{Rt004+@cd*0V_5{ob!I|!vngY#NJ&SY8gm&Q&!gDo0hAHAqi8UwpCnmWn zhieq#>565k*=dyubav-J)K4ihTBw-yz*zH62m}{oNg=RNo<&2{?2=l9SrgnRe$@#8>{itNsDsC~9qqX3 z-o>m`w&xutD1tlD9?;;~Cl^eddo`lJ*vqjH1yMZHDGQP)IjIYfC<5t?c8f8@e5NWO z-~y`B^cGrVx=pGPjH)Cm5!;pH;p{+4(kxC$#zmzP(U3w|i}o$)M7>@;=((sgbtp)6 zHTkWr*88%L;aOu8^6*Rv18&t3B5FwPTbaZ^xlV8MFznLm8)AStf7+w88X(lV^1vfW z(mT{(JXUV?<=0c4N4;tm1%d2ZiKz?}Wo={jT|TT?5%R?RRWU#dgn1~((OSOVmnP%UH#$xv010SL89%rXQDV21QV z96~M}E{YW>85&uT2faA(8*p{fHEo)htq_TehAM=Y&Z|sq*oE}%hR>^n^rPl5EwZSG?r;PYGCk;iT)zvf^QznFhVEmDD)&ZG3of9t;yLYMSw%kDEPwX`0ZqWqVc^F}S)wQ39of6f_G4pK3dD78=<0q9bCt0G&Ey z=7D~gY~F~CxWq&bxlYIMuB<9VF`xM0{?tM2DdKyU znT%KImV1@fcrw$d0#kT~1YHTp^;ve=Sy@`h0`;NnQS#&@4XUBG2&Z}{biPG6H?pWi zCM7T^o4G(HW?kw_nj)ps%LEfL2yxl;fY9}B4noRXO+pm6Bj7j-2Su&IzQui>%Z!2p zxdY*x2x5TUh>Yx86yS)kXpweJID%Q7NJN9TKNMMN2B3yLnXf-(gvD6eiT1#ZjYlJr z^2iRvi3?`STqCH&^$damn^4p{)|^=a^eU!|?wl4lNjPvZYd;1AwizcCBBzq_<}hI{ zXoKOMB0liVYYar+TzTTPZGs~nnZTFmH4}=I0n;rdFfBNBMCRTgwUEPeptsZl&?^Z% zcLrmZ-bOMGUBNdFwMLWLaXgG&h)q92(bQ@2=eMxw>z z;);gUzr&E}&5cL!smfRze=^#5qhSXcnQGiBFGKs>Xu=32FNEtIJiVIEQW`C zaR|z8N)8PQi56ud1(j%B7A(r{6Q_|>bRET1p6G-K3fu$P&}~~k< zk!Z<@JngAaxRnXYK?#8LI=eCc6vVthcoiQ|14buK>wW7@9gg{45Ja+fCHpPq-IoLXl6+#Ft*}FV;_nH z(!2?D=EdJ!3p*9Y;VsLFMPM2`Z5z|P>0z1=U zufg+LiSA9e;U?-&YN9wW$AED{ffhlEy+S*1mXj(HE{5UxiYDrQqjsdZMnF{E902p;MhI?j`iKw~|+=%kpl8|B!Qi57Ba~g~)*?ydfriA;X2I<#IyMF_8|7 zUj?@@SvHw-)TOf!(KvPmFH0eZSw%@{S*b|#K?oj~HRn_=7Oaq)=$cJvk zz_Z+=2)w0@QweDq1!GhTtJfq9v8f7)^WjuR>0m`#?5E|8QZ{ml0fE6%iHPr4bjGQT zQ;MptOQ#iF6DY)rJ=5ocCL*CWNCK*gs;VS-eqoqiWbc3ef}Ek@0fFpCm$1&rWa>a0 z-B3Dy4uJ?Uu(D#c8#0E=y!4l0eVf@N3Nk}zTxkL1)Sp(0eiZI?{A5TB4YG0XXpoD;gxl1)YxoMW> z$T2WB+}rwOhU{glXvW;f$tAU=%w!}s&QT{&f*=~qH8TrlG9DozF4(BVhK}GZ+O^sG z3G6F!4M@0(j6=DiJjihNIghGaGGz4*3t0(2+W zVYKD$Le(s zQIXxrJXxi;#3}2@)To%!pW1lo5x+@lhk1M)^ zSEPa4ALddpru3mvi6YFlx1}9ByJGhe!f~3>srKe|$5tSle|c?(I!uXxH0s6*aPWaG zK%a6OW|cA=iL%@EEz=2gay}BbAz4OSf>G=>Mh{xJ7iU5);07JY#|Xc;ATOkyJ&L#7 zi%9k@xsQ~n$}V++NJDViAkbyuM`#+zxo&7h>PcxR>;XxTN#yrileU45Krb_Jb#51o zZ1TVNRD|SAgq_b~-Q3(QDHhv&R5tHGNaj=oA{dg-Mga)K6S=+APEo>0#w3sn$lyq_ z>pfC6e3q;xUAqv8dekjBSQJ`?YLxl_sSBVyN6ghfm1=1m!!&qOil{Ey1O_Sb!3PYe z2X|XA7=6PaCiGsuoV}v2}J=bwk{!&5NdbB7{rRJ^&O&-xX9`Eie!OVgtz-NJy5% zHBgMB)MKTwA2Y3?KL+GjoQO@fYZsJXHc^~9BukD*=8BJ~j@y{lSrPL(8)i;oUN3s{ zLNYHn8XjWOQQdcVG@>kj`(%rfbo7aH9uyDHxfYk=`Bv1XI;|nh1tKH*b2>f#B?4{< zB=)D4#jz&Rp*Sa>OjKr_gYe{(3$aQ>8FI%9J) z;Zu}NeMr+xqkvveHH3Pka3|%L-d!@)x3<8@26CYRgbbyS8Ga=YDG>&iW<^R{bzllI zoZ4={5cNV*o`N&S@WpA{mgI@Ku>R_-rQS-ZSE2|Zm_Xga0gzTlL_-^kqI8(Z)Q)K8 znHwvf2gCa2Drz9H5%5_x=Pyp!24xlKDQm9ojVSX_W zavf6@2#U!^Y{-bwg4Hgq$<{~rAptRJq9Z3bV3>RssV+zc!4lnElN);-a$tmZ=PALw zBEaD68-O*9l*Zv3lZezP z!~u%b4MmUMWi7qln8>h^6ojxJnn*@N^)~))#AXJa@W*mEJ{yDatuN+m?*}o+luRoY zCs4#6m-3?s7Tt)|%$Kr>+;JL-e(&!oVHvXlgvE7!WMgTA`jE)Flq4$|YKcNoHKdp+ zi-^Vj^+1penj{m9P1Hw->N>v(Qk+X`k4L`CYCycDA_?Ra-iL6p>_cFD@t_=*^`0TGadUhfAkBUlrWSDG+p7e$!{E7xcK>0R?4~&$jq360@M*!F$WF7DdQr;ly zpltbmR5ralI;SsiZ<=3u9Y;uofXpVGiwjXyssj1pzzh&UpS23T5LH3Tgmj5^j1_(; zt8iohOjB`i6?g&#QE_i!-;3m#UZqHc=J*9%RS-V)s7dYGh-M2qpy(BTB3#jtW{AZ+ zKf{UAIh7N+;EOOs$gly=VmeJrg;J}Vts>biQWe94l8l+kBMvTvw>JnDV52x4Q8?x7 z*;QS5s*aMnNlhsCi>j&sRb$=|A;DcB5s{%q#-2eI4mqP89^?swL|w`ou&+rI9luKG z>H?dQv2Uqq-H7t$7#t7cO{zjbIVUFm6%E)n?1xOEXiP>wjUC;&4akwGLRdW|6)8qN zGK)?h&lLhC(pU*%$24}g2SK#OPiio8BY2$1Q`-{JIR^&Di&UvF8Sz>o9v26d(Ak@5 z-H8lvA1V zT9jiUs?*&!f4Ee3;tp~q-RsJXLDJ;Sla?V-f%xzahDZm}6I`k}R&??m!E_l$ZDU0s zNh1`;;+ewnAOw(3nG%^Y4Z*QgRb946gQiD56&;Ang*dn`7Y+s}On8bdali#c69hEV znlUpXII<2piMj-lK~juiJJ7OWL77j$PGQbTg*Z_)g^F-1flNmcO1;n{am0Y>CBggK zY0GzTJZwq96JVSa`6i9IuA^Q_q*FKx#0silKs?h(zhbQ*JKcGnVB3GH^hPjgR<0e& zgy1y2zzt=tHxI05ANW)vK$g_JGL69slXkP0la9o8<%V5Tb zOCNgN-6K?d-7G537L#8%5fO`pEnX7>22AOpUo)YOX37vEksuz+mASn$;wP{IXrZVu zI%D3~Q;DeCAd|?g!VTx9KII4+$PH6H`j?p)>z?z7wjYT$;Tp*RcDAJ_tVhI^857eP zQ?OvSWw=g(Ai!M@FQv)T$ObV z6lg&i)z#C=M1k_zIDjo^+^tfF>5|%j@T0XRFp=pv6!B&?Axb8@d$nsZ3bA;bOX}}Y z-G^3+gNrE7#{wL`)n%YdP#sVIDwO#||!cqF7p}7#! z5gL0Gw+ZSrZqAh%NVKEV9%@7Clb+QvXzUQ(>Kl>yg+dXKb05{OBAzzfd6U?qv<)R4 zMD|q@M~1N-xUKOjlE8bNLQ*df4BMGa8dAK}1n&W5Ey3W02We3fAQ{QdWCgb!vC;)f zlws2tHli^qLb3n=214p4ap^JCT}N_avyr3N3n&R44;B&2Hct~nKuv0B$WL(S+l`_s zWiR+p)p9yEjr5@r)WfDAFaw%n7|(J=wxk9>bbQISqvlEgox%eM$oPc5`VjRHBT1y< zwQWyUL?$zH%y1ZxY=nyk797)xgd{_9RFfiiAp!vTgQ)-x)J8ZWM9hyR(H7=vn5dl4 z<-peI?&vaOf?~P?6iic#CkBW)x^#5m>{S$6wHMkwYHY%SP` zmI;Cmn;sfsS5-m`EV53k!~t|7L34m8MU0{?ny4W2;!~ak=9ns~g^#&htyA2T4it6z zB<>sxg22m_D4QIXsRJ@aBh6|6Nw6Qhts;6iC02S<52b#n^0x&HpeXoDO4I61MFhoI z{{XtYxt^m)y_f2SiD-~8nDBoZ5j8t^!Fgi4d!Uh+kX4Jt>T5#0(erru6dqJPJ>>{) z^zKB^i-`(V{#6}Oa~e!!bBdD}F*HAQ)Ug4`b+b5DG7lq?iIJm<_9#pb{{W&|ds?VQ zSF_|DNP+S}KD(H`Zht_ST}1|%R& zOhHhD1zE>osx|46dFJiX6z8N$TbFzkVrm0A5EhI{h~`rhSPd_emnu7&0NaQpL}ofR zlkXQWj=6T66pTw51~Mb447UFO5;4+1G&qtG-QF?;jXO|?bMN_7wHR$}#@l4G{xYMt z>4WV8OGGqZ@hwpxk@DDlR-+dfgIy=?p=i!FZ{|kSH;Hu60L$+tnH`bAUPrWW5{#!4 ziH$%XdMZ)#!Fo;U>Ws{jaN z-kyFK{KDOftZUkk>xr{1t=+IpT8&#^RQAIgOn0qp&IYKPgeD6ZNCHL&Kf<+cM@)b* z+#ItDcWZF#+ks-=G$ohbiiF~I)1|SG%P+?TIsCH>TJQBC=VZmdY7>5kG15K-Z*1Cd zG##T-hSuQIs**VnqZrrAkdq8XN)m4jkC)!0B79XHfGHiq0VYw8@|6=m1iE)6^po>X z>5P3%7oD6^7Q{Dj9I$Q9e8_*x(ukJ@pDgZo+hS2E%xjDL+okXNHastS0fJ=f& z$q;EN76{ZZx|`Zk*IsISpD83h6j75jB3;fbw|9-C;sdgX#?qJVLvO?&Mxt{I$b_r8 zCy|Ny2Q?CUJv0ky&67f2UeJ(o2B}LZ#c0?d&hnI|00WXT^;_b6qZb471P;KCI$>eMlCnB(5q<}QHws-l z(Ew3eqG*_4P$M-U&s1h{Oo0P|O~9^ZiphduT$fNUi*RlgOp7PMR@_FL=72l04~5i3 z!2;ds_!mX5_BEhR1N; zD{eBla}(H660jsq9jHd2>V8Ga+}zw98?cC;OLNo4Bu*kbn^O9s-BFV3>5k+`FNA_( zTlBh)L|H&&B4yf56lMeYQ7_w)5c#BBqD9R+0P#tAn}lhIam$%gpNpxY1}WSV;y>y} z0K9g?YRr?WNCFx-#YA{xA|eNR+mZ1+i0w@)$Qf$nleX2z{c>BFGA_nR*`N1gbT8bQd07A{0aAMtZ0(r9UvfgfhAGTUw48sL5tY$liC!73+tD-73Bj!SPUo;lmQ|#(hRBYWf3_EL`T3)00P>+Ri{{k z(a7ngI#4m)FP@)vn(0DYd*9|FZyCfN_6j0$^5pf~Ac((YmZ7BY3TlATf4Yv@iHz}Z zT^}xlJI4ZqqIBjmRhyAkWDM_%Q3}(KK>+iyMmk8FVg9u;2r&9&EBtA;ixX3=$M=^` zwf3NBx(Urx?z7}t4Of70Rdo`**Lm1C=K1YO0*E6TX8#PGdN0EAY&FC zybofcDc*rF1;TP!J_(b{+<}y2Mnc`?YLF3z8)H1wV*2vh6mPvRx`5pMX%k37fO4;i zRQYhei+oVq)%aTDzvfbilS(YW zI4JIe_hvtt5J80F@|#KiXrN38fxmyv`B5;b+Lue>AHuxY0jL)+{#B?C)18muh`A3U zfQ=e}p(7IH+|hy?+w$ifyno?DiVM|FAM(I+_>YMhkxBqUleRGv!f}k~H+D z&D+6&V}V991U)qX@jdHa(K3a}jYvdp$oYb?m^~-*e6mN)uA38e(%>s5Wa`QCPaKz0 z2FLx$Zo_Imx?(`J|EwS%mfXd@^CP$- zBPm=9AHK8@1SR1j&ZP-dj|t2#G9j3UQrX_scWM49+ zF-VJdDQ-l(jUxCI&@MshmS+{JF?XSjf{&H91B%G5vSlXdfUr5EJ^B(|$#Pehx71|X zd9N;!GOU{n(AyM%9xmIJGr1i*laQNa4IvdBxrdBt5rL6Me?gz9lldEB(I=z*lm_np zR;j$M|3@mXHV|QJoe)m;kmbu;)l4` zPjqfd#-+c7Hr!>})JK5=J?W$kEYwT|)@r1n0iek|jnjbd;UcS4!JNL9r+VvcucTw{ zUSwiq2hA#;Z5M$Rc$~iqJA1Dgn~`T886hv2Tnk8j z)OWT-j?9bh?fk&~Q9zk7HH?BWnA9cWIQ`inyF($BTz(XB$?TkUM;u%J73IE}1Jcw+ zCZbe5s7538BjG0=NWLLT8-PkSmTiM6?aU5}0C9*>^>ZHr>E-{{XY812ND%SvNnaOiZ+AFk9}$ zZSHOjFDHy`B3+}>ytxw{y^?iA&^Y?m{I%(p-6N%mf0XqjE$yOZx*B3HEx8Z;fAZwC z$8DkVkD5P(uDw&0hANx0+mwxq=|WoJxDGHiO}hx0i%d|IA}=2eEAK`wV#TBmW0G(T zpl$*%B#566B^d8RR6+3y6&M-HF{2WkNE3q_r#4iZe!`pw%yP^l8G)2!f0;-&k}e3R z61H1JwvT{E%Bu1d;* z5@?-KF+GWfC{p)Nz!07yq*ZE`;CHEK{^|p6PQiy_BnmFnXp2CV)Z1e@7E+8)nIrR9 zCiw6K(2E{OFnJqQRM32^=lGM#LNUY;>NLK7DfAIzc*TTnh!pTi;k)Kp^s z0ILAN;UV)xBeNp#yiP1|O(Cyteh@G@tx+u@fx7c{4v=pXhv8Z`L1ReBV4ji@r*C>K zCgegSZU#6Tg{}u&NwEE>>*z z)LpPZN5DfD!9*Lz?1|hofFvi4A{YbMnpdEK{mNStqf#8%PjUK5cF3`jCvj#FG&CpZHMPE_Sxq?rsV>{{W>Q#J2Z_ z*oiV4q~0B|{{RYov8pmAQHT|hk?D@57;iSG^Cz%t)%8rR$v_=a8eu4nQLoRZ?xP?< z6R(*YNb*j0jJF5;-LXzllwyw8s)2v1(g+Vvx+5c+5#EW6$Hyr}0%Yugoe39Y@gT(6 zteh}7f|#vPi{hc?Zb(7$w|WQD5xxLvmtqq^n1>N}AqFiWXTat8k=)ykjloHS`wGxx z?P9H>Sc77I=|Hm-YPqg>@v^2U6H2;dO5~Y2<5XDHI-;&T7-EALVvcT|IpTeS3M0K8 zNXUDaQN2@eVuWE-_>^Ud!3h*Oq1d@5RHD*k*)1?6!agRlU;rw%Dwaf|AtDhICRw6v z1%lN&Mm22LPe>L4Jj+HkB?*Uy%h!HXkq7948(yA0B8E*;exiY6J=16myHOK_cGzl& zi7p3fsFvhp4@}+2`INLsfgZ9Mhyv@(+Z}OdI7s0?%!ceMBY5HWl}jyPzGoCXnH^s? z*(~Vona^l_Bs+h+nsTVIv-bsXd97NjW<-hUApt3Bh_O+LnCijQoY6=-F=Z#0P_#{m zQIuJwE~sEZ@hL_H;Ci5MOp(?Yz6wS-R;cdg55%a-f*&yFflz@jq=%Pk6AtY{r5MTN zV%6k~YkD1%_38B0VYK$O|oVe}Ch zrxR&u`9BKgEa+0Wq9F%kb?5I`R6_>{k~D#czvWSp4Yd(Sp|!FfRUFs_v$jsTZ4VY@ zYnXCx{{Rx3IVJVIUHlQ-`WCNCu^CMD7)W({;=H}Vz3FEYd$Jo|sKeAs4hHcWcPsUM zN%pr)M61#nyFxM5=YShd<;UTEr0`+MkL^ssuuKB#CL=Lq1jR5W#SUvzn4VTbLTxNQ zlKG<&bnd{`BHgi4fn&HR2{b3~qd0=_>6}f$AIPV&@k0Np|xmuH6NXo*x} z5z<@~$+2Cj)9?V4?vEO9x^HIbkOMSpimGJk)1v9Pxekx+RiQwGR4Rp46QcMh3nVaZ zHcSV4FVr#}xTx)cM@T|h;B={d{HiExr9CZt*m5MhG3bT=<2%3OYuiA*xUx-q44r*f>-V z+TT))<7rfKIIU4SCzvcf$OAadjyGfs%W_WiM&vG#08S*yiSI?tFe*`$sMP-e;bN$4 zwG;1A0)hLgq5*R1W?)pYZcEBzk?cS~up{F3M{jcUJ&nydu|D9M4O^U7U`Lt^F~aJx z1omLIc7#BYA|Mamr7gNdSybAgP?8Xcu^~nJ=Ws3ua`NFCk>bE7Hx|hm3Y=_ z%|~t`Qj`ms8#-;WVG{2@!cSi`H_E1w+Q^Q05`rV{D#p(Az1s>Mv34- zR62(u>a>;tHLF)wT4Wp{7wk=V;{O1-KG~)*H&8or5!{KTS;k5`(VVt@PS|TZ9Hk!= z_R*Y3{{T0%D&84^^0NN`F}KsIlG`@~UsM$+{wNzC!n!Ul4bdeyw+gBi)rxQt1fb@N zNV4oqw7rviHeAsh1ESgR?Olnm_K9aXC(Q|B5KY0j2*#DFVuGd#lkg+o*n~DAu2gpa z0BUG$nVIen5jb}(QZGku-m>J49S}t`!4|qYPimrNKiX%~~XF^l?t;w@ZZpB{-u|yg&I!p6K?t#Fes-ox{0##~X z6%Ea}X$7b)On)+z_aID71h6VBDDDp|69`H#f@#3%2dQ*zohKlU+ZtRcJ>zhiLE6D+ z^!sF!K+0Q$>GctC38@8WJyIpzG(jzNJ}4N<_c$R49aBCdnx(OkpzKvTMT4VZ2;OGP zOnf+%^%HiwG2MYMI1;x%%MBBdVYedWkEidcZR~AF%-XK!r_{~)CZzyp>E|O*91nMf zNcr&U!*E`mh&;6%-HI~QNH`Th>4>6yl@6JN0X0tIL+Td@JIkI*GZT=5F&MTMLvi$P z!~STakGz51kLpg_ptGDpAk<)UPOjlM8u8T-h6ORcXrdm8#~W~}w;I(2E&&y(GGya} zp#gao;L4mI1n2?rMsgJa3Zfp?S*q}LnvF=5i;)@+!zCT4ZRv?;5v@M-ihaU4H~TzI zX_7QjYzow)dk$vbJ5)5N*#2PHd=*f!`>40#LsoH=BCFM)4!#G4rP1FWlk%V@3$R_Pq?M)COe(e*I z!sL|aszo-xEONRmWJy$@KNQoJmidKBj)=>0c8moh_@(!d-QOcww%D*ja6iQC^E7 zz$W8V=Yy5a1ZlyrKrJlcJ31hS7;z-oFtD;VZ+ z<~4klsTV8-D4e1=Qk|+%8M??O5+Pii(UgdYyW#?jiEW_d3p(4}k$uK^16W5F)ALC& zCSF+7XCp^H%zzRANYJ&ErM)MFMm4>p6t>KCv|RHGdtx%wCIHbVf5ij@PMye-1ju3t z{{X2JC%X+hjpK=Ul8j|Qazx0}G|>Q@!g83|YTGL+GINU8M#%}5B$s6pM^-AK`;&o1 zW*8vR?Pa9cn|7u>#Zd|_Tqm?ZE(jf}Q-pPGlVMiuMmkJnF6yj1Q3Qjsr$ ztxiETfWc&e2G&68GMNrtIH1g^nbeF?h=GWY%yBAGBUHfYI0Rgc1ipz72mbnHZPPT=w?qCVx}k8;tu32PX{{{XEW zB$%I+qZ|bV*DCuL>X&@MIkU!hz5`Yze;PnTBg+=jXg>fCxy9B>bQ2@A%Z4G zq^?SfsM91&%WcM_oPO#XyRD)>;cd11V!!cbif3ZsQk+l5KNM9aiz9ahLH4i>to?@VeFU_FLSID^`}&5|}|Ib9gT z^`n6lPwHbWQMfi@m_dgmS|k+akD9%B4Oo-_49(ReL)^8QIEZh0j9~&|esWj^`jZkS z(21g?bc}_EB`&WnbRor99}?W2%uMCAgWvBcLu_iB3Sa^x<7CM4MbwS|0BR(F1^|>*P>~Oi zLNF1D6{cuPb2__fp|br_8sri61d+WVUO{lh&_MZ@gX@< zRyd)!1Lc@nbv9u-%TDBW=M#f5{v`WPrYoyF)6iIlD~RjYnQaf$d>BtM|*T@AaumkpUdG= z+uQ#D5sA_R06Fe`NbQmfVruJ1>N!L|>_F`77{vVhq;1Few;6x?f7Cq`Brh0l`GkM? z;)k96P_x&{^z)GXvL+}u^$$=WK4a3wOOU>UbIr2yXaJiYy?mY@=2+$5iJ zB$0`#21Z(?XvD2MdzwRaN%cyfEL8n&%+Bnj1ZA};2|=>@;Xl0v37xrmec%sNWZ<8w zG9S#fZWAd1UBCc-)uz0TD)i&3Ju*X6K4BP;@z3|5Km<6)1tA*(Sv13p@k4So0}N4t zVsR%DhN2@Q@lMrE8bp@oL5m30*?UxvCyDMKM6KFJuYs zLz3o(O1uJQ!uW zlLn|bOY)*a6DD~AnrEQZ8SThwyJobJ(l56me#xUKK<&xSp>UiVV9KB-w=$Sln5bHd z5Ns@mmjk$5JY!OX}Aq_o?k{9detGFb4$x0Mpc^x$CzFq=bNw z6Y)q75QKhT%z*xh=)9q~7Txtz1AcCPVX5J?jQv*j=KkrU$e)B!BnS*ay4SjCgbfKF zbST7*i1;8Z6YcHFLIbITtrCeJDGg*iw}Brq{{ZD5Gfv_9he`hGBhl`kPa;3WkZ?Yr zHFT|ek&fsis#erLAIteskCYgWiOMcRUOS`aKud4SnEB#f1a^F~RpC@M(mAe~Q21Ru z3lw3v7}@YmxMLtiKsGt*L_3j=d>$vc4cM9=mnEoNA=KlQ893@A#B(cD=mKQvl@lr~ z&?F~B1m}Td>rub3A}+N`nG8{cfedj)tmq)C!naPWh~UDxm1O6OnsY`3(5G5+x($e` zuMktH?T*@VP~4mV-`oR4FQPb0@yQpHMl>TK^?qd%1qgB8iR7Rj0HMuC%LK{=$TmHQ%gfU|WN{fPLqx!BOI#00x|W8JZlPj5%8d}# zD@rg5rD2THi*p~sFS#4 zO=MUFW4XFT3_zj$%3f@}R<{Owf-#-GQ7ImT-k18Z^n1e5+}ob7%`K-=*gmJze^J~e z{Qa$4ThS($(W7(lBfsdQA6d1hSDCnDx1*l)C*ntY*LsJejlWFW`;z;kOth&^U()v@ z2)K^_0IP6+{4nU$j@&2Y?V5ReWrKGphX9)C7fl})0!YRir)~UXRnQx=8|AzX#I-y~ zJY+UNB0i6LRDAV#w$Y#T?d{2LR$IQ3Ro#FW@_+vTPRHJgX-`(+lNk)wo68||@bvN` z;GUX(WJj#^n|_!9W;2uJrU;1a?s}&A_ZksDmK|$Mb|JP6fWPq+B=Un@nqRdD`H!bD zq^HDpI*iaYq=xN&+S3d_yEjv~GbIphA8lfpmD%>64xm zLW@WMoB_xvZ^V4x1xqLPAJ}vSlklRoR+6=h?wnmdD82LcTI>Aw^J*l=W+xJU>N~`| zexu@)jiM7Cm}=^f#Pa*GT{fgga;149&I@1>CCJ4Qwg}VWBLqk!xw+v|SLuQdYhL$^ zS+V>`M{S-(pCm`8kc1#>DSXngLPVG~C#13bQAQh+^CBoDd`-GK7W_PHuQm@GgeLPw8#k zb4FW|(-DBlX#}yx#)^80jDVpzAa@09dmid#+9iu(K{ihrnLv|3HCftQ12|R%L zjJ{N1{XcK~*6e6b?1VPY;?w$)UQXWb$D;J@{tuI#$53bP|_&>{;ZV~lqlys;^)CtGe&LRVLaPqddNbtvc zBm6ztf7Mg;arC=>lt8eI;_{{&B@zd4*MKt7noIk{9$E{HS+iReSR*Bt5!Y$J7_ zvKs_(k+w*gfMx#x*HRi>wg%3DA&J3iv)FLtp9JMmF-IOQ6O_W-a$3Re$wy`-1m7G?&depDnj-)>aWMM(To zQ;ytW9iV(t0sbhDN$I7zw@~pQhwkcH{ZJcwv8ad|NOc*)QjDWx1El;A+m_(ph9hz! zNBECg*#4FMXJ~A0L#NXE{i993Tf1}dZK#cz!_^JS!Y~FiBjucS=i-7pbBToFT2YW< z4|=M)FqL&%wGazRG?ad>XqeM6lBvD;7q1&|=BeIMY?Bg75|n@j!pfFiMu;0lF!v`Y zlusAI6zh^@LYdenG-G6YQ}Cj4YJdroCqQ7G5yZM=!z!Q!3aLdfPlgJsT*j@!e@+J9 zj$eg&yRSXTUPkCkUzNL{$8D;5k6HVw9-1Zgy);CpE_3RX7RYTo-k=jtln!FLx!VQqs)Q9;|4^C)AzODo|nV zu%j(Rm*jzwXy5^4^CloxF&LqwYK%>+D2(Jl3F75efpSjCh-yD~u)_>^(#eVumL(U> z(*;#T0?j?0Ft%$+T$7YsCkcurQ5T{(m8TZq??lY4RQGI3X!wl}AhzJ=Ni{hE)ad;) zo~b>_d{L2LHlrPlLWG&0m-49{+RJVvztRs(WBe=B`jg<)+i&+rQrz;T*CV7gR-kHb zN~wQjHBJC8Fdx;!=%#B_zs8X25yQDWcL2x=AK{XB>gJk65!jCCxU(EnOzQ&{Vv9Q` z0J!F!A-*Zd4en8LMr>+>ZF#zdj0z987)$X)5!&T{ixPP%u{e?z9>(s>L&`#dN(2$0 zkyt%lT0AKdVMh{{T;Ba{a8EcX!n9kz);k7yc=) zFJyYb^t>Ro#vZ7+W2#YN4GyMFx3#+022+&Z7HGv8gR{35Xv1@6L{ponou5y!zW)I8 zc6NaSv5NrEAYD6C z6iX{maT3Wodz21qp9-TIy5tiME}JecRGR3>*0oxN6QL41P?Rj05OLun^ukBrql+_h zC2VRTwgL4Xq3mP_>4m#D1I&rsGUvO~w-u^i;%JeN;yqeaVpzx%DV}9&H0FpG%~y{| zV~R5snV>m#Ex%8;A9rxO#?f#%NkR&%LLD!S&Aa}OK zx`71#c*;mdGMA*?7)(4Hxed*!jqR6haI1GFk!b|xObr~Nx_2cnENIDxi%>^qTl##@ z36z`zn*RVAJFpK(#K0fS5!#eUad|O3Uk(jNQu5z!GXzKjLOzttwSvYa$7oXC^%Faj zkPjnr)E>osm_>k`9nN6zw+gBing(Ij$#f+rd;`@2eaYU7Q@ZU;P4Po+08my+;POpi z6=()3lnGUxcRX9~`~MSr6U5$I>`m3)1VKV;)vQsqX-jL5B=)RTC00aHdsS=K?AwTz z8nug}v{sG2zxjN=zwaNv{B@t?&LiiXN6vkX`?_Awc^jwQ-Re&h7AlT&IFex7P}(Gu zEM#>a-aB?m7T!TsL#6W;ig)}?ASc#4b{;`XP1PgA@+7iX)tfjJV+2#H%<6mV2eHPl z_Ow5LTaPk+p7O|>HH5)3yf*sToob3>k9DiDQU2cjZC=Y}zlXljbl=d%mUF19La3sb zTqW<|?>${zL3fYPg7yO7*^fV!sZUhOyN%}XA1r+pRE_7>O63>;FnvHY>e*^VghOO- zx%5NJb(HnuDJ*?D?_!iZ^rEyN%AhVX*rA2jCJ)=bx+|Kfo$wq zwL~6=FR3=On;bj4ICRP;F9f*{I|oFXh3<}xecm}5b>&?e^I1(ZPp2rE9$+&uKGL@; zK!o%O2>roUJLxX!o8!zq9`eoc9reUhbBT< zMD%l4r9vO0A|GBoQBF|~T29t~X}GgEwl#E4avkc+Ef4<7tbD8gyan`EQbO?bM{@F| zZT;da@DI|#+n-hblmSLi3Ot?hAANr}03Z2{_P+hQL8jl2FEP=OJ9?aLOHENpV+EfE zug9&IPOC8Yvb8O>AbgA48M77KAFMjQ8{7E7HSKd%Zo>!t{H@imgPG!QO`O@$LP^5+2Q;XT3(XOCXZzAs=V)Q(OOV_)m=UYs+5Q!l{ znJ<+Eb+6`yR9vokzwUiz3SSp1$~Y1UMqC@QhiA6tDYwDyu8R#PMA`&p(Clf!oi8@D zJ~SkwY*36%pJX}TS@+^U0dg-mK1OCNyu+5J^(swvSZ`Q2B-X+!h^ZxtO{Dam$-a1Qo?x!y=n>tuHgA%lBuPR|n z#pC*!yppDx4wSrLkO27Z^p`Kncs@xjM)e#Lc8}EX^D22cH&|Bwx2mt2;zy&p{H9Sg zgnh!2Gf<^FQ2g=GZ*EPu`TdIhUi~r-PCbjy)4WgoIr1BgcH7&(etdE!e0!A3``jpuCdQ3^=LMqBTOVjTS1!*Djq$Mq!R*%vw zOjh?+*l$2dBS*_3|7`y}EhIjLO8356!R2Vs_H-;=uomuJ@b}VSB0j`Hkz_D|?1Ued zHTwAK@K94*b=SIo0cphTl=ypW==h7x4Rd+zmR@={`9;RUG+}ZX!Ul9f`pA}-T(woE zi_V{y`vH#+flo7-UVL^6F5MatdaD=U?#n=T)%Q%>`DmQt4iOqhx1~TP}AGbSkX6TSa4H!AHXGqb1;8g@C&rB z;RhyFexmnUaFWAL0ccO`<`;^D5Gu&x(VAtYJ`$mze#L0HEv#hY@~f|G=jCivG&Tyg z{+7xV^b9+x>-Tbw3`G|B%Pz`gKl#@gR2x+~!QPtl2J9e3Vf%-eZY2{8LV} z#+p=Pm+Tz!LJJ&pq;LFLwi?_!2vXL9qPY{9vFwly?0=HDYLl}9yC3nR0axwZ46V_d z=&VJp?K866^39hU1bA*3`{;h*yEK!Z+EN0=kzrYEx%wuYISp?rvdKe2M@T{y;Ef8w zHYlz2-sp32)5wb#MWq$b#{8X132oR}YeH7uUN`ZfT)Za#b9f^e=(OX0nXRRi@aQ^}c17gXY+K_R*M^^`~+f)n}DVVtukJ zS~4rCxwkTPOX}*^o;4j3>5^_Y^I%m<{U)}WJPWS_Nd2H0-7BDIOIR6(0pAeOuW>|=w8mNu&l+;UFQ?h?H;?8>Er!cWgl#NM!G{>I{6 z=h35I82(BQ;*?zd4(i)QUr+30VF+4?2VPe;^X7fI{@uNp$pafo7& zJEPXgp^ts|)IL3#826}En zlY1e8Mmue;{eP+O>1N}gJI3;#1J(EeVWz>su=$0`Ko&yMD_41q#E<^a=D;TmowgCa22Z^`@A+DGbvU?(I$CxM%)RC=`By*q`BU_C6rChuBM?Ad* zKK3fA8sHn`{m*_}(idFA);`T6Mg_*J_dm%8V3i?mwVZxghHUy%-gnDL%G^v?y2ti>1DE} zP77SE$2jh63>e4rm(Kq+Gncq^6%S2Ypg*gj5y?a6eDl6$SI-@UTBlLubKmIW^Oe^> zc{|(uTT%dQ^n(m2@mG25EERl0%bGE~M`=i3zw_Hxm+ z`s@X_){`$7c&)$4&$Md4Ws|S9u(Ry-eb=()1NRA=Bg?fk6$~0J#M93Piy!U%vY?f> z@Md{6M&wYL%95<0;yur;!M83^u0iC%O18zU9rqzOhe2j1I>pRVq=-uOt0 zO-lF8hx(=Yapy>08bS~JD#SV^mGsx@J%kyIRVWXFjb@`^OJCXq*5XD)_X*FZ;l$4? zMw5gMO1X^NpXDm;TV)=48;Gk19HOWTj3|KxWg6Ry0}bGQF2v%Asrk#kWZ9<v zg@|B_&lZy@+}4*wsg;h!d4h5ESt6s1TyCsC%q&(OUFR2k96vod_CqIMWvpfgppzD9 z1R&3{%f;RFfRXis+gX$p|4i~om+k)9aZop(<{mR=5ahwT14Q`^C)!OCFbbnGq5)?j z;(TDPhv0pi&(M?>V)&hR^~sNIc?UjBTxn0Isi;gX{ehiv&i*>DFg7Gq39r7su5B+l zq%sS?&$-L{4B$B{c}@&qIM8XR8yfMB9na4IgDpCzPl(Te`%8Bkt6vWszJ6kUg;q z-ejmKduL2>4pAd&)zyu`lMGvWMF>PX#MV&S3cc9wy453?=JR(Br3KEa;||gAs0~3# zq@OCCq;$cIS`%hAlJJx_hal#lbFkmTvXak4b9?gIz+XUg7JIL9T6EMDovMHm%wCM^ zRtQ-H^c5DfK=NmXeB=k*o6PvVcNvQ9VG<@DL5yave$Y!Xf+dO<@tDZcB?e>DO^rx} zB5fov+896BWlcXBHVyr^_{Vm)>yF>6dcF>4LM}8IW@zY9+){*1>kQ>o<%@piyISJP zX00BW`HbmbfVIiRLlXK~&Mw~sH_hL6y|VkN&2#OCg=KpFqL~7#r*34mJ9+{SokOZ5 z$yW?8{B3j?r_^l%4XH&$o~(~rS5D6_Ii**@Lf^H8deo85`%{HCj2%|G&7ug1R5CCb z5kqAUOvSmGh{RbL%J0W+!xCh;DD@r70DQqZyYDRaKKO=Z*T>=5rSw~B$ja4~ScP05 zRI@RuCVVTD3c!dOLB}AKTKgE+C!~cG(D&bl_*}ewsY+LWfPQ6Nf7T=HyBE}*yTeVXO%;Dle_QEREyCNy2FO; z;^|Jq?qU7LEjyY$!bj|j?G7wxVZVi-?k*TN!=>jV&4PPKqu^l9s8asetYM=77hhk3 zLUXHa!PNV=pah7YIdi2Ma#m1lpejUseL-pOUp-m9~!%xw(Yq=r1tS`4pj$ z7b1QP7#Am%VdIjlif#sGK{}v1MBp`3oaPot{4!+94h(A@9P&%xt68 z*7i*+U*|m$Eq*-mLVIKJc7E&X9!mAukr|O@o}fatNRgqK3n{=!QLJrhhJE?h+6;}R zTdQ_x20H)7^Xk{`68Sdj#GJd&5r2MOYvD1iR?o(#!?E>ZkTJg4QGim~j7;a?)H*Y8 zd&*|=0=LNcIIbT2R6$GV;1f!-xY_ZSqb#hSi&({;w9wcBTBqy+sMeC_Boe5@`3rE;wRYl( zfqJi|N6_dL5^e1yIabOw-sW9eAxgL6VDm{n3j9LeEV0Mfd&eB0{)s*Zc;0y!Cz4Ex zh1Dw@iMcVM6ALbnJ@H%*OBD~yWf%i%pXY!Lit<07lAP~w1ZD>6zCPpK&HB>L<4{`k zi|)Z`6hRuYAH~d9A-(?3!F{=A7u)QQrGkqQkQ)#m&ukpYOFstsYC%9TIrIjoip?~k zJhqa~$r8ylbX^G~@?+jXzrZKQT45{#vW%&-_7V3BN?sJoJJH4%L7p*sDFk6HD9 zukFpO{kWH&Kd5k44)l&83b7f466bh?uj#Dt*K|r)R%^KkOZt)vW0x0m_-MzR9pDd& z9^XOsQ>h5+c<68pHB>P9JblK1L^WC1^duM?G$a}+vePO;A;Hez+P>$kj*6gPMLkkh zDGz0#1>}?;$+`E}e;EaX!AKnNFu&se3B)V(5{waw2=f{?OcqdL>y|tOzKWHb{K22#wSr zSQ>0IyL6pkbvmx#KC7Y~q8l&%nr@Y@vZhPI@f<>Rhg``StP_klPB2nbu=L)r%`)E` z?gx9n-g&*?!wTWMEVQiH&WbX{x!@Lec+HiBGQ}4HZr7k zVAKXYAa$Jf?ZV_qg<+>%K&vYatE2OLT0P93lIB|T{k>Rsy-*K5E&?cHawnF1TS1N& zjAhZSpTsuNG9<?gH2X|-F|n}I6b?U6d?#cF=v z7`GKGLa0uY$Hc=-SerWh)t79*WNYhUX4}?hni+Pre_zWVXFjuF%j&I+eij}nRaL=<-8H<4^u*1oQ>+50*;_zQx$X%tkzVu$nFx8kBXbL zaXn++iEC&-2kyVzKT7wyQT=TpMHw$ckiIE5<9E+#&NQO$Xp9?}xaL0CVFz2Nim)Wmo;V2iWO$frr?m)l0=01# zU%Qvebm^Jxvx)u$NfEYmcV1;9ncFKE`umsFF;CSjAdpg54~p^A>9_pyFF+^F{D}sq z)5%w@@v4q4CicF9PaYLF~b-?7|t~ ztrCETdEAhB-q)hCDGEZT3WY9Cq^`(aC*FyX*moEhg6`<}`v9+hbfxQ5v4PT3gspsC=IpU~<012)mSTKh zh~JJ4A`xR_x_Fxs8p@9#;;#x8h;nvxY>I4lbuC>=#yiG%6-dfL) zr7!w^H7Z(`moy-D9WzGGvAdvzW9W7>{4D4EM|D;sE$8^N(7=&@YMU1Gc4fdpb)dgh zxp9FU_ww-Ku@-)>5HOpxmzG_2D1AprU_CvVy=5UBk-jBx2>X`{0>Tf2B z*>LXE{nQi2!rf|h3#NDK(1hSxy2Sck>U_a}*NxdAKidqtwRZ$-MhTaehp%gy{AjQE z(9&fx5Jn;%(YQDB5_>9&eY;O6zRH;4tvlEoy2CuOdpLQo*}MD%noKuPb}s(A8G~JO zd#%Vo0l|a8;5!|?!X)<_-+q(N6=h=AZ$i-+1uWokK>%`` zLYi-aSCIZjR}eqjZ!5L?N-@ohk)lfL4tz}V!Nr=)=5Zcw7v1~|4+{{}WS>g-S=%YX z1#U-k$$Yzlx-rl02@agD0c&}9%+3WhmZ(Nl?@WsUa5anc0+y|owlht->LGF~JW_?T zj;FV#DZ=*(EQ|Kf#p=#;H!TR_IDmEM+O!tC8gi;D3)5O;)Fasv2#Z+nZOKQPfjJrc z(hz6*sW7qLPhI&&B}+Hw3Xil5hotWlGCf^vwGjb`_HY6mL(*uTx0NK=P;|galhh}# z2+39_PXR2t#7Y!z=eMtrIc)+<4hv?t-himUA*mfk zGE*oMTQH4VWbmcUQ~eUpxl^dvyVvi=y00Q)Uy(#ujlp{G*^9c`g9s$;RPRK5^mtWI z%T?*i>zT$ly@RBpd4>gCSsbC$XnG4~Z%jfGIwM=F`?TH&3`H=6As)CSc&ya5B$QFR z>~a}V&C9E8#XN#z{Oy#WR+L{0*Sakv;5At4ECUw0vnNnuX_jr-Qbzz+eNK%+nlc}9A6 zqP0ZR#3$^c)H1&BI1@EL=TF3QN@qKOv5XYWZuiQ`+qj+V!1f6n`>hfe4UYDk!cWLa z$CL&@N*M$(@gMq!vG7K^r-JO{qQq#zqGBcF?BM>#Dm5My=¨NtYG)aI=#l4qq% z)4DiM@B2i;tl>FV*#d+kv2scA)AvL_Y#gb4;}a@@Kjk2|+Uy#l6pRyEy6@|*&Q(>i z%>2D_kJF$(xgjm81~6WrR%&i1Jw)BG%6o-b!H%(z^vqPz8s&u@jm$3H{iB* z5^OcEAwRyKP3XRw1|DVj2EAhT%$o*w6Z7D?m8oh!f_Nk3`a~(mqsU{`92Ub3fN@6_ zOw&&=x7NxYVUJrwv=1^VmB#Yxy&MCI#dYC)t)3Z*;gvj06=i(ClGsT?slHl|8HJjr zRqK&k7?^1VGXaq_%wH31OYqPJ??|gBPx|GQcCZC?izG*SJ2g{0%sOxtC9GV! z7{fj0NE-oX#xU~AEg!7h8c=u5QfGRFgdo0x^P+1 zeO*JN@b(zJiLZYJFzm^a3@I%6{7<`c2$(AS7$PICnUviU9}6>+H%Zi)*5s^WBdN)X z=8NjE3I=xcop)+(Kj6){FY`+ZCgq{U;G>?Xy8^%3R=H6~Lqn{93GSc6; z(_Xc~HJ$qc?PPaQ8#Zr~!IF3RVe#`~ovMC{2lf%#4*?^!9vcp!K0n5l+p9n%sTfg# zzEnPTMThnkR!D>COnTZGiX)oO7jI5Mzk4OOJf*Pd_3!_|i8nFdwO8Cq0J@+o6s#f> z=-`IVThtd4z=f2jMX{7wj<6IvrJ4<_ubi$$Bsawo-Ps&`99>C1i3y`C%42){uI4V$ zOdEqXy@0u0!G`Lq&gJuZhoz7H%ABx|CxU0Lfj-*H`fZyPzUC7EiP^zUafzOfL~$8!;gh(Vzn$!V&< zAS9I5zlB;LwL+ITvDNs8debfLIllM7>B2`ZeDn-0o8`XNlOQwqDDr@C(ke<=BAlHF z9t&R+6x6K?C39paR82Fn=vvjS$mzrblIEe?+&XWNNI)@v^OX>}4HfTX%<^knKym6c zQmTU!h?ZM=n*UyFD4~|AkS;Pm-pzMaAU!?DOU1HXLV_uD_Pyk*QK(}%V+5&KD?0YK zNU8}N81KEeAwIraNUrRckA3NWVyJtQU)~`^t(2LSpWjJ@H8aU=%CfHj2uTK6dK~a7 zAua7xt2Iwg=+rP2bEZQzv3y&_fp&5O`A`F!mhPw3dFYzA#;7|MMh8lM6V8h=QnhNl zFZY}nXS&kyNPHXd%ED2(i%)z@iH z-|B7jnBwt9e#v7#f=%$2sI4c-FyMIp`n&glhPUoqi+16GmOmV|?iTGGA8nbBHE)g;`nMkxkta2w?E8C-@M) z|M4OCLzT6#GrKCs{{j}738qJw@~Ma9Q@evdC%xdggt>Byi)Q!G3;5dVh=B^r+S&z# zMI7bcpSL%Fn$jguGWpp}&n1kVUCXju7no?6mIGhboH@9MJbXZCoq|W@zA&m5#YgDb z$Z-rF9Yw#7CE-X|WHy5f5Y4^@`j*FM6o^e=ejP6L$V->uF!>sR2$PjBBXIdn~b@VVUw2>}$X6MisCfwHn9vzsrh(kF&D5i+`M?|woB)ksypK8|y&-3BKpMZfP*+!+ z_l8l)3Z+MmxC;ccOP3juZ9bCB;3Obki7pUa5Qn14iuXqF8^Md}g7RS^A$ld#xYq1p zVmBtvyoHPa7V&g2VkGsokPbbV-!;^@ECkL;0}or+C<6SQpJRjxfnx7wmZ}w3=dXL1 z?FXMZ$pY_1J|01rKLl$IUF!^=z~NN&Py__ToSH;AjPMzVhua-la!AORVH&?aI^ACX z{g`PsdP5gfkp5RaSi||OYIzIAO5tKYPaGdQFlwj-Xo7E!IjPShY&ZH) ze%dwE)~5-%J6#?%w2wT;@q^MO%xb7Dh5j+JCMG$TYmp)!Qn2;-NXlSa!08f02+1Nt zd4_E={NE2AV;VawA7VNs`=ENDxG$}dITlh=cQ8J~R?6bYYrUQh4Xd0-OUitnr}Z-1 z0rV{Rv}hT8ibGLV9zv`D7Z?;e96C?MU`17R4U+z0g_KjKP^I%1%$_x{pvEa1N)HLL zBMxj58#QSf8X8a{=zYTL%#Lb*Up>-w>qS3pJ?^S7j*^wR`wL_gPe|r#aJmi|5Z7J> zF;#vXQrk(IiJd<)99LlS9@R1yss;(A6f&rv)@O^IP(v$lf~(z*;T2ztwkgN^X$g)K zn8kpF%GwumP7S-@?_wOrG35HkMx-PN?Ld7QgydA2kA;W5TSw2RaYeWFPi<~Vs?R0* zTjQsKp?b{68e`R*MYpIJ+&Q>eP-k>XAN!a4U zQ`^qCObTfg2zF?`QY`PDIA6ktDg?L^+^MriM+{&MkBs#>oqT4Q zQfgJYcoo%nds$sBMUmBnk}wp85%p7FJDxufB+t8L%ntM=_>AsOr$Mkof!&yDcYv~oGR#@l~>6r67 zZ=Vr@NboSqJpKw(DLvu-!2F}O-0l!i}`i+YBTd_*zuq%bZv zJQXK*x4uOp#|yGN^rqn=Y0fvEod%`Ywl}Q%vZ_pRPfJf-%(`=*&5(nqlr{^YxBUK; zEJ%{L)J;J~zsu_7<2FKqjE;6%MYVDEPPz%w{jj46zJ{Ckz~VHLJI&ec_EgT77TaE+ z@emU8EG9ZKNFF)KiDVJ{KGz(_@zdL~63WyKXL(kN$5;DVk7r)4T9qyj^*5oSz?-^_ zBZ+yI1h6>e-JrOhtYbSngbEz8e~f>d9hWm>xw5Xj$6B z-?SU|%&ycplYpnTD&cR;YX1F1qQgm>#k0nno#;e(Ce=H{5@$EX@fPR(RXpeU0?B3W z4)Snh2;X4*m=XxNL$kfhEOZ*`{eXmtpX2B}X(l0>R3o`;ZZ?E4Zsil4Ys@+^ISj<#srYPW6iD$shD+mdoye%kDo>Vow96s2%U)-yHcLm9M|Q zli4=n{WMCtV(lpG*A1*=nCH_YYPa1>t#_EopT7tO?}eV`ll9v=dNl3r(v98KXI^mf zul-Ry( zd`Xz!__FuQ#4_Q;G?Teg$o%&#|8GP%eEtFa>jwV7{AB(*f1Y|NP;AR$oV+$ zd)K$ty}$16wW?)x?XF$>>0Q;&?)vx5zc*;?Y94`Zo@gF%XlQ8Ym}q2}7??QEhyPdp zJwQuBqaYz6AtRwCBco=aB&THIWT2*I;N+nFPdLbl{zv|w-T$8ae@k>KN=h1f8b(@L zMyh93RR0wuRR3Fr{Qr-Fe_x}K<6&MrZ|EFo806@f7<;G(1bzpp>rhJlHNO@@Z?pP*x5U|{~w9q1TnnB-U#*ktUKY~mbLVmNBlsuG;) zl3X-W(&W$m@mvW98xIc?7ZVd5A03SxgB?>`jl$HG1B)zJLfwp#E#57JQ!KwWp`cFE zyu0T$O2ZvYb&^=9I#oZt$%Xy3SIR`w;)4hB{kPlDq#0@owC6{oVc=onU}OJpv&qoO z+1bR!FjP&{OkF8BB!bo5zUCv%5_;l83hGd`-BYhmZhQaTMht={0)?K$Y zyP=i_#FXIGBcBUF+Dc9V^C0~lU#}5cr!nlvWHjn*@smQ^WLs?{nXsdEQy2ZgM_||F zoI;k_QZmwF0^D|4>@N%I#5KyARzF$&c;>s`A?+b*A7psR*0cYiv4DABI)xMfQwe%m zja{RMb*66D#}gk)Q(Qt7&lWFLRKfh7X561&W)f1^;!2X>BDKj^j+^!k4igC!2@K9< z(!G2z5Vc%+X*lHtX1ZUV+F{tU=4}+OmtTwdL{|zjAS8t1HU@&3Y*A26QmXNSQ$SXj z&!1PK=oecnUE_=s?yjC^v&xVeW7(r!!aL2Paj29~KE%Z~aX)Hx7B2kuerSOujuTD5 zFK9`Ih$6d7&|@)Z&*@vt&BxNRkQfT2v}CwsYDq^yMA;vSvxf0K^~!Y0phRsrUDvys zCVu?A4tL@~(2}6pq}4a{y$q5@Zb&xS6`59MToX5MvGD8(@OrP!>X67Wgv-1Qg{Iyl zgS+b;3JV?mEO9qvr|em3*kPVDH>5|;Kj+CIRFvuy9LY`>vehoyX)+M4W|2l!8jVjI zEz@?L(7O6Bn(XykeU54UvD?4n`s+)ZAQ?M#x0`~`@!NtIruUL81Xaa$m4D7AO86-O4cr3etmjlmk6Zr5Ez?SecItc=vlpcIx`>B9_j z*r{)8Q$|Bcm=5Pqa?~hWkk!z8nVj0QYF8PkSyn%Xv60p}X`i-h?#9=ij`>o`);)tV zGrzbgHh{(#1`pQA^!G;F$OPi{olVR8*EU$`!;qdG+X7k5oA7*|xHLNuP*9qJ;Fibx z=1!D>mf2^#ar>3hX!btT!4aFd#-pMQ{xjv2s;8VT`er#z^0N@)J!P@9oRcn1wr)XJIvA+$EOf% z`tvWHQpNa_**=u?Fn(`O?;Mn&Y^)Mz=H`60w0i^GZ!APY`DMsPlZNWq;_T*H)hc?X zcmFSg902VliFcurjd(RZXcVb%aej@#%MnTwbTG8q+=d#QmfYk!lG@^wHFnfriW$0X z_Rr!~^9UtR(ItFhW4SmQCRheq-rp^9tWfJ%pb%==7h86o_Nd3GrQ2R=>e?IWlpR}R zeJ(BUqjKP*rFFBK&&Zl&F@Q!nucBBQuvB3h=ZQaP=CC@;;cMBoC&0xr+|w!DF(;U~ zIicYOnUbtMXUQxo9RWGJ4(joZ&yMU?3C_?AvW1>%LfFt@$2^c*FtRGk8s==v~LpW)>2bv z<-3owTHW@dEo)&#E7w~-l)e=B7ft2hG9Dv5o} z&AhRnqn7}~y{;$Bw4H=ef&_pf-3C~R`V==rlecyAC=7Wup>UR z*X1XKYo?s2!Y{{K9PG zwU@;cSN@&>Q4e7oj^kui{TiZ)c4J_KK_C3c!v57uYUTW_ZF4Mn)tb%6aAjzmFvAIWqkzl>H4hhtFjsDm=$otJ|WoRE7>3cPr!u1@dHy>!7we z0TREgQ_rq+qv!n4mt|J%Y+D_8_>z2@#PBz4Ngfl% zW}qjdnJVn~I%eU|ON%i_0^x{Z=*l_$1*la|ov_l^LW^fpiMWaSyi>zEW;^J%EH-4eYaLOQPkj3Gi z0PLY*YiB^X`r{wQ9$~})o95KtoN$6Hi`}k@b_u}-N`-Nq6q|8nD=))H!CQ`AXa(qE z9W9GB4bX&!Ka_pyKRo(FyHkcsJ$lqjfMH1&2Kh%y&V?%;BD?o;;b3L=9TVCpE`s-_ z#R{$f>>BOQcLC>`K529!B^B*|>{w%kD}HjVnQL^->TV5$ol~IJFHy09S@(JzKD7jx z=29N7cr8R}N?OG&2dZ4;-U+xlP$l|cUXD_HXxsSnCEp?=Yih&+AKj7RwB6s&w|a8yFd4(PluyO;L7>0bx~;5dUXWf z{g1m*M;fgT28*(Ly_I?=-fU}qU@jafvqL>@yM%rh3|zjjh1cp0IhYGk55E-po;o}_m-2H@>-nCur(}L(+yg>-8SmDQn&^CsD`!Q%5#Ay|Q6va2 znwPqWa<)7tFz$5h3p;d#R?Ocaj6tHt6`sm1Fl$N-!xsi;I$llRRNhT9#^81G<^_KVj^qk<@Be;MHmIsr56X$i;=OcTDE{ zzrCDDFl@%!2{1v{q1wmkEN81?>&v58hV3@q+Bb83v9d-0bz8yx$qfVkKD~i4-<=DB zvD=$$G|W2ny#=aHYJ??gV1i2?JBf?~$;9-h!z^O;B6Dv$?0;}CZd(4M961m|{)2SO z_i)*Cw_m5^BVs)LR+oHTe5ZFx^qzL~vnRbU zYetpcn59@b<*?={Sv&Si#hhg${CZCZFpK0HFP5 z%BPg%6Us7*wrmuOc`1n`!u6e_L>n2vkMV^;Q(=9Ap zsvy{W)O2@5f|$&A!nP6{Zxs6Ma9u#iS}24HVz#1!-+iK#LndXqmtgCAbuZ$Sqn~!C zNL*5qlK8V-p$mN_$c0UK4R&(LwSS?yba_Lr6TL&O$IIW5VBE@nyV&S0`clipkF2US zI=8IF$K%0YU#_XYZb#%U(y6E;)qZGahe<|1CC-RB8O2YyAW%8&P3{_<5j4=9l3v>n z+UQr^Cr&VzE7SFB*-4DLF?4d9Dj1$>ibYO|ABv1~K`116Mc8X9B|_j*&1NHon$nK3hGHt?WS`r0 z4NiEN;m0d0?%mN%G-8UW1V%*Jv9SLu-o~!f^I8*b zp7YfsV7{ztK1=VRyXK>1M~Yhw%aSVxxpd1r_I6oE`uA^*Qo=gM8}ov{h$&*XQjuI7 zj|-y9pv5nlC-h~ceQ3qlgYe_n(j@y;y493sCcKvB7l6M88)EmH1G28`r&HA!%d~1t zcjD(mF@<=J?X_r|;!;w4+j72UpdP5*Tsxq7-TXl5M(EE@yA7)@&+L-R;=UM(lo2dI zM3Ym3nV!70WD|6XX3fc_HDa}7c!tBCx39bZFiYn%w%G$U& zc669`xTv9Pp2|Ol%ej|I>4ZWeR;#Z~XmWssu-YD!yh3a7X&6H-Ei~IZ0j&CZ>AnN? z@Tpx*E@PayS_1hp&Gr*RH+l`j>%6OZY@jb~%M5nQX$r@8N&+hOd>@5w#!7UfolGak ze@rY>-o;1B>|X*Ln%s3YWA*?XiNHxpoHX&BfV4 zR-ZiYVdgGmfL07sh@c=riUbs%GMg&eNR>7&EE>HAb$HWj-;3%s#EE8iw+a#0h}CxK z6JLy?{mFMEbDcJKv^q{hUtcL!c`KcAVy7x!|*x$w^b1F$dj^l&|GVYUZij99P}wga#SwVBH4=8LmdHu-2e=!U*?WQL|@^+75ki zf2DdLS?Kw>M=dSOQy9OoF|-2Rsd=O-j%?b!D2W zRea}$Mx~c!N&?KaZMi-2zxao8AVdkdSLPxmP;|Nb%kr*_{MpqilF*lmQd&Lo2*N&v zf~}Elfv8xbH~aVfl?lScZJ%_rA(>6w@KrwvfQ>msmlhfI zrv^D~X_e~^IS(4?~MHK0bRB5|~?)e1tm*&DK!+a+1#o4%Y~-qZE@ z;2BB7I(AiYfeAmjZ6BSQJcuk@K0GCLw}YNaWjCj7CZ<@l{xR0g%K8^gFJ5wN0#sp& zpYky|gulCiLF-Ccn(19*MvJV`60e2^v+Cjzfd#JCik@aVv;@&wr1p6MT57EB2&@BA z10p!=cH4_ z$K+~ToGoRSJTU-iaP8*^+1YKmB7(US^=II6pe*X^%ldS%pX5vWg%~Feree_ul~KRy z3!%?WdIl!`#Tg(1HzFwFKXe*vV%@7c!`YRYn2%m(P|%qI11?89FGsA?i6@8{Rum62yb8eln7tKZewj^Ya%f z7^r=Soo7Z}hq>+?iHxR-8kD3m*Az4RJE66pc3L;_BO)Y#TvpIt8Rc{6gFyNb#rbId zvpd9v{=?q!O}bY^kWK`yW>9FDmQOSG6IU$q*(gfn>xv_R{yVMMh&x5^mBT)lql~k_`?p^3^@J@{GFdrfG~@6yw?Bd0cP$zDd^SLiEIiH+q`qDxXJ z+9gp=l&(P@lxWy%l4F^A<=C4$d|$@<-I?}ra;FN6;lBy?JztXLNc2b=u6>C@1mW7uOPSN)d^qKh=^J;uj zt3nXJb8!)c#RqKl1qmt;c;1+yyMKEAd&0)3|5Mbx13Gy1{IYQ!_r z{wGw(Db~Xz?XiA<73(nEhd6q~`SYr)hwsZ&M5gX>l&lek%NVzx#ix|FnG9VBje4D5 zEdFp}!rjD!rQ?r_FpRzhBl+cN-y=MTQ|yu#-1Rq#ZBUD5jNm{IU8Kgf{Rl*Ok#9IEDwb~i!^<^NFl3`gI zYuL^D*uC@&))R(X%g>nCX!3>I#NTCdIkhdzNnu}q+&avVss)tKEm!1kGJq()eE+@b zRU=E)F_TgBv$LjT_b zRqz3%t3Si*lm-~2r+ZQLhjnr5(Fj_)$wbpH#PG6k`BTFJziq+2N;7q%;c`yWPNf38 zK50=TUT>3UhBtDWID{RUZIrSSTu8VmQd!Ekp?RQA$*rN?%I)AR_aW7C*JO=P-PUZ* zFX`ez@kMdT-s!j~HV+hTc1qps(5yY_LqsZrKDIGgxT1pAjqDmDgkI{eqewIAX=wkz(YV=twbUCb28RCCT0DvjgF15w$Hl zkn0Kn%k1lJ^*2_N$3N^i557z9^36!5u9Ru>ANYKGDJ`!;>K(AigJkazF<<>ec>zIo zNEeWfsJW0ays+6swE-wzx!|gRwj$8^ss}YPpMV z^;m`VuOuy{Ws%FMEt2r16{h4yLjpqsF1r~)dEr_9=JjxppgpgKueT=kG$#Gt>HM|Q z;dKzfRjW^zgh%KVtq$@Sb7|8dlKYjRPi3PvLxx#hDrFa~bO$!mdth>Thj;PHI*j<0 z@riOr*+s!1JO}z;M7bbIPchL{6%!+=G_3oIjQtmF^~dJ-LHx!8`1HjMs5k)ogP2O~ zsZSJSzm}4TQKfh3sn0383BGk%?+}hD;*EOg+wqpeQJjDOzKhjB)=0wpadw@zPysUl zGsrr*m5}|`>7Y;3sx&^sCJfgWqQ*$yOql0XJ1d{#w^=FRebI_W2>}NpaZr(^jXp3h zq4!XiPaQ#OSHG|7o>)0PU3rED=?t2@i23W- zQ>y^cKlR(vnxxdxG3Dv+RfTUdu6`<5nDIL=c{!F4<0RSHeET=G+0j0$9Q5pFadh#+ zwe+|qk@afOqIDY5%N)Phr zIlID>yns$|2SSU$A$1Ag#Xv%k!z^h4S4|co=rc~RhXCw`x>9ei8HU|dleX7IId$8; zZ3a|IT0vav4Wqt+@GeABD_myyo+?ZmP|6A*Zx3I&z-34yQ`Gm}%S|LHrxmntgtAu) zm)Nm*oH46T@r^dv5Qj!Vfw}f=Dy^H<9-mPqgs0`=zo|wDN-@Z=Z zh(>E5IJBpdTDUp~B5ezAv-*(kQ#8a#_8e?BzV6LYfX@iu2up+=;g)KWJnZmVv?*R| z0!y_juh*}81YPNUje@@b9L}nB3~5Q0y7|l2w@$)SZIjy0WY59tQ6oE1hZj+~S~xJ1 zjn$H^wdC}W2)W3SJtDz-Ap*f_ZdS{Dh0`)Y>AK3k$`r0J&w4bPLU{%sXI}rPDA?OV z5xc+Z>yTKNoyJt@evRj021!(Am#}iHvb)qvim)rEvc`=5QXNYaxDu*z^Zo&X0by^w z3TomkgWk}os?G{mHuZRSb7vG1t%`A+`?DDOez9;hw-#{*0Jdac8cqT4gtE5qat%Zd z*jp~#P9PF!Ox8lz+USARiX*fRmKW}!lI1~BNW#9`y{4a3(v5x8Oiiy7Z?&U8nZTxM z1NKC{mKA5PX-^@$Y5k%lg9!O%`}uLB>z`|dsxP2}L5~xS?+V1Ox+6$6Yb*abt0um} z&RE798D4Ib_mQt;ekF7Lk1gHf7DULm2E!55^gIH#erJG=TG)Xkt4 zxv)IrRcB7nFWNMQuD+liej97!N3M~O)?F^!YGSA7!38*S$V{$ZCtiosmi0K5u8Wx% zzljRe(Y^s9Djr2_f;UhrM5u~bICup(BU}1G@C{~Z-4c58I@*wuqk&eg6MlD=Gyr$*;6Y41z$Y>$TGe6T?5@Z zH7_wszJ*^(7Z1wsdw=V-!(Xp5ZNUP-7K@v{t#~XUY7hU|)rz3%C-Nj_V@5^JW{8uo z2Y#3IjV~?;?F>swU8(}4rUQE9fS4AnLKSBqcIns^VgkF{a*_bKdfmu>(fr^_!pbbs zwk9s|!on%#wpm-BScMXy8rlw9_oXup$zuWkq7^ne-558?3zAuZ)EB|_BQJG;>i&Gc z5;o9az89M^ba7`lTOdF8v7ZGlAfWe5x6caEQg`U~ZUDQYKg9}9sKCd)?5r|7vO4KI zqN*xd%_~EwsAfm`UOokbId$WkKT=o$69*%q)!Z4}s%(sYl28SAlniOTfim-KpvzOB zV#~gbN-|rcOM=oe0zFD0vA+TrP9G>H6!R;7)xC|s>=q9t!%}P!cdycUu6XjDGpDTC zLa%4M0Hnds&|$jGTPr059CP0LEjc2>@JCpA`VBd7(3i&}km3*ZSPq-j%ZiCtT4*4>akC(R zZKWmu#te*h#$zk$UD?30y6a^&2ZovuO2>j?X(MK5WiZt(O9BN>&6&%R@QWsdZqhg0 z-G^+vi{LHmDDk{GC9YxPwANc^t(z2BJ$)l1SQ4f%>G*n6KHkn-WJ&5Wq=vJrMy-kR zV_&VoMq32H(I%;hv#=`17bqvZo!4D56rEjN_4M)G%*GvNlt(fc^C|;GO&I1%D>23} zqT-c5kRc{P*dTujHQFaULJW36TlSL}f(ZKxbUEe}H#cN;z^C)K3X=*Zy-zoV_kk`} z0xs6gptq&X!cF^uUcUEEfpv@b9jZ)+;1_Nst`b%^S>SztVwTyQjw|tc!mDoUbbr6U zF?amd@@#-!rT)}^M`q&gq_z-3KYP4j|b68MWah+=LPsXg+aV0jIooQSJc`55;Wr|9lCtS= zjnIBvyxcJ{bMf;9J0wL)8vhpUAu1ynPQ7WM70IK3i`+=RIF=<^XqX|ICAJHNeZ-xE zGJx~*KPjJy7`No|W8D9865MKE`B-oACT8*MoHe4U%crjbS-ej}@7hKG>FAG`GHq5< zLYw9rVzaq9!>e%IGv{>VNY;v@?26RGY+<0M(xw;NcJg^`e9iQNG;>L`Fwi=pt*kcb z51O9q2w|odqZ)l9Kv-Y@kQim$TuE$Cwky47!N4ex%w)8(wH%Z5fxgkR)SUa{ozCS~ z0T^=f-ZwcUe&`_n(4VT?$ql`VzWvT>$yWgMnl->84l3@xMtDB`+jPJMlMZu52O_L0yB!yn+2TFIqQ);^*|Ur;QXLB3v$%`nD*ygMmeG8 zLpF#aeTT0uBg*Nc>kbXHkg(H$#cqUU{Dth|FCxE}ETzD&dm|w%W*bV>JRisF7 zTADw5NQW&#Cr3xiFQBx~lt%>CkDy@0dBFElci%0}PI~6MSu*o#v1zjwcF=SA!&@CW zU{8_hpF)_0TZTigZCH{hT;T)#LXVVSK4;D1D9OPD4fg^g2<~F=WO)Ve#O&Bn(;&5` z54ATUUwrk44(I4I!IbzvMW=myi=HRK%+OIkOw%eV{|H^CHvbAy@icFo&omcq^rakf z^r@E-=gV1=R$bA5<~&#CY@$bsiQK&Z@gJuH01yNk3FBq7;h$wD1QH&Cp`mxXdo!?Mf$nJeT44-HZ=q#M%JO@{BDVt&$pBt} zxMl<|^Z$}$l+wpHmn0=vm?iyGH3xING8lb)r$~-5#7E zK_k`ePw%snmMK;`Y!i`N9rqYRH~v!YbS!ZLLUpT;u{2pk`~{ocp}{BK4Mwll_+OD_ zYSBXs#&^IDo7CQ)9T2OxHeBw(5EkFlUZ_Bv#X8WD7cKnMUGb{3_6#<@S70a&UZAkh zP10rD=NFSpPl||;{LfSQM=0j`1M|X;U2V%=LCRphS|gsn{5v8U*Oen>k^sDClGc5nZW{9R7 zU*TA){>FqC|C)DNy1tSUYLA{AFmzfu}-!*x0{EGilS~0Z3%Af>6leNGpe8jE}Lk{84PvO;F7jl zbzyN2=^BF06kxkTyWKk9w9EKs2!&2Q_p}U0J&O07t8(h2YB{86)?#_TQ-cyrf5qTb z=B3Pp?MaV`jxty)LD!4`J9=-b>Xpwt)jc9s9)F)II~R{v>zCmSBO(3FVv9FP#yaR> z-He7ts8tVRu;b1M_DViibM0`>GrdKUQ!?VVcz~n~>G|r{I9ObotMn7#<9UhDYHX}M zw|%qHFzy1@JXmu3U_S46r!neE)i4V*>!{VB8wplR>F+S8j9jUu@1g!If(bk;G}^{k zx)tl%#{hjCl9??88g5=Pp<|$?<&gbx`$mGc@ee>;=EG+=MFi(5{J=zSPmM5OI}SJZ zUAWcoA&wQP$LSmU{mbfXTx{*;_}V(Y-b+8Q!H{Z;q$)^slUmvL&c5MpeT8%XQBM@ED5?v@8NB)yoY9WKl}wp zi$=~q;HI~2a8k7?)A>|>k@sn3Iq08mZ^bV&CsyL!E?t5PI(kWzx9fZCjeYjCQTO#Z zW9vE6*DJ&WPo8ym)`sn*g&mR#H@jlC@qAr0n`}4bL5H;*8XIEexyq5f|47cf@!a&A z8yr?1fXxWchbB25kWi^we%;ErmrC1(SKm)R3x?m?4|d# zk~x%0lu--g1?yiKEZ(Gyj;Jd($OeUfbGugFWk#|e9X|7KZ$}rNDne!l{;7^Iq#C*3 zoKWIzu9RsI?=ofIzs6uv>;E6(Uiq0Wo_q{xG#4mJaGGVb80#X~_4MeAeUY)3zVofB zMCYU)Ql@0cXDB{b_6g^z!!>Ig@Ts<+*BWm3yQwxV>25D6OK0$R*_(ouM|^Hmu?d*u|9*@ zR;JRI%8*m}857Ks0}l1r*KqnWah7-LqBV3CihlA4&lcAjdxUwZ zf^2EEW&5a6DCk^v0MA3H_k3kG&!z62(h~y}Tk#^T3m1cl!n9O7Kj?6+O$OIy{XkoR zI-pocyX&y5uhDh7pPQFx%l7C}e#dueBi5ieY^r5Y;x+Tesz7K6b`8*8L(E{ubv?pB zkS2+u*w;sHP?`}MjsHP&+VQ`{O3qD6lY3z$(}%`OaJbHLb=9bCrDC8n^{j)ne_HF1 zv{uC*h~D>xyZ8D2D#ewt6q0fsjc@|&@pLJoj{~b`zAP;hUYp8%Hv>XvHE2iv(ESem z*~-fb@xC^BkocvjR=_rFGp`-RvfTw%Bi4bnR?*?igVO}Pp(ewB*C!r~G*n&qf3o>51 zxX(}624^x2QfFzvS=8$^R`Dqcq+IJFdHhTQqFU2NX{ zDXnfkHPXj_BJnC%|1lGG`$sf(sxW{2ydb27R=BTcoFIxNwh22|J|DSr)h8c(>t^E3 zI8rE1ouAy1iP%4RknnQowQ4&tyiT$<{MZa~Y=&N|jy>VDuc86xLt2M|nSE;!fK`&U ztxcgJm&UMks}#v(8oU5ie8M526NqW0={frV*D3JZBYy9;0$BU6&6novIw6i_L$-Nu z0SmE(66a|A0;J@$9n2CVQLknOvQV?Wo!0;Dd+?PTk3jT)X*BZxa)#PD(bod5=AUqq zn^o8dusYTI6h9;cT@8~R_yi9DqPcP_PB7yP3zQG< zCD4xN%z!lOz~_P>@7ULD{Q%3ebYdbeu4^HqJFIe`?G|>^IODh5U8HmdtoK-mc&1H6 zR_%)wYWjHFrq;r|D)w>L8x^!ZU;Db;KoMtuAg-tBwLE)@IdxU;w$tVd_kH36Lnif# zUK=xYvYEM^G)$e^c`j8`NQs@cR`yBn{K+!v(@ zzAW^%?q-WM#imKdMs@uM$$E)7e=n-c)-aG@=;zA2;TE7Bg%Ckgt3(=26}||1A-tbhe=6cpTu$-|yu`(q*~-NBf|kzGw)*g@GW`aA_Mz zsRI&>jBi4zd2;v2LWgu!61XOJynk7f?W*GWX7w%BmtazWFGQ`Le===^nFnVXbnf?T zOTr~3u6v*)1IfM)rGh0^VlY%esFc~9S=qJf_d$Rw@6giaon3$^VQ;MFHH(Q#zg>HY zqm>VzyPTP=$Etjb>bK>Q(@#rdcL6SaSK9$uR}-1lrciG7BjK^R-J7c_w{-)XdP2ub zw4!mPuuL&tP^0{*FR92#z-?Z}&VW16a65qGtY|Efa+Qy^Lu@vKqt*u|XtN4^fmxJj zgaZ!1udo|v<+-+SC!e)s17?&AkF<0dTojFSSk(saxo$_2nz3D#;)v)qjheH6-OI4n zQ!>128>Y7|xpr+a&gq@W7w7VNdz;kzhNoC#t&%;Vs`p& zl!BjVr_Uuj{L)$&s6URi;OPmXwt^Hs-a9-gdeCPB`%!nTjiH>59O5rm)Rb%7W}CnN zhV2b;e^MY-^_!a*FfSB`+PXEZMH?>jaccAu5YT(=Z>uWM84~L*L+SUg@54neV!pJ7 zG_fMqEOU;yQ2Q-IxHB6Yg2b(^f4251LP>)_`4w_IW{Qb%JP73dCLVR_f;EHx;6sMDTJct$(I zR6ev}mdVF=XoY=K2Sls|TIF^~#xrW2L@0eP!I*_ONe;Yw?cHe%twUUjSNEIH1$8KW z3A|PMWmo5jCjz{UaygvU8@uZd-zVvC8sR@(@yX;wTI8T5*tx^ais0};{mBc5EHc;i z{c)(BF~_W;YMsfIrBrHLxsK&5HKA`XHcLIDyHbkYryyOTxuaG{AU`~KaJII@?w78P zIj3kPsYZ-)ga|hSwwG3q`K~E5H$Crm&2+^gvx8+lzt!EMxb-NV!mcT8j@5w7jqt*v z(rE^9dh(JHrJpUee?U~G^ltG%cVQAaxbi|Wg~!s% zC}7=>zloxeIY7&kvjk3j;zgcIJa)^3W4Q?yfAE~UWIYgNTfrgKPXTigj;7x?Oxh`g zf-e%8qAAsFYXsUhae^)#T(&CS>7_3VWjw9D3#2qQ;_@Q?Sjdy(h&wLdP%@cu(fc`O z*Y4xpVB>|9l6j5~mR1|RBct2;{qJ1)>z!!n3b2vGBc1)5gZpk7{+G%Bh+dS0=EU<{ zO5Pc6>xzYB1|6*WJ|MEyX-f^qY}-$6P8~%!hx9`?94+}w+uv~5Ed%OQ8bq+SqfC<# zrgPLu8#opez0%;q4qEkhcYbCM;jj0jYRo^*?f-@cDiQuu*8#hEiv*Kk0{&!G3X9+T z<0LPy@bfX;_V3FO6`i_;qLfZ(rSy;UZwG{2-GBa(#5dg7CgL8OG;_K28cMVPg};!L zg)}ZV^DFlQ-wb&094^!|CMWe{eZ98=RnaD)B1Sf(fS+sDz5D@M>>UOUh6IVeK9XjW z7a=teijIP^nq?1y2jCBLx9+7s9t)t*GK=?MhKEngeR?%rc0cVkpmj1{o%rxToi)5reC6wOq?Q z*F==l=X#+-g<)}v%&ksR5_LmGCz$GzW3f@Hg6hUvxiZRoF~2?}r*PMZ5zW-jAt|~4 zur}LSu}SjGx9@x{`xD&&@K$v}pk+ss77Q;71*zqWG}+V(FQn+@L@_qxOO=lc;Q4zH z(Dsz0RyQrf)ugER8w_OBvO75gix3fNLcp)?{_bcht(T4UTBoe+BhxK5I9pab1DRVc zO;bdl*gkP@p{`y=y|6z^xy(Q;IlL^^X#W?Dsu-*VS`ii%X5}8Fwei{^+HW_4^oVFz zz09NyuOwaH-FT#!3Jx^I6^=gMEQ#_TXtj_Lg8In5m5X8eEXR{Wq}4~@ zg4@IP0<8->aqv8}ee&iESo%t7F*nTD3NG`Vdi8YhLEWfE9S-5M%IUW^o5CTggMHo% zDOK2np>YUS;aUjU(#o0Pv;@(HdBd_zlO12yOUHH?tZSWsJr7$zb|&Gd%WE9?sy!0> z9&m~sxCC)Oj2a&YL}YJrOBCT;wB;y>T14FLS*)$G)$A6b5X!t!oARN;Ku=tXyHM1! z72K182tl;2Rq~$T)3#lAO_U_blkSrJ#_BSxTa>d-df6K4B7xtr!QZ0PlmpM1^@Jsv z?EC|}-}fw0Zs*Sad0LXFQx8eBH@3_w`HOYw&0Z)lVleFd!jVs6-Lqbqf6eEBrfJMc zt51?eifnC_LDcT>Yz;N=!=ux@5xQnqfAxrP^)6TRr!RAg(5z%>KhN96|85v2u&piK zUGw!PYl!%Et-V1LFcse>P(!cV zWfRn<#o4S+4{s#cUt5SJ+U*?O{fwa2GUG*S*z14VHD(Nt;np%W>?Ke|> z5dGxduHA9SVi)vif;FqIa;lo2DLg;JB#K?LDj5Wh+}B2+G9{SR)|C*S`$)Fac^wVO=bw2dx1b+Y<#oWlGU-Rs)<5;g;K z+RG$8i%s->w=z$Gm=+;-7N>2c1tJz+LG&q7!L2=}?xdCCTUUsy5L}JPk`q;Xjdl2L zn`sv`Tmh}#X(C#=s(0<&`&|8eXU*+;#T`&>BH@u{#&o{6%Vu}8%>X5Q961tCmN^BoqibdbLA&a&k?zm3V zmCe>w<+Eg)I?_XCf@~R)<}d55{Qa?0Q(KTcAjf z1z3i9B9T6)4<&a-d*DM~;BrS*yWPuw<7%4<7%Z64_}pN(wr8Kl5d1Zqr;q+MKV!*k zuv3)*9=mnR-zfB0w@V+2f66|L zS(BpHVdsjN%kLpLF@xs!-_D|(+ephl1>GN=LBt=hZev*5RY3Wl!u|&C@79SzuR8-o z8Miw-LDQk{Q08A9MfZGve`uRXYfd@-E)*F6+HO66d^?w0a-w|TtZ<@|Uj89U)R{%6 zPFXnS+>0gWuUBF~(68A0^4IU5JiCV1yQh~v-5nrrV;xyOzhK|&Y}KKJE5lvy)??m+ z3hyu7?v&AEcKIGkZ<)L7{YaUC1<7cl5KUud_ z3m?2N+0DM#MeeSHof7_n3m&q6=7|5m`Wkb70DPRNs*X%8IB)KtExB0U8@>Bm{$s>h z?0WfnDW-RX_X_lTv*-`HG4-4E4Xxu}=375XPMe=sz2ZHu&qc>Zd*OZ+LWfn&L!O>5 z`9($Y1CIL?SsLyxtWc6@X2!Yi$Hx^wBqfj68W8w03s#5@`!>&(z)qQ7K{U0(2E}b( zSvdhx%B7(tW2z1%#bT{#<2>~q|CCgbM5mOLl;nUs#U(C28RviexW+@DFD{PtFPa>T ztzsqehw>!QU9C_I$Vsa1KKHEqky9)yiY!u+ENVSYCGjET$34-?N`=QHo>Fdw2L6mW z-K0xU3@h%ZcG4_umKx~FdP0U(x$*oT#iv^Mn-p1Cr=p^Dx@<>vkNgmF3 zUjXT|0I-=f92bNx;mqB4ta)TEdNbcZ!ketP*S0Q$T(M9H>zw8c=A&*BIo6CIUA#Py zR1+O94gfF`i~R#QAyT#SFm@6rZLMa*BuKN8Xc3r<7gC*?!f9F%V`8mmb$UyCuNu;9LftdJS;dGf%T211g`0%`ASccK`mBM~M>cYwVhQ z$`B5TvT8325$kInmBI)7w6{&*<34jjbn2-rNsKhzqy+r$WOQ&mvMl4MAL2Ls*+k?o z=ci}gkJCCNPJ+o&mU5lM@m9$G=zO7yQZe3AQeq8}z<}cLfbo7TULqIBYhjz|7Q)oNBU zuiEE!!kEM2n1%;qk`VAP&jLRFhrZ(HtASWl2+_*qdd12g+>80dA5ZJwf}!(SQ>09O zb@67UaR4A0+DCAFicOn~CFO7RV1`4PX`6amSjksiwd4|39r@O7>fV3Ru+qOjWxbe% zRSRkQK9Xi(458uG;2i_-cr^b-lqz6yK#kzlwP#dXl#FKpz_M27=9Gq z3lHq*WtJG2D&_@-o16QZ_;1)(*S|!7NEJGd|HIL_$20xDe|(tpFow2>IW#%8Swu01 zP&uE?DKkttl*5OV)0~GnM_a9kay|={WSBX$P&s#!Q{@!OA)?>DzxV#xU++C0_v60r z>vcV^*L6*P6P6t=lx#%si8EKFSH@{P7WcZP*SUuWTm;T{6N*pqdOvtz#32UU{Vh_m zH}DTYyttLbM*(ljzfEw9)APpU5~T^D49Kdk^rjF72)r%(@Ro(DF5u}`QkU(%&0om2 zWH&Dct*O@{QLJU8mC1uOZ=3N zTK)s@Z6QMc;`x^)_KniMBm5z%h2b2mxpIlQOWa+eA|)ZCa3Vl`$LO~(JQVG5Rzpx% zN6(R>H_z1DPwu9~h?E&@^uLQW4i_lU9HP?{p#Abpc_nSNaLzgP423BMoONo7qnQP$ zv*yhd*#5$3pKjr>G!%;4Xm?$m#PMmu5|^5EeNAiOZ!EF_d;b2m8}6+UlZ=KpMs zIC%gZ%8s9HqEZw2J_nKMeHohWn_Rkdh&&Q9N$&Zz0Il6hCK^Erx=TP1KMQFY#ZsUr z8^#_k#Z2kn!NO2jy;pe)hp$@rHspJk#!os|4gq_IWP$AHe7=iIUiZICM5_#p0UK{2K`NqX)%t zhC_QYVl0KSB}Yvgpl&9he#4I9 z@^M>;INF6-%U&ROkI~=tl%<|uZQ@NS<7hMrX*@Em8Zb&x6f==7t?b^`l`dGJr;0q| zyuDe>De{;VWnct0+UAQ91?*ey0s~QkO#tlW+T6m8lk{X{WaBKF%H$&Tk?jq2z|n9E zelEq4Ta;fEBM6dAe?~b{yYP)XrvyvHUnnNZaH7)a@PF_a~H#aq;bex$~_&54hv83IQNc92GDf0T|5aHZ4b zIF7N0IZTL^joJz1#u}nUs;I@%K!eMh{huv$iN#z$G~}AYk=jNv+>NsHES5S&4Ly7KDz~jk(2wONrm*HL z`3*#M3i#;0x+;bslRw0a>HQ{@6sNKf(qGK`w=NW{0E8C<4j^-e+E_0a z3-y;cDB^sVljqPZIxl%79F3Ia>YwpPul?2{N*8b%mP9EaPstn?0}u#$m!K@I0!}yh zY1DT*ia(nlidVgYMt~L$DD0&)M88YitY_Kf+RmO=zTxLS_6LyAqNxzON81nudZTse z3N(%#wthk=+)nM*bW`C#A>+z+Zv+82#fgRRB$O^cmBMx%!;pmY_?STrrmAeV2^>5i zPmY=mJ0^odqXhDuLZp~-9{Ku4Nqqg}y?|aL`}CYymeLg0cNKUDJg5h*d*l&a!~>t0 zKDEHBDlwvg2YaIt%20Lm%118r6s`2|RiZa_<03IW6OI0Rs8+(lB~fC*@pp3}_l5&? znF6?4Q&CEusJkK#1O+bU_XD(X*s{0mcnxCkZf#sl*x65K3=?oyyLR?T$)Xa?!WXxE z(DT>;@F=`I24@!1Qlx;=Y?VKTTqFUw_~!*7yq}CxWqH}!D)yx)TTAw%Nuu`^`5syh zBVu4gZ=hi~=8FXvG3Ktvq0aAb9IAM8=~6y3J&_9-9UHO{b|dZ=MB)3$R{OL)^0&T7 z8F>(l#DuPkcwL4fnYB>C$Ic)<)uI>{=VbG2!uN#`rT0hzG<4w&P}T zpYY|>e;G03>_{QF_a70-u3)9f&hnrS;H3; zeioO>x=j7?>Uw+g&~{7a8*FE@xQ6s_DTbigH$Y7l3HzFRrzztxNXBv*VR9p!gaqqJ z)Ajs^O;{t@N^sVp>pmZgxBKAqhwo4QuCiZiGD1E)nJirbOP&iuFYxwdJD>ccNHI^W zfY0)Frq^9IxJG6&ochUZsXIp?NOv!gfbQ;~)@r{PiWs zLW#y`eIK?QM<(%vE)clTvj^0jICkUIuE3EknOPS?;=+e=0^ppt)0i<}g-0}q`i%vV zlV_)3o>3k9cB_=wZtOq;0aLK7>Wjx+NZ{i7hD;|B^yWuC`@EJ7+>S4{J^TgT%6JAr z$^t2`Ud}QX1w4G@I3(P7oV?QA{2ar)LkOKABp(QG3g?of;bDnJkh70J8Pu*i8p zO%n}A|CEPz-GEZkz_elq%rEupa>oFN-@0H79s=t31&TOPz85oUhOJmW7;(Uqq>Ae7 zVZOnY*~7`z$G<3G%V9hL$@R9U}$%iY7;iI8xC~J7Abg!_* z7LvfB>pL$a9GYWsm3l#^;CM4mM{ysJFUdjT5%wNTi6w;n)-M3EFkeN(bHc%Vj@F8n zlvoyjKSYPZbDS|}GMJEU1fv(vc7NYbR_0KWW{<#NNcV)J(+4a`{XL2Y#*CCJ2w8`9 zoxW`}S049U%9#Q1nIY=3Y&Z`A$GX5lA0c;Bdb#7KsT76`O5YFxNDZZI7*Z(IgmQ$^ zhyF^25llbF0h7VP18}vJ%EmlROGFq+UmGBqqcLs9+vx$7OXeKm}ecP8U!8iM77L5D3 zK4MCAi{dBl4B`^Vfe6NgGC!yTQ>i#v|=1L6)R9=->S8jcGE6fTL06 zN)7S^|6w5eK;@LNGwnm(Rgb0yKnvjbfQ;X>d;pKX5*iJoMKNssMUe;u5Wcm~h918~ zhUuFH6=Dsuc8T;+*034jYzhXIqkufa0$cL&lJ~rS&*AaVFYyX% zD@-(sWRt3-usT$CJVD@v#52n4_1ev5`mwvo5M8*LP9hp2qQFUIhJwka2gNTT;C&VR zfufop1qs#P1D|2B1`{e2_m#WGR$9jz}nc?#>>o7 zNS8lP>gQaE1@sNc>n)sUsphXt!9|kSl@#(n7amp)oNAf>U1Dh*%f?i8xfg4rnBHJt zSrZ>R@?axL(Sb6UFb}S^U*TTJsX!ybsMJmBO83@ta6Z9?MhV9(h?t@U@RUno?o{Os6Pa*HCT?0`aKqMgAni}ol8goTF&)XlnVz8Q zmoGcSEf2WdUkf7&8wopVQNXUk{CoY982()<`(X?tp$Y(RA_>#~GpDquTgd>M_D55` zdE_sv$x-|%Bg97x9@cvjdv;Fhi{@o?Udp4QE+_yOj(I=*@GEPV56y!Q-K2z~R*k(k zQe}bo*smEOVexYWWe5>@ICL;9-$XNn+4BI0@iZ}LVq4D+r+U!LlRm>*Lh)kU-?X>X zVqnz323vlh^oqlEbdd8LO}qu$@i`0MbP1YOYVvMILqtzYJDuSRFfqF39@2S^n_QAW z3~QmJtu5R_F)qy3GDSY~fmkq9&A`;t{OZ zg{$k9yF4+lbkfE5zCt8nldJksJLctjjomtMP*3^kg5M~{$V6PyEJ@_iu3Q)ZIri)f zIkX#{mq5}>UbrQJY-kamZSkG$F-S-&ZF0;3J?<8@Ff56Mvb-E*`bXnZ57_WxUnF67 z7zf@2>n%cAb7;_(j8hnPpNYc`pmwM_gZmafuVcH0WNnRHUW!!aaEd6=F(O}n$l7gz zG!Wqi-O2mvlpgs+C~i(!5VmF4Y$Vzw9lark`GADTqT=3;RDyVs(ndQ+9K54AH1hDm zvVk=pmBocBK#qmI*y9D{t&4(;_{dYz>8NaZ8qLxyb7z1m%3>%?Z8XO7Z5BU1l;?|2 z0^^TCK}rLq=aVEzsT1&#jmsIIvrqbir!hw;SbAZO`pzzB+h7ym%>j{~v%y>BH;8Y&{4?ztGDtsy~tNKf<}tU%a`U z^V}!jlBJtEpv1+uNzEfe`uPTrI-ZxMMCIt{Z%;PPx_&t1Wlfk5{pw8I0GcPabY%HxQXa_WGw?K zN|xbGG?1_tOd?t$L^wqh_IUq>p#}DNbLpsRur8R^yUeF=c9lJj#~+OW$;9rlwo#lp z3Io3=Hh;P14H!0EPyYcHL^1{Pvl#notY<(oc=vKN8f?sV?D)<`lVg{B*+4^+V967P zvbZ(!H)$IJ0xp2)ax(fWcT7% z3MH9Mb9;d*Ax5l4j^ZVs?!kw1IaU&lQK{UP4je@s$yZ%aHjvqR@5^+SsmxKqLOkDw zchYN6HZp;1tarQHE{~cD$xB0ovc957_$83p!5kiyig#v;eous>f5Ob~#H5P45g_^_ zw%kg9*oy<~oty+xw3uA3xlCg-3}%D$w7D&p%! ziV5V>$<|j;vxohY=_I{x-@#EQA4L``#w?YF^ClvEG>7G5{CVHLhL-fm@kDPj`HQ)j zm{eHi_XQ|0Vq=i%Rfb~!bT4n7NUq#gTuBU-0Sy$5hfc#$z-XB=ZTakrDIqD@JkGEq zn+8z0e)rBiinm}Me9_f)TjbfGq)xZlb}?`cGHdw|MiI2x zP}s2Jel$`=v0+rs@FG)5T5Q3zTRlP8Qjm!B#;}|L2?I^GsCi`fvTrY{7)sfw%PC`X zAS{wvJiCpQb9z`JSU`?OH1$~!`3WkI`MTJhW!r)<5S$b)dML$_^3_N8HaNEAfFQ!- z|8Wp}KxOBLA<{-Xa$Mc}xA=mK<&d$|CT`+CfTU6-8~%$?0U-iGlPtSwASZd`jx|CU86(ocPcsY`pUo5RA;idX6H^C^(2y|!oDB6c z`3W9tV&gHH0Q#MtR5$DDenB7dwdUpLw3VgolvIkL^QEPVFZwqV)6c=Wy$bK z7#2ojLw2-KAm(H-G(m}jpYaVgn88oOdGw3|xTN>MjafNRYVx=5{i!m!(ffPp^eCgX zBeLZ!6%cN~t{1emZ=xXE^K&g50D~)&#SLR1yp`c-v~IjyLg-&Z(}U-`zkNrT5|}95 zx8;Kzpi!<7?}+9los*fCHpX1}p_=}PgwAj8eZB)Z&5OSxFZ_<*+=o5u79q#N^K`j% zmVCaCY;b%nM%nV2KzR|IAbhE?6(0ci)U8R7k0F`|VDX;_aR9Jx^^eBH9;N=qWX6rt zD&-}qXo##W0`ik-f`m~NzzU71Kt%i)E*IFHfMwdk!L&XjKDv)StDOu>;xV!WAEobv z|26EL6$#$mm-bb9gd`yJWGm2UFoyv&6c{PbJ2(eF)pBpdb{Z@zj0Y&-Pf=*>y-YS# zLMoUu{J`XFtkRd9O_6tTkG7!(s|Xxg9t{v`p`z9NpXYmva6TPiUQ*#84=K2==Hu)H zNrp&1MqhpGBi)C|2yT`R7@0KXLwTL70yQrdSwoo1-|7Mv&X(r|;&*Zqc{-bRF30 zbfXEwC(oaRLMKr!5%(C31Is^La<6Km1{P1GXFMZ9IN5K8fYTXJTT5fmAdN(2K}ZzV z9*c>qwduq8xpHfZ)r# z(U8Z^8*c*|!P+qdsEMy`f7uSZ+{B(sEx^CG$*))h?Tyk+Y2xB923Z>=l)%`*%Jlc8 zEgUKv9v_puNCD9)Tug)l+CI|oy+rFRg({Q&`GEFE-N)!s7eS)T*|(r5+;Xi zqLt(8{rOjb2RQjkCKAK2B&)L3F#8IzA@k=dDHrPyd*EoW1ql$_Kbnql2uJZ#sl-FrzPe25VXe%aT&HHxX;MS38-ZhfGC$1Z2GG*5Db>F*HZKY3FMByr zb_{>{Lh6lICZMLK6gm~ci9+_yr`2+B7NAMiKE*~;hHqy1>BH!pL^+vosiS9^G#_-@ z%OhvKlVnhjpcDs_i%XLbBRMEYG9qKG&BD;sOkcIelxvB6xC_?CAkHb=kkX%Frx~S> zl(<=DsIIjGjOMMA5PJ-qM=>!(1E`EeT|@M|JQAHdvZn;h>0c1FzWAJ?FIJ&vL+tT2 zMBXC3&m#_7Va%v3IBb(?X{`j^LU`{;G&-i^sk@mwSM##v(z*+Cu9VxH#P=7}icmrf zLuZ#xe>BE{an@K`@`BtyKouATjwTz3Ap`yaYLH+g86aJxF}>|G$4#Ls5c^FS?oNe z&pQGQjvXhaa^x>_0L`w3lY3_&^P+)>Fh{a3i0&*$g24x9C2S~w)KQEE)2yRt zEok+*-c>VfC=ncnYXRh1;G()Ioca4MZ&C5m!;VS z;f7jD4D27E|Nn^x$Nz|jx*xe_W9Urcc8+5Uu}^K@fXsW=^iswp;+9(=@MIu{2k9tIXpADOG#r~8Z|t&| zY>Lagbn1Iz|BoFL{VRJyRnKZK)GmawC|D*7_Qo>Q2Uc(l*(9)fxt3O>1J5VDeZMJ} zWUDWlvU53yCe~dg;-gVGP`K&_4SHJBd?h0C9#QCpv{q~G4*yoI8ua>AL4mIv<=HP> zJL1;TcRx2|ZXdb~JZS6wHVe=rM0lR~(EW7#jIBrT)56Iku3@6({wAayk8z9|sA1KZ!O5VbcI zv#bMd&kW4GKR(;kC$(l*CpbwkuEX^i_J-p-9hxFeSb9f{SmB!kcdul8Ix#Ns>A|>b zI#hv~{e0$Q8>8yxPD=Tw-0JIX-_Dzg{DdB~&F}1-^jta&cQ}x#ON(3Ybw0hLzJ1~b z;il7$xbwkzZL&xS3Jk`j*F+56&RTVCh*kQ|BND#KrOf-*azXy-9MN9f9p<1Ftz0JK%Ta(<9C%S4cRKgvUKbQ@5 zZdRq#Q3qZZgX-~~-`zk$6y7<;B1m1Ax@D$YAJk3~;7Vu_Vu=yJv-qtM-9lAY$eeC}KynZC&v4+Pdck{G7nOtBagl?Hr(_NnR&x21mY<3|;+DBWl@3NfgpPoJCnI_iR4vZ9b z{&c+^_vzi5+U?c465@O}`lomw(!^R0+t+r!vtj{Vq=Zsgb@%d}8nP-lMKA5?v#RHt zazvKqg=t)@P%XzGNjYV%sWmNP1+EZ1b+!~8$ z>fZK#ec9=-lb$knyS=oVKM!{VwclQ@a(kvbKIGbIUN^H}7;${f=Zh+Dz^7P&!qXEI z6IS=M)dE!>7ua~bZQcGlrrVq;CgJ_9UgzW974>7f4Hchn9d%h(j*LxM9|m}ZEu~Bi zcVIDDRY`-RPfEyD89S3Ac$Sn`-7D8|kEZ4SknsLTJUAU%w_udFXO8%EU|fd^R?yW& z4BC?R_}k^ILI0z$^&qP!xHi{oeSKtghPL$Iz+B+ol&Kev4=_Y^-FY~R;}p&3$f~Nl z+VW`B`zT{{Y`gP;+V1YX&FD`_r`G>tOn;CZZ6!WPa1Nl5!uVksI=y%6A~MaN(()RD zLjwZ2-5Olasd-NTrri=&s5-|dwu;71isxF~aqCjmXntx#X^Mtw-AqZlYSvpNEtFMQ z-;eFg9+PsnbhvpW!(kjXOp-Lg7( zBI2&^)jQ*roUbOmJ3Crg+vioCDwkJY1W&ctKX&IX{GhRXX6@KNz?DwZ8TY5;+`tzM zt+x%WYRkDNf`!~xC9++1qKj3(gy+XSQ0d=)kj%Qh^|)C*n%mF02BctCLTbm7JHDxP z#%}I6`^|JnG?x?~hdLxXCk$#&xnQ%NIE25*xdZVj0!ge*O_#1=Chr8=rNdQDntEYH z{cuuO;UV(pH&m|+dwV->n zxNEOkm7CzI@XU*kz1u~)arKKrWw-BC_uV(x1wO3{MV+}i*C8HL)i!~zIqu{){y^$G z&wP#51RA^I05R4ms%dI|Q~DuVzi7xaw)TX-OPOc?&8Duv8v$S5J?U>bb2nH^qjgk& z#p)a@)3DL-{hbyE=joeHm!%GGT(ecjX+=hB3E?#+ZoF&mee`(r_iyMY1@9LR#zmgA zJ*tU#;1Is76JC{ia(mC&Pb(-%VnAqfMr}Mo_OztMC#z_df_7{35a>TdGgF<6=zNX; z)L0!j16Y~1cyUy7@!k7Q?Q^?(=R)cOssPa;Fx%v|@v}Uy-HaC91*;LvisXyZcXj^G zT7c$B!Y8l4%jA%l3Xu`C1c88t95e!SxYxUKTmtUSBXNs=C*yI-96$PkUx4|JU1?+g zDJZ39>AqH5a9y`=tDlAUDbAdh?P0+kEuN1vjbl>RnT1pQ!gQC8h#NLROcz3D^-S}p zb06a?#^3C7NZkJiaQg??FT5!8IwR55%k%k@J1zQVuN~cMD-6r5GH+RB#<9KIZA2L@ zEvud2Pru3^w%+z>9_Q&uSRA`5)S0>SF7@mar@uv|<>mhX11g<%H{{jnc{nN$=yj`B z!`=UGKmP6}%iT6^sT5$hJSx)q;mw)bQD&aAjq)ftzxDyvRQ@HU4lPfe|O!`NC}4E)YSoA(s-exn^sYxK<9h`xJuH z!WjAd3Ac>tMtKGbT8CpnO@8`WnX0R~&DU*tDg+@fINKZMLW2cdPgVInEjsP$I+ktI z^?LW_wpZ4@^(41atKvV&AdfD-cGrw{?ZCntZq6O5&$J|{?PD%M_a-S9UaP^ zOA_&0-2Iw}>ZEg1##r*f*X!$_4+kTLby~%znHRuD#*@IT5HL(fq-EQu2M_A6mk5chro1aHwvX3Wj6KwWlJ-77MwvHzgC5c~$hf zNBHj|4#}RAHy^47^0oQaJ-o{UXcG7T+I+Kc(?e`}hYNg>om50Zvz;9cqzoPL&um#U z`@=SF$Dgh#y9dtNCiw8ycb}?DfAv^!oT&EF<>0`=sc@ZBh!UP(IMcC(A~q#TFEkQ^!Dln8*i(3pThEJiEX|-y`KL7Pki30uI{Hlbo-loh4sVJ zZ@BBqrPW&A+Z|*eY$riA>NmEz_suvJznTZRrqu7ydB@$uMOE3`+X{_uw+QEM2)7ho zypz35`!GEou1^}qGG<3MoUuh2eg3RbJrGD5y~m=uL)&zd@ZHBke>rb^T>ET3bbV~8 z-BL=8>&5f6>HdWpz_#x{K+6^SP@1-zxuEJ&o8tus8?m;^d)eh3Q#w;BMJInqE*5^c zeD?+V>N|<{>ehCrTG!3G_dX3x7Zl2B8UTLaftjjiehTnH_sh=H}TS|DnhOI z^IfAWKfP|+b(xmcujRQne8&F0^JE8Z{hH9^65CcsIBH#sko3Du``@SnC|?o;z^wI| zqY3k&rxGgkOs6AcpT5<0%pIcJxn|iGH$`PW`+I;`vR!h1uAN)s_n2V&E0&t)F!;_>q@$sN4cXH4}A7gkoaYNoEAw_Q3O^xw>@ zx)ZXWbZRrSDz$^WrXnvlG?x9)zG!z!WVDt=hoqoMaMqR`AbsJiYC$Y!V^nROeYw%vFic=DWZfjJEwh`m1E z?l!Ki{dxV(WMuSyr`t+B`-_wodLm~N-aOgZgESZ`@A7Ie>y#+`bS*Hl)b-|7rJp-m zh3jV}&)x`>Sgn@ttbVAk?mKyy6Zqi1_w_&rk13kU&!R0Y0a6$cRgk4La$)D1T?1X-)F;JL9cGs?+W(JLAofo_*Xd zw$Sn9CkfoVv#*m%i{T*`_P}CrAFijuxS(N^|FNe#YL{FXOK{)foD^Lph;>8W&VO13 zt|l)=f<9B5ptE*Y+`VpY1bPSthpW8gZZ&N%-%#t4tNnY;TT;Bv_`wrrDZA>k*xs*i z-!%xfVXSJd+N*P0ceXjsoPQm_&SUO>TyBE>avC0YX|ht;%}Tmn@zvEw6&jzXdM4=f ziOc}}xHiXAuYi*`p0`{tuW+ri(Ei-JBLy0()@leq{sRo!(Kg&Ji1mAgHY_ifGXz4< zYHI~JSp+w|2(GZ`s{0u_vl4wGj88P+Nzd{PxjzTz{9?=dY+QCugsZMSRZAUG%i7bQ zRDBaY)8by2*=hNo*i2SX9Vy#nqVbHk_<|K4bJ6&Thvw=Eq$=>XhyUA5SHzM>DJ51~ zSZ-+O#-9_{|D!hGuG@}lGznQf%9uh1{sXMs@UPXG`D88SW%$cQ151pSc$9<9;OCWR zgB^?!cZKw9Yt&>b{`8v{bdT4318T-G1vviEgN?jo=n zZB+Y}J+Mcdd0! z5>WzgQeUf-gRblsCcD(P5K;sqX9ADbhu?hj^@`Va@$iL;f}HkNLJ6k658D>Y0(}na zhaT3kR$t7RV#d1~c2}i1v194&7cT^TSk|(-6R~F8PKbE-`BOcI7;}923*P-O-p^;d zvrg&TN5jr+L~*dW)b4*l4f$BLTI8HlLZN{mbV<2eVTJ~2dub5t^rh*hMFF?XLfIGi zF;eD^a>2E3wC9`X6RncR$zS#14gi_(cNtCvO`JfH|$^ z`MQ4JXxe8NVXM+qcFqLeX-nDoz3q12tvygPu=9^g9Kyf)y)w#t6O ziBrHXTDupiHO5s09F8>) zATxM%@r`6rPxl8h)HipJB$p+h{#|z-+B(kfQggl#(N*?WM3BZsd7cpxiKXsjiHz#; zb;VWF3nv@JCxg!m#yvBeZV9d(Z)R$zC~xHfgjFhCR6l-IpNVX~e`NaH{L_u4^oTpt z2ScXLsOKlx%>#|C;zCdU95a8++3CycZG4yet~(akE)en4KgUhd5czParz(P;MXG#$ zliv2Cp5N(GRe;}*@)~W8=3N29iLuGa@BaYwgNnl&pc=2UX>K1Syd|i7Gi@4E$SZYR zOEE;AX*;zKO;1h`)(6WTRdFAm`OV|x_ad8n`}@sYkv3M#rwxo z?hj*+j-<@r+>4Taamu_N z`M5LN$!@RyNx8bTKey;5ZV!@^n|3pFrdGj)Bu=Vdk@BKrE zfxvie(>w3O)+4sdUUk&!LLPq@_!)uKKPI^B2RdCjF*~FosZ`tOfyV#@eOkMZ5C^10 zg;VygbB&I{;p7Vjb#7j@M}3zn+JV&CHMSr`HNQ`MJSXz9;U6Fv1=t4~_=(y)+8KbX zZyu4YyDDZr-8QKC;iB^6T&v`7&+v?MjLD1d8#DXH-Mc@-BLt2nx9g-Hm|v-xSPhg? zm2}?EkWh6hTICW=NN5sG{viFt+)B&w&wubc+<_6MZhue$Ju}ldhrvG*6Wnr5dkarB zE@&l}ceoZuffB5T9!sgcJwfjBt5$9rjtQ_(S^VDMT-};vdMe>3WI-K9p}+M4eFrn z4NWa>b(vc0_Vw-dYIO;`_QPuM11g!>**jU+67^}k-FZ`-!mp@&(nzdN@!YJaLG?3s z3zf6k*3dt|&seV$;|CWxefqEbA;0wWF9T3mf<&N)DVOLk0iw-R?;=#y_2W-1!Ut&i zo9_W@2jchKqn+&z?#j1yH=a1A%&LmC*HY7aUYQkZsO7ekWKpx`cMF;zy*h^nAn0?xKde$JrR?C3nR~LPQ7cK zn7V&&v7%MWbxWS&xZtmU{b6~}lgl1Q{uJqR45-OdEybq=SGXu0{ievoD*c-y}z0_1-KlVV4WqN<-eaO(;5> zzrT6pJnq0#<9TT>ua7rg4^%I$Ru?|^sY6v?W=LS$+{BPGle$wC zU(_d27j5mh=5{h#DyW~ZoVcTMY3fO~E%Rcp{9=uoswK+JDIVm<(Gk)*HZ_#`mprLb z{~Q}F^V4F?@!DIEA5lM2f#+|s+w{;xV9H+G=NC7wc{(^BimJygrMzf(G$oYcg}=U% z`CY~Fdr*wz9Sb{^7WORpejQj*G4p;#r~i1_gPU^hIP`_U`Sz#QMrn2KeMvHH+s~$} z{3T^?F9%A_44oZ`ENA9;Us;TOBfs7#8x-b9;E*8s`~#TD%tW-^>kwJBd9bOnt0{0w z?J^)HVKlaNa#8lCKQWDz^r2i#`1h5NMVP#S|6-yu5N&PS~zqiy%4iyvYjw zO2u+s6uYyv$n&tCM9%lKK~?9Lsy@a{;TBP6Q^RePo`}W zn=>mJyAObl-QKy=O4aR)E76Cd=Q`qqduDpAj$XY9Ja;r037!G^Z%;dNmj^_iuD5tn znWQ;1Y>u0vO0KkNY2Y}xZw9KetBItF%$U+cBMRBBs?puERnY!bMrP&ISO@#Qop&0> zu7S$i`tD5}ji1+#hXas0v{~JR&S!s%*U14bFTxJ<>5q>eW}6(V@8NMWZ)WJUMBEbj zh57ZD42J6zF9KfQp1SIt$`_ZQ{132NDe4*7d?J%rxBGMGIx@nBVz;HRiap+Bqf5vf zdNBI;aW5|F3|-J%FIb{~c_z=4G<41RiBtLLw-AWusPO>fR8h`g)x3{b&%rCLnuw_f zEJ5vCX;WA9;?LC!`b9c#)?e@u12j!(Wr#gm?~05_ljNaQ+_dv1^>Wb+;#VfN-bGkI z8pmQi?|pQ2P@zt3XIcEX?$l!Lp8i6Y2dcwk9~r@XwXMz5UVx;>-l?r0Kh0lz?<1@O z>-^0q_1J_ao$!!gTT5I8x5oeQqasT~nOZ!yztwh=w)Qa|cUf1^9%$P3|K+?x! zn;2oCK9cf*?X82M&NvVCn#`b>p>eDGhEhMkVlvJX@L-Q$jbfTJSF-=_wbGIu9o ztu6wFeP15jaCK>d%Cx;#!wKDemlv(L(6RpZXPAp0wOZhPE`VG(gSMP_FTrd=j1 zw-#M?{rn$J$sA_pb51n1tu`lEsZe_<$ccsQwF;BDFVPt1cE7gcoArpTz`MT^&l%%0 z+L%6mT+5}ZGVA_Wji+!Y0k%g5G2F%6-w{~-&m9twdFH1GH7;-Ol_tFQO|_6U@bY?? z4%cP~mz(}eh*QW^UzaF6Ux?z={10&B+DQ#>4%jyDqazN4%7KES63oi@MZHevqpV$n zONrP10e1DNfFsM<*Vx8L_=={A>$T?_CEGir-1}o6+}R|yRX_5At5Bw;cC*P-onx+A zL&)ZBEJv6Mo^#f-ZlL*2u#3j%Kfq+%R_gsVT;)>OQ}s41mdks#I_t7`+=Htx&daQF z1bvvknWBb0#eGrb^>oqXj7Ap{I~>{8UO6dY=X&{i$CT`qv3tPr_JcF|LDkr8YL&MW z2+=Q=O5eyCj6eBhLZ#_MUh0%)dHo-@s#$8^eDUi2C8vnUXU8Yi%6B6t$D)PK4caCR zqDHZ8kLtGT{{(&-^Q^WrZ)>uaT%K5oNVo1T@p7LSn1vEF#TPu1Uvt5H#fZkPIb`h% zXVKB2_6x-@lEp(%K-Gh?{EuSaR2-BDolnMPP}qrFB$!2Fd?`d=6E4;P>OXjyP-QHf>RTv|{5 zyOR+pJ{jP~5RW@|yX6NsXfI05Xko#@wO^Mtxvzrt*0wL)AR)*S{nM=%_#?#dX7)(cw=Y!~D z_7kh+L;m2ibNqegAl3BSxj~{luDs#~zn?U>raUM(6~*=>p+yIjlXMPHmkU*;+BzS6 z*;oD@_jA|4Fln8T(6>&9_y*-`3aLH*yqvq*^?Ry(n0q-|O;+LEv4HzRBw8Nbxe@BF*+~CT zyHj8~p7!kV;n@9i?9Amsgi1z)TKP#opK79piR^hj7kRDZC*#>F@zM`ZAs$Jr{6hF^ zE;!#vE%kE2=DfRTP}AyRp}nFWo2BhQQ6**DRk`)uEOD^vv!dOY92;F43v{tfBpl<^ zdh1FKR2ssCT~uzuA~G+tXOI~}ieK%tdH(g4pDkpmlDY!)jV)fi-P!Qft`afw5MV+U z9GrLm0eo333`e0mZ!%l^G$Ph2Lvqc&*wRLvFU@Vt-|UzhQsYE8gUm> zgC+@{q74>D0RU~@-bw2cr;J#-gN;ca^LG<+Rq3-Q(|_>51#=3-zO-;9f`ycNW=$#g z74PeP8m4*`>xkz5gtlFgUX)<=7P^A!hs9lg`ng{SpQ#EWUlyelo~Yg&FiEuM7M zE&PUiWhKo;(mB=Ln7g^@wc8W=q^+l>tOx%!&c6TZg#y>k?OSFM0g#AW>c@S_;_{Q# zPd`cO$lLjIs#%6RlNbFWd93PeubUU*_sO9WOR-6b%cwdRWVd5YG%d^(4?nl`?#Xq8 z%k>Kl9a7_$nIEGS#%qRhk-KAVlv>B-qM7Q4uL`}1w=9Pps70>aEpQfDl9mpntdF9y zr)A|XwQVKicE)qJQ;;etc%%s&cO;k7VX)jQ#}uY4yw1INcpu2lvck8OcO7mY^#3Ka zdIx^|nTD0Qx7F2fy0f;o{#eNeWm^4DNl|kJg{FTE;`JVZ1 z&tg9^tIrcV(w8Ij2=O@?R$QB?-}`EQV~SccDBb0UA%@CsXPss>3_B_n#_G}aG`HZf z`X7_)+hxAaUmYmR;5C83IaoU6pmqqvi;^ElRbHw{&n`&R!`)AwP_l{vZr?g^h+ue% z<@QvmjdmMO7`l=ZucxTdV6U*cx8Rnn=>8>N|Z^JRhuV%U84tw zMs8hhgCwr9aK9|o_f?MTB79MZgYP${XIw_Pz3u$m0Af3oqB*tW2Uw<;R8M5pgDV4X zRTx-pxX*#NvAtJGpMMM!%T2C)N*g_D;*sFl?RR=uP=-0TP@DXTUZp+o16&Qb0n*w- zJ%oG~IFMmRxZLD%sH-dufhKJ5<%&c8=3}~z1^ECXkdmoo`1^8bNL>u{H&D-p0lCG4 z`w-8bzqZT0T+OZ5U9z3;sq%ndnPu6UcC>HeucOuTEWn z_1ayyR_{VRI>ufgR3@6eq9=pzz=d_y@1I@IvfNeJjT$?NXmsc}kY(+Vs*EUOU4e0V$*qWa+bYk5-Zm^#SfRygrWxsx^~!--5o@+sd7#HOm>CN6x zC721;pWaCRvSsKP`Kgi*WaU_O`uB}Lb-Z&yLHFkEer)6F)TCz=uC`p;y|%rxOd95G z%K6r!NPNsK;%e3yCa0^PPTE~!cZ7TUtw=_VbLOM?;?Qw)FV@H)>!ejK1@DO_(w}Xl z9|K`EMF-dOVuRh_MmDP9$C^)9r=o#PoFhYSHaQ78O@4md&0*TL8Em2i@tW%IjGxWHv{7LK4 z`Qz7rHj+t911=d`?*bhL#30LPF8-mmQ;8=q29xc5p1qB}ygYM|Nu@B|b1rj%N($en^M(kqj+|f-Pk_wj@P*})G>f-XyMExZS(jcSq zPew9;N$6d-8)B`oYSDpi^?2P8+_AZURx2Bc-UE;CpkIDkow@39@v4cobmzhmX^i>b zj4dkQH?y@XIWUHA+denRJ&ZF$F`amRwi6n$>h8Nr<4;JWbUJI8eG8b@;MA$AA5PCW zqrM!uOPvel1$HRqHIO^l?gA5b-P5AUVt=85(l^g1Q{huTzkHMJd}0A0+w6{4Vklhs zbmY;yk>U3k`pTavobT7S5gB7vx-cY`=Zgv@NX)wpPh_gf6$uqm-uWm${p<;Cm`tq^ zvmO>Q?kf}A0|k3q8F-tE$tU0)A1?$C&3_*H?0fLZPVJ+T5QqABFDkuXBJnLk8jjC1 zet!dX$gzIDINpzK7Ntg$bq~yv$>Ao7g5qe;b`5zq9SKL=awtU|2drxonWNEdf@<Y|A7c&@@&LxBr{=`_7pH7ewyHhjH7yU=JK}uZ!amUe+0<=W9yxyM!jd zr9X#XAoF99b3y?1+_wwRwbRn&e)N$B9;`G9Pfzhbbmf&hmO%bV|5T3KF)AB=>~xcK zT*0Fz*YmbGFVZjrB0X{PDYfAo&`RS4C#_~aiesgrq~i3|3>YaEcf2DWQ~h&cnx%9| z;}E9)3_mK{?H^4G-IJcN4xy!kqIh$_#ptb~(C9y)^(Wy(ScAan%%R?>>Hj1wM*N5l zvO9lCL`p;`!#JeVx`)rR6LhCDF9(SIgVNf;f*2{!erJ&V;YR$ND(Kp(*#2_tp)Jk? zp`F5KtPAACe{r`#A`dVOP)}L#5t1JRh=TK za7+CmIfOkj7=%Hd^>^>}^p^WJ82+mK0PqeDku{`?d&Ox*gB5p=79;0CW4MOrLpY;J zQlW3Zyk3Xe@(;K0mior$!58h!`165in=ry?Q3#*b1dOH(aBR2?a~416X+tk*2M6g( zCp;ab-0jZt95||Sr~keNAyB$yHONIWMGgM!*IucXM?MlB3enxDa2gw`a_MWplcZT~ z_XW*{!#Zkg-d(Vsv&50rp_tC%(inE@RDG4 zr(qiUk~%G5F{h>!ks3p?Ve&`fPOw^6>}nO-6?a>rK!y5lhBt*VRt!=I^+xilTLA{>*2o?_yDD1{|A!M|gHr9jH`}T-`=SPD2bg{>guJ zWDgFyYZxSZGF@TQ1XhED(tOA69E^!{$?9nTfLBx{`nQ6dt1Hn;y?>x&-S;?#O6K**!GU zo*qaj5$H{MmyoKg&w2WdXK2SL$)$jbJ#+m-;vxnKP5k@B_{;}iE*^XoiR~G*GnDdL zu!%jL5O5sOcbItIoqUtn2aM1WG;4t3Qofy+AE&6VmMG#o@0r}Le2C}nvX~nL{?H)# zV^bPorx5+_Aq65C)myHf)XUh^lKzrk2PKBWdtROH2=d68QDHlg4m^x)+6Zf(JX(6l zSG4q>MD2ePX<{4sg~Fd@HGaP3Vt!1K!Uy9y-h~g(c?5c+)Z5seQ&S4E#+g_3dpG4rN+#}1GJD3u{i z7BbobVC?>RR00Lr7)mK=sR*GYIbEOiQ)&Fe=0Hob?Z5H= zZkv2HF|l+Eho$qrqZrsWTXI;#7u+rW*nyOIl)U?a~LH1wy_3hTnKkB!YR|DcC|7sLI zl(g8|^`zv>&g06-GS5V;%Y9CS&y{#v(1d9S#K!lO-G#OmsflSYUai{m2K&aU!l4me zhE1nXHuw20K>t>G=g;ojmp@sPp_*e0^sb-lt zH`LWR!Oa0jc-gh=dyIx8nJ3uyDT2f)D!F;w=bwifcl~`+sdt8d#$LMmP5Ii}onDj`@piJH2f>cq0vJXj_o1THQ=f4UB7p$qcJq_@;aGyuokLKTv>BVzmAv@t4?Wwv5cQ(HRECS(z6Ah~9oALy4U_ z!q;pLLh;QRdzcT1Ew>?1EN(IwqkwJ;#)pg{cr|jd7dyE53OqwG?mqZ^*U%A{j&eVS z64^IzZ$Gl^Ah7R179?v#`;F2L;XuUweFtMLtqPJbAOCIu8<>`2Hu2_}Zomb(eyhk|p`ooqEH7jGMPSQD{_^C<&~0hon?OVHPxWhd_4rZ0UXtpKU2vzPOZ|{%CAZ$f zNmb>4SF}HPFcs$Z`;h;+8fOT8H!ywc!(;g2mqydkcq)rjNw^WZH7LE+YcPH9s=Mpe z_$=ixeg6ogp~7X5`^>E06nbHuz@%z!{k<8`5Qog4zE#=0(}f+(J9kN!#jIK;Q}pVt zxZ-G=yHI(beAfrmRi2(HwGLE%RS|?b^Wup_$4$4>eqYtHuAKR7rCo8dg3tGd-nV>W z>9Ts0&9Tdw*{+mHux2d>lOQoe;J@IF5^!P0_ zFlOpDK}}68pSYbgZ7sL}`CH=L%xIddDqabn5|2S!o#bCEFA zbPHjiRj2tJg!XLWdqJyrD6YbB>sq3HqVID8_Y){q+nd$$q=I^5jJuC(whJjz+fN#f z52O!b25xbkZff=qH$lE6ui&2I1V*2CG~GTscSIXCnP+E8P!n#<6}7*J*I4y0Q4&C! z8w*Q@V0Ocon3FLy{bQlFN!@n}l#ApfEn3?Kf(ai?j5Z~I$ z^rj@>lJ-McNqggu1{?)iZaAT5+2API1?ksGlvmHqS#o=taFz->=~w%`8{)u}zi-pP zHwkzLUKqKUb+TW9+4gkS088bQG?^h?s8TZ|p2ad9os`B|5Y+TkDD$xt=6yuPL+2Zk z%6UG|{63adVqKXo)?e-Jw_9&S{tYMDKd+~-}fx`Rh zFt-#Ms zJhLy-sDn_@g_26D5&kv$g-6nKUTHKWp04L4q6##7n2x(&^?vyVvaT+S54zUvT$=A| zC4{xB)s;Rjde&k0>?X#!GuR{OA2%4p&4*ncm-{}ia20aN?(2z?@woJ3WZ#ynf0&9X zX}%`~ueWm35WZX&=~12U2jaJ)EDmLxb&{d?Jvn=&u0Fo6_wXb4tZZKq{|oQZ^YKY$ z^^Pa6x1==+#kSo?TQ5`W6|Ah{WAD=&J*D;6>*(*yH{o3y8mR|6)pp1Vx6+``w2oLM z1s#N~n%<4x`8?y%v8I-(i2V&TdDTfB=7IzOBf4N&G;%MaCmRDmA?9;$H#MR$%!-7! zX2~Qh4~=QOV}wq&&rk(>hbvgC%BQ&IHvIHzO=o!Ee_bou=y^%cBdG&GvbLQ|A@G#P!B?zxD58-sGfC-z*(iz5B2+ zdew+Ns4(OX05%!b|xvl@2@d@pG5LUez9I{rg)wnz_h?6)9F`pwc*4Zt zqEFX`LAhr$tA5Enm-`@O3p;7X)`4Gf{WgroXmSz((EC~z*N*#PVIrhy?PuVkdws)f z1cm4GPv?XFEauPJvYgKUB+B|$jZ1epN}JWG>9#%3lLDh|(Pqvkqp8}ZgS%9Vp`VdB z3Fy%K>t?{J?2+xclwb;$rCo}Zf{8wxe4xP?E?Ttbn~;HMBfnd{DA`j!=F>V(zOQlm zqVc!ib|PZ(Jun-Y3OGXD&N~D&{tfc{HH60=JO0I-tw8*cf@Y&>-7wu$li5!D5ukEl zkYLAZTt?r84?i@wWW#l9{B1kLDGlYrMJbeMb2H?!WOu5~$rJNkBgE~UseYv1OLGOU zt>U8}6`3^`|D)7gQ1mqu{C3BewrTB}kGqiB`S* znH2q=I<`EpaEKA7ya`>=L*C<6ABQUEPIjDS518mg;8;b?NAKGV-`A(kJJwbk`x!i^ z)l%)g+2#A@&*84$VK}avH2ZKg>e!mn^LZ82HSb^xOaAaxgr4KEA16IJ zJY9S$vd%wD-{WJOakQB7w<*+pY9`@6;xJEr=QC#|vg_N1p_UV+Ho61;@bK=WH$!jGMH)(bn37T24@kTRaj6ZJL@G+sOm?<(BpNg8)DjK+Yj-@IqeD--p3x z51UmGM+fE_rjDCar=lj%#R{0>Aanh%6va$0KDL&Y>AEz)yJx?{y-qGVJUne*pDwsof+UHu zuzFZq>t-DBg>bD2=#J$z*jD}(M7{XS1)Y|)d|1%^j1e*AIF<6Dpb7e^4)b#>mi!xE zo1`1yn{n*0hH6^kM2eB@k2s9vuexh{4b#M2nu8ra{!bCf`aE~x?8kdd%VK3!*a?F? zPA!;qd;9Hvwf3I)lw$m(DV_v0Jykiioz@#S{KMhIdj(D03$jZ{@psK$ll=!v6Sn`(*VmW4vI3DQ;P zdy>y!2VS4$5y-DFB7T$p*#T@DVOR@zNW$duMr~8jx0zO&E9NnwMUQD_N$dLG(;ZBxc7f(Nd zGh1Q+C-sz!)ifH4P{9|9YQHSI`+DhKobyw-vXJX>7Y>Nva;X_6=JfttPWq{xT9uyf za>>fw5NZ9PJ1l(!))?tE~P?|%|aM$lHcTEH}K0Fvb{d-5z5gy z8MY&5HkDv-13494Xjc(oc#9Afkg!|hKMAcf5k&)@mdAiXZZDR&v&!1gK&pK3;Ac}G zO>xTzEqc0pvWIXyaY6sDUx{WJ`)H;fzBHj{+$}yDXW&uHHA9+z(9I7Ftz3y-MC<+A z)|yhy_!&Ck^s(8&p*<^!TFLE4U4HL>HuDt1e&YGdVAJmyGT^gnXXx>xw}Iiz^4!-0 z7^FSWsN2cg`r($eiG)Sy`*d#uR%} zVvfCF)_ckgsrg(a`a~;>^WYw&ORJsJZtY~F(>8Y1YzsNL0lwF*I=3`Cq{yWz9+a0a zEn?O5tiFV-&KdfFUk%i{GS2RZGytSD0pm3Lm{WmC4?G**=8Ky+lFzp8irG$yjI6%; ziJh&lob%#P$1m2qrq6TkJfq~}-dv!}iOVB~abB(Yz5N)_e+ zI*klR5tKbS3iTo4i7>q`z6sM)(IlXJ^}1oW{=?VO)?b6+ze^Aj-PGa75F2S8f&;dx zSm2v7@@ibcz=1ftxC+FQcwz#{`~ z2r&Img;3(kj@=jmi2+0_N$h(yzVw9wy7A>PzI+e}`qGL?az3%=4Uh_4tY$!v{Fgae z3kV5!f3Oj&^;kp`JKUBw8}VjiImSq`R% z>aj`Eex`&W@Ci?1?ku~g`DX}fxz)8agQ`O2!)|S63=zEEVQNdj6U~$AsLAra%X_Nl zQIr+Y1gb@;|M^mt-&Am1;!)Pr?|g8hbm{Hw+_PKITi&l8_T}EJC6o?BAF##Fk%P9MJIy9Hi76yk_?=lp!_UhRIL6S@p|13?KMA45AzBL zq?jG+?tUaEcrQ(=PUSOsk?)p-3WXT5*k}2RRm2aVont;n4`43Y-X^9`N2N`%S)>=W zwggE;n`Z?Y<0K znnmTF2fmr%7v=X~2ctr!iR*MAP=>^+O;_B{90r9Q!SF2wlpaiKG`FGKo(34!=1cR` zWtP@TXI*m%*OHaSuZH-O?%J)xdF>+#9^@m6lY67Klh?i+(e+&SFO1Api_5YKN*;Al zpCpFyevGw{Xzkq7zqVDY)c2w)@Gzj|uSpic?QzxFuLFd*{G@bxVlHM^ME~oMXQ)Issj+_schSmLMVp^`@Gh7k>Vgod#}6ZS^IYVQ~f@c{-uKpoQB7!dE_l| z?EO_>cW_!orULMQ`=THN;1AP2n4GB+c;gTOgs&Z{Ycr|B57FSm4P*NBp3oXZLDby~ zHCIOoEj6yyv#!fwXq+>L6RYHz7h}6?@0LbkVVX%m!#j>kns4%slE>odB+b2fazl*X zc=)J4@M+4!{vU``6{8D#Dtrvp+2Av_h<_zp-Y7=MbLzIDH~E4X+?u`Y$=^HmFN_+Y z5JLtat7~Rn;1pWmPQHXz9=of!FqPgweLNvup9!gn7w{ZEw#4xN%|;_#ho9!Ob$!*O z@r@dzij6){4y*vvjc-mce6qr*ovZk;g2vK^yNBw3xeiD)XsWvMsz;h#AikdT4Ac8$ zH9Cqi)luDPl5|j<5p$rT-bVve-V(mR4OF~84#%hscKi2<@9^Do{}`4k zBF7+mdw#^JT)6pYz{Fm*S7m=NGn0`}HK+@)qNeZ+{Tn&3k ziP1uoU&NE7gC~#S{9lYVSy#^)Gt!s(Enoul%Y1kk_m2epwf#3v?bPk}RYlEe6MhyS z-n(rtX{`Qi1TWiPUPNjnx?D*eC8mKxB2p@Fmv;5xWWdFFwqBbfaMB>Gkb@ zSKXE*wG35!lIvsG;-x5)zJeOfb$-HK_Kv>-v$v)i#4237eJ(>ZHr|asv|0yeKbTD2 zf5+bO%`_#{OJR4d4$Uea{*?`X+%v`(h(a7jx_<4RF3aw|`MN$xe^G!6zIkQ;pMdTQ zg*(+$N-ms!c1g#CGvUykA)xq&@lYMc{;1czezRw<$^$x z!gh@vyqxGQ_}=!hOJ(VAj%@bx1AC`JpwG^oRb3*f?%X>;Ui>x4U1; z;EO{SO@Ggjv@MLZ1SzKiyzbri@q~e=wnV0I$nyU5#YYa+G;9ao;Z0y5t9ayP8K2ik z--H=Bx8<6D2M_P8QGA53kBJL373|mBZnGX3^P`_VfUCBlM>R8uyBTs`%?t2vbl9$G zP;G*_!^xGEfYLBTYs?wDt(|^bpDq`-iyZpy{ob zLUH88(#)RCYsl7kdTqk@Sj{e}&pgW;cH>b|{oHQ`3hTQ)LX)P|E_G$2G>^`Ewd>d4 zn430*D1~83Vi^m|ZF|!r)8b+*cI*2<@@g&fCeU}TF=%2*XQe z&UO=0pA&cdAiKmd_b>d2bg`Aie-b)|aHSU@g-~F2d=u;x3p)WXapM1FqJ`LiMsA98 zL$h9|p;St0J&WeWn2vtC;oG@Xcwl=O5l!`+p(y(Ms0~et&o^%*@OpfJWYV6v(|MA) zR;2A~64FO){jQYAU(pEqXU!Sso_KI*Xk#IH>*2B+BVc7vndXxxP}ixx3BTr35$Wm~ zDdh62Y@$D}s>B!U^0A}K9&kKrfWb{mxe<4EnWo zX@8#q@ZO%r*36Aq+l8>rH&Q8W8x%g9&g+9iJJ>8j2pl@P8+Eodk_1k{UgRf-+BK%a zvy=)oqb(e4{7GDOuF#7Z3}Uit+ww|LWha&p)(^1R~GvW_BBN+fUBoDIQ(9O`JL`O=ii` zn92nzA;49`ViA^;O~&S0#y5@{B;Fh7fSOGFaE91AUAjC*9p|#ZnWS3~_+#YQ81$Ut zZqcOx#ED-^X`kTLaj)dB*1i4rOseCJ`#WSidJ#|VgFt|HN{^u-8Nv?))dz(om9eES zi^9elaS1r{EZMhx{U3sB*<5=90d!8ybUGJyF=RT@e4ON39cN?LX%wCteNH}Pb8F}E zy{K8X^{*w5n9Wmlw0+A(|4B>pNCm`7B9Q0hV*$m8e%!3j90r&=H zqm?oZgq))yV}a|xcTWc|#heNf2rZV`vNIf`(5Pz@^0;FmuC*iO{3NRnQB-dTQ2SDdl+IA6OVKD%iz6Hff)hy@B`VKCpI4=uYXFhf}D z9@GiIgIA~XwwYq`O4XOw#R4d`#;->w2Y0LNLP?f_b&YQka7bn-vVxrvkT>w*Vc%&h zHX@{IK$90=K=tX?Xi%mQRd;S==3;E&n3BSI%!$QQg4(_BveKa8A9U`&7 zUEaa+SiP$NpmyRe$0mv@8XxaL8iJeayZ@7Tq^74nV;6l0(fgz>;p6Y)`}+2veI3j< zt9FYtYdpB2$TUe3GaQZ1bbR85#*7wgezSm9ZRKLi1fjjdLT-9|rbEQ5e20`K&sy)7 zk83|NnH3=eJpTuwJlv7H2>cj!T}zEW;l*6#AS~1b9)|R1m~0=GmkTB&*L{+ZKblE? zBOnZPr7wE%j9G54Bck!yg*T)jUaggW&#z<>tQ|K`!0Gt1n(Ac=%j>y!@Hh0KjA*yO zJNjbfF%wZAr2$?c+t$LGVe#gPC*VZ!a2M$a+J@>`Fw%Kw%#f{Ko_#Qff{!+@Y9vNi1-99@@ zu}-mTja!z1KQ*6=_!!mXH7=G0H(W{*9P%nojS(>;B zgoZ4kd7m}Bcvn3Mk-BBywfM?}!ZMVJ1XF6=;6|#R-iG^9VuTJk^fn9Vdi(=x3Ys7s zsrs}4o~%bVaoZ^v*RA{rsEWzjmw*5{oA+*Ot*<@+E>y3A?po|r*{VwSZhX_|e?cQa z<(b6OfIDa9SPk9ru?es;nw4nX$dJ}R;Zx^*Po9Wpi4ASNBTG~oijA+NkY9ZVy^FqQ z?-_FSM6sy()JV121hH$<9R+&CgG$$@*hW^2!tYdK$15!CFyP+Z8kc^=9>N%-xttbB zw5NW}my2Oe5XEB5+VCOfb6Y{OE(?LEh01GP3_pmT51Af|D&q3dETpGjJq$V!t&iui z;Q$WD7Z=)hF7}W1hYm>etH zMXmGB?b{Pt0#*dv$Sm8+TMNFfk#S#Ay*%4AhnQ%sr6xUhkceW+NrZQF->!5dQ|@(bDM z0MKP*=AT6$+kBA>)9W43&5vQH$@VHut*E-bzV8%ww4^v z`l6Ie;4Hu;5gk@ zC^AN5Ol9n#_HVG@L`+6NqIA4QnQp0nGC(20S^y8!g$J-*OqwQ=d3 z@o{<;4dy~0nH7qn4>IVL*)wHN5eO5@j%vx@1q!vbemAnNTm7lv< zPeM|)&_Qb*P#RtI5GH#lzw^gq?n}^JN%QF%M0-RYDrTS%#X_UnBc}I>VM7N;H(66Q zF5VFAb?jz#QKm{Gk{ueYC9^OKEp~$rT!RN-qPi;hOWJgxvl%m25{6zz2`FNBynjOv z6%dhsRJB$=VVrq{1V$j_{3{&-ERVa8p)AZBBB#N)!*w%=>>)0L{5O|k9ONwtQ^ z@u9UJ@@U6(6mmmgW^*$ibelI8H^b$yy0o~K(ABncA^?*8>5-+4J3JCJ!7*e_RI5>voCR(! zLz-4Q91Jwl^Upq@wEQBUI9lW8?dE+Oes%yvJ{Lt<&kuA81d+wOVwx9ScJfF`_rP8n)r4zQc=U9sWI=a%$CZ4a)`X zZxv@r9$$1?=lN$ZW3RX%Ic5$SNtxPwSW%N1v1+5ZlXjya8RXbQ2U4FPn|Lq~8MJ}n z=#Nfe!QTf-W(A@P@t0Z&a620i+9X5pW)+Fr<2IFxJsF9dPjS+ES%2pWwRSEy>u z4CG_=yRv&K(8YiX2^_Qm0Ih>zVN0|khy7=V?;soaAZRTL#j=5y=Hkz`4KLUG6m5ny z`a4+J9xd_6`=Lr#w9UyN1ASfK=7EEcj;H7lI4Vn7UuB+kbA}h+VaVvu|ePff;rA zP2Sm;H(xAIyKb{pU<`bkg0hjcw89aX_wz+JAGE6foBq%Zna3VKy%ks#g69$~3^4*8 zXd=SI$d2)qGUk4JbHJ|sLa1)vw>4-{B4OgzY)zyKYD^G4Q^tp4DQiurm*RaLOG`#` zGfz+bdnWLEYn4cdxF@tj%`Wkua2HRr1i44$wrdndJ6Tf!wu25K{ z8j4i^P=7q8fB^)?vStx)1d8Q7g`PICEIk7MfF)7875SwtWJ+Cg!@4M2bICq-&(7B$ zTFM|hnp)tjsm?;@T9A!k{Spj#ZZ~xh9KvEV>!QzHgm?y=D|;T+bQRo6FyF5Y;Gthk zQ>JD_fc){$-QLjmcTs<{>zQrQ!jMTi?wC9}hM`m4n?t&lN+_a*2?pvoVcE6d<9E<- zC7xZ3ov?2cimLigf(#9K?;G7|#11L!DVUmKBQX4_osG)1owy0ofROJY{g2wt=tGYi z08B;S!Yi6UNZmns34b;JsBK=svc>%Ocht^I~@Labz#x*=tH?-e5lC29WRnI z8vRh4dsJA8E@q+=;oJ&z-BOq)C^H%(Tw-CPH}a>0<2#T7s$p^iE(jfZspZwmtZyAR z{$65CM?EICW^?9UNX@3fKLTpYOmPr`w)kQzW0ERN(crzmFJ{<==)lbI_IUCW_h}XT zT05r|7w3j0k3;lHc*J?LpN5foKpeT6v~Q9CT$i$MWWeA4vycrBbCYgFMZIF$Q|oN< z{oZYwbLU0WFldOq^W=S=|8cG5PZW;Hq%8upFB*xE{kJs6>m98Qje)}E-K~M9{DuJR z@S)+MB3`c6u{Gd;U~x42&vExpOn%t!;n-rAG2t@dCqSzD@!Tzj=A$UYZA;D8k~Kp+ zN-gizR;(V)OY~}tOYI5tIGxb=oMNerYxJ&c-{cS49E=8{1vz)oD1m?mGE|e)x2DYUi}I54meJbzkE;V|77KT1Y(7XidTkADYim0;YpNA1-LNB` zX`FHNvlTor*s&m~i3rS4EZ&%sk8K5smqI>V{{+#(FUiI0vOCI-8VQ~(*Z|_``#~V> z^6S!cT;ikHL_txlKDCtFqMCb)*qRUd@x;*Pi4g>PL01PwPto$2aGIc z*c*Ys&8!@efgIx6kv9utl$C%{mQ6Y8`1TgcEKEyHv_G;&2X+lpi?WCn7_toy4I2=v z8^22oT~tFZ6T3K2C=HVECic1m3Kh}n5Pcl8X=&D0F6Oh5D2+wmY}CGgb2i@Y@iMa@ zFyxof|9}l{oY)GA&+$214jukAvXbUu*9B&Z;zQ|Tmrk%vacQ22?+{-sM32)T)aoTb zRW0LqV%@qI;XN?MAq@#Mr7UrgJajA0oL95p#ca#g;!TIxn}s!r)!K7$+Br6u$s;kN z7xo3UK1G=k9@LFTHP*7Uhfk%je!_VX9`2p$ppx0wOB70z7=yO11Ij$9j*s+`lpudb z+4d9ZD&7tvr7pVgC#D4_Z7x+0au9ivGH|OBStbx)bi^Pg#C8n-f=SsA&CoZaFz(j$ zfwUXF`)_!45Y#+fr<<|RuQ3@RcUp8AKHS>-%vZWq_;#uO+R@qu`X57=#vyZ^+E#Oq zNxl%;k?1qU5)Te4#T)mL~$|Ta9`LLrxwx0VgJmMw1$V*`;Kwy2+NZ0-I3V$HntJsEmf)oRD3xQ=>(q>eif;o z)l9yqC|C{@EO=?2-C)XmAtXw~s}#f?F&!0AfI*mF-kZh*s-SIvHq;#!1fm;Z2VSB= zfTvnxuzjVWqnUqdYFt4axFi$(*Wkv< z8S|A-!sKvdr*Q-XvkVS+geq^E{)uu3I zvjqDIvxX(BkX}lpy0QsrOHu6n6lEr<~5}>Ohw+(kt%=|*Ea@W6XM?9$vK=Rdm`vQGYzLtFs z>n;J+JPVRK;XA7`*4rPv92-!z=!fQe)`$JQeYwX*R(e5S$>%2g_qi8kxi^)hUfTqO zgFCf;p1d#?gC$P?^5yvNCFh?BO20Y598|A?UkkQY+`;8A=M>Z}-%ZvFzj6uKtzWCP;%l@_H9+pq*daH-qC+J89#QjZmM4`$abE3_&Yt9!NM(N$R4t5Z^v8He_y2A3xSYzhJJhAiL8QOahAlR0-i3rl z@GxWyp|%y+u`5T`vg2qfYrd?~L|6Xcj#xj!V>M3_b)v5jn<9C^pwpB6%=NhaeswR^ z;=^5=7gVbmrjaaN_Avynw)<_vX zCz$nLTAsygwN2LrwTKLx63*FPa{7tcFk2BMk`~Q3ycuJkdCPw1;-k>nJbtd57cIfz zzO)^&7xUBpzHf`^H#Qk@%>oFDFf$$`+)t=L6sX2aELt_Vr3soN&) z(9YaF4T^u_Esnfs3PfA-Ij{Pzgf<&{NbN^a9&hYA$3N``d(<^25^5du_Wnc&bF@%* zac7z5ad7pw*sZ~J)BXutq4>$lFeHH1I9pA%IT8m>+ns6;>&BKXq2MhE{D6 zt1#^^B2useZk3Hu3_Q0k@uA#SEY7yHlR)adzwKSBDw@z*8#o~8kjiGya? z+$Rg%j(fWVH~Paw)ToWu)q^Vroh#&h%khw}k%I>0f^RdUNIk=P!gGQ>)|n!SX@;j7 zZOT?~2!_@=dgqZYh^ulALbinqBDhunU3g7_dFuq(DJo)G`Fw+F^Tl7WIptf#`HZ`d ztcmL^^vx7Ja;BJDEaE{KvPSXuEMfb#-R@OC_;zp7d0H)1m+vC~p&O(6yqyCt1}+A| zhoe|mgwooO_86t-_I`&Om_}JO_3=;WVR}k_mq9Rq0hIbVRcygdTadediT{2(UrG|lxq!aT%c#$G5rs~ z43T2FU~&k5zTQGk#aGs^rl{?>sA$SYP~Y&?E)f;VylkcgCoK1Nh8qOfBC2W910Ke0 z;<8xq5MEuO<+4 zX)468({?y&DiGz6%W1yeQ1S275)SH^*NSTQgD2#J6&xP7rKqLaa_bz5>Ax=~;V$vk zFO5*5^^wjHzvuR=5k@Y*)H7T_R4w546~BQ|&zP@Jx`4>reTZ4=41gM*+bXv0TC9i| zFyG>(wUsuVgpP7Wp4TllPmupCV6a!>KN5TNG zQWP{YU_FE#MB~d(y|4M@?a_JC8-n%-tIwvmz?V{(#Y#c?js9d@AmpiF+PZd=Yd;Hfm+AJ+jw9r>$Uq}Ot6T|+UZ-jox= zG`bcuUH#qdjvTgI(2fUBgr>0)%|$=+z*FrRBXX{?S1y_D-O&3EtEeGQX`YMYd0VA3 zN(piSJjQw(u|BABEtQT0{i4=;+FS{{>s7E-d9)J5sU53+z@CB=n@aEJB9VZt;yVpr zN&n;M+~b*E{6DS>U8Ji^k<*pjMi-ajND@Mn+s0$~ucFgWYO3F_EQ6+d z!F+w;^lh#s9OT)eO`1E4%5OLpHt~Jx_L=GwUK*C(DvqnPVMId_=qm65Vuf*c`z-=8 zY@VsXVrA&0j@PQrQWP2vN}D*~&j4WWVwHsbn0duv$ENoU)~wo^2|Ov0H+|#IGX0@r z0zd4-n7mg5mqWBk#sJJKYhKnOtI1X~c8r>&vv2o@_wV_nUdR=V>zK;WF}$M2od+Ba z_Y;flk8VBd2uc-_zIR|;{L@?C;RG^8W<Ch?RnCIij z%kq7FI2arUoY*J`&VXrp!W=Dl<2_^`@%+-3YpstqnjO~8ir5NUmE-$#k(ViM{qR?< znPJQhjt2fmq|52UU(L*$lI+_{EeKw=HmzpGgKo>q*JwB3jpjVc!^OU#R-^LAh-R>h zf%%yjUL)1KA|Ys3*{y>etIWi8wsR}_4tfjrrjrQv~fFNYeTbtm|z$sAZw;W!gwPAFzUutQ9F{w<8EQIl1%t$o` zKYL1F5Twtz!vRR4)DBMZZfxa-F$3-OM4-A4ro(6OkrCKFxq~+IVogbyEV6m@B@CMI z?t*x+@ZNaqPkqkiwh5bKu@sT;Mo%q>uay(eZgT|l0H zu#GvRkHFpmtZgteI`S*iXF8TVdF1xizQ~QgY^17L*uUQ(_o%G9n1G;NUnM{qaZ@+_ z6Z38yFFP(Yq4-Ayv^ic~d!0TprX!95O(ht~NJUhJrvt!kk8D0sAIlo$a$?GkGJbE{ zn4dO4vQ5Ak(o{es)?t6Gu)K~rRZ)pro-b!_`7~`YS=sav6*4(hJ;V6rrc#mv9w4~+ z$qTUebKG8gU9F;hwvZnCW3P`+KqzOJ6Jh?>a_TF|CKID{CYou3Wi0_-ib{KDRWQPz z*Is8gN@tecBh=H3`vudsX08rgJNZ^R+7OA17YiOD^?g$;IB)mzs)gX(Bk9~<*ePm9 zAS88>W&c7eKI!0JwYa6FJp%I$G-74?>yhc3=GD;COWwr>Fd|G~+sDYWVb~7&Q%G-3 zrE;rbj=b!0_3EinhGM)U6u$>D|K#Yc5Dl z1`}L;Pk6gyk8j(u{gdz4lWnfZZ7uA={)F+xNB^oVqcQp`fW=e%;XsY+iF$tDwCBq1 zHYE3jn2z{c3W$|PJoU<;b?euMoe3{G{^Q}!q9J4^sraWE&EMHVu^uF9zVBstPR4fK za3h#r9}y@*T)=*vXS9y?W^ZqIF5R#sB4E4#Sk0L`($A`KQ}r7)CwJs5Vb-bQG>6 zS<4$80?c%d2#uIl9H}K4rxU}C0;}HD#1;O$bN8P7ZL`H15YijULfqZ?wL2O>-EQ0~ zLPWUH-I`IAkLwz2Y1`^?+iYH(&SxS^#!tq;%mN|)Q3L5VW4!OFjPnD?jL!0D381!; zgvKPmUj^|~)w7;F`SDQym(BTzONovVD{cONPW%`a!R6N421`(bL^)e`@C|gOMO%^D zdvA&J%UvWAX7MM|P$-DG6CcsIthoxg)nj`1>!IK+^IwcowNNjn?7wJoQKn8V_1jX% zTl(I46Nut}^~G#JXkLcdG>T%s(&Ee&Kc!POJ=JN8J@i10#%2WZ)+FgGyEH{P%?=g0igQotL*;C&@HW*%r+;nXJB-2A=oo-ZC~;=_tjV_Q-VIU z(J8D5+?F?~gf}Rfa-ro^^wty0(Z*QsE5FG8WDagv;+Q>7JXO=qEfo}8kDdPfbJ%T5 z-rKTtrqt;aR;UX%vt(YpG+LXa@Fqt4_UP^KW9$DY_i=(`-yk;Spk}K5N}8u--7M4N z$8?@%-*JX!vtD-K<4V(F;n!)E)Vs2W?`R_Q1~)MpqhYHQuN4dY*x(eO;^jSgs_hw@ zg$|>P)0F_3OkA@Y16ywTL! zNL07y&Cq|?^TBZ9Q;cC+E;3OhN4qs~iR)SKOBd%1SHy7C^ss>QxF(<&iAtp{w<_My z8g3I~JLllSNZ1sx$jCX^b4TKeuLpJA`hWToVJOdveFMd-EJ)qPZOdGPhuuNg!gKqc zd_q`$tv%CVufvdb?!E~>YURmgr0=KOkGB0)b0XYHPBBey?NntMKqz%sBUfoZpPZy8 z<(h@EK8f@#ZCH~yd&nul%W-xp)9>3Gvc^tF-Lc%$`+8Il&>eMPE9@lF1inw0vN!Kwdva{;h_fp_iu_1oLo}ygv z(Z{V-d=Kh)xF$*M#`IM%urUFn2W_ky6=OO?8^y~-DG66Afkop}Tvtv_bFhg%AHo0_!zV27*nJ!Y-#6o>Nk;$Nk zC6Z4i$Wn0Daj1t&7^CU2*dF|8p5>yoRUd~QNxc;W#7 zNt{ouvEa&?jd^5c=@s`#Rn=`LL#(dyIjp-a;qq?(IL|mwb|s^t+Rrl+ZR*u?qb%@vT!(K)kdbB`iO4uMFz`V zeW9M>`*6OGC&^u1Pa|v(ZrH$ADC3!!Mt!gcJ+-SwLR&olVkC9j9DM+e^axIeIrrl8 z-J>~Y9P{!72_*Ji5&uHBso872G}M}!H&OU>%4F?Gug^hX7IYIP3-W*9bmX(d`lD~5r}fbe$2|vx%=E(`ufh7!q!|K&>JrjGe6^$_?QKE(C#hgT52%MaFfd>1D zUfPz!u(i^b1@WNe=y^WuBpusp(?e}!HjME_wadss>r@A2;>5I*tIw4L)SO4=x@da| zuPYpmf^@{zieIWtjXGj@fj>#_P`x96&@b*0M=O<W6Wt-Np+ey92b(HNJp`QF4_SC&0Zm?;%Ev+ zRsI~`8wR$|kU`GMB(BcIRi7~W%_Go#=F}>dkqz*|B`Eu(7wYHBG1a9@=AXSn+)IG4 z=>g?NmR>~88kw}!%&^5^bx8&kDhS!gyRIQGZ=EL9X0b{xEGJ-q$OnA4ne^VJKVT z;&Y#WOS}k^^{JVHrDupj4Z}+&WA>~i8+eGd3G*>($uk)TcI6mL6k+q@lAs%5igWly z2_R^NU9Y;$#lyDor+v%yU+296OH{}tsE!yLT(egLOB|=5CSuJOqh(MCJ~F)n|NesQ zKu}ud=a~DgAHi@FM=)X{8ZrK27SOQ#V>~E_R5MY(0 zZ`a>Sp6>@eC5KJ&hq>Z3XHnzcwISl5i*ezj54MJVJt zH3TMCc}giFY@yGaD86;$Cb@0bh6`}RdcT$~K>F~;6<7EMw%-8yYY{3S^y~tP0fB4B z%Mt!B*`jCwu~RboySh+7%JE<8iNXt5`GI{jm3}6FWJ6Z5|6jE-YH#nhUmu1zE>9~d z6hfZ_rXUunLL|UEmjg#7tZdvhPhaGiI~YC}DN>OKDO0b(BoQZ!3StsQhMI_2ti2{( zA{W`P6uFty-UZFwW*mfcs&X>MZU^Aalpw>1d5}-au|bSSEK1SI^PzVg(z?JoEB{fO z@d~!hnk99&+LitXu$94ECA6<>g9CRRYlsJypL3|}X?Lh(&Gu#YT%5%o*@#okTA@s^ zdY6aVn{VnW!~q)|h|!kICTbU>Z3hx-DR`Kn8A{$P9Z{L_?f)DuunjHBb(8_bVlsgi zhqC9LnI;E^g}fT_2y$PEtGHa8s{1ghEbz*eOW>VJ^1R-Y9#JTnpi5NCqbn< zm)*!<jjykW-FDyftFvu+3jJs+>)} z=A6*ie7-oh*J{=6Gn5N$=Lw|_AQL7@K7?1;i_(EOES@i-mV3gkbGpvQjDG<8!cNHS zWhc{!eu-qzPcA=${Qj0qwT zKU$c8?n+1X3!g3YtUE_vMt@~%FP^Dxb0^8@e*7^Z?Mg%*@$DFIWE2n153Tv8+z7up zgVtk0D{)1pP`r3)A&2do#TTg@ z=aOx)Sonu~q=CZt>uCP)d&4QpC@O$&qcX!Ni|Kku9u@Eq_0E11(IxZU6{w`P1>Lql z#UAx(l!cZvqN~=gc9gZ`!_tR= z`__G0EBOdY+jII;z^S!sM*Lmz0z{c9Q3Cv|gg2iqDx&FKBB<*cw*%~7X~ZzNfd(ze zp@p%6?a0G6QLAMtJ>;64Bi<=ICT(x+>UA*ZONE;li1~NJ;EXN&;34hG4vq@DrzQZQMHbSOGcn$`7CF2q~+MHAdWz$P* zM@1CNFG=F^)<5`-)2&bzzb5d`T(Ee24heS7evij!Viyg#0Xx_#PV7N^K~ z{xJvjatAM~k3>|FS26xRk5YKOOQjVj*knv~GQ=7dUaL^Nc_fs>oi_PT+iT>DFLL!Jt>1L<}Ui)@%|`-o8VlXfv5&nR_m zH1{M3_>p+x)dXgAvb-{)IHD7#;Q6PxPFXnmH&f8WFiQ@ejc*Wx-;on@0h&xrnG-b7 z)(|#7Z);1!pC`&R{;HwqjoV8t^jf`5sE;i8qm>YR&C>W%V0{g{ZaMp}n%`?s_f)`S zzTsp@BLRT`c|#l^DZx;kn|159dg{Sdmy@J{?Rb~j9T!#G^5#)r;z|dKA+>h5y=(WN z3_EFOs@|RSGIQ;*RY%ur7MmQoePmU%IqSDWs#}PN?-U7bG`ddoWCBPh^aAKYa^xT5 zN>03O{sm|iEok1ZktcIFr3dT^JgEPUod~|bO&2PgBq_ZM zC?*L;xwe}Bh{!@hz*5)Alr965_=+rmLe$A|=-dLFPZ9qx9*D=cJi6WzJfo{XtR3c6 zD-oC-U|LV^*{mI@ofMma;iX^2g#9WN)0qIYlA(Lx zwx%Y-pk7knU9bk^;!iA`V@14^+c(y{Q*IG(N9Ce=pRRXaH)=&CbE+c-2cT@AEc&Bd zo1S0#Nl=te_o}S)xi4T(oFGR+$aTi=66gm6gDCX&$Na!rh|#JS54%ySi92Q#&$P9< zaA$;23(1KXp`_jSds#9bzoD+ydAFv7Siu(%VV;6&YlsCb4s5z{37q58h4l1(_xvpu zTdx-x&SNkJyMM)kj!|gI%ZIiG)X_(%T&w=rLFZ-W(R;Rmg?>ED=Jgb=E_-|LtF6Y) zJmO1kbkT_LL@U3JOU(=WHpm!%)fPK1GfJK*d_VuZ^&1Y3;5Xd5)2wDPOib3@!afW{ zZ2ipdcuF-V@CVL24WJ`CWZlSlGhPOz4A%aDh7^y!MXSRW-F%L31XOKdqM zFis=%r#!ee2w4S_pe{p2V@jnZojvrw-` z5|ie>UK~NZV|%!~hXx%sF*OtJhKrLbXk}@V>==5RWB7M-Byy*pBB97+&Yxh!1~9~xP=9v*)i#8@trOvoCr*x#E+5ZluSaScWbzG3AytMJ=2RT7V47;$#M{US+32&lL zjSpXbQ^&Z08>QNL&~xm^+i7GvYPDo-ba(`t;GLV>dlyUk>|tqm11dD)WUM1N9JKHl z;ZabD7D}AWLjeoLdKLAEK3xCS=tfC&?bWfiKrbzv%^TGh@4SiI0XZx0l)74t-){n(#KB+>g@CvW+*GjhpXd5*o=xSc!W@h63ejfzb1U!Kg=QfRIZZ9^$m zTJG4n3;kwsOWB+oVSMruppcR)PJx_89*r*%Di~WGgD9!e$7F5_W{=m03fzhpeCM{Z z0EIF|nlQ{-a{IJ>yaa}uG2f}aTb&lQf=lv3-GD?eaGdU_a-t^HW+26Q9)!4VvBbUd zS%AnriEEL^1-dBwNtVdvD~(Epywq05y?eK28UL@ED`giT@MOKuIUWg2Z$u#c4gHad ze+i`LF};ZTo7X|5hq}VE=P9y;(KyouhBCCiTm&I-LrZ{*B3|zEr-rJNb93^=Tz^8y z?b^o8IO_K77JzfKere_?@wR`{owBs&;i-Gja`={dYI=I(k>JqYfGlPMm2K zmZCrzp#pGS8CPwGOtuZa75(b*TeI(unMH=V)~MW<3PFIujfON&RPrxNW)q(XZiG7j zRht&*QE`Dkk8GY2+o|>(n%Zn$;54ZcdMC_BplTS%Rd$JwBsY%`L*rZ4HUvfnOoJsU zeP$z@`25p|8PTudIPZZ&6kQksO0cn$2?DMm%@L`D>XKEcIN8%gK0km>66#FNK+c3* z9xI<7^Lt>0j;{^(8W9}O=+X*(xWaN9_B;{^Y}O21Gyf~jTIlZ(R1B)bU>;1(&!D7B zwmIhMExJ+@8DUj!>Rj!t`HTcOSjweEWWx!tX@jS&0M9w4cVDwG*t^6R__Yq!71OtI zJ{8TSu^0%%D0?TL7HO5xet`4X=V5lmWo21iofiMzlx%jg)h7P_%hrC$!$$3q5y_Z07 zLN_nn4<8F&`na9+8j#8T7C5SEB!2$Wu24PCCL;qX>QHD?`UcaOYW4NQF6fER->KE6 zklZv45}PO9GL1N92Tey*7k*d6zFHNqBPf%)u~Y~6>w=OS0Vv{AIQV+%e7s_`kDNN- z)4n)~#%oQ|r-u~kd7S4Q0)b(9o8zE~+AF~{YLsR3h*;3I!*8E~Yu;X7EBg_KOp9SJ za-1(xoQzJH&z{oO%F!mSEUH+^>e9+6*l1zRcnwKpiINaEulK;yED1bw&(bp8 znd^4`t&7tQVp99!Q{|S8u;GcZCN>}4e1;sK2RJ1@SnSytA`h&kdG{(H7ua^d_y*^? zT&+@j@)fCL# zl&`?pw=9MCl4g3|aYO@LY1<4C?z_D@sK2q#+h>>)MQa?PI>jM(Cq&m$IQ!Ashy$aW zn}%4tziP2iP3S1NL>e;tLn_Thdhh;+u7IUtd^4ZFFM;d7f3Fl-rnuVHiOCc)7 zCogIwDtce7uUG5KqG@{LR|23h@(b@+(bns3t5mZusWAaKPxi+HC;pjk#WUG% zB5!2-MoEayb+kgO3&0iy#48fkK+rk3w43;tarS84GLxQvz@TezJZu`d2 z_J-X>;5{eAWVh0(@z}jgU1XVI3Xg|}eLyp&5|s{gce$yN)o6SMxPw+K)~UwK*)Ro# zkpE=6e)^i(BQ1ti^s6#T#t%t}MhPl@rxWKai;Bh29Av^AaA=!ZvY32-8NO>yB?o?5 z9gP5r93Mtmup?wKP5N^}P9Tyg2qfC70BD6p8(W?WM9zc;FXXy;>EEkhj?B<`gCjni zw1S=ITE4cgJG?!Dp5c3C9jP%m(&&L8+V(oWlAhAlBdET$Z{jocq}D*xnHVa59zniL z0Y&ucFJtaSSJ;|KkuZe{DwiQ&8uodMfxsfTLng+e0t7$`9@lz2`FjfSE;U&vuZ`Ylo=j~kxK zpmXbhS;2sMw|B2wbW3p!O>Upk16Kc+CiDNJMZcmws+&DEAtA}A>raxG57ooD;&K(A zzmM8Sl8j;Gkxv&N>{K9SK^!8{Db`Kcu!xD{#N@0R`4KjPvQOo|&pXGAAf0nlkG`uj z;|t82Vc*WT7^>(&KcClqIMNEKPrVh*-6Kjk{XODO47jnF*2to96X-c9(XPKB37itJ zO(=zUc0KbST1b0N^HSFwCpK4*S9r(GDld>&Lau%L0bz9`>IG%F=K%_?YL^ zKCMFf1@-;(P=}IZ3}Q+Jm)>5QZex42MN29ruy9kUy1%}B`r=Am@p}&3+h#)q(COft z@Ci@XcdohM6_D`^du2u9%QlneuCxwYH}+6p=A0783SJI{lD>cTYVc&3CvBRUOK7Q? z;n8YiT^j-eho++b-Q*t=`X+V$5vao6@itU6>wymnYO7D`mSS|yKZMm@oVl(5`;TcD znKdV)6KW%onFHa^POatooM#3zREv&%f52N09SDY4^nvQd8^w_-##E?75shkw`}CHV z*M|s6HD!@BJ@M#T;S*a0C=JO5Wpd!?wvZYk>{*Zv9*T)j$91YifH^=QQwT-DTA#P0 z^}$Qnxwvd%r5~$%DE9Jckdqu9mY6fTdOrk}#F1$SvYpG`T!<4_FmV`v|6_f1wZ6!4 zL9>aImJ+oHDVq7K7TWAq5>^Ap*ppfk&<>bdh<|9UJ$Mv9qkvG5=09Omtpk$00W`q- z$F|0<99ZwHi%waEU|4~*(}4*^X?ax9fJup^zGIv(w2M>lWnM1-R5?zi@4XY!d`D%nrq;o zZDYs^F$ntd{*))xH6Y^-5WKXv@k!OrBYHb+>EL@Q>kxu}@UB%$57umaioc9QW5pyh z7;{s|Ml=DJKKQ)drd%=KiU--|J)_(oo+59KWBLLZgEBC z8wnH@lX9s&-nc%MP{T5~eju1xOAm`2k^9Cl>XK;*#E4}DfmK$VOQzTI4pKQ&uk~AqNe^`n4UIb*r<@JbK-3SzbhL+27643i8uhi8AMVE@u>7w zkn6xt!)W-Ad~mqi2=70Zdx}nC8|=#u*wSN9ap@#c^SzM!i0BA{#@?Mb7vE+t+`GDANjn%kl2mHo z9}`AQCq(ztPgp@3&Go@EZz7YhPkU+lKjGCYj!;jfDKcs~lPlD2eGx@I4~e`hdCxVx zCr2WTZBCa#;_$UR;AG&T3uB!E?TN@HR>60*3f$4sz19$AU9p)kv}h3RC&HVzNI|eq zt_%tVI(nU=s`)U+C;RHW0Pd{6YJzp(*Zw9q8?Z~tk?qo%OiLibml)4>d*fvmU~|Hw z8l~e);rh3Tp`kq5MuDeI3IG{eHZP8|^hVvf@>A!}OEOsIiYEuAF4=0h3+eG9W?-9e zI|GfqM!rIdp~0cSxY`&b$VFr})B)3bAwxgF<0KoGt)jG`G3`=lJ#}50GsI+g6MyVD zpO2vOM1lL22I$tJdp^r$-i}PB*B#@HV^)eN0B6$optp_QrV;nZXW7 zKq=f7Fx%#0$nx;AhL(om8~y8wN==}-k>#`yYf(c?(yIk?1}ZX?o&N8e3IxaA836tI zVLL$g&x@CL!s_N%1#!I&clG#_X%Tw_l^0Jtdf-DjM9NDkioSZ?#P1!C);z+_1zrGx zAt=`zvIxiFG*%ao)L&bw;?y8!hHNXk~DXAd}d?9n?8WUc1?+U9m~bkhyj!yRD|pw&=(V8)l$$NrysnYO17g>ge6F}YFf1TjKLynFYNJFMT}ycbgG00}g@ zb9C644v#dbb6H9;uP&S=j&e4!&*_00m%TQBOeKuIk6dA7gq@yoA~VWEr`z8RkNhTy z6hw|~DV3Q?Fa3g?#W(%7&_T_H({RWCs)29jdOx>@WoI)oD^I>lywuvKZA+7EejLzu zo;)myelKI56UailRKjtgM9%U-ms&#pf1QO~uNJd4Rj?GyFMes|;eP zPr^$^fbGclbE`_s13{iBX5)d@4lEN%>cA6}Xfz@a>3E+nm{0*r?$gR5)6cg+xcF{| z-={d0KW~sdU!oNV@B-dNCj8}fnNo!i#%nx$CO znftYg^G{73>~OhDJ71Oi8vE?*q_J`a!i5<9I8-4yjdxWL1vjPz^B#>jX1th-wsBZI zrb2$iUfUH8C*m4NSi~-2-G~G^?B5zLTh4geAPH66BXF;P0-=b{uu_0bcfJ`t6~&Ct zOsWe&+{q)w4(HBXyGpe@TmY8Qf|3f&xhi$Mp zGvMXr7$2RvDB+rp!vnZ3NHBk9n(=BBZjT$X@gJT=W>51>)`Wx(c!|1Ni(s=~?ob9$ z=B5Aq;ZqXeUI{7s*b9;bLEtT3kgwYAjH-YjYzud8V{RENTY|grbHC=fPsCk2&tdTC z(ob>_;^b$OZM|Tt;nX*J$39=}722H0s3#L6y-GZ9V$8?8WO0;a)`<77WQ+}mWq=xN z^RPVNkH`6Qc<^SC)PgE0182?1PGef%0w5LLw$qRJs86c$UcHhb`0WtW_8GCg7^p`3 zT%?5FkCr0K0=47P`UVI4gBu8jZmVqFpYtUpf_Qmq;$AIKQYYUlh+F{93S(_&jx?mfp3>{xxR}(GDf=0C-@OCpp@Ie)UXfiJ*wQ%^ z!yZ{2%yC;-Lz{RB8Xt|}EV89NIFiM%FNMQj#sAZ$I~zV{wFDTrY52o)x8p}20Lzzr z!FHOS^q3L>ioD<-!!V7G!r%9PU=XsScO`|0-Y3~0ZOh!Z>8r-A~ z5!;NnSZtudfYx0#bIpSC&mm-Z;>MTnjcnr&UR-*33RVJRbrkZyH1d9c ze%6%T+AZlQzPD#oKtWqOoSIv^oe`)P%!0U6Nk1d2(VF&-v5c|=cCf}w!99g_Bv?Nb zt3*Qa#L)L6RW@aJMZ=1`(#!MK2``U|&!U+RA*6YJTPyAX57BgCU0CQ-IP1_}IW&;< z>w(7BAOgM6yotYYq~T>DyP#&o!kf^mjHH6;h@?zTtbvS3QvNAsR3FFh$|4|Ptj-I) zI4jB-`h&=)-bU}#3I3cu@XGvd*(^UbGh98Nalx7H21>cW=vr*aTd8(t4^mWK->_=n ztis@-rGjlxp-Y)(=CSjBk($C5-BO?SL!NB%k4lvoJC3lCZX9CTeePwuX*d$F>{gY) z`oNPS074SS`wL% z$%b6*ttWKWl5t?5z?AlAl$L||g8<7SWY$n`W-)FQ`7L}^tXgdQg7q$b{HQrQeTcI1 z3gJtyr~H^5u8>+|zgqi}!6m?jscX>~9qeORUqZrrz<~xElKF;|Bh6&;l0Kv_bnbGr50KwL3&TWYrB6%1dKmy>_sep}tiY*}D{eJVO$aDz70l??b3C+t#$KCl5tgrT ztI8-W*|!no))fDPdlsJ7fY~Cm7{5@%()4_MN7!W1rE9!G;vje@Re}?^fSGNt$!_y( z(s-cL(KDh+T`V&$?NH4?uKM=gvMt+Vu%mqFdx4STW^rz=AsCqJkbd~}!{!+y9{VS@zClic&Cxj+qLp%~j%mWygfu31|}M_a|QN-bZkCxt)ldX4!< zBp<%AJqHJ>&!5*hQyDnx0?Dy8n4e!CgvU0jkX?4&?`decpO_Mvv!;20aaMo836xLv zy-7ow`9}A}4oEq4s>z8qcuZqj4sspv5ej81YltWP9Es`WDnz2d1Y<;MTdM4h5^z}ihB*x}%A3s|))fqf2Z5AR?cf24O zlDL$=BDM9aR%=;XnbLdJ_*R#%?Y zEMkva{zqG|`nl*AIrmqszy-qy-odkZv9xV0?I0)Z;IBgG@#T%;y5P2>?S5$+Hdwe+ z#TDN;@`+aX>1$^Cqi>xBP!lR!W8vTmc{F|ETi2HneQUwBCk0nahF?Zi4|znLue9%~ z$=P?Hdhk;4wVv3^AK-T-0bP%k!E|jqU1_% z@jI2~<#~x#D2*LJNKwa;R)CmVffoEJ@{u^ zZ^v)ur(=f)ZkLQICox`0yK?oqv%rzl(zHi|E-d7s_32vZ;+2#iwkZW5X1zs?RdoK_ z@%$%wVNXh0W8Dq^c$i03=a(LTKOV22S!|Ab3e5`MT}ah)Ii`4LWyk!ow;&bgX+pj4 z*1kTeu6)krdA#LcHS_s{u@=XUI#~7vWdw=3b+~%2C%5O#oc7pMTy*C@?H)H)aoN5l ztsQ+nolaINX(8!kx?h3E<(GNGo@e(xVSk-5K723oO}%``1Cgd)_Q0!f<#AF$r+u&m zuqVqsAwdVFV^e(Y!=T$c|KINS%yLZbS+t$E%GyA3d_1i5dZ>|G?6v*N5o$=q?ux#D z-(Byqe8d3hKHr;H6y&URw zj{EIgO!Y=kvpte)|D~p{bv|6|Vt}`G0DbvTvBoo{TEFHd*i# z5V_H#H`w%0y-DHo&KH-x36Z9Kib>CDYB3O1@g*LV?!Hgb-FqBkb(>n8L$bJ7*%@=U zx~9Ofiv2rhUseu8?$K_qwSAIcP(2i~q2Sq2;^d;g$C@h#X&&iC7^7#DCxtT+_dGp5 zjvwX~H^lFxo`8&eDn4)i+nedsId3;coz|M(Cnt?QwL0$UUDY)3it2add-vN!-`%=< zt15pP&Fw-L9GrakBy1rSg z`2ExTWFf!LgPQlVSKo7R!^d}Dmjw@;5WSr3g=rzwTf8z1G%X?lm zDweX7!e+c*GpWCupVg!%io$9?V4gj__NTMWuJm`o_{O$dii+^8Ta(M~`J6K+pWTNk z|GTczLi)VuX;psivkK~1&KKX=lUQnERP?1nFrt6k<7&&22YRLAU&^FW_SJ0tumShl z!vCC_qu+i4t@#!0s!HMS(!&zt7clQ0UQTwwB#~2k-To;#(EmQ{gjJMX?=`kt#D=rV z8;du5j@_;DlxcoQ-1#vjX%euxlf9jNOWcpCEh&u_7gZF5jdgzhb3D1wOMW+F^OuFv z1V2gS`o5-`15Cq852_w~%y)jr22)FqjSJ%zji+xkjC3R!?qdTn7~Y=so{EFjiZ^F= z^qsk){os72y{G4eP!2Y~Rx|-xvI{)Co^3x98$IAlu)&8!f zmJ7DiMh74L?ysiG8-<(G&b|)~XuGBS`B+rg87E|P_u1X4#6LprGeS&Aowg*Mhie^L zv~WK1S8bTPGUo6dJ!A3WR}fUs>x_fblRJ%BK@q^R^zqKS<bmbA z6{PIxrQat1A!`3IF&eqn2lg<9n`;%-%mrJ#SX*HohBqVa{-`l5GwU`qy1y^C>PYmX zK7&JNYJYT783n=1uFsE8hsMqoeiiO_dV9O&ubOt_aQeC%KdX!G6n?So@iJncue(!4 zf7~w%3bpOo>|q|?0^p5}JIZ$YUR9hV_gfcV{RsEkej4lvH>~P5u%ISYvCRs8RCbx# z`tc6G89m?cJdxRVHYy&NG7J*`_LnYhTp9i2h+i>MC9j*`62C*d9x}f;?Md?Ujd%Ig zi~jBD<)mE?a15(9QDrjOd~wN))9T?7#qwlz;f3+UQNJ7iOO9*)KRNd1{y0z{bs?ja(KjS=(al4bTrEc+AiOG<;$G@8GIR>2q#rx^Z^#7tG zH}!{^T`h?wFfIxBt3^kjgAP=Hhpu*19UUpPGL?G+s=r$~d`7FQ)mxNA1C< zJ2w3xvCVS`A6m0SXv&TQzt4L7_6sjg8i9Wbw)m2qnVuU(+p0K7J$C*{F+g4mu8fx& z*JZ5fJ>b?4xVhm+w})-lcg}-D&yb-?f)v8fkA;C%b1!EiO3i@;S|x zQoDE7n*HH)$H&c&9?1}oWGE_y)=^YO41!Q|%2UL|MJsN{d>&Bb* zUH)j91>ZYSx;|p8ro;89`??Tw^IMf6X*wP;_hX`{AWF_x&jP=kK_2jEt9mQ2G=rkB zQ!RCt?tlHdQgFXu*{wM%!C&;azwm3r^FhsmZzUUhj}1jz9FbBcO3s|Hn5>T9w-33n zwbSc{skrc>;SP`JXs4sYugjOlTCJ{VzwINK8+;J;B^ww$HN8J|DfDt@BguNwqTvs+ zG-31b&Jr(=ggl|B;|DU=;6sOAZnF0!c`-F8Z~Wh*UZ2(VMxlc9fXkB7NuwXtjXYh@ zp3gZuUR8)^j9xt4jjsr~Z^D`k?#=m3yJXrF_=?g(@?z&x-$vwIN%DF$%J9AR&Nc7u zIp{@GQ`f>mNU@7?&Y`MSu(fZ3i#DXD$D&46L_Mpj0JRJqcRtvr>0Na3(3`tQU3BAf zW8)q6HVzq8>sx*=*-KLYXsB;+mjNXn(clqAFI(xqmMm%)J=6D!Zu9lN!|HK-rcpH9 z+xOk>6`|4rV(W6^-?8zC20B%9C}e%ovGS)zsLWjbUP^wGEh5tZ_YHheWsx3*^YzuvpcOe4DP9*`s{d>ImigzaeI5{1%XGs&CaP-gW_V} zO)f^dT}5|}0&TrcG;*q+Kil~Le6+wI2T-Bwtxfvv={ZvMx(?huoRu&PS*W^r)r;Vf z;8p1IU9vZS!M^<;!4Fd_jj+}C_p#R}LF%fA{Of1iANp&bZ)97U zdPMBbG98I`ojgkIWH!y8aQ&QKSWq{v?`0iVv&R%s--Gdfl)u)&+0w;s_1oZfU0G)y zd}?A3to97M`1LiljUwu+vj3IS=;36EXDd(15u+XWpT&voY-m4dBy4npGf?$w){ zq!m8;LjU>r!FSV>BtzZe!eG}QM90vH5`9%6Ix_Xo(fAR++?Sk)qFdI#P!sn)04Esp0ge35x}PkrJAXT-^a8Fs zCr9hnWsh60|Am7O9o*5t-Ih8U<%9S?impAL>F@u~-R3$Q=6*MG%aB~6nVGq*F}F}@ z?h7H8(#=?G2y-Vi!!S{bTuQm-F4sbmdx%drUFh=t`ThI;^PJb?@qV4>l_pNYD`$_O zByK?(g=#+X#*7>CB-{|2A<7XC(n7gKdneQS&EviHSDU=%?3m|8`i?T69C4S+(4NmAY+G6It}9zoPWt{?R6yqBZR}Dh|Oi5~mvs`jOi${)j=~wayJWlb3f1vi{(Q zDnZ7rvaf!6n3J@)r$Z7Bb4}R&uAoHTRd4(xM#=g0m$bjNdh)Z6Lu4gPnJ?PbKOOuj z>+Mw`hW$I+Hx+uK&Nu(_6a@f?;&bu|VaR}A3xKY*2A}9(KSM~+H;38P=B)z0OFKtP zWPY`>u9Af0X%D`Y+R;{X6EtG3rpTq%2;>zY!sq06IfXuTIIjvWEEqw zS*|RZ)S%IX)FIX|NY*NHO*E;Z$7?IAMCqUaXQ5IV=l3r1mN%z({c|AD%Y7!2saX9u zxMFgQc_6;PsdayvDKq(4XSr+K=2?iO)(PH?!KLOUx`IO^et(&#T`-L+(TT|AQtXBe zxM%PojgdS)R@oSR{Iwn3ArY&d$PY!l8m8cK_d+~;RpsGK7OW%=JVvZ zZ`(D4tmfu3{8B&47ys7NiZZoQzsCroP98;uTt%TZKbX9iSm^;|nl7NXqAZfB@~N9D za{ex+gAGB3xzivOi)Sa=Q$jw--CEmUni-AD-2S=u19A{WU1CE8{>P|u-q1ig|3o9vOzENgVL`67w(75p=~Fh6 z_3B&H-O}Y|>QE0SY*Gh@Pe7R}f6nhMw~q1(n(HUr4Zh-g;&$Am>w3WN38?x;+G=V8 zxf%8e<&Z{eGmqAh-E{}aGyh!s21zplDePH9)1l;G2 zs$fNyps%#L^@3 z_ZZ187)o3~c376bnR7PHj^7X033gl_#5Q-;sIcbuoGgwJb0p#gc!51)-|EbmZp(nR zb{?Y#pgc9Tmw@0=P1U=e{5&&7NiybsgjXyMy(8*K?FE6)Jrsy?wjgBh*j8#mD0Ek; zOGQOxzci)$9zB4lPTrZzn1}gp3}(Nyo_u2<>fN38%P^xlst={bg>xIcMp(%K^NOf) ztZU;|i)IZ0wYzR!5)Zwd5_(iR7z~9#E0qh=RUCf23#d}!wsvNzB)O~uanBYcQXuJO z&MzzpSN%OfXB^3%98(QTTtr%?SLPd?ZP&60mcLxyol5d7|kM|u`qQuDs2V7%?}a!n9i z9Va$ZY_V@>o}BzxUC^8rtxd@!r$+}nrP^8oacjCno=LAH5qYZeg{tyNIk0@gSmYD0 z$&7$&>26t&Y<`z7V-fY2++c7^Rl8F9{PO*15`vDDZiBp6wj7d{3<*()>df}`XgYz^%m>;8{?Sh z>zzTTZ=PK`w;<8{B?EVdW_N6)m2Z9ecli;kk{fL{7~FqjMv5)quxp{=EX+5ta2z_65#W^my@RRpljUz{qUs;5 zzOG$rLy*(1C!aT5r7K*po6AexL-an=HdjL2EKlXjbx6Hr<79pQ90hgyt4q{REUel(Yk!cWlTXQ`YK3oFjR~0w#ol5fC*mAUwmq^4&kse zDk~@|`k!j^OAIzPoCdtUUFiBfM*MQXc&TKT9e3zji@4RzG|nnQ~R)xh+1vnMm1vh=Qmw&`$Rwk!t-?uYaYwg)Jza*C7||9hqa>{|wch zyQ@fEwsJcCnRU(g&G%D?ThXTjDJWEQ4}Q> z{)C3l^A6vl8yJ4J4FOAB_YL=M07t0*a8~=STJ9>5Jy(?CwNI*ju2(y_N3Nmrgn8^e zz8V3T<36FP4p+7JcgUM$d@tOdE1S$%k>P3PvxwwI#{1D1h#4THRbxxN^~sYTF#+Mv?|cF=vBxk-7}wKzUa%lOVwV7zyOKDbF!)=nWm4JM=#T*<)i?yn#X-4 z#h5)bG3VJf9-H&tpSz2A{|B%`Ez;iYS3Gkks%yM%3n{r}&JiVh)Xp3&bFx;JqBk7L zV@c!m9|)}JHti5154zp~bE`39^So}dW;{nP(#`@gWs z%wethW<}RmDtq>JxsPJe8vptJ8M;Y2`5yVpVAtkb@M3D)?cY-u3r(p8s#YJOfz_(0 zE{TdD$-KG_q{7qO^YdRd_OLmfOQkR9QRGr;^*gd)hd5tXw}m@smhgT5l3Q*pr~l<# zp?igbp&U;gdRHj5b}aIx&Be*ca=mm=5S=ou`$S(`e!$K<=DN!%^v$%#%7NfJE(UoG z_exdOPJ~U1!xiS(*;8>bo*#VB0Ya-8_5cnQWo|!aLC&K2)n(cmC(eY{u6)m1SKGwj zc)pA2%=2@!--_3(3~`gx!I12ady_PzN7SwR*rnw9G(Rt42+q_3{9wG+T}rel1zRua zz{X4IXxCW6_Fd_HwCmYg?M@M7yV`&_&)ffv(}e1)Bey>-*9xCI!-cd>)CU^tMLYg z|8>mD_H7z$1qcpdZMeY#=Bki@Vhb19qhNhb`^L~3q)Op_iTk#gO^X$>zlX6<>+b)$ zDR^-Cc6a8k6ic48LWEkYs$cMkF{KD3Kw6Z2O|it*y6sw0J@;z93_QPM1vCA`SQyb) z&XtOSXd=z(1kXmQ;y!-9pAg5n?JEfysKnVs-w$ak!xgR>!ac&khTzE`w_yY+|msbb0Nlp4W&jWyP&y4fudpem`2pWH-}cg8^u17N8Ag z|8&H0r?ss8!guMr0ZpG!8+$QD5U-V)gyYx$ie=Y1%NJ1W(5{}FLZKAEovILg>)$`fY6eXvVdLb8n}r=AnzSBFnI;OBE=?zZ80%$c!;i!Z(wyU4f`f$` z;*bX8rHqD54X@cqAN2^lA3C&eS^$4|?mo32)%^w4+Q4NP^E2(2D^iUY!@-g3&|=+o zXY#tFbgp>o$>#JkvW$q4(9a=IW z8XZ}|*CX4s)h}^zCb5gp?TgAZbd4=7MxOurVfGIq=4!$J020ULrf0zRm~|I5$%pJy zr?6}2TH2O+tYTF^nXH+qrm_B#(XOnqvpT&s$)zpPf_XF@Nfroqxpz#@&~oZQbZZKO z?A`kthmxPb7`7+dvH*(@_m5*;0WR;!tNNw4y|wKL`<{~ntmiZm@#%kvb)Fio5U8?Q z^myXb7=0m)V>}b5z3TEdhg$W@{NGNX>(T4nS#1ZS8&b;K1wUmNW!ov3Jxr9r+^Z0% z3k<_N%djsWy+RbmhXmPkB>rkitD$_;1RphDtCLJ0u} zZ0(K4{lw`vkUC@~NqphipzDoKp>BFn-}l3TJrb{YCVnSqHF0Lfp)`4yJlw%wh*9ES z#5g(SuB1Z+t3f>kKiepOt+4-B^!uc|aNKxM;)-}i**f`eNFvXkQ?*^fDNM!31BQV3 z4S5gk3WDckzyP~EIW#5$z`G_9AZZhZRui6n#N@oJciBzaFpdh6V;oECpB`P#NRfeS zQi!5iQ>6tmdTLBq-OlHuW3?cz^O#5Fgsimy$=v{F;tj2?l%>*Asch6z+xpkhvD|=WySq1PYr%znk4M_~)MI^iN4o)~**uCCU4Y^R!+*%)zRybX|xz)m#mJr)u+^ z02X!N_w_WC;k6&bivHxi-7fqp0i?OLM)NaF?P$AWl4qrCXq=(u>LC6t{B7RD#O%Dg z3)5Xi=ZkJ+JXe@BG|ayAoZcRCG%^*&-9jtx#8wJ{-VNM0rasafUn>9|BuD+8t7|}- z-Cef?7I@**)$&;osvx}u!>>T!hM|WJArrCGs@nD2MnZMB$b%4M`xJ4?vXx{TD3)N zn7m6uBY0vRj;Kzbgk=b)PS9mQaMPo`tYa2!QBdsH*)*?XUjcPenn_eH8yub!mm}u7 zeap#XDdwRrH0NP7YN_@WGWX8{8D0&pVnXL7T)x2mNw&n7Zfd@cg;*qas{GME19V)8 z_V^XvW4ZZR8897TS(KS-cO}4<`Ss9KoD(SGVwA~rQp;#=@YqB5;kg6XhYp^8W0=0w zU{zYV3b=E8VLM7C_juB>5aUCuq2%q6gZKn1brydQ+A1og;o;tm88LXA>ys#mEYAD( znhZ~3ufx>8fL4~q$b$MNdX_0NDMU+;f|=Hxj`(cBW2k!4Wzp^Spjp}~g;yD_r@ZIZ zDKPilSKYBr+=bp30}Si8dT*01g|7yLD)&VoI<%VyPM*ds(J~dJj(#iI$3D7ll5`}1 zR-(}J>G%lkSZfSr*H*v%M0W$TwgIv6ISXVm#&dE>vF+Fmv&Fo zl1+o}v+mddi4`t$z#V%~n@aH!r(#%N7m+hakXqQX!G2Bmv5FE{~v*(?B>jx^=6 zbJD)2l9qKW+HF%YKI0lz;qLXu7b^$%`cO;U?`ZtDYp2ZZf&>A^Z?l3dECm#V(cK-{ zr9!Zo*xTd;R&dqD^cly;`p_N}vVlkkD$KI-M=lBCo|mOQOTObo6`9}WO1zQQVlfMF z*N@|!TmiC#>2oc^=JC*AIg-r@(ZqI0gpV!3pon1siflPf%m&@J_7q%tLV zNHYdHgsP+mHo>w9fWDP99&+uDJJb~34KUp2e}!uEkM6E%R{JL@;{JBWm4}B=eO6P) zDYL!Pnv`_len;5w8+2<=^XImokwa%k^_Ggl) ztW_WDDNks|llIs?xjMq3h^)xa8n@bNzEmSA;Z=(9?~xTB@cR-{x?g4+cU+4{PAIO6 zsWE>As-P>LOlVi)ThEF1D9GZa^Act4{t?jvA24~SzBELJFGqRJdNe~)UEPE#*iRl; zcl@ybUO(065HO!LIL#~=Bet)mg9UPuMp$Kdfl!Dy38uoPAP0>1-0YPV(#h437ZvW_9o@5PZiPHwZFCwV&=+|RfQ#LV=mUrm_kcModLDPnV7EC^+#3!sctq>fmZ5J=7kCdH-&seg`B)To_PTx<|QvEwrX zj^3*arIM+Quh*Zxc-S&(ve5%r;zZ%i$SBQX_@}+l&56^>EGPm8yrLk|n zylnWiFC`{J5|?du49HFEyorLo0~;_X2{_fScYc+AE&|R!%EMm(4m{STVo4R^&GH(D z339A@CzOmPS&h8qPkRqp+E!ziKfn9qn}mc%W-4m;Y>k{|yxhfi zAu%pyzW7o?BlmXmqzSN7B4G0hU#(w@DnUS_D+G$!sCTI#b}TdwJTjkoFhc7}QJCxa zRp|9w1s~Opa5#Pp*LpSlRi$KlZJ*mv;^CsUaI8W4i@N*aA4L^@C@%_U;D>XL0Aq_XiuWbLQP=|-@(7UqGoHE{wf-hMp zt+%6m+D0?Sv;!;jM>f<)#u97y?o_@kRZ;o)aqC8>>{c*OKZK@3)j2u1hxP{k2%M7f zWPa*Cfob>`AQYRaLBg*m8UK4@t)|`J4-~vt0qOMX_PfWrHe{9DFj&gqF+r8g76?Pa zlr$wdo{k*>P|~ZW>mE0Da$+6i)BHWU*jCjX{-;(&}%3~w*E#RZ&`Iok6^t^K;Wt+8E$YXH1>hUj3`lB_RtLq|Xk9SzBH4RsM9d4kqN z#WP3~cpILiu-qaUg}TAuL9N~k> z?=wC%3Jc*T6liU0dcX^B^AuK8EpLg_T;*QjmMUXnLJGrfsKm1TDcvx6vF$A;d&at7 z=ykdylA$;?gFr51_9vM20ke*7VzBXf>ej{wc}&-DUmmY8Fw3ZuY8@VlRcj8}cIPut%B zdvTrrQRGTJGOzXnF-POrV3t`R$Vi($iT>!yQ>AcQJQ|oL0i$XnP4CJuGF4iTz3M|xv{y>L zUG?9bGtN7y!NbE|Gp^8<#*{kn8K-OVLRWJ20^R-sQ?&E30=&3aq>xlh&bSdqypO<9 zN>=-#vi$GHivt?vOs1CP!wht+%3UJM&qJEzA1#B!g*rv;Q#)dD?PmLCqP=`k4>h{x zT7rwLcjmO?IZ&Ipkji=iP97FaGVc`F4)EA@`+FYF)rkq{WoAqYh?n~3$kKOEer?sh zXSjs(r5g0-seN#bT<%g=o8#=zHCwm98kel`J0D#QN*@Y5a#44id35Ps1+{ouVx6P6 zMa#$?``IyNVAkM@xc53G=L=l2U(=2EJDV#v&-Tf~8l*a+Q;BPxpC3N3DgL6Vd>GKH6hT5#51o|JT%pG$xr_3|!|* zx!NT8>E>67V6!@(9zxeO%h|tu2w+^-;J;v-a924oY}?wsDm2?FQ4W0x>M;FqAM4z# zA|aDo*QsuEy_Z#Y=)<=2v#UEXWYfUYZp7q|(JAK1phf_px)K?bBn)<2L_wn=8Bb zLPL64JR*L5f8IgQMOFnnzIYsKT8qe5xWc=bDIppU?eRODRkw=4)JwqXyz-q{U3-*2 zW-$_2P!V4(fvt6xJ1OVye-y1%NjqoT%+iSnRmt%nb)tjA!b z(bm6RDzFMGUS(7@8Ghv^xtR)vlPcm{-F#Pn(^AHDFNLihYudhX-j^ZIstEL%zN8|* z@?^Yl<_!L3`hFqeN6pyhG4FKb6!3e)Eja%|(MC?x1OT;fis^)G8Z4H$Ms?cXc8COt zw!F&*4+d=y@bd9&Ha`^}UeV_l!ks?()ft@1dVA>9r7jGSUThvw*pe5sflhM#6t79K zOyl4*&a}$M)B59fWVw8EE}2GokCLW-E~%L)t#Rw0_s^KNr&WCyZSEJ!7K1!&9Qv-k zgWpqYCDlq`jVQb;U6@#wu;>c=qGA#Fa(&u~+V8_Ogp?yHt-p#| z*zox(q(yp`3IDdQ(n9O9Omo9(D z*a&wjx5C}J0Z`Ovn)0s}V$jNSyubxaau)MtV_n(>TifKk-L~8kt`hUeApx@Xqexg# zYo)4!fvN?-*<3P>#hUWC?^D6_5}qDE$10{ER-{mhY4BkD?!3Ol8YwiAeJ8#F)6@?e z6j&gh<_?+T41Ca$9#3Oj?ld~;A2H4d#KZU&CM-~(lGi~&O8wJ6$9%N zhwX%;_2sQSYNKJi>s;s}VAqc6RK#Fp^JBvmrz2ekhHT-XVW5#~DT&HS7}(b(*O>@G zRO#djd}gz~`VsQwTW5etOO?)(4p9t&#xI$(v6!`SS5q^UB`Uew-(2ZtF8OPC~o~d?@CdOYr|4jR#fJy!IS$DE26cz&w_~l}@_1dW4 z0h^t8RcYSMDlrxU_HZH$xz%jlsHw`#GH~lvY^12kt6*gtb(@Pv)({l$1d zwmUK0%_%7&ZU4wplp45ATV@-IUt(GBHl*1TjN8-^$3Zu7&9G5 zPMY%RI?_|z7d@|rYQeNv?X@W$38S!rjHgc-r5ihF9$`rCxjRuX&8&uOFmnvqZBXYt zUtsH2;{*cB_`jL#Md={q)u5BRnS*9j<#Pu2#FuWj6G}lL1w1lpriD}Ett#z4JO2aR z!XmP8%H{&hJ0q{}&Y<%w+Ky%jdn+}b`k&?epL8*NE%h%rEd9^WD=#OJ1norBp=heh z=_cMYiZ2)}QW1e`;->R7*H#CBs>FwSi0TNQsy%Cv3qR#B>pb4($cw2_Yau1%%h!mj z!}Cu_&Xfj}g$fpf!trX=V2LkrPRO^D?#yf@G61a7!c9J^=Nu(H5vdiN*+Y%d1!wPS z^e-F_d`(qFKBz;F)g}d=aoW7|yGe5*ShZpA8jUNG3lQU;Svl{#L?1%0Wh%>G`h@5j zJeh3BXWm*no4Yaf7f@bcYwFrz;qj7blKB&7crtRO(7R^Hb84{+eNE{4R>ZzJLuTDQ zL1>wtVGlC*yTU=_>`Iv2p7R@TIoJP?7<5=erxC-TY@UzC( z6rNk*P%{uMQ52p`3LO-guX$o4JS`9atYXj{L2Pmuk+yp}Q|alXR;i{{c9Xl?etgUqO%VMr$OT6ly%~tPI3Z^o;5b=KNJ!#E zF$28ek^^vhp9-wYKxD^Bl@DfO4d}`Diox}q48s3rRP(gfSr87xQo6Hj5c2Y3ka;O?)8(>lmwc(5BpXEsN#prVb4oA|7mB+U zwj35S#=9Xke{H2(1v`(+KH^{h9ps7ByBi=_o4daEdO!3=PE!Zg^q*1<3| zsQHX}HO1_PJI8j&soXZD-}yfORORS5p0(^R^q#rD`#U_{VLab6U!Xdzh0y`RO(6AU zB;rehN4;MjwuH)?qJw1yAc14t z%b;Co-Vdb@NcJWLbNJDz+?6`f?64P|n>;V!E0~|Of2uXMYmc0)9*8+ow_j>KSFQ^W z5+rw#sJy>2bt{DkSDm9o)V49L6061vvY_bVt+tD5iEifD4cF5)j0sU3AaQ=WDrT8e zTW0=%b(SmL^pIc}1tHk&)V{QSO9pc%o)~tpQXZ(SgelG{`Pn&R-_g1X)ugsB5Vy>> zN=~Wo)OBFu*zSR!FEn0}DG|IED1)P!Sh&y_dUjgFc$Azc?*_Q7%~@6b3|;j_XjGX? z7ZT%&lws52`tp0#>LbT23i^gIxj#GuAQY9zf!7%p1JDn=FEJ+8Mr|+ z%y5z<{zP-DR!a|5%o6n*3-EuTC@x~swPx?;XRU0m)R9?nl8i%Km)xnoYNP}<5ZMpV zG1Q`yKTHT0tPkHhX1nRY$Ry{wH8C&6X-L+n9I*B8bYxo91E2F_=#j3Hc^y}{c zhBkUTof2xzV~XvXR_Os0f!YU#R~{-E`YA>KF%+L!))r0G0#gX)9Q4#~hKb%FB#GyP zucZhB6X5Q!$QtqJ^QN4bDw@&TK1A+or)2jUmRrxBn%e>MnubXOF4N(HS)iV$1h!Iz z=~jOlZ%N&sHVHlXVYuK+M&_%CTcdcqkSt$MiMt!CntzrPe3h|Kj?^@ha#HrWh&U-P zzG%bi^wjQ=m@~vtKJx-A zbI|m=tnIC(yAuhj3=pv?@WH_}~F2oD%_~$Faxhqwuo9GT4lY+_- z(_)pmbZ^%wcs$Q8Ta*XW(Rzvff~R?wd^O&AP92Gzk~9s)M^mmsK54Uq=5C)U$T_?1 z7U2VZ#pFriORnL{<6Hfw80PO3tu-s(Q!-BlK-Kp`BWUBudOS5)Gj+o zMIFqNtOBhu>ymEX&By2J3j)KnDJP2(X3dpYX?5#f;V4Xpdck5DY)GnmI&VJTmD3^{ z<|b)9qCneq{e~=1zzuVCYs=3vr;@Q-Kdu!*F)Z~92=1HvZVe{I{y@=P2K8le*K*u# zp$=2pDi0z)pHE6z_pBmmBBXp#ApRhIu+1i`*Ww0sK5xC_avA`x-3IMG#hlmX#ANSp zPu}26jtp4%7zh(6QMhABj;8Ut6DkG*L$r<8#^lXiURA?Qoy4@?%l~F=NEiB z)JA=(M=%m6qdqB^kg(Sl}(z-=R0aXIbV16PTw7m^)*TN&Qp#tZzkDF>W1pmB(eViCaPq{ z?Hg@umaXh_K{YA_cNt3JE${GSdetPkI``ZW;b|;BF2+q?a-O~69bx@LZ4$G==hhYD zQOELPs}~EOwu$lEzUvf?-spMz*lt94HW1a8-0=o5I92Z`6XCtd_b?Kc|Lb-N4kM-U zzP@5y$6k<=qFe_JSXgsl(*^YyI#0}VU)O!QuY&Pn^`jVQ*)4og87AS?D(w;E+i=Jc2H?%2C;;?JshwSl{j>t zBIF;Wf$Y-@+rrr4U@8q(S4CJz$&Al>B1l6o&1 z&X9>%X9b$D!%EeA<+77xq^DW&^sq?SmBx`KvASRM?&$x{awSZ>*J@C8xy-0=Mbdh1 zo8BjstyRg2He&?g+rG+8W72&IdJ{Y4>DSA7zpF$?!u=vK_> zP#)v2B?{{izRW}7ndDD8tI|Jgxix5vv6(>5Q;0(+w~1E36-t+w?7s!%HIGH0;-90g zzFYrFG)2Y6r9~dAG{dD%PZR-RSL|6*jdLOYg8Gd6Xjq2avye#GVy4)VknM=wW@O7k zvSTNt1U_8o`Ui1+y296dBj z*go7~Xxu>_>C?Y)y;FPv*l&|*QJFJpW|4a+vZ&8CUW8xcq~o3E$%lsRG78Q~$|+X& z_;p+GxZ@5D3gXv}l;`)K=ozUC)Z%xj<*EIbe0rvw#km0A5nr982JmK&%eE6yR(BJ8m7h;g@<7vW5LD#+z|xh6exThQXCw#zD`-^jSC> z2VQAkYB4I~B-POYH9W4&fR)hL+Bbd+?xF2gl{2N;Px14@xg)%j@@i7PfTXAxf6Y8+ z!Nuoli^yYg{wj;BHhR|cZUsl{ivMI*MN-#&F}x0Gp|%_&+I(*9UQli4W8HGEO7>{1 zyf8gi`!7sc60&c23k9GKh1)N@*aNo-gt>j%5;-F;Rfl<^eGE59&az?hoAC;d?8^N^ zB`2LQZr7~-=C+@5s13DQt)&KN-;1734@$wwJFo<;7cm6*v5BFLv?*-@tV4AU(t)&` z%WuTwTt+Wo|!AtfXmw0-AWF+s-!_0 zTPb@{VP$peW~s-~Wjjl0Wh*r4d=VJ0Q`5lZb#2Qv`MRpd?oR@_FnxZ-;>A{d!(}RY zJ{wI=JNMKyGwZOG4M^9bQ{uSLolh*@ba|V}Gj~kY&YyS8AON0aP1F-7ZM_qWPpVe! z3epB~;fmQRBu=^Qd*v31F*iNb7JpVrxb+VfSEP~mxlXZ?Rm;#XY_08Xa)BcT*_G$7 z`Zb|qL(yno&n%|}?@dCUbEmMD-b={tm~dji5l+9@1zNGHfM^F2QB_V`4Kbw>M*Az| z+z)$Ga zbnq|NJ9BNCHhp`p=u;w4v6a$+Y8G7e(L4 zS+Pf$TZrhT`j=e*{kKVhFQm2aQ9TyJcTsC7qCncYeD?#Eblx z9D!}Ou7=Bc;zi3^|5501S_-H0ZxSlWCsj>9?W>s%-VsT{6+S3oT`tg|2u3P zFc0aUrs}3PNlUycKYngU?u^@?krP|2o5gm~dFrsF8_MR2!86I87gaBCDIBouFLwe- z&JLQo#C>rp?xTfNQBD3i%8}0hMp9Ww>5zA|YXRX< zfl0Fnj(9xtJcyGm+V}NXn`W!DUvu`)$bhcU1WL)Y@);59&^spiYQ&RiQ}zQ~ErJJy zfD!G+=V|NDR_3AiJnwLe#A{bZEJsF~^@o;`t{Dyf4{*mN3_)>1PbG-d7vT)^i`V2u z=YDDyEqfG4VbK-j#M!WABSFCwYeUu)Jx7+vHf!+OC#$fGoVkxvFF^XNH>KX!8; zV5L@NC>h5x;9dXiCxOSpy=Fs=m*_Cfga-AtLqbN>m|Uw3C2@hyeH$VxAvZ(kV`MO# z(EMD1>7Ybe|0Xz%WE-VU4CUpC z2j*`?CuV*hg)y*rt^TIHaA0~p+B2_VqUJILmG3;yM!Ulk)CrtfccN&Ls`v~NjqC!| zz~H95-K~7947ONc&Ncd)DjT#j0c%<-P#7jp1CGw+(M0Rm-;+2d^jYERFvm zcC}-r`D+J}d>1K(kM)P<|2$1=tI3jlywJ+{QL)_iTr9|%cB+;h#5<#D${wtU=2`K* zAMe;HN&%M3r-fN}hgw$T7!_!R)_UUw8J^Bbb(J+k28h18*6MYljEYt+F)8TX*MQQz z1}q@k!j3oKGNFyEY6~jVM+ioFwU=t>BchL-p9Rg?_C5JA-)bJ0oCuVjOzHj_uDs)^ z2a+utIlVQP@@NYnbA$t>Rg>x`F$z|@{UhRGUlLn0atD;?f6gXOo~L6xQ;!u zfA`*arCZ2$3Btfia1*j?&jq&TCsRIOgKR~X8-gvDwMZ| zU33$8gRaEwqP46-D8G2+qM8C2$06Nb04w@zAihg=AF*l=<0NC0 zlyX>l-&nvJXON%6RW}x@e&$c5B#N8G*>Uq#hilBbOC~5dsB{TTXQY_14E~R!^KfVD zf8ThBJu~bX2@T_UGKNn4-9Igg*1XuV%q*<8sF(R8K zS6)TePU{|rVG(TwD^8yLvOJVER(SVh?qmovav!5xp!+}&4!|OyU-%(Hp?MiGo>q&% zbQ$g7+!+U!j-F`CKhB1QQ)0zq7Ku^o>!~)eusH`ka=6F#qli%J zr~UVV9RUW~>a`r&SAB%_-aXMWoq}0D60{{of|hGf=hwSW)?!P8HXXdJYbC{PN0{rwZrX&YyY8J=HoK;=)V4ojuGs3hv;1;n6hK z#^i3P9h^oCGQm6iR?w4r`6z#JzRb)g|HAEbl$h07QN`i^SXfs$k#P`qB1rv~!48MGhlE>8_UyeL>R4*grer+;jvl;8n7*pNyzqed2=M{LBwUq@WRxTOqed zx=j1^+-{L~;#_b-fTmI^Z)K%TcX{iLwU)YLoV=zCP=5tU4RufL$$uG#emB`5+-I z!b4)>1J{N;9M4mxu#Ua|ss5rtAcX%^$ga+at3D{7weU%1GP^Tnm@^apU7yX%b$3Gq zxb-C{{Q;IyxVd_;0Qrj{Vz5f3tRQMqi22DT8IF(~Xmo7|8)pii&HhY3^Ua0m9*UP0 z7I)2O0vFLJ`p>6OBaT}l#c!8iT=k|<=37EoR7du&Dp(;LRol&m?AQTdW*nm;(Q};d zDG(JSV^aC58ztlX_z8;*^>5(s&iDHlm|CcRl@}3UsqYBt8U&XtLb#7n1Ddz2z)>T~ z`MkJWlP=4kKTBWGes6e-VUcEu;TL;r7f>$)gbbZWGK`@=gisX^QeVU;Zvv> zQY1jA(GUAc;LbQ|wFfGi0&*bB&Rz@64w#b5@L*AK9Jp(O*SJSK$LXnG-6EgEF&+$2 zX?^>Zwg|9xDAzM9*2Ffl+dg1OC@O@k0PB01H5LJ7a+ERR7#P6rS(<(MA0P#W6O}G@ zc8k714({h5k?{bXnmhNN0Aa!THea+NU2Qhg9ux^nW&FxfSNJmP;@##>tcCoPEWlW^ zKO}(CSb_wa>?*LuAr2pNzrg(y7Qvwon`0Pu-La!nFg0FQOn_xnY-mweRu?9@*Jfr#aE-rqLj0Y2aYR7S+v7zK}lbss_0VNk3KZ?p?P7gi(+)7gv6Ul9 zcCRS+kYf&aHP7xkOr7tY0}D_5wxHPIgAI63DRlZBQ*f5E+UnQ(e%O+*+XJaxJ zA&1(B^;aSYcd?PV+%~D^A#=n)nsZqe!1v*)BNm*FJaqgf>zzG(#4rf7ihni`v)~(R8iB zV9aW4fd1O+K>?Faz}TIWA@ll3NX$H693pYSKbv^*^_->F5>%W747MwBj+zyfz|uhE z{y~u(PvvXYzfr4uW&OweKkN7|e7NuZjaQD9+d-?Ba%zrA_ve0{SgFk9VOv*uEl)+U z7_~w>19g1jMCLzd>N2QeL~F&W4Z&1+Jp&@{Ugh-4?U8Jxws6H}b<59dEoY+Prau0d zs2!__l6|3FRbDmu1X7iCR)(2V?2v0R5SiQG0htwHDqNH?YtWTnYSF~dT}+15QSO}~ zw$A=CRkS0@1;39`194{!=%7RqBvaL;_(c*fmbyH#$tw+!2j9RL-1DN7FCj2!m(jC{ zx&ig^C%emefe$Nf-T|>X3M+!>rk?L*cf%s(nU%bJ*KI80$THac$6KXY(R4MgJLX9t>SxB=y}a7T*UD-uxggw`D|4H zFl}r?9YT4S2iw9EXDNu3GXn>PW*~CV9dfvpR;IUTBMv{I;L=s8C+T!8YL5SzCOks7 zfu_yk%O^Z-Sn5VG>X&p(!Yia;syUTBGuS2N zAuhe&v*{A+*XSo7ET0oWY#whOd0B7wuhL&gIR4`PyRq4ONsagJd?pq4+a>RP9fd9S zITKtm?CU-ETVrnr19q{}NRRZ!HoZa?PARWQh+NW&hkvx2{4Ksai$9RWX4Xc!y_k^?vVzh2$e{=a!#tf}Kc?qH}7gdL!0eDL2M|nK5H2x|$fzNBit3n|8m5 zf$#~QndI8PX*X^plT;RcDm+7UDJGEJ;DVh~OBbEPj zK?ZP=3ke)_8-Q;I))UnQT+TC^jYIf*xSTQ(VU5* z#ncX~vaMDQUysiI;w9;sq9g>z-xG9pq1*w8$_}1dsJW)sWi8=?S|p@>=QQOi3$@18 z%7QNE!jiD zD{*q)Vzf958!oOv%U-)T2rIvI5Czv<^w%WLq=~AhX2mo6bit+qKi6t{!Co&w)!0dB zdJuVrbK38A2s%#Ia_K5k0Hu4?foY>KHXI*u-^#@|9@?oDtjfbpGrwVVjR``(;n=ug zPWI|jU2ofvdKteRbWXeUW|_$~Y~-=$b_#?g=@{d4I>V-=c!r&2;k~080HRn6+ID#0 z0C{l5nc*Xxd%(+2cpVOp6gkF-StC7=%Q;-hf%hKuA4@K&nmCi>PsOB>UZH6!lGpM>f;4%0Wmu(8ive zF1IHHrk8QM$_?IYXAzTYGb_6YuZL_ueaJ;YIOy^whm^Cf?G(zq@DJ9E!mAWU`2bE( zpEdFs_TQc#x-18 z96A>$H}W%s>|$6os7uG6E2;XJ>|5Z`EKI3H%WF1szRI_@UW|vk+;=DyzDA%G$dVE3 z@^#i6g_-UB#srdR1bMjz;+I9z-!~8(XVw zzFQ2>EaK1GCI^X$Rpog#w8HUzvM9fdkvT8 zTWnx?eln!uAhxZM@xBSE+{R#j?H$dQigYlBD(diq-SM z*@fnEA}e|4&X=F@0C#(yQb|OaN(OUNbmMdK;0(Sil>n?St5kKY=2J`xA>v|I$s8u2N z%Lf9>`W0`zZ63`~@D}C7I%cPXSw-!zIkK~tG%0l}=OJH2>MiqoS^A%ZxNtjnzOwo8 ztVz?vbczk3gRxp-4_2)MI0N@=QOt2*ZxGeBGG+F2C2BF^HC&k2?+Sq&IaG>|8}L1F zVM;~p4L#mWyAl;PKb7BH@0U8_TF>7*@Y|r7m>S zHIpOMWJqv#txi|+UXZ4`-GVc=TiI=-Q}#!S_3;^voW4<%^*7prZp9^0w?78S*iD%O zj{cithn<#p2@qNAlcHH_4FXz#$L<%jJmr|sKwd%>SZ?y-L!8i|v63?e>h{f;4~lIt z30Y7Fq5%^^&4^)tObw8$(PbxqDaZ@~@jn1cPNHG;b&i5zIcR@a^=H9~ zFrU(7Ky+5x1Ui|GY~$ygj7TAawsQ1E)yJJGy3Zu303k8CPl;QP*}*Iy6ORtDSdX0r z5iC>RSpdMM0LX7x@l*9YxvChi0PMO|#|HW>Z@xIhD6`~cVblb|QcC%6 z1BfNAFiOcb)hpo*L3wg7}c)S9~260=0N`kEgX73w`qRX!ZdwU%|D?*>ECWFxcX{U-i`JcHB97(OoTFl=^r2;L53NwRrBW z{)>+9ocr%&VDBwnV%(L1YZsci7rQX^y@;09qH17B@iXit+q?mLJk&g|PUl-f^a_js zN`)?2luslZx|<^ki<)2JV?P2J4NRxuLGJx6-M@ZLtB>GUNLzZ|I{Y~tjHZsmpbzaN zp9QoIzq4{}u~Qdr&rpH|Y!4oVcFKSg1A|={mw;`4HOF(D^#)a#nUbTDWEY*2Lz^6K zb~tAhKj`q3S^pE_L{@L)k;LxyulQ*JdpL33IbO&dv8u{Td6bh1_es4R3+Vz;5y;ru zVYhzCe5dF(w4C!|r<^Ds#DO+H+#2c4>q%k;l9R5OfDEnN zm=?hU{BAdND^+x^+nK4=tq3wDHA>P`-nVCI}} zqj!XIRqM>97X%x)HqS;$s4ve}FDVL__-oljm&8w6TYQjEsJ#3d?$@0hJ6+d^G}iM1 z@fplb9-g5@?zhR+7qlLl+Xu=H<43jl(F?&3jm}D<)_3I^hBr{7yey|pmdt934%Bv0 zHAVDd<6}RyQzJ>nQZ+4KcLnYF!bJE%5`}Z%H=KB$ws?mnjqFX)ps>Td%EG2vE9KP} z@N!Hla9yTC9hO3bU|pmo=xlhxiHcYjtzoN8eg#keHgD7teTlDxP1s@r%~`_}z-@&D zHB$i{dNHH^&i6CdBtwL%m{#pyd4vN+Lp5zryrc5?wg8HiaB0iB6>7itgBN-U&_fT) zA)&J~M@`_f=b9@F%)m;SSK>(T>!ZE3*JkCnvWV}_iET#vIi2s2=~Ice!F>jHd~K?n z^fE-WI6p|2;j$IaJx&@<=fL`5??BClqhTLpE_IqvhT^M|SBlR}xliL&_F*(*FTy-(7cQ+PQ95efxl4 z`Z$_?u{&uX=WSR~{@vOzt@{z}=vff|s=Xfc)ppXALOE0@^ zS1Lh2H?2{~abkWfw03WqNxLGMUCz6F-wIX>)qAWHn19O!Jpq%D3NGSB<%^xO$NX=t zpHlreXyt6=TLlLCCimY8s-#4qc6qRj8jVcfBAtWkLaz-u2FL2ZDzLEjbuRO>5Rbs}S4l81g+e-Vi8)ms zAcofi1+J4{un>)le^ZoTGxI_y ze}VUwQOWeumtfBnlKW135y`{fG2b+APbvIj!cE=MiN}Q|n>RH?q)n<(FP655Vv&fp zk@l&pQTXWroD`WD$MjtC!JRqs8(}Da_pds-ALe*Er1R(nGei6KAfYJk>D10Y{)0r= zj*E}T*BFm5tdl`P?Or&y>7bMz{j1wRNAm@xqKgbvD{rKjKiJ|=@&kRgih`n=G@-^! zZ|#~i74t{hzpgmmuHAF&nVj+}KDun-TKYWxq?x{Vm(M=qq^i)AOS0Ti(n~mYtDF7r zhcuCmIB{7e=zHr*zEqvQ(H;^3(p7{r&@qxfsK3(U{{e1s3q?kTJDf$hVODa~*+EZ5 zS%z;}lwLHY^5Z$@A8k<#a(kEnil5<%QkT^ND>=+t#_S+$2HCY|r%dD^MDIOn^{&^V zp0^t8+DvQc57*UgFWqT??yi_5f-r+C`{n(dbBR_Y`<~4zE`7KL{%G3U1@4Stjs3&9 z@6^|<2Vgs=s&`0S!I+voA#FHYb}&IH)11#66uT*u=L0uzYsT35@G&ZS#Q~X!08p=j zW#TUQ)oeV@jGfwc4ES0h%}bfgYGDPdf~kmHcP69r=y1IqMr|OhFyapRj6LE4L5c#uVWtnC<6CQ?ynGC4iSht!K1UznnFs% z$ZFdjRZ5vK6^vLelc>2!A=y_k!4w+zjsAq#*_IytGU;dVzzW%x8Nw2iwTVZqmIk<2 zPFw*x+k2OMyx`w#U=A+QzuVPebk^T9n6ueJs6R(Z$JEN$!cN^IP$vAib%a;?NACQM zcn7Wd+f{x3 zFJ_pmq(9=#)lPBY=^Zj!I`u{Ti5YoGpEzsaHl5Rp$7cJ~#km<6n$ zV*cnyMGA2zj+gw9G;WCPm=sxEuFIAqZ4^Ifw@jWs!7b;{kuE~Kc^gv>Bu!?de8S!c z)~MaEa43D{(H!|+z;^N+I3rc`c02DpobRfi;kaR&cZWxTc#3xH>+DBA8x#A*fpA|Y zzqr|>gqBLwjnP{y^u%(rPpUyw5% z(8L)hikhejzf+;wP~l?;-M`<9Dk+k#6bQ28rmEZcnw?PwU49K4|aALU@(C(QP z8$$ZoXp=h1T)c#e9U=S#Q~8BwsPBcBG_Yu%QJ&mnVHgpN)RYg-+W+bt#e{lc3klJ> zX)%Mpd9K0@L_|eP6Rn+rK-VSLgc5Ec5yd2mpg1<*66pco2Yyq?oFFw3RvEIZIzX0? zHOdYQQXd-q1{DKU>I!9B$5$A3#>C8IH+|xf)xPdjE3!lVJu>HMS@F3FH+3maCO&5} zdL*9^S9&aelGNl4XEV8MR|p&1VN>`);dapDTKmPl>B)>O+#famC{sVV9ciO`U%imj zkJM5u*0(+AkN4u!`}G+1R6wJu_8-9ZmS>XP4IDp7p-gjiGWfK5UevWk{=Wa<9D=FA z9KZll{(dTOsQd;sIr_V!&e_l0^ft!Janhr-Nw8= zer}Yy2SiO7Z{Swvqxd;}ztM%2UnI0^Xjw2p@W9Yi{uEI`-346t*8T|_*20QCp= zUU&toF~X)#9Y61}w20YRVM*7iY3ISUCd-)sbLSUKA@`PXyxCOqSD4RUpXf$Qzp=4x zoGcWvjZru2o~Sf*QN ziAnGTZpIK`Tv_q(#Rbske12?FqZj3s%3mWkGH|v4`KBz(U1*3+F5p%As%)x0|}L z@^!?y(lw7zjHh^H_LFywmm1Pj*j&uL?V}GCw=x8tZijo!^9%d~kWi_47L@m+iN@Ee zX~2s7I&(Cff4Z@9vO_{iK@5)=)|3eip&y?&s(gQk@po)51ZFTIl8Wmh4uYP3x$Jo} z*DoY8oBy-$Ko3eAAa*m6+c~Hyyx2Vv^eyM^ zsI{?sN(4cQAB!0%XXBM}tN@b;Hd057eofKz9Q8sxQ_`w^^vBdaf9H{aeZuwlcT5UJ z3>Kmm9lS2U-R|P&D(C#WF)k6U3AjnHZG4eG!CZPMBhFYkd+d8@L**MoKpo0z(cdfx z+F!Q&nLIrs?u<#Vx96IrvB~TRl>xJTbUzhIv%Hf^)I^1`it@es;%7Z!HnFJBM6jCo zcKDV85^NTql|2NuU%QFA?`wXBEj3UR^*C2jNeGsIpVn9)rV7L1;`a)egYp%CO^?8q#`+*yszbA52#;6m_tgIa{(++g;WM8>F32~}kKDagdFq$dZjOxkvjO?rDg9Gkeq z7=cbbZaN{$db?qYHfO_T%=jK^7mQX}J776Tg_a7r<*#o&OJTfAxvx zWzy2=ba=TORTpD-NB*bPUa_xS=v%FNTJ>c z{(KisyrKfvje|9Nk>)J%6HvjLb*VcvXy(|dGtS=@5R!`P$DU9%oQN+EolI5N&l?TmD?^#jg%RVFj1-xZtZ;TWm0ZQnm9Y$#-x0W#OI~Znkk_fJEtF7RCuw)WlL_{8bf>!1V(HHyO}i-&I)i)L zff8FZvM`gDOvn7TEE#7=0}AGKVoF5|lM_3sOa#~bf%rbp+e35o`7H_cU(b$;eg5#e zr5fSPEb&551^D1bl^<+F+(lBOBljBLh#!3XY z%YgNsn;*2U~xIEUpVgt9FEW{>VKRV&zd0wH&O zL?F9Z#g{g8SJ|KYaqWPz zECn$Az0m+(JcEA-U^_mx`dgjo#iToUEUR^FJGistXWkkPvKDj6J77T`%j zlUZtAfe`1)kpFV3Xt-YIPiBVA(k`I;y~4z2JuSM=DAFqj)BrC=Cv&&`aLQY!iJ)q4 zaZ8wGayJu=IWjW6(6 zUa-w_kf?b{6l*mlp%G5kBCgiFy05kjcK#+G5Gfg=UXy@f;}WiscJvl%FznpvPA=-? zBJ*GpxOS@b#@}$!9G{JOfqRt0CZ!vy6@bky^o@*}692ldHi~RPqXmmfWi-Y_CQc*; z65&A+w3BRQAsQ;-9wcao;T-Cv6@}Wf+#ckPcm#V*Ftd(?$8|_?ER2SzoqU;O=3uN> zDkYgxKVecNQsxyc{-6frYLr5MT@Y2U+G)=>Na%4u(ys4;y$1t-fq{U!gdNVGf^R{a z4w9M2icKXns)_6-Gi0yEuz$!SL*V&od%0gifIPidG2=5Bk~Oqo;f)1rh80xmw(G?z zDAyjH{|Bh64&Si}gp44|NBvF{cl4fme%xKjNW6G-BK3^soA@=&i3(oI&vjO9J)reS zdbF%j(%4Kidj0`k`(Fx;-tI)&OeEH_xO+>y=OCRTc1O`nq~wvKRs%4z?LyJm)00A{ zuc%f*8AERqgWS%*v6{ea{<)@jZk#jX0bXsz*6RC}E3_2Z5`kBn{PaL9t6XDs9mLF#Hs+%W2s@sCR{cAEfNn$&}f zYlPwv+Ng~+JZxx)%L*wn9;_IVuMzjaKbHTvoU}fuu+MhAtFuyuiTjO6;pl%06uN%8 zql9X=tQ;^w{$)ine!s5%BDsBbN3!IC?2e=H_@X3#Epl|`)?m9Ew9lD?oNuHD?&eK=`nokelRj;is#_YAxvJ&XNXklHD?oIm#W{C|_B z_yRC@t2E!9?~iMDv(p+|=XG~~IUcCz3Dy%L8N%hNp)1n8l)u&Z)c4s=+MuH8BWY0u z{1tY^GUPjjddd5F7x?s2}H1 zN$CiA1v1~$kDlIN6h`f{Z$3})MXndFpZs-sQ=P>6Jn!VaiYO0=E+*4??vKmz>oWq+ zB7ICSAgF|D^T8RYqTuA}%5IL4qD15uo2E}#=bn)Bsd3_c57O!1b2GwU|FK4aq-2@i zeL`^e$?|b0bcFt0?_l)K7TzS<`?ao_&<(K%FTrgqc6^Oa59HMle`<>PCRyWLQcl`i zhG%GSVL`cvAH$eqY@8_!_ z-xD+9L4v4rZ5Dg@AgieKIf}4tYe6*L@M(ocx8=h&)@Hv>WDHzWC4C5o+&g~0uPg?XfJ}C#iO?nE9y?ge*rpgCa}ah4^nMGz}L%t>UTT z76bMUw6mwvv6Eg8IXcauU!kivf=6s-@dhRp^+(;q5vPgVQ$3thMPmw;L}d+l{lkx5 zJq%V~PS>je^VADn5b?!k>pCxYZ~M}<2gbHN!Ch_%JFI7JbUh* z5@)nMxe40;3NqpN4-i;(b;RpCHfZow2!ps!YHS!I|BR1tl_9via~GP{tcylc{QS{2 zyCw@1pkTsnZbxP9Y=8|&J~ag+Sx|3rmgX)k@Z?^FD@88@b88h`BFSO$uy!@GV0PromU57WOkZR72C`}pB zT{y;E&6VR9)Q6&6=_~bvA~SV2(hM@MrR(3)r*-;>@fI|>hK;C&41Hp`1Aqp|ItVIE z@L1;$c|%RN60+~!7M;qpTdcB3KH_3svH#3yWF@TmJx;t4;t6o$;9D>)9wv|XW9oVj zm2iRT8|fFO^72(C@SbRQH_bPH5+++E%{8)b*N;cux>*Ui%ZL0C@u6IZN6kO}ipi53 z5v2u&LzloG&f}P6m}sN#EcG-UfXQY%cfza)E0#Y3S(ED!@EX0`v`LJ#;T)PhUi(MJ z{uv4{0r<$8>ooAa0^Ui;gvHz7^JE}dq{a*WG2Vvwre zZGq+Xl`Y&)L!W`qDmK>3(?C?q*`l!ikfJdvW`z-vMMrs_i81BQd-iq-`%?R)Lkvr7 z>tf6?=F-OK285#*Mg?b7aoh?vPvE6^YHz+-}2ortv9afEVi1S zErtoz`lYDTgflPNO*o%;4Q5RFJAwf1RNl;&2Ls*OH>Mo5N0Na9l9RA^xI0mREET(b z(hVU4s9B2k=6h=h`QsU@#Rc0JGpilJEMat0$L(CttO5;Zv*ih48~1M}grb&Y5_Jse zxRu`Dab9`RCq=#EvJJunC+q7VmLw1!;vHE&9C;qNz^Kb(%A8h(<+WQlEIdlmD3!cG zpJ8Dyzt^xLV0FY$TzkV;G?W%?Th`O0NRFZ-;%A$&2oT%CysU7jWm5DcMwBJ5(t}s;ZBoz6>g$}+J>cGeRRn77OyX0VnZh?>>VBBaAO@o+x9`V zVzPE5L(2y*w-zulvv6NWC~|wkUb)bBIw1xG(hkoF?28P3h70-Z?Vs zuJ~f}4}dPa^5jxK-={zKL=v#K16S~F9QjMt5kiX>moJ39)$s4Ql=ec;Nep*AY?(mb zIFc3RSZ%Ag8GyTcE@YA8iRrlBGxjBD>WJxb#rJrE-a_-Bz!2NGB-0Cv?!i5qAR(=_ zQr6-4ErGS9e9YQA8(Zjb72kqt@=#pdl){9U{mKF*gP)+@3s&$HslcNnoqhH=eu=yB zd}|eDs8$4~Nfxs@fJM2MKT2puoi|US=myPe^T4-T@^}?!Wls=KzK9! zzLux!^IV#=SLOHUr+o__*3anWg;p5kZv8RF*sW0m7>s+Jv9_34`|obXjX#8qw<@Ps zl17rhw@MEs?XUcvmIS1L9UNZG?p27px&hvM{U!Qh)uj=t*7+&{$A~j5zcLAJ5t^Is zf4`M;)0sVqAshtZ@7tE4?dgY7Hc1zJB0rv=*a`ky5S_M8L z3y&FWb9E6EDYwBkT-$VjJY)$JgYicxobOgbEdR;IbkvCBgH~urG*h50FF2&_5XTvE z{|#sp5$z^+M|A!@DnXJkp-u;zn&oHFWphniL3kbt)F|>#HnP#JO%K^al=`UZke19u zx~O(qm9`OZ=jvPMEC;UL-;ZV}`i1vqPIl~GV*ve^3E71(qSf(PXFIt=b7DwL9-pOEqWHSo@h zG>`Ns6h|w%lEs-}dFsnr70v1YdSY44od*Dp{K#M|Hm~OgV(Nb_2D+Csd zx2?v5jf#8E+QE;a3gTc!S<64TeZy;h*`r7itXJLV6uiW~v9^~C|&Q&^v+g{Jxi!y@E^!OpZ_35O> z!}6i0`V#5a9a{3)FnU`3)?ep;0CL)UJa+wu>#eF61<*IU26@IcX}xr{qF@xmK*x;6 zA1RTHDH2n6HOl^rg?348WS%5_RjJLkF?l{1Lv;V_)$VzFl0h{q<67M2cD7^>!d&1# z=3mEa_Jq3+E!@72xRSfgT=V++&C>k)0cR6f_&>=?yyt>+K+M`q`w*mE5$DSv>#ul? z3Qi`ulKXE&ikOLKR{2uXpetG9AE$HN-|NW;p$#y$)p0|GW3|hNc2E$gbV|Dd!c$dicLVLax8RI~eBP1j-r}z}Ucrj_?D8M<(OT z9Q|(SzFzMPf!vA`lX<}S7~{5E*V=D5B7*rG=Jo>5CRAlXml?4Y8DJ5#z6@$CNj36< z%aX`zX5(N73Fj!RFtR{*s8T#0_FgGDfB2V-C{vc6h>Ds%(J08u=RMPq)%7vnse%%N zHUg&qZ(o#Q?GpGb*-Mg^Fvw1WOZ2KmO;>G9bB{cEyrYUtN3;j_S-+wvYV|d_j=P}X5LSQ*cl&(7d=btVA?Ys=7TjJFr&;&1QZRQ{k4Ix z^F9qXPP;s~&u~;n8*d}R!(c1ry03&`H-($VxoJcG9e1w;e%xJ6jNbhRuskd2Y?`_6 zD>}wc_LFw_XuV1(RXGFSa~xsSdgsSeqXhl>L}Wct9rly97s}Qa{lFmWgXzqM{#DFd z1FTd;^nl~=r7m^uy|QnCe}b+V*HHsR(odoEJFJPqzr(4tXp3>%bF+eJ~wM z5PET6`;*bzaq-NIe9K7a^4&1*qNx(2rEk9PS8%ySm9ct)HLS5;S(wg-&8PZxjNOP> zjbFg`_laJZs|1HJ$M7(xRy?QIQS25TvoCm1@Ug#^QHjOxZtI!(u9wp{H+rEN^ z+@^J2zw@)`N|sLD7YNB+#r4=?G?+(#y2F$vLbRVS>Y?|_t}W7J?^41`L%1&@NjjZ_ zZN@EBRos9n$TWr#b@_5X{r>fn-y`xvFG-9yT^1mnW%i{YXX^DvIl~7y|X(-!Uyru99OYTUpDD=N9C_ za6=760^5JGEo87xfFc^QOBshZunvKK3B2_G8N^9eMA z1cXASH0Y_wW=ziBy8e<+du^JIpFmoj)n3U);UjRq8fh+qOuWiO_n7jK-HyPFv8_O{ zvvg6o*SU0Nb^bG*#Zxb@Pt-3z=Bw)&6S&&$lZ~*d&E}M}b_0c`^5qGjJO#oiOWAOs z%YQl^pQp+&Y12{6CNt>PJUey1%m;n2~eAUrTI2bCF0|GUayR3E7yHXTDW_oaH+cMeD5}_upjfFjD7o8F>KZ;y>9Q%`!5DP-#Sywd`3k7WS44-3h!_ z;C<=Pq0gTnj(Jcm&9#7$u3aab3{G;Sx2vv^!xB7_5J6RdA2vUZWqtOlg2c271 z_y=~zE*#5|*Bze+RmdS!$l2V7*)4z`LQnboxbyUsU_7B0wwrV~4mvAHPZs%8g@jxM z6RtI>!Fc9*1j2AG>%i+-%9ZiS0A_QJkK_+tuSMlfIIPCC?*ql8AGEL2y1(~fa|Xk9 zNaG@JXPGbuR+mv!MMn5KYdo=t*x{`suPP3xvql#CInXqxEv6#ASt%D3<%|rv>tpyP zT_!6;9Ys#&a>mU?C*q&7aD-k_rZ!m#6EQwCTT&R)dTi}0Tb(8*EY4xGoW!DAf2JzF zS}(7JHw~2U{iEV4kPgt7hrp2lhpn4*O@%3(UDyF951#8kqwuyQ)=#Ca#`ametUQ$r zGQX4lQ@iIM4=G;-!?87M!?!8Xv))HnJba9R%N3u`*f{}aEV7y{#=g#%9ckK_kx;<+ zEn}>#ux_R!jky=?`PV_x{K~^rUgRkM+C?_({v(5nK49H^I-SRjMGm>dd68`Tz0mjp z^~{9OF4M=|I?I=FhVMPWF|_j`y%8H(9|xGzwtMwz+5=&RVZW5xzSN(466DpP>}@6g zQp4afIS{O3d_hsaC*NxF6X;WEuVlzpm*2z~(j-Hg|JtKCUs2pX_E~ScE`t&Q?)ppv zQ;?*apVOLR^9mTp3aap};0;M*;);*+cB(}I@!tCo;(A|)*yI{xB4816U;T09&k~lv zpu21MkI5mqhK>$hQ}k;EJg4QuLu8d9%tR~~t=qi#di35+b-cQ4@dEogo{)vxWV1$> z{BHM%U5K9i5kcq)O^Bq3(@$R9uc^Ti>bPGG?Z*`{Iv0@sS0=In42FQ1@G;=WB~Ksq zl?(UT-Ci8{8(fh#Nv0{CUth2f*^{)~6A9?>Yo~1mlWiWy-!SV^-o7SZu*GeuCE8#E zDFQL}uIU{SHiA}V#NfEf9ETVH>x-e!T4M;k8OZaqmpbZFkcNH>2`SVYK|ExPTsU+CJRd@;;XZM#I6CI#Kf)<#fSec`+08a->zg z{92P$H|8QC7EA5%b9w?Z{0J)j%E83KKLFP^9(rfYcJ$EcZSiQlMz-&1N^$pTm?ryz z*jYP=k6vfGZSLN)cw*u6j_}U0OTzEmKR|F$*LT;E{L7!#^S;JNz&M^qa)sLzUZ}27 zqf=C^5u7c#-ZU_?=|a8Zi9>6s9{pjP^tRW}WKv&s8_Hb_M%0j|XnwQk)G4gZ8Pwxy zd;7;_cWn;N&PiKD{(PHygY;W(v{kK=v`!4#OqagfpVm{X(T^*6{bi@a=4}B>vV@Jv zyyl#TFUWG5L%TpM(mTN&o$dy|inj<^4E;uV;M%KW_G(0>s$AsoTEvTIA8a&vYP6v5 zgD-Ne8cA&ZE&K=QdbZt_IPHpCVw2@W`CmWMc8exnzueqrA)m{<|WklKx}BLcC3bHZF?OWV?hjFxMQ}XNlB! z0Qh+@;nn>=0OUX$ztCysgQ8OYntOAVnPbKYA8spLXA7I+c6-SAk`~8~M194>BXr_q zb3V#zV1a-G6*x)7+2O0Yw3DtQGn#Quq|qL<^E!{vw8y9f)pA$F9C@k&+g6zXLBu1b zELN1`pomnXIfmSXv5Vk|m(|>xO#%7W1xd*pGPf3&WasvgfDI9dZ@i2%PR5TL<)>#89(UGG@IFPu^r6Y)J{pA~cxG)&XWM%In zBXi6+I02n8CAXs9fZQw5s~ff`cf-+*!W>wPuAiznRm-HlsT1`4pD}ATjEAV{m+P)u zmS=g1OnV`75N3>SOGx%f)defZ@h_8Z*64rt7k}lgh%k3U%ZZUVfC>a!F&UX~{<8S| z#j9Oq$^_q?=>u58KfZ_dkT#nM$nge7ce#s`+$6}%+cG}UbVNf0M(&5p(r84;?hdtH zBGhwyRN-n`=WuDcEXe$*dkDtR5gV~YnHs5Fsf=+u6%pk5*m?Y{_SR7L;-h>202^{w zZOrhP_T%hcwvYEQ5u7U$cx5B(#XZnqU>ICI*717^R|#@^qX$o*FwM}TN;#k-A|T8I zynf2^uUiGnM1!?BbM!7FO^QMU5PMLSVp9>Ijg<;E?3cMZovgIpwkK(;s8D@%!!0|p zT4OO>%@obEs&Q~6+#dQi1dYp)?a0zDd@3;7n#=~FjQh&CuOd7;yJy)#%pb~$Xj_P( zKGL5v%#mQ2m|E9KYM1eJAsJZnu0k<{0WA$bnKU&tK<>meR+uyV%2AUEU@JnJ3?gz8 z3MOplAVK|9APsZtT7;fkLWrBp3?qbB5Vscbm-~x`pydd#`^k)m8!-J;4c;;^cV(6v zl`-g|)hrW56{v?J1a(G-2@I5Bd@*JhKUY%LC8`-5kdAmRB;yog97ZUJmMEEGR)T4o z5doqQF+y|goH~}c&PH1b5`?sAtU#(Woyw5J0LYVBL{w zllx1{k+r1Ih`ft?YTTp9#_gNI<9JAzSKUL&vlx@%4yATd7PNk>Pm9hB3U50Z_ZPSLq~RWk0vKGN6DT?XR^ zw`Yi!@z1clzx;nI9$q+kggwUot;8qIUpFKRfC-gUuN!hDfYjL!c^a_FAfmy6+NgeD zm;ynia7I^1Bw?zNf2^e|w}k%ymuOKuEU_Lh+(qO>a}lE9h$m@5F^j7X={QMqX8%*A$Gf%L5)G9U*fImKXv--zUd zb4`HShY&{Itw#{S*=v-9PDKWfnHH(Ss|JkH~88 z#**9YNjFa^$@h$W*^Tr600+t&zm^^L9^$j*mxQ1s4&7b9zK1kxCN`~%^-r}wscj3E zB4ao(xf7T36%oxePMDldOotUpFt>J4cEHgS8`6=;I(xAoZ6vbO0Axv^iUv^EeAY)8 zz;|K9s%jNhHC0*;BeMB>%CTdyd?*5T$goZ3+Y#`B$lSjP-`#9D`p9y&Y@2z?L}<|w z+nvoyc6o1!2vb?i_mu==>U75GZOjvkk|%j|+|7+IjP3BB^!o_%_AeWjNPi-Is2}A? zYK&uU5zJzQBTEr99x^-gP~~lDw-^`mCOdNeU-Hyvcndtn5awWB)X#?UeI#9uW3wgX z-yxmY)W%*osMx#l?ZdI=2Ql|2_=Bqq%fULwI6+_(i?R|E(VGl}h-$FhvU)j4A5j2}Eu7CD(iaja1{ZAM`K z020+he=-aH*o%+dJ)pGQiGkidFm@*~i~)|2BZCgkwGb#|bo$BCr)piRaxpz&TeTXQ z$PcAVN3m_89;opAq+&qxIjL*vgg9;e^&VRACeE%23}LIaAL*-7fG3uKVLU;cxTc`P zP?-^wP7;i`B@}F0fsd3-OLv&?%|>@H2LXy`L7z3mWf>jm;VY6O5_=zAOH{zpHlc3p zOs#_MOAoTHTD2hv2@w{E1-1!d_;C|{q5Ry4sz--XGB)LIM-v14 z%iwb)ZkYY3C$}hE8;c*jj=o+@GXqfjsizYYyfTcelWsQaJDEm3<^8>qbVapG#(T5)4%lMRUP3OwsOn-&_Rp(f1yu7^He3wTa}}!i50{4E-1>}L?729 z?_84M<_sb-U{{wJ%2H~uc^}+DL>PltpZL(YZ|RW4KCilrFUV^p-aAKmN0)@m%yTrN z^KMApxC2HJ4&L$?i$FQ;7W<3Mo!Hs4Q09U_cogi}wXc_B9a+ftXQRM`xx1%G!Xkma z$e|k$4sa{VvYDO{IoWy)dePjVF)I%xe1;!8{7BDC5J24@t~p)_@+-)b+} z7zF46l5BHYf&ufzIma~d?6RP8wH!~)(8C1LAZ;gNGbN$}9pPxVQian|933%F1|%WE z+X+u{ve*`!Askkbb4kn+MaCY(NHnCd7_15MbM6OAMhwa!8u#jE0q z-eiOS0Bvql`qKDT)TU}=T>PAEMy6y$ddfDfm$hIEQ0jn2Bkv>S#Pbpd zw0JH50F7ue7VHX<4j2S2!jTyJNZ!5K9g*Ta)hWvv5>q-l1H|`@Blee>&0O;?3h0gt z%H<=V93v3&l9w@dZy3l%Lm5T?0EoBkRbTekOhLvciU;Z~Jk6NT+O~Ad>`5MN&8r~k zdxK9BdlzAG@{EC14beiXn&cxtcmj>PaTeh!$VQH(8&^%EF7Xb&4B>cn=|6c(;LJ-Y z4lmU}^H#jdwhcHhGTD1_q^-}qKC+iJ%75|6_`JNhqZjahDG68_Bbjue%Sb)8 z!!N7uBW~alwFv0~S+MRdq{zxvwO!iFf+9U=G|PJ-YM40Glx|O$7UTo8Yc3V0R^@Y7 z@cNhS+MVgpImX*a5<(TsJ!&-!DJ*n`mFt$vyVf#z|=KMs4Fo(pS+~7^h^7^He z?$_+vf)EBkO9;SdA9Y*h?#P)q)fiS%pYx`!a2Gyq-IPcHUg@~MFXFt`o6{iA) z$&K*`HRao>!Ny1zVu6uzr3nT(OC^Y|3x3tNVrKOq325vKtpSKB*|v0P7_K1YFaQM6 zs7)xZGF%V=n`u&%N0(?0)Rjw*(zM4Nh|YKgD}73N-_1_S9#3=8+SCsr#s0w{5is5|G%4oMU!3I7pZ5D8kSN_2EsB_SQP0agIw= zK!6z{2?37P3C`!03y#FQQ3;M}CQSNM29%WpC(RN8A|d2_ApvgT(mS%KT6xPFCV!1@ zhwVWba#y2t3}8-~BO9=k2--RXqF939j6qEir<;hVz+d-2o`7=+0(THcbwd5H(t&HS z8~ZmFlf!eIlaZg79>>aF=8Xl2aG&N+;oFKbIGU9ZLr;Ax|Z<9Kw%5y{#X{{WeDj_3V2Eq7#YOB$7=VV=MjlWVduIg-5Fl!SBJ?4VCz zQX{&^n_-Sc4>n$H(Tqh=hc6uv$}q)wmu>^Ax(!sO0~kg{{aR3k!|<;}!hKTg-TweC zYIbE=K$y7AyWJAI-H8L|QML;}R^;-82X1OSt&Sd7ml|Oo6Mq_B`qdI*c7n^{^Aa%` zPmF)WO?utCqiA>(*R5Q%c8)l&hs?pw5s}d^vo)#&NNolBDpej^S#VoEv|iL}n^kIJ z4$w+A<;4;hgP17F+u0|y0eF?dM21YQ-q)l<5w&#+1t`h{84^Lu>ZcTXqsy5aJjK4m z#uJa_%?Oy!n7$}R7US>D6L#2yHb77)!qveGVzNFWq$e>mAROYF?ZZO)VjT}I!6G4q zr#21sSHO^f89<-gT4i=xa)_h*+E8(nGvvP9s2O4v{pAqeI|Iv(JfVNv}Dw<|fyA;x?3H!Sf3QP{PZkJ)1LQ<*7FYT>VJI+~qsiH&e zQYX~3cSbwVg{n_-?6+EgoV3Cf0S+QbJ5-6k8OlAVNMj!sr5A{d-v^;bS1k=-;tdbI z1O|qP?0^}ROv6MKGZ^qfk$N&VuLBtz4qo(AT&1eclVhR$6 zVpWe)^2lH+f5NdQEzm5G*Ljwl_gWu?53fF#i7`7tWG!9;nY4SC2CuU;e2Jb?_kvY| zGSjXZP8>l4Q7D%OFhUxd-#ZpDo)Nw5-2Q471->v0XA&tWHo0r}6|2_k%a9`$?n)!@ zr}CEVcds8de(z$+jl3g_n$emNh-zyr$QL^VC>p)O)xZQY+uy@c;K(}x4C&vimtN40QSA?L9mFH%c``UgFxpcze$T!)V7IjV`t0N)?@-j`5}j*&9jwk-ge0fdnRbtc(@ z4dMxhXjPUBRCI@)&56xPO%mWBNM>9yGEbqltxcpIuFK%SOy*@sp(~d}A(SAg#k{7w zCnR=ai4l~baHAj&abaFEbApZAa&C31L^h;o4C?0I7|Er?vbD3}oOjA{?o zOanFu#$al06R~+Qc^DScyTqk#cK#?Jf`ObO2GfENwhVzR1YHYUpcB7v>7Bz}KEIJaJvuXDa^ z)d_K1o>Gs=LezW4yEOOJ%eN$9ZUkGk7>AJqhq2&-nT_G5=m1HKWFk)&!$iR9<{#n5 zA1)*K_rIBQ_;%Nue3^*6&qdARdvHM4(wN!BdT690(kt$ z`NW7n4Y9kuGz6kR%&&wDhGltnL9-HGQjcz5T}C#BvITV?zM^Mn+%mx7RjxzI_GGmP zlNZIPfCEGTXUwf~kkEV1>r8>U904D)fVMP5VVwT}z=tu(-h0f~EM6>S{VB_WYEsdr zZN&_&_>i2QMCOf|$lAMwshz$QV|Vd!uFqvkGXTdiwMOSJ9ai1sMP(0v*ik&emD$}I zp==T@0pgsy>roAG5=2{sMH|HEwmaa|0ojU;pgiXsaFs@%eMWXi_HyPUU06l`0AlLY zNSQ9g7l|JFwHZf?TxfdL!vf$&B16L=+~{OtL+(&5!U_elLtsyu8;B}o9Pma*C9YCn z^iR1zYDXo-O@TB6o#~Ghu_t0CK(N}FL`!l6hdvR4CPz!;j9?6>gRUac8Kwi)g{wWI zNivPw$6Uog$|jGQ7{vC6L~dJ=w|X6!57d$~k<}u2 zwyifwm3t!uE!l{Ks%U!1~ZJa7M&I z8fTdwx|sAh4mONBP0Lpd0K{!8%IB?#63ucvBp-cwQ=gk=G`D~FmL>e(${$La2p9G_y9$bc@}kk=fIxqlIAgr#;-EMo(KLIgeXAIk;E zf6bC^Uf6tt87*AQEZUAQIT*|AaiciHS&wT@A9+`mExt@uEYD7L2kfCCYE2;E7DUR> zk6jT20!|}hZ!RQmHCl`bj>A;?Oi1+rxNg92#Qo*?vN1n0t-~&~!a=qbs^1pin9(^E zL5oT!hGabtZr$GU@po;cSmK_h)NNx;vX3^tLE&b3%+3lj1$=ABR$56 z*h#h(sKki`?4u#nRVmW_8l_Mch+=3FGaxo8;!I!;Cuq&Kl+eDcBUs2oJ*o%Nwqhr-?t-f)XYeucS!I~2y*<-!z1vo4%L|+?0@~89#@z@_jcp(E)K=$4(`AGq(KZuN7Id% zQ&uwY#CGPiTcw*rMX|a){{XIpU^Md_;W8xmY_^Mk>Y_=h#XPa`4a(W^y#9U&{;I5B zX4gmixq$#*M|sC*Oi{DQN63Muq49H3RjKjP6>EsYJ%_6=dVHA8}ojXBinCF*t!9*PH!C zDMrb=aMa4-Xur7*HKa^ix8SuA2>q_wH-rS(w6zzKc6gj(D4{DZVKC$q+4tnX?j|nY z9oSR?v~pI3pzI{he)6|&^o`4aJ=wo__hSU`I)VAs3w0(1z*vw%A_4@O>Xbx?M_i2M zkbMy!R@!LIfXMNFve9zXNop}?Wf%(=0m8EsdUTU0ahYiR97P>knKgs7qtY@m2(D!Ds(?a zoY5l?yt5KxiUKD}9Ujt8&^4DBEFm+NOwVB03n%4;+VO7~YxX?uk4kM+1A&QIsVMmax?W zsr{rQXNu?=mRQnziIM^+${_K+9Bk23&t zfOPq>A9Ku8i4zTsOoRPoK#AQ^_E6OV(F_!A>6WP$hyh2J<;$0TVqvlTtzN=mcI1g{ z4w!qZUo{eWdyWZ@Za>Yl`HB+dp&Qm_9yf(gxpQCP+(JUoHDehjy9-3%VHoy)x_L?> z1mH6IsmT(AgGw!umDx?L;@`5oxJ+B++K`Kaf2|};rLIQ^L}YBy})ExI6z#v5v%Kdsxtr(F=WxW(tHr| z$Rtug{>sy&i35uxVtQWdX+9f+v85Y#2AK*>9G0j=8-k#NnkMWRcV(IA@j_j@XR|Pm zH)I_H@gOTJXWF;Mt$3BAYAUO>(DZ^w6RI0CLpSG%JVZD-M?Vlg=UVGVGtU zqbOY$B1jG2iX!O(take=98EeMIK`?I^w(7e=Ri0&4oT)wHb9eNUXF~V5kFt-qjE7L zdD@BEjal%Dqv9CoWJMs7Js?&CT7NXBlDSM$25b=p~@BaXS8*zu30&rv+W)trSN{*Tr z#3JPIjv{@zEmxr{=2#i()AbQ+3^Ng`VQ9pkQ>TsCMZ8kMMm)^naGD9*FDU!y&KIGl zY>Na4oNfYJhE9YuokX)OBP>qRv`-DY5OG6ToQUWEVrV^TQUiuIfc40J`iyP|y=-ZW zyih&6H+_UILbj*b_D|(Y(S^Js?=!LkifhZVI6={BJ{U)ZL;nCB%NMn1>bo)gTfO=J z07@UiQZ~dyq!>(rDMR{N`{?txhhXoZ@625|0oimT%v=^_An~yRArL>CH7_fUQK@41uQKF>?v1B{9{1)W zEk@*yGe5HwHDHS~QIM2Ik_&<{1sj@2od7(>{viCtq(p-F2X*FOoUII_sOp!wZFyXn zkZR8L!`O-u<=mJaE3!X`eZ&JmosIg+P@ExlGp2$-)8Un`%S{VJ7@i^s!`(t4O7+2r z0mMCy6{cX@mTnWki4CD^msCd5fFgqpiPc?D(NK<;5fR(yGZZ%p3;6d>oC8@Q~#B26dP?fW$V=SfY$rzEM^l^#(mF8QPnF=`lrK->t zd4r9!OILrV`FBv%2$o?=dUe{lOP02e6g{;owojYM^A!iZZezK5G@IQh$^e@YE&%OY z_U-=wDF|F3i4Gfs348tps#s#^#?dwJZbEX}SngO?plU2mSg(g}Qr^rJV;GI)XtHK! zKEz{4{{T|u#jBgSE0TB8Fm%SqZMmbSG&t0!6Ob7oM5QCSYmZTN!XD|R7OSP@v{tKz zawjO7MR_{t)dK~oo;y$s=Kul%B!@lSh>L*ACn8Xlpkhg=$-RXgfg5(>#v`fLnK>l@ z)kKQKmm5d2wFvP>pt6iE0|>i`3?JL*HuOg11vrLlRKYz+!V`>i(WU#Si&tE293e_J zxv56vJ67(%;q@04t6a!A#D_m=5zTVOz?BFig#N4_4A3qEo@3;T|K>5i)TW3*AV%31!` zm&k#~A>x`KMpL+U)!hqsz!DfRc{jJEF=@cm0s)B#&dmYIk0yna#_`E#hkqdp$=ZozqTfsMPN zr5vJvGStz8u*Nd+N4Y+mg3B3N0#b=AOi9teu7N1rmEa*6!n%A`)UC!K-k9Q-dIR!S zi*PcLHQ2RtavaApFtr-SC;5|F0nun-ZZv=&wuRtz#xeq6V-Ks~nrr}|2uR|uTh3PG z2U{V`^ITlahr+ete)2c39YpQPLUEV|buk^PcWxne!Jr&=iNo897o|uN&%F(NClLAc z0T4JBGoX(LpA>Wr=25xWyf?H<~PH)c_kM-9*`68(#kij0g6xsc{3 zKjvCqUCt^iP33kc=Y(c3t9V-pbPfpbpR(A#O3(c3*eel9b-0;)sOh7tI@{<3KfM zM?yky)0>wF!cdHZv_K0~U;$dCP%=Ug<)b}@MUfgrBn|;Tn8ldHVx_ zUzaOz_heol<;qA(#rR&2TNVER zabU_mA;dn^5ipRZ(^O;22!bmq$o2>3Emv-t-hLZ~qwGmY6Fy>5?m&LB5imBOXd))(Wagm@ZxC7}fcjB{Wgji+azp?^Y4cXx&Mt$PbUG( z+rYQtPujV05>6<^u&p5E#CLmnL|mk@-Y}W{v;EzJIZ;>REKgkueBr$W+543 zBij1YLr}Xp$T`pfuk|L4&RQll0w!C3h_Azwlt(b<%@OZ6fkHByNaYYXN4GR1aLPo~ z7y8_4IxT@AQ6W%AjYcO=nJPrHDa`S}hL*!0;&7vC7y+s$Y)@qaBcKWy0@8nB3osMB z4hCaK_|O+^2;8ABHi-I*RAp$u2c&}lh1>8gY|=oqSa+1P7jVuoG4MrcvjYg7gZ9vr zhv0@cVqb6o7xQi|sg2xTZ5RB$Ph|Hleo*-nmEPYFg zAg_gTR^#FA53-2hy&F}42T7D=;Z4j~xoBia)&Bqy5pH$PI2EeY$RN@s02s`MoEu4) zJ~$#{WVr1>9yt9)u2*Cva*)I9tyV6=MAA?ulOLC&YDboLg>G`PFCpTWDz#g#(U|De zFg*tJj=6VVx;>tNeM=A%3WVvwd29-}X08~P4ADLih__D^&LtSywr3-r{{RvC--qL-E(6DT^DQpo)3Ic2^3ftDhD5xSY>7lq zPsP*S{;HOaVS%(J?Mj}^&^E8eQ2~Y_!YO#edsg{@4-O zj7RJC(DUYRP9)XaQ1+Ie{C_J57W_&|UwLchC3bII^!?QG1U?Vw3zno|A%O^gaS>+h zspiA8BmC<~O4P`K?1$BEK}>(-ZI3g%dI6+;^kvKQ0E0sdfwcqO@Gm!(h?cBv#zIVg zl$vodBNq&uDC)*IOEe4!9Lxv$fOwaXMww=1K@dCAUPmeg%Y=h+@$V7cwd_Gk<>K<6 zXaL_?2!ds%fMQ$OR{3@Zm~D5Ziy3w-f0xkrNfvF5cMq~Vs9cA`g{Fm}8fk=9rlGIS zpJI%fjup#NmZy`DF^_o|c~hLm)1@>3fI9_LC6IG=A=Z5kdfsBJ|EvTbA(G#4z)4(GeQGXHch2`n;tVI49#$MIbCp#t+iK!fR=0B!{ zxp@yRSTs8zSd|}|>wH8vd$VM@GO7FA$q-7}?T{x4qXwqUq z$YtWxv|>rrQ$jky?6nWMp>Yf;QB{cv$)-Uyvc%=Jd`Q!1HCU<_KuSP9)vr*~mzhvmXD>w>5>v#VEsJz=8G8CP&pS1;|<+hI1P~NqQl-S9(FTs)Zufmh2 z4;dP9Py)3XQ0DYa0AxXQ^Pt!iY{vTyI0 zp{^u7LVS_yp!=HY~fbr})t(?z0?XC|lWHTL{V{hIvXq%vX8X zM8Z)Te7s>Rhp{_q-RYk;9KaGX2AYrXxq;UT!Q=;V*z zMm8-8OY?4GKgw3_%CaG)?*zjIs1}gTYSCDJBu8!|^KE`*wIA()yUboJ3w|6=f8Rrs z)B zLRPJVP~=PYP`e5FQJghMMz22xge~3DVg&%r(2p+$h`^zDo8~?qCrxYlR{sE+h4@+! zaNND4Q~v<>7rnG~KDn19=OTf)MX+;`RJdh$3TfN!p$pNKlO7>H;eXDyh2C!AyAJGP zX(EIG^6kcg5uy^yas~yC49klX`V<^$1b}*!iNX$I{ag#oo-AZ#_mz491&F}|Dw9H1 zxYQLwLSRW7>VS~aFx3%L^-vaW2+l4@G}h<~{j3PyoJbhtjC`q2doV^O*)cR?jIDwT zxC_lT*t_T*mr=JMKow@At{B8@qT!IP3sDTbt4Bx~_mp8dk`6Nclq0R-FIpxxEHmI8*T$@(xhr=G{ zjY{Rqf>R?fKv!XUEh54a-nD6_+(L~Gky52Q{1L(=p;DJ3udrAuCxfKD9R;} ztz7|P*PAZv$l)JdNY3t?ggkVLMnkfNsdCzGM6`==%l6Z0jjisr^B3V3%@NBJASxv& z=>)$NRlZxXrev4w`y5ExxorOcZ>i;~f+TnCtCw!xFE(gKydep4Hiffr>5QkvIeVIv zp?>Yda|r^%Ap1HMK5~T5biHsw6Op&D!KD-aRFRF?!t07&v<<6~Kg^HeO;Y0xg#c$= zhM_#|#zWaoUfhnv_Bnt~!$k3zBWQG3lj~17833B?fi1)++z-oh1G6}aKl39L*NC-v z)69qA-1@1lTpL5?gssP0VJH_`d5dPpHy-f~8-w|(H#y09*T`{ZJ%|2cl=)YLMN9`2 zB0a<}TY+%SU<(88#c2du=7l8#K|g9L9J~&@lOim{X(}9zkHF=4TzBXETd3P5iR_)~ zfB7rp;unW41DyW=JN&gPmkz1OyouVsySv3OTGg$x@!U|oa5Se7N}?Oklp1#D2@K`o-h9Z$Na?CBdXb`BA@!urDRE=cCn1IlImKG0*v5VW3eIns_Erm zC3s9l%dVSyv?hAO~JlHf@-mlJ6NHnQaxeTo3R8o05H@$>Md?ZNX7(yq?6m= zoI6mEVh4}hQiPbqVqS$GE+oj+qM8vpW$Gv%U~7;XL{uUq+K&e=J5GDv;?vA=o(8ONC$zjb1~EDNh>Ez{oQegf6@zS2Gwm=0J}eX=XqelDy;oe)oLZEm2Mg1w z3a}&uUBCv1Y7#s^;827C6Y%1nO**knrH;o^gy2QS;o1j|3Sw_48i`7k?ye97g&k?S zHQ7HreWWJ>B?Fq`hzDq*CduQ-6hz=-q_KRH0aJo6h{Xbc6)~Y&F3U{Z=(;PV&+$-= z*hf$mBMEzPM%+M&ZWFN^M~3tZNO*z-hB0vBFwd2yAczCbR`K1M{#nrFpT3v5r@A7bnI7uwTx#wQfYaT(U)mXN0^5# zqFPLF!`hM;BYxf449uN1+1iw4j+8~1#%AN`G7in!im@B_P1xNhRkzconzSj%fuP9f zZqx&#e%S=T8*zqF2Znc4Wjm0%`2ywrsL4bT!4s-P;K6DZ#7^a}mV{;OfyACxX88>& zLgy_ZGla-5AF8!Z5&-7ZPZB|2-SaasxBNSO>!$@a_86Rg{bzYAP21JQMCcXZXzS9JSn!Ol^Zo; zVDXvCB<^i!N(9D32?`bQs!dFW)Ce>c1bx&cKg5kBT9WqB8&IbjR)}$oGS)7G1KMKuOS*nHy6WNLT9dAuV24H4%^S?jg)v<~*~J zfYelDAv?OO8Axd-{WSJ%@y@WB7z6Dn-Mv4@tfX<$6fN@Bnj}84#VJd(+Co>MDPPJm z93dOjluvXK!~R+U-e~_Yhz`%zPZR3NT9~#WAT*f{{{V$VvlD?C6`}Ra8Xd?9 zNseKBD`GSoa<#4Bu$*Se!Skzo=IxsTMi zyz4=^n-seN-;d_OF)qBzCsnz2pim&O@T#YFn=qCjj|If`_)JU4ccA_F{&j_~$&Jg5 z@*Xd4+A@WZB>Yj8bd-Fk=sAeyh3jp0%(IRo*|`PfUj|lk+XN&F#NbuFa?xl?mqDP9 zhG@u8%ru9!-BJL)nkhhr2lS`QqER8i28#WH0n^9xB2_X#%M(-%Qcr-E3NeX!NU#du znQ}L7$h57DE&EfNjKOxOS}w#+62wZXQRH4@!La)?MiYsQh9HJq*aEVk`H(kAw+|G* zW-1n=3C-&f*_Yf>yBkIk9lJ!npg-m#C`p8f8V_kk++H?fCu|a8s^lF(wbecPMHGQAu>q0a5))i?k*Yp=n!0lCSqhhly$;Ue-LG+nqLfnU6Udz4|I@( zZt5Z!2v84o6K2HBMRgn*4M6dBL|A>~{$0swI2V6)aB`-&2QeRcYVyUm+aQ6LGYB6E zI+63{FKuB@w6$4;B65>eK9Wx1`^!}28*N-z{aq_nyV8`!8befm`f@~1vKe{A3%1{i zK)HzU`w624edtO679O?28dxyj-C8FCM|cpPc7#m(Lj?A(O55!sOrv@n#4ihOKxQu7 z2)KC0G15g}#=9n)l@~7%47j9e)Q+2SEx8G3(laGo7fy??W{;0oeOt8#_(Osb4HDI(-)BNsS;f4AfW+CT#%2bpa5}%Mqm!g!~il~pjO_ZyC>*a6(AxdF6~*6fFsN_ z7S4X|rK;pK+=y-9A!b8V22$ctF_xC@;G!&z5s)&Bi%K$n+;m)4XCZFwyRjZ_4l$zhd3kwPC0W3Z0(^G=04+RBXB@ziGK@%ZP#2fXLQ{ls zn{mhbpP02S7a?i}S zzB2qob;dH6P+T^~fl~oVQ^JT9w>}w1IG`m_Y=J*&-|(+}L|d>;O4RO7-JXTZBs4CG zMZH>;$PmIZAT37-hbfd|D8WGI8Ge-Du^1F}RqD^#ijgf3z&?cbenRBkrbyI)&P|-kZ^>^j9_~@R^7-- z-M}s$%w@t~z?$d`+);d0d6FZXBngoFD^0mdLoy<2h?S|y=+dHle6E-r0Uxy`=Nl1? zvRWclGWTtuBW^IZ^=OAvjhp(d;0qzfbR13ygLQ~LB36@Z^EVyAl_I#J*2iW8gb%W{ zT$r{q5gMiUlL2!1%Td!`sDb7ANLd-kj6cl0h0br9iEiu$w)PhdksxSH&kOsJMs`n^ zEWoS?@J`x|yZ-?8{KUYJ)7TEO@y=k`OqY*P`g5ZWVUk9B8Y&22w&zZ~~-!VQR zZ^iw+kIYuBLfJBRmDzTQJ~VD=)p!Psa*y*DnO#KH@F4dlr5CQEesvgSXc{>&m|Y zJ7;b&J~$z4)B(V8q<;!)&AD(!PENP2JP1PV-!y?)EmnwQ;i5kj%zjznk_O~u8`1|z z4Kp9+PqU!|K4LMMu9=NB_m-h{+=d)7TB65gRX`)Hk(A;<7?45+-m?+fF~I>FG8qO5 zl9Si7Q>H zX^_uvvnA$A1~&~fj`8*vn|4MK#6)C>w=hWUl4|LqeA!6hiUh<1xj9trWD00V3I6~P z$ME%dG_6yQ4EUn~T}RkMndb{!k>Ip3j6da}1K9=iTuRWjB1Os4X@u+qFEdk*-Dm054Yp zChghCo#l~;FB6PSLL@|qA6!wpe6_w@Ymhzv00pnMr+$5_h0j-ZDrQY`&|xcb&HM-( z28Z!2pAUmXyJ}YD5&$KE@TPMoeO20wJ2Dl7yfPi0B=?~ii%~vdEcE;}FEZSr3xYx< zhL5nQnm|ZwAsf2k%V@?&zu>b7MM&BtM%dLk643+SLOOvY0>HjLB?#TO+Zv({ER1j2 z({nRe$qgq*TGwXab6OR4T6GDKVnur^Nw!9AT}Bro(2-*yj9+pZU)!VEy9&Vm;?n~y zQLIbghno`-BU!V<>uFD--NE9BM8Z0DfKTeJ7bB=|Ly3r}6R7_HQ5&77J*;BiRYp;e zn3fuHQfrqY040QqEqux`6phM-0dzFlkr-7*2VG?52#8n}VSYS*o#?+B4r@553u0ghm%k!cRhmlZC6 z#N{#j>&VtYv`a<%rEy4}v(Qbj3Gv)V>C4y$l42OM_3w|>%$ zt=uysDxzjZyOaU2Z*5Wsh%DO%+1n^vR%aa)Jl)~6WimB~npv60erNPA2E-t2XLT+2({*VslEp!TJa z!k?FMJ)%#tdqFbuIKQ@qC|Tv~>Qc2CIg=R?l)G9He$efH_eDz_nvzUTpetNg#ff9y z39LM=k;QO+qPc$A;D-gl@2Fe3dQ!0>B1dR3O@k*;hc_ugB=F8GL}cH@f)=eASmoep z599;+k;>tv$;^CvFM?;bQ>NknL&kNq!`N$hcB6CS-Or5yjz0&5WX5fMw>+bF*X` zL%nuh3wGAKvyH!G_LD^FeTmHo zC|t2{%3?HZh(KJN!{FkDPVA+s5jc^715~uBUbx0Y(;7#+6>P?la%9CAMnpbt>xOmS zqd44&F)jg7xGm_BHPhH~k4N_S-Q~%hi4N0$AC~_Brku#rCRM5Vs-b=o#*>Ne%yy!O zFLY>>gTS85-@=|OLSY0ac0e3h6eVu4jxafo6jZIsk{}hN;uB|)UVL#Kp|mQEB;nMg zc}AVkT2LLmJ?$|-Nx%h&J5V-;t5abp^`UmrYtW<)Za@2bzw+0WbZF*bfjd#Rb_Np| z{_>)-80wrgPv_xNg+a7k#ws!+TpqF=sK1E{;E080*dYtyoYbhof|NNjYnQF)IESyi zm|u~W$i>twTZX$}@jX9PM&+pkY>GBW*t8xXfKJowAv?I@$N(4&5jiaYh@G-pBut48 zXdnK@g1ogB0Z_7+pQi1rtmc$p2S52~j-gGGqM{8LDY_ZE&; z0`1Hteqx{GTT=qD1*6ob{{WEYsmt=%56rpzD^OCqe4mI+q#ps@Mpk|#B5N695Dw6h z{5dqC^8^hmd8hclKQ z;y4(N@Te)z%z*-k+KgBOT2!MT@FTiM43UN2Tnjf369(dswG<@@JRXNO_PxX>67iKm zOrk9q>IKfFZsqcPaM_F3S^yN4kLR5%VkpouP>iI^v)-v4{@+N@7LK zfzo%yF{amAA#NmQWCWJD0~r&<)Q{C#b(p;a;DPM}Q7GV4e85^k;s`)qfRu(0h)*Fn zld@F%=n-^;)+Z@TEC^aWxQXQ!BUjl%cWy$}!YODIDHHuPWhe<;A+A8;N3x?k1eg&1 zMCLIiL~;Fj#Tbba20Ig9g-?1b;K)m0G?HUOu%}Kj8KZ9W ze8@}0J3iIBh}pJC*QkWc#*I6{P@M` z8;9FoRrs5CVUfa3;ZVJ7>V)ZXJ3vNqjAuUNu%FjYHjJWpE)UdAcI1qPBKxCGi=T`0pf{iMUXb9HZ@|@Ps_qY zj`M!fYxiO<3}YTSebuJKr4cGIpfBv97CvC1u4qs|H4Lgz^CzsInM&=z@P%YETI}hV zA8J1GHm)!tu$aaYL`;W)U`8JD^P$Hn<1_mx+j1U?F^(e=sDU&n2_K0ZCAZm3aYsam zd}_4F2@TXuCl5xJ1=sz!)yv%)phtE4R^+^q8C=vs5(1SRaYSHqMg}Hh2J1lcGBT9A z9K1ggU)21&Ys!f7Id95_f7PFTTe=4gxd|}Y-%ypv>I6I_F#z}xHh=CdWH7rC`)Js| z%g9SRr-rDQ$IrYw5S3#T_asHJNWwIL9u*JnelM99h=ojj`)m7~fb(&WV$3*4`O4hx zv}{{EHex;;WBoNF8E}$XnJC+dXU0dn5!;%#%5wZXr!7bi%%AZ&o1fHDF~&nl>RO~S zW4tcQeF>sf3#SFxjvGBq1AvT!3i8+xDistpZl!borUiUz`0j z1Unc2nG1w!(S&j*x#lTcnKv&_8H_#1BdZLDM%5V(?156N%A|VAh3L?N@4P8a*mzB* z552^@B~&ufF;TaBj!(lAT?5-f=$D%r_wf3@!h5mmeANMb7$GY%E$vZ|5WO3xJr2=6 z`UGmG44;{_#MBuED^#FkJ+UmKrh1|cM7DiYeA{rlWcGZr(}&txpnUsY6HAf(^zl7l z;tz&uQ{|-#cYD^{G>Xx1d1o-yQn|`YRBh0l9Q!en^6$e~80y7*Ir(z%OZSyU4t*L? z&`q)Y%n7#unnhmK&c!-{)j zK^ew~*BsS)Dz>T45PO8_5%y4w;A0|$q)4v}hxkwuEpsE3(a4NKAm+T8;7mCKT0NX8&S5%(=YTeES6$?)nF zKGK|Q#$3(F_&cjXCKjR?!UI71+EA2;2NuRh?5-H-5baA~O=1RByUar(LJi-VjlNWyl80*i zsQ%55<))38LPH6`#J|jsV(M;U%tWP$G_Q&}P%^W6+-2Fh2DxskT%fGP>9m6Leeyhp6X*tuSd_E+ssn?Auy_)FoFa zLNPOm#vnZk^5-rV)hLcFBif7;@|Lw-J(=D=`*zovca5MesDBdEt>iEx0<$1woX5+$JLC42j%0aP0ci6U>He*> zHE8CncV_i(eG5X_wgJNseXVguGawQe9Mr8sar0wP0qq2B%Zeiynyh@eY&alIVYDVa zlUSYI6&|&b8}a)~gf`G~T4!onhT0%uLk<}Oi14ft099$b18>^y6bS?aPktSH3QjFIz&W^=@=tL=0s|UI%e3z-cqwTJaLWQ>-+Liv#7B;NsndJ ztn$+wz2NUK2OqJo_-MxLBxjs~GJmY1d6p0r;x%4XvNkkof%1%F0Kv}Q^5VH5a~7R{ z-9}LpEG2~gyc5~Gsw6pn)|DL-9!z_&4=aTU7LB++24PWY8K@0LFpe#nXa*Y*C7%)# z43RVVyr`s}vbO8YwqG|+YH&_0ytb057>$%fi8oRYBO)ptYl$LV;1wk=oiLgC(bI5& znRO^yChW#DmtZe4U|g5Afc~5lT(=mUrZHY{4**z{@mmUM4RDq*6#B{S$PNNdM5!PrsPzkAY(D}CY6}pjwBpMyQN0_XLFXfGpEa$7cJ@@&*h_f z)PycdrcpZKKlzr^U4=4`tWVZd4b?;=Ygv)QNs@*Ba`D$1i*Y)n{@l|>QUT~60Ew3# zY)S$)#j%@an%R@eg`Gn0arF@X|luPQ`J zFIZv%5|1lx>wMs$E8qC`Cld5P+vN;o?@2o&a{a5Mj5QEqJ*6As69m^#i^uCqD9f3g zthEU-c>Sc-`MX}X3xk@ldgcBUblkE**&vLDi-!QxizhH79fTDNbF^GY$hJPp5WNEl zOz_0epZ!5{?Gi+72o?_7SDAIK+8C0*x8)#X1KO8~#Xkln+KVECiiPCVCO+=ka<$>kOU8`ZVzlsT?GZHZp!3aKrO9C9ks9Ku|HmNrC z<_MDk*pKZXZdx#GY6B*_DT#+-Ku%`lWs2Yh_TY=+qi)@2BEW54Tn?FHZOrgX-PD0& z%3nDe*^&4E0EV>Qwn)lCvG=1mr<{1U=Dr{=^rzNiEuoXy=0Sn($SB5CI%sxDHl0Iq ztVk)gA{ZCUqT(`VA~XG&sYcn*XGvR@(b>^1hf}-EHP`$bEs616{#rLDd`9u?!0_7x z@@yjxVmc3U=N1GW@B+xnj7lb! zqS$t)RV&h&l;cAig}%%X5d%@gL@zANh-IrKBMY&3)YAa8j%viG)JaCvX7itFaic^h z7UC?|MrHp1Hb~r*^_hn|x@er@j`UWi40hm0f@)Ib&%?QmuxSZ<3C&0~3Dcf{B3P>8 zC{w^KOv6ZInB&p2 zcujZW$8UHp2?y|+Njh;w@ggn(fRJ=eh{7^G$p-%bPV&(*V)WB0gy!t5&&3|k*;;3E z*E4T@B%^r9+CA$ay7IYb2t;h{6k}`>b9E48Mvx5a+L7*kMN{)91Ls<;9WvU=-K&z} zwp@YVm*!kf&NeMqC8T6CD8xX<4Eo-oyX)b>_R3kWG9{!k$ouO+pOego;3whF-9|U# zD^eZd`p9uc{B}rj1+KxJsLQu(op$B*KZcwb-M2U4kkuCOF3haIfa(M4+)oZ<05Y6N zKXnsrt`A_j`ziF?h#QDRO61&kT(XtjQi+KVEk`wK61r!^4NK|!NZ zX*UJwWHbT_^KMwPX4w8E!R7w|jYbi*YnvxA83s`Ul7Z4Ni^VtglC&|@BpdIh4mP_K z@VsNanGj_Q@EAs7zg0@$vTV*lh0fNGZrU=Do%`gH1c{O(N0n+!Vn#6oz=SUww~_$<2}hQYaag?Dv%JpSS^5-yh+xoNaBYwz_o8Pw}7n%g1O-v(__*bQ9*ux&>Z?=LRn#0t2~`*B!22WCMz!3%SikZ>Zk_K}f44?YT+j=5UrYQ1aT`ZVE^aDj;n=$jCL=Na{Y+puf81B=L~rjxrv~ z1~Js>MIkQHjv+YHD7gTMR-VGSDF$Y%)1b`I zL^)3q@Om;4a1Kf0svBZt3$r=vZt9fc=TUO$oGg)@?vrgdwbxS;$1kw3t-%0!dy+Hay!RgDa5!oRPoF$ltO{ z0};g^>qDc2JWeqIE8^+F(arZj;TY#Ay2%F=W5(1pO^z5uB8j+$b`E>GIefT zwyTo_OGI;T7Nsd(vvLx-0RjV7BZ+9UYBtHEnQ{H4=3db`=dOB()n5xh-3XGLMq_0d zCK`v)MADfF{Lc-`v;g~Z0WNn*~I)fIWO#sb`XjVPw zV-h{MFE^Bz_T0uI**p^VZT?#Fd9Z9#fVT4lxbp((_~ zXR(zIE6tcNxh~O#H6Lw!8>};YQ_5y-+y@k8I|h8Yq0>?VH-GW>UFvMglLy&aZt8@F zM-%{Z`5|>(?$@^1I13fq^rqqw!gLQ%+!(}Xi)95%;vcXj*fCe*auh;5<4Z` zlN$Pvj3V+mAueQ#*V~57aF6v+o!I~fX2|SCWSUGutRd?ZDdQVuaY7!X`XlqF5Fa|E z52&p-VJct~+j9-#y&P2vg;Z5OXhuWKhWvig;59hJ(DXXfbZP?w-V0s6J9|bW`jAuw zV5pp|$=E8mW2Lb&$q8MNzjGmOk^1+PKx-C>0LB=Y`*&9X8^G-;nH~#}ln>TYj3Ooh zp?pH(wIS+PJj;U5mfE#@NJ?DJv$7$R+w)1nO{Yx0s(XB_%;lD*4x9TZjk?LABG)Tw zJK5B3=!k(dXte^5o6UJfaS{wTebn;iuH1hPA5~J_%ZF@=Yh1m!zAhaIT$Q^K@kRkg z;N}g%EE$0Nin5fC{=3PAxb@43ty=V*k)qLN?1;z~m+K%d+_o~}5r4Xmnj<-Fh+3?` zT7%+Ee)`pN@toU;3c4egwJB^SfoPrKX3%4zI4E3&>z6!srug{n&KwBnfF2U+FqnCP z#ULuL#P$WbNP3eI~;8Z z9buDR2AN==`=e@p8OEP(+SO)pzY*Y;qxNK)aF98;C>h(1v70oIjo3zQTI96y;uiFt zBGvkZKwL|oG-XTkydNrM5oYZHhkv?OYX z1@fr$ar~>+cLgnjbb)Ig^Ps zB{dZQE3yh=D2xrTPE}R3es$Q8LNOAFHPlID1ZHFgah8-dSu)y5s(H1znzV>t1Zo>b z3CM1Smk{2#3m*)>c{GD9>7~X#l+iD^_JZM0hV;TQmBD}@aTG-PScwcl0n?ih>R_1i z@gt}v79s}5Tvh?wL?Ud&$0(4g%G*6m=m}YrNTZQPQI)G8LUoC7$LlROELS2;KGJyG z1M32f$ygGUMPwTUBXR?TFccJ2w0%W9t&jkzbZp*?u33<3o;drHUS)}i+uGG0{7?q) z>Mokx<2f1t9cctMI$RV*%X)~;C-sny5;+xIR*~5y+b2;M;LVtzD4SfowC7!4d1)4i z>cwl8t3x7WV}tnI)${ku@@j)PBsuPtQ0L+A~-4>i+Xn|3G{~+I66N&IKY0;*l7!>7+_e4F{{Y*{&CQaIxc#N8m7StIBZU3rXFEEZsD)Td zmc53O9@PH;aZcR9)rhi7+L0f`w1I?8pbrU%Uzk`9Q68kFU-d^lEh)yI#z_A|nxh^>!W!^pSuAI3G&AtRd!Y35Z zrWiQ=SS?bd2V^Dzf|R+Lfe@IF zD_z5=ovT!bsXHbG%--w>EkCJD&(^f|+#H^c?IW!qIHCPp$4kd1WY=rd|LW%XBO zRfgvYzupx~de)I}YS`)^T@jHyITuZK=!Af*WWN_t*dj_c?OK%L^e2V=k#H9~x?;d0r?(s4f2dMbcCQSi#EhFTf3tNZ6dbo30FFbF%R|wA#P_098xN-~u{`*sa@? z#O&OxE=RrmwJFMgeCgSv8nPnz7VNzubw+agDfP}V+++Jo;9i_0G1SC;)uwTYn82|R zm(^M>OpI{c6TN0z_a_(Mtr!Ifk_##X?;}07Z`=C z$4qvf5%wf7Bi@v185qX~qR|jZ6(=A)rJ_evr9ql8 z)+G-Np9I?FVq3|S7Y)r}+2T;QsKQu-B0}u#=`2CWr`OrQVr$?)YX$=ty;OXHB{y>90rk~l0g=ej zc^tXbe+k?eN7%N6n^T&`Sjk(Cff5^pUH!mPiKtr;nn)QIYQ&6P7{*O8v%Eco^2d6hWFwkvlSkH8 zG*rWC;;Imh6XdYy7Vbk%nnY?65eg(+)I0z@o8Ch}&?88Jl3^)o2A&p8&R7>csPN4< zEzEgxA@toJcg$L&4>les9YGlBYPoU52tqu~zdAd6Hv{*r-9pKb^C9?Cl_uYnbwWJE ze7M{!^nZB+YFdFT$7<6kkV0lr<*!8PyTW7(7?apU^7GVuG7Ns&61Y6L_oXwYAr`$E z$WRCakG!=^7(uu}t;v(&f~F%8(SL2eTP?kkboR^8mc1 z_acNPc*LBc2LAxwRh`KJ4l> z71T>M#iR*c%s~|j0uzvws(O=O zGLZG@Q3hINAV>sY%gv40+P5Cm-RbPx@2TA7nm1)h+aaQVpJ%{F8(Gn+RB53z3IVrnwdn&7z!_^pb4nW*y|yrpu)s3grG zM{@R+s9Ti4#exnC93P~@PGUoKA#irGBNeXEluJGssmDOKeSwyq%)QleR>`+^BxyT+ znW81|xw~|nN8#RY)S4FV+90?X;EV(q#|dH)1c7p2rD>SQ1s (b s) ...') + max_s = seqlen + cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, + device=qkv.device) + output = flash_attn_varlen_qkvpacked_func( + qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0, + softmax_scale=self.softmax_scale, causal=causal + ) + output = rearrange(output, '(b s) ... -> b s ...', b=batch_size) + else: + nheads = qkv.shape[-2] + x = rearrange(qkv, 'b s three h d -> b s (three h d)') + x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask) + x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads) + output_unpad = flash_attn_varlen_qkvpacked_func( + x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0, + softmax_scale=self.softmax_scale, causal=causal + ) + output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'), + indices, batch_size, seqlen), + 'b s (h d) -> b s h d', h=nheads) + else: + assert max_s is not None + output = flash_attn_varlen_qkvpacked_func( + qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0, + softmax_scale=self.softmax_scale, causal=causal + ) + + return output, None + + +class InternRMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + +try: + from apex.normalization import FusedRMSNorm + + InternRMSNorm = FusedRMSNorm # noqa + + logger.info('Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm') +except ImportError: + # using the normal InternRMSNorm + pass +except Exception: + logger.warning('discovered apex but it failed to load, falling back to InternRMSNorm') + pass + + +NORM2FN = { + 'rms_norm': InternRMSNorm, + 'layer_norm': nn.LayerNorm, +} + + +class InternVisionEmbeddings(nn.Module): + def __init__(self, config: InternVisionConfig): + super().__init__() + self.config = config + self.embed_dim = config.hidden_size + self.image_size = config.image_size + self.patch_size = config.patch_size + + self.class_embedding = nn.Parameter( + torch.randn(1, 1, self.embed_dim), + ) + + self.patch_embedding = nn.Conv2d( + in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size + ) + + self.num_patches = (self.image_size // self.patch_size) ** 2 + self.num_positions = self.num_patches + 1 + + self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim)) + + def _get_pos_embed(self, pos_embed, H, W): + target_dtype = pos_embed.dtype + pos_embed = pos_embed.float().reshape( + 1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2) + pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False). \ + reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype) + return pos_embed + + def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor: + target_dtype = self.patch_embedding.weight.dtype + patch_embeds = self.patch_embedding(pixel_values) # shape = [*, channel, width, height] + batch_size, _, height, width = patch_embeds.shape + patch_embeds = patch_embeds.flatten(2).transpose(1, 2) + class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype) + embeddings = torch.cat([class_embeds, patch_embeds], dim=1) + position_embedding = torch.cat([ + self.position_embedding[:, :1, :], + self._get_pos_embed(self.position_embedding[:, 1:, :], height, width) + ], dim=1) + embeddings = embeddings + position_embedding.to(target_dtype) + return embeddings + + +class InternAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__(self, config: InternVisionConfig): + super().__init__() + self.config = config + self.embed_dim = config.hidden_size + self.num_heads = config.num_attention_heads + self.use_flash_attn = config.use_flash_attn and has_flash_attn + if config.use_flash_attn and not has_flash_attn: + print('Warning: Flash Attention is not available, use_flash_attn is set to False.') + self.head_dim = self.embed_dim // self.num_heads + if self.head_dim * self.num_heads != self.embed_dim: + raise ValueError( + f'embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:' + f' {self.num_heads}).' + ) + + self.scale = self.head_dim ** -0.5 + self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=config.qkv_bias) + self.attn_drop = nn.Dropout(config.attention_dropout) + self.proj_drop = nn.Dropout(config.dropout) + + self.qk_normalization = config.qk_normalization + + if self.qk_normalization: + self.q_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps) + self.k_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps) + + if self.use_flash_attn: + self.inner_attn = FlashAttention(attention_dropout=config.attention_dropout) + self.proj = nn.Linear(self.embed_dim, self.embed_dim) + + def _naive_attn(self, x): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv.unbind(0) # make torchscript happy (cannot use tensor as tuple) + + if self.qk_normalization: + B_, H_, N_, D_ = q.shape + q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2) + k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2) + + attn = ((q * self.scale) @ k.transpose(-2, -1)) + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + def _flash_attn(self, x, key_padding_mask=None, need_weights=False): + qkv = self.qkv(x) + qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.num_heads) + + if self.qk_normalization: + q, k, v = qkv.unbind(2) + q = self.q_norm(q.flatten(-2, -1)).view(q.shape) + k = self.k_norm(k.flatten(-2, -1)).view(k.shape) + qkv = torch.stack([q, k, v], dim=2) + + context, _ = self.inner_attn( + qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=False + ) + outs = self.proj(rearrange(context, 'b s h d -> b s (h d)')) + outs = self.proj_drop(outs) + return outs + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + x = self._naive_attn(hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states) + return x + + +class InternMLP(nn.Module): + def __init__(self, config: InternVisionConfig): + super().__init__() + self.config = config + self.act = ACT2FN[config.hidden_act] + self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size) + self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = self.fc1(hidden_states) + hidden_states = self.act(hidden_states) + hidden_states = self.fc2(hidden_states) + return hidden_states + + +class InternVisionEncoderLayer(nn.Module): + def __init__(self, config: InternVisionConfig, drop_path_rate: float): + super().__init__() + self.embed_dim = config.hidden_size + self.intermediate_size = config.intermediate_size + self.norm_type = config.norm_type + + self.attn = InternAttention(config) + self.mlp = InternMLP(config) + self.norm1 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps) + self.norm2 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps) + + self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim)) + self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim)) + self.drop_path1 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity() + self.drop_path2 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity() + + def forward( + self, + hidden_states: torch.Tensor, + ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Tuple[torch.FloatTensor]]]: + """ + Args: + hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)` + """ + hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states).to(hidden_states.dtype)) * self.ls1) + + hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states).to(hidden_states.dtype)) * self.ls2) + + return hidden_states + + +class InternVisionEncoder(nn.Module): + """ + Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a + [`InternEncoderLayer`]. + + Args: + config (`InternConfig`): + The corresponding vision configuration for the `InternEncoder`. + """ + + def __init__(self, config: InternVisionConfig): + super().__init__() + self.config = config + # stochastic depth decay rule + dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)] + self.layers = nn.ModuleList([ + InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)]) + self.gradient_checkpointing = True + + def forward( + self, + inputs_embeds, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutput]: + r""" + Args: + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): + Embedded representation of the inputs. Should be float, not int tokens. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + """ + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + encoder_states = () if output_hidden_states else None + hidden_states = inputs_embeds + + for idx, encoder_layer in enumerate(self.layers): + if output_hidden_states: + encoder_states = encoder_states + (hidden_states,) + if self.gradient_checkpointing and self.training: + layer_outputs = torch.utils.checkpoint.checkpoint( + encoder_layer, + hidden_states) + else: + layer_outputs = encoder_layer( + hidden_states, + ) + hidden_states = layer_outputs + + if output_hidden_states: + encoder_states = encoder_states + (hidden_states,) + + if not return_dict: + return tuple(v for v in [hidden_states, encoder_states] if v is not None) + return BaseModelOutput( + last_hidden_state=hidden_states, hidden_states=encoder_states + ) + + +class InternVisionModel(PreTrainedModel): + main_input_name = 'pixel_values' + _supports_flash_attn_2 = True + config_class = InternVisionConfig + _no_split_modules = ['InternVisionEncoderLayer'] + + def __init__(self, config: InternVisionConfig): + super().__init__(config) + self.config = config + + self.embeddings = InternVisionEmbeddings(config) + self.encoder = InternVisionEncoder(config) + + def resize_pos_embeddings(self, old_size, new_size, patch_size): + pos_emb = self.embeddings.position_embedding + _, num_positions, embed_dim = pos_emb.shape + cls_emb = pos_emb[:, :1, :] + pos_emb = pos_emb[:, 1:, :].reshape(1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2) + pos_emb = F.interpolate(pos_emb.float(), size=new_size // patch_size, mode='bicubic', align_corners=False) + pos_emb = pos_emb.to(cls_emb.dtype).reshape(1, embed_dim, -1).permute(0, 2, 1) + pos_emb = torch.cat([cls_emb, pos_emb], dim=1) + self.embeddings.position_embedding = nn.Parameter(pos_emb) + self.embeddings.image_size = new_size + logger.info('Resized position embeddings from {} to {}'.format(old_size, new_size)) + + def get_input_embeddings(self): + return self.embeddings + + def forward( + self, + pixel_values: Optional[torch.FloatTensor] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + pixel_embeds: Optional[torch.FloatTensor] = None, + ) -> Union[Tuple, BaseModelOutputWithPooling]: + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if pixel_values is None and pixel_embeds is None: + raise ValueError('You have to specify pixel_values or pixel_embeds') + + if pixel_embeds is not None: + hidden_states = pixel_embeds + else: + if len(pixel_values.shape) == 4: + hidden_states = self.embeddings(pixel_values) + else: + raise ValueError(f'wrong pixel_values size: {pixel_values.shape}') + encoder_outputs = self.encoder( + inputs_embeds=hidden_states, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + last_hidden_state = encoder_outputs.last_hidden_state + pooled_output = last_hidden_state[:, 0, :] + + if not return_dict: + return (last_hidden_state, pooled_output) + encoder_outputs[1:] + + return BaseModelOutputWithPooling( + last_hidden_state=last_hidden_state, + pooler_output=pooled_output, + hidden_states=encoder_outputs.hidden_states, + attentions=encoder_outputs.attentions, + ) diff --git a/modeling_internvl_chat.py b/modeling_internvl_chat.py new file mode 100644 index 0000000..55b5871 --- /dev/null +++ b/modeling_internvl_chat.py @@ -0,0 +1,349 @@ +# -------------------------------------------------------- +# InternVL +# Copyright (c) 2024 OpenGVLab +# Licensed under The MIT License [see LICENSE for details] +# -------------------------------------------------------- + +import warnings +from typing import Any, List, Optional, Tuple, Union + +import torch.utils.checkpoint +import transformers +from torch import nn +from torch.nn import CrossEntropyLoss +from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM, + LlamaTokenizer) +from transformers.modeling_outputs import CausalLMOutputWithPast +from transformers.modeling_utils import PreTrainedModel +from transformers.utils import ModelOutput, logging + +from .configuration_internvl_chat import InternVLChatConfig +from .conversation import get_conv_template +from .modeling_intern_vit import InternVisionModel, has_flash_attn +from .modeling_phi3 import Phi3ForCausalLM + +logger = logging.get_logger(__name__) + + +def version_cmp(v1, v2, op='eq'): + import operator + + from packaging import version + op_func = getattr(operator, op) + return op_func(version.parse(v1), version.parse(v2)) + + +class InternVLChatModel(PreTrainedModel): + config_class = InternVLChatConfig + main_input_name = 'pixel_values' + base_model_prefix = 'language_model' + _supports_flash_attn_2 = True + _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'Phi3DecoderLayer'] + + def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None, use_flash_attn=True): + super().__init__(config) + + assert version_cmp(transformers.__version__, '4.37.0', 'ge') + image_size = config.force_image_size or config.vision_config.image_size + patch_size = config.vision_config.patch_size + self.patch_size = patch_size + self.select_layer = config.select_layer + self.template = config.template + self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2)) + self.downsample_ratio = config.downsample_ratio + self.ps_version = config.ps_version + use_flash_attn = use_flash_attn if has_flash_attn else False + config.vision_config.use_flash_attn = True if use_flash_attn else False + config.llm_config._attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager' + + logger.info(f'num_image_token: {self.num_image_token}') + logger.info(f'ps_version: {self.ps_version}') + if vision_model is not None: + self.vision_model = vision_model + else: + self.vision_model = InternVisionModel(config.vision_config) + if language_model is not None: + self.language_model = language_model + else: + if config.llm_config.architectures[0] == 'LlamaForCausalLM': + self.language_model = LlamaForCausalLM(config.llm_config) + elif config.llm_config.architectures[0] == 'Phi3ForCausalLM': + self.language_model = Phi3ForCausalLM(config.llm_config) + else: + raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.') + + vit_hidden_size = config.vision_config.hidden_size + llm_hidden_size = config.llm_config.hidden_size + + self.mlp1 = nn.Sequential( + nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2), + nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size), + nn.GELU(), + nn.Linear(llm_hidden_size, llm_hidden_size) + ) + + self.img_context_token_id = None + self.conv_template = get_conv_template(self.template) + self.system_message = self.conv_template.system_message + + def forward( + self, + pixel_values: torch.FloatTensor, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + image_flags: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, CausalLMOutputWithPast]: + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + image_flags = image_flags.squeeze(-1) + input_embeds = self.language_model.get_input_embeddings()(input_ids).clone() + + vit_embeds = self.extract_feature(pixel_values) + vit_embeds = vit_embeds[image_flags == 1] + vit_batch_size = pixel_values.shape[0] + + B, N, C = input_embeds.shape + input_embeds = input_embeds.reshape(B * N, C) + + if torch.distributed.is_initialized() and torch.distributed.get_rank() == 0: + print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}') + + input_ids = input_ids.reshape(B * N) + selected = (input_ids == self.img_context_token_id) + try: + input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C) + except Exception as e: + vit_embeds = vit_embeds.reshape(-1, C) + print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, ' + f'vit_embeds.shape={vit_embeds.shape}') + n_token = selected.sum() + input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token] + + input_embeds = input_embeds.reshape(B, N, C) + + outputs = self.language_model( + inputs_embeds=input_embeds, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + logits = outputs.logits + + loss = None + if labels is not None: + # Shift so that tokens < n predict n + shift_logits = logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss() + shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size) + shift_labels = shift_labels.view(-1) + # Enable model parallelism + shift_labels = shift_labels.to(shift_logits.device) + loss = loss_fct(shift_logits, shift_labels) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + def pixel_shuffle(self, x, scale_factor=0.5): + n, w, h, c = x.size() + # N, W, H, C --> N, W, H * scale, C // scale + x = x.view(n, w, int(h * scale_factor), int(c / scale_factor)) + # N, W, H * scale, C // scale --> N, H * scale, W, C // scale + x = x.permute(0, 2, 1, 3).contiguous() + # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2) + x = x.view(n, int(h * scale_factor), int(w * scale_factor), + int(c / (scale_factor * scale_factor))) + if self.ps_version == 'v1': + warnings.warn("In ps_version 'v1', the height and width have not been swapped back, " + 'which results in a transposed image.') + else: + x = x.permute(0, 2, 1, 3).contiguous() + return x + + def extract_feature(self, pixel_values): + if self.select_layer == -1: + vit_embeds = self.vision_model( + pixel_values=pixel_values, + output_hidden_states=False, + return_dict=True).last_hidden_state + else: + vit_embeds = self.vision_model( + pixel_values=pixel_values, + output_hidden_states=True, + return_dict=True).hidden_states[self.select_layer] + vit_embeds = vit_embeds[:, 1:, :] + + h = w = int(vit_embeds.shape[1] ** 0.5) + vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1) + vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio) + vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1]) + vit_embeds = self.mlp1(vit_embeds) + return vit_embeds + + def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None, + history=None, return_history=False, IMG_START_TOKEN='', IMG_END_TOKEN='', + IMG_CONTEXT_TOKEN='', verbose=False, image_counts=None): + if history is not None or return_history: + print('Now multi-turn chat is not supported in batch_chat.') + raise NotImplementedError + + if image_counts is not None: + num_patches_list = image_counts + print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.') + + img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) + self.img_context_token_id = img_context_token_id + + if verbose and pixel_values is not None: + image_bs = pixel_values.shape[0] + print(f'dynamic ViT batch size: {image_bs}') + + queries = [] + for idx, num_patches in enumerate(num_patches_list): + question = questions[idx] + if pixel_values is not None and '' not in question: + question = '\n' + question + template = get_conv_template(self.template) + template.system_message = self.system_message + template.append_message(template.roles[0], question) + template.append_message(template.roles[1], None) + query = template.get_prompt() + + image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN + query = query.replace('', image_tokens, 1) + queries.append(query) + + tokenizer.padding_side = 'left' + model_inputs = tokenizer(queries, return_tensors='pt', padding=True) + input_ids = model_inputs['input_ids'].to(self.device) + attention_mask = model_inputs['attention_mask'].to(self.device) + eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip()) + generation_config['eos_token_id'] = eos_token_id + generation_output = self.generate( + pixel_values=pixel_values, + input_ids=input_ids, + attention_mask=attention_mask, + **generation_config + ) + responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True) + responses = [response.split(template.sep.strip())[0].strip() for response in responses] + return responses + + def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False, + num_patches_list=None, IMG_START_TOKEN='', IMG_END_TOKEN='', IMG_CONTEXT_TOKEN='', + verbose=False): + + if history is None and pixel_values is not None and '' not in question: + question = '\n' + question + + if num_patches_list is None: + num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else [] + assert pixel_values is None or len(pixel_values) == sum(num_patches_list) + + img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) + self.img_context_token_id = img_context_token_id + + template = get_conv_template(self.template) + template.system_message = self.system_message + eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip()) + + history = [] if history is None else history + for (old_question, old_answer) in history: + template.append_message(template.roles[0], old_question) + template.append_message(template.roles[1], old_answer) + template.append_message(template.roles[0], question) + template.append_message(template.roles[1], None) + query = template.get_prompt() + + if verbose and pixel_values is not None: + image_bs = pixel_values.shape[0] + print(f'dynamic ViT batch size: {image_bs}') + + for num_patches in num_patches_list: + image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN + query = query.replace('', image_tokens, 1) + + model_inputs = tokenizer(query, return_tensors='pt') + input_ids = model_inputs['input_ids'].to(self.device) + attention_mask = model_inputs['attention_mask'].to(self.device) + generation_config['eos_token_id'] = eos_token_id + generation_output = self.generate( + pixel_values=pixel_values, + input_ids=input_ids, + attention_mask=attention_mask, + **generation_config + ) + response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0] + response = response.split(template.sep.strip())[0].strip() + history.append((question, response)) + if return_history: + return response, history + else: + query_to_print = query.replace(IMG_CONTEXT_TOKEN, '') + query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '') + if verbose: + print(query_to_print, response) + return response + + @torch.no_grad() + def generate( + self, + pixel_values: Optional[torch.FloatTensor] = None, + input_ids: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.LongTensor] = None, + visual_features: Optional[torch.FloatTensor] = None, + generation_config: Optional[GenerationConfig] = None, + output_hidden_states: Optional[bool] = None, + **generate_kwargs, + ) -> torch.LongTensor: + + assert self.img_context_token_id is not None + if pixel_values is not None: + if visual_features is not None: + vit_embeds = visual_features + else: + vit_embeds = self.extract_feature(pixel_values) + input_embeds = self.language_model.get_input_embeddings()(input_ids) + B, N, C = input_embeds.shape + input_embeds = input_embeds.reshape(B * N, C) + + input_ids = input_ids.reshape(B * N) + selected = (input_ids == self.img_context_token_id) + assert selected.sum() != 0 + input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device) + + input_embeds = input_embeds.reshape(B, N, C) + else: + input_embeds = self.language_model.get_input_embeddings()(input_ids) + + outputs = self.language_model.generate( + inputs_embeds=input_embeds, + attention_mask=attention_mask, + generation_config=generation_config, + output_hidden_states=output_hidden_states, + use_cache=True, + **generate_kwargs, + ) + + return outputs diff --git a/modeling_phi3.py b/modeling_phi3.py new file mode 100644 index 0000000..8bb5dc0 --- /dev/null +++ b/modeling_phi3.py @@ -0,0 +1,1610 @@ +# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" PyTorch Phi-3 model.""" + +import inspect +import math +import warnings +from typing import List, Optional, Tuple, Union + +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +from torch import nn +from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss +from transformers.activations import ACT2FN +from transformers.cache_utils import Cache, DynamicCache +from transformers.modeling_attn_mask_utils import \ + _prepare_4d_causal_attention_mask +from transformers.modeling_outputs import (BaseModelOutputWithPast, + CausalLMOutputWithPast, + SequenceClassifierOutputWithPast, + TokenClassifierOutput) +from transformers.modeling_utils import PreTrainedModel +from transformers.utils import (add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + is_flash_attn_2_available, + is_flash_attn_greater_or_equal_2_10, logging, + replace_return_docstrings) + +from .configuration_phi3 import Phi3Config + +logger = logging.get_logger(__name__) + +# Transformers scans dependencies in the modeling file, causing issues on conditional loading. The regex only ignores try/catch blocks, but not if statements +# if is_flash_attn_2_available(): +_flash_supports_window_size = False +try: + from flash_attn import flash_attn_func, flash_attn_varlen_func + from flash_attn.bert_padding import (index_first_axis, pad_input, # noqa + unpad_input) + + _flash_supports_window_size = 'window_size' in list(inspect.signature(flash_attn_func).parameters) + has_flash_attn = True +except ImportError as error: + logger.warning( + f'`flash-attention` package not found, consider installing for better performance: {error}.' + ) + if not _flash_supports_window_size: + logger.warning( + "Current `flash-attenton` does not support `window_size`. Either upgrade or use `attn_implementation='eager'`." + ) + has_flash_attn = False + +_CHECKPOINT_FOR_DOC = 'microsoft/Phi-3-mini-4k-instruct' +_CONFIG_FOR_DOC = 'Phi3Config' + +PHI3_PRETRAINED_MODEL_ARCHIVE_LIST = [ + 'microsoft/Phi-3-mini-4k-instruct', + 'microsoft/Phi-3-mini-128k-instruct', + # See all Phi-3 models at https://huggingface.co/models?filter=Phi-3 +] + + +# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3 +class Phi3RMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + Phi3RMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + +# Copied from transformers.models.llama.modeling_llama._get_unpad_data +def _get_unpad_data(attention_mask): + seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) + indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() + max_seqlen_in_batch = seqlens_in_batch.max().item() + cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0)) + return ( + indices, + cu_seqlens, + max_seqlen_in_batch, + ) + + +# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3 +class Phi3RotaryEmbedding(nn.Module): + def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None): + super().__init__() + + self.dim = dim + self.max_position_embeddings = max_position_embeddings + self.base = base + self.register_buffer('inv_freq', None, persistent=False) + + @torch.no_grad() + def forward(self, x, position_ids, seq_len=None): + # x: [bs, num_attention_heads, seq_len, head_size] + if self.inv_freq is None: + self.inv_freq = 1.0 / ( + self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim) + ) + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) + position_ids_expanded = position_ids[:, None, :].float() + # Force float32 since bfloat16 loses precision on long contexts + # See https://github.com/huggingface/transformers/pull/29285 + device_type = x.device.type + device_type = device_type if isinstance(device_type, str) and device_type != 'mps' else 'cpu' + with torch.autocast(device_type=device_type, enabled=False): + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() + sin = emb.sin() + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding): + def __init__(self, dim, config, device=None): + super().__init__(dim, config.max_position_embeddings, config.rope_theta, device) + + self.short_factor = config.rope_scaling['short_factor'] + self.long_factor = config.rope_scaling['long_factor'] + self.original_max_position_embeddings = config.original_max_position_embeddings + + @torch.no_grad() + def forward(self, x, position_ids, seq_len=None): + seq_len = torch.max(position_ids) + 1 + if seq_len > self.original_max_position_embeddings: + ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device) + else: + ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device) + + inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim + self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape) + + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) + position_ids_expanded = position_ids[:, None, :].float() + + # Force float32 since bfloat16 loses precision on long contexts + # See https://github.com/huggingface/transformers/pull/29285 + device_type = x.device.type + device_type = device_type if isinstance(device_type, str) and device_type != 'mps' else 'cpu' + with torch.autocast(device_type=device_type, enabled=False): + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + + scale = self.max_position_embeddings / self.original_max_position_embeddings + if scale <= 1.0: + scaling_factor = 1.0 + else: + scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings)) + + cos = emb.cos() * scaling_factor + sin = emb.sin() * scaling_factor + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding): + def __init__(self, dim, config, device=None): + super().__init__(dim, config.max_position_embeddings, config.rope_theta, device) + + self.short_factor = config.rope_scaling['short_factor'] + self.long_factor = config.rope_scaling['long_factor'] + self.original_max_position_embeddings = config.original_max_position_embeddings + + @torch.no_grad() + def forward(self, x, position_ids, seq_len=None): + seq_len = torch.max(position_ids) + 1 + if seq_len > self.original_max_position_embeddings: + ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device) + else: + ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device) + + inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim + self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape) + + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) + position_ids_expanded = position_ids[:, None, :].float() + + # Force float32 since bfloat16 loses precision on long contexts + # See https://github.com/huggingface/transformers/pull/29285 + device_type = x.device.type + device_type = device_type if isinstance(device_type, str) and device_type != 'mps' else 'cpu' + with torch.autocast(device_type=device_type, enabled=False): + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + + scale = self.max_position_embeddings / self.original_max_position_embeddings + if scale <= 1.0: + scaling_factor = 1.0 + else: + scaling_factor = 0.1 * math.log(scale) + 1.0 + + cos = emb.cos() * scaling_factor + sin = emb.sin() * scaling_factor + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +# Copied from transformers.models.llama.modeling_llama.rotate_half +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`, *optional*): + Deprecated and unused. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + + +class Phi3MLP(nn.Module): + def __init__(self, config): + super().__init__() + + self.config = config + self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False) + self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) + + self.activation_fn = ACT2FN[config.hidden_act] + + def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor: + up_states = self.gate_up_proj(hidden_states) + + gate, up_states = up_states.chunk(2, dim=-1) + up_states = up_states * self.activation_fn(gate) + + return self.down_proj(up_states) + + +# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi +def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: + """ + This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, + num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) + """ + batch, num_key_value_heads, slen, head_dim = hidden_states.shape + if n_rep == 1: + return hidden_states + hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) + return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) + + +class Phi3Attention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None): + super().__init__() + self.config = config + self.layer_idx = layer_idx + if layer_idx is None: + logger.warning_once( + f'Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will ' + 'lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` ' + 'when creating this class.' + ) + + self.attention_dropout = config.attention_dropout + self.hidden_size = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = self.hidden_size // self.num_heads + self.num_key_value_heads = config.num_key_value_heads + self.num_key_value_groups = self.num_heads // self.num_key_value_heads + self.max_position_embeddings = config.max_position_embeddings + self.original_max_position_embeddings = config.original_max_position_embeddings + self.rope_theta = config.rope_theta + self.rope_scaling = config.rope_scaling + self.is_causal = True + + if (self.head_dim * self.num_heads) != self.hidden_size: + raise ValueError( + f'hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}' + f' and `num_heads`: {self.num_heads}).' + ) + + op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim) + self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) + self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False) + self._init_rope() + + def _init_rope(self): + if self.rope_scaling is None: + self.rotary_emb = Phi3RotaryEmbedding( + self.head_dim, + max_position_embeddings=self.max_position_embeddings, + base=self.rope_theta, + ) + else: + scaling_type = self.config.rope_scaling['type'] + if scaling_type == 'su': + self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config) + elif scaling_type == 'yarn': + self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config) + else: + raise ValueError(f'Unknown RoPE scaling type {scaling_type}') + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + logger.warning_once('You are not running the flash-attention implementation, expect numerical differences.') + + bsz, q_len, _ = hidden_states.size() + + qkv = self.qkv_proj(hidden_states) + query_pos = self.num_heads * self.head_dim + query_states = qkv[..., :query_pos] + key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim] + value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :] + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + if self.layer_idx is None: + raise ValueError( + f'The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} ' + 'for auto-regressive decoding with k/v caching, please make sure to initialize the attention class ' + 'with a layer index.' + ) + kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) + cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len) + + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + cache_kwargs = {'sin': sin, 'cos': cos} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + # repeat k/v heads if n_kv_heads < n_heads + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) + + if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len): + raise ValueError( + f'Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is' + f' {attn_weights.size()}' + ) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): + raise ValueError( + f'Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}' + ) + attn_weights = attn_weights + attention_mask + + # upcast attention to fp32 + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) + + attn_output = torch.matmul(attn_weights, value_states) + + if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): + raise ValueError( + f'`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is' + f' {attn_output.size()}' + ) + + attn_output = attn_output.transpose(1, 2).contiguous() + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) + + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + +class Phi3FlashAttention2(Phi3Attention): + """ + Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays + untouched. The only required change would be on the forward pass where it needs to correctly call the public API of + flash attention and deal with padding tokens in case the input contains any of them. + """ + + # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__ + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. + # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. + # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). + self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + **kwargs, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + # Phi3FlashAttention2 attention does not support output_attentions + + if not _flash_supports_window_size: + logger.warning_once( + "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library." + ) + raise ValueError('The current flash attention version does not support sliding window attention.') + + output_attentions = False + + if 'padding_mask' in kwargs: + warnings.warn( + 'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`' + ) + + # overwrite attention_mask with padding_mask + attention_mask = kwargs.pop('padding_mask') + + bsz, q_len, _ = hidden_states.size() + + qkv = self.qkv_proj(hidden_states) + query_pos = self.num_heads * self.head_dim + query_states = qkv[..., :query_pos] + key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim] + value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :] + + # Flash attention requires the input to have the shape + # batch_size x seq_length x head_dim x hidden_dim + # therefore we just need to keep the original shape + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + if self.layer_idx is None: + raise ValueError( + f'The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} ' + 'for auto-regressive decoding with k/v caching, please make sure to initialize the attention class ' + 'with a layer index.' + ) + kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) + + # Because the input can be padded, the absolute sequence length depends on the max position id. + rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1 + cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len) + + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + use_sliding_windows = ( + _flash_supports_window_size + and getattr(self.config, 'sliding_window', None) is not None + and kv_seq_len > self.config.sliding_window + ) + + if past_key_value is not None: + # Activate slicing cache only if the config has a value `sliding_windows` attribute + cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0 + if ( + getattr(self.config, 'sliding_window', None) is not None + and kv_seq_len > self.config.sliding_window + and cache_has_contents + ): + slicing_tokens = 1 - self.config.sliding_window + + past_key = past_key_value[self.layer_idx][0] + past_value = past_key_value[self.layer_idx][1] + + past_key = past_key[:, :, slicing_tokens:, :].contiguous() + past_value = past_value[:, :, slicing_tokens:, :].contiguous() + + if past_key.shape[-2] != self.config.sliding_window - 1: + raise ValueError( + f'past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got' + f' {past_key.shape}' + ) + + if attention_mask is not None: + attention_mask = attention_mask[:, slicing_tokens:] + attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) + + cache_kwargs = {'sin': sin, 'cos': cos} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + # repeat k/v heads if n_kv_heads < n_heads + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + attn_dropout = self.attention_dropout if self.training else 0.0 + + # In PEFT, usually we cast the layer norms in float32 for training stability reasons + # therefore the input hidden states gets silently casted in float32. Hence, we need + # cast them back in the correct dtype just to be sure everything works as expected. + # This might slowdown training & inference so it is recommended to not cast the LayerNorms + # in fp32. + + if query_states.dtype == torch.float32: + if torch.is_autocast_enabled(): + target_dtype = torch.get_autocast_gpu_dtype() + # Handle the case where the model is quantized + elif hasattr(self.config, '_pre_quantization_dtype'): + target_dtype = self.config._pre_quantization_dtype + else: + target_dtype = self.qkv_proj.weight.dtype + + logger.warning_once( + f'The input hidden states seems to be silently casted in float32, this might be related to' + f' the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in' + f' {target_dtype}.' + ) + + query_states = query_states.to(target_dtype) + key_states = key_states.to(target_dtype) + value_states = value_states.to(target_dtype) + + # Reashape to the expected shape for Flash Attention + query_states = query_states.transpose(1, 2) + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + + attn_output = self._flash_attention_forward( + query_states, + key_states, + value_states, + attention_mask, + q_len, + dropout=attn_dropout, + use_sliding_windows=use_sliding_windows, + ) + + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous() + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward + def _flash_attention_forward( + self, + query_states, + key_states, + value_states, + attention_mask, + query_length, + dropout=0.0, + softmax_scale=None, + use_sliding_windows=False, + ): + """ + Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token + first unpad the input, then computes the attention scores and pad the final attention scores. + + Args: + query_states (`torch.Tensor`): + Input query states to be passed to Flash Attention API + key_states (`torch.Tensor`): + Input key states to be passed to Flash Attention API + value_states (`torch.Tensor`): + Input value states to be passed to Flash Attention API + attention_mask (`torch.Tensor`): + The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the + position of padding tokens and 1 for the position of non-padding tokens. + dropout (`float`): + Attention dropout + softmax_scale (`float`, *optional*): + The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) + use_sliding_windows (`bool`, *optional*): + Whether to activate sliding window attention. + """ + if not self._flash_attn_uses_top_left_mask: + causal = self.is_causal + else: + # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__. + causal = self.is_causal and query_length != 1 + + # Contains at least one padding token in the sequence + if attention_mask is not None: + batch_size = query_states.shape[0] + query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( + query_states, key_states, value_states, attention_mask, query_length + ) + + cu_seqlens_q, cu_seqlens_k = cu_seq_lens + max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens + + if not use_sliding_windows: + attn_output_unpad = flash_attn_varlen_func( + query_states, + key_states, + value_states, + cu_seqlens_q=cu_seqlens_q, + cu_seqlens_k=cu_seqlens_k, + max_seqlen_q=max_seqlen_in_batch_q, + max_seqlen_k=max_seqlen_in_batch_k, + dropout_p=dropout, + softmax_scale=softmax_scale, + causal=causal, + ) + else: + attn_output_unpad = flash_attn_varlen_func( + query_states, + key_states, + value_states, + cu_seqlens_q=cu_seqlens_q, + cu_seqlens_k=cu_seqlens_k, + max_seqlen_q=max_seqlen_in_batch_q, + max_seqlen_k=max_seqlen_in_batch_k, + dropout_p=dropout, + softmax_scale=softmax_scale, + causal=causal, + window_size=(self.config.sliding_window, self.config.sliding_window), + ) + + attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) + else: + if not use_sliding_windows: + attn_output = flash_attn_func( + query_states, + key_states, + value_states, + dropout, + softmax_scale=softmax_scale, + causal=causal, + ) + else: + attn_output = flash_attn_func( + query_states, + key_states, + value_states, + dropout, + softmax_scale=softmax_scale, + causal=causal, + window_size=(self.config.sliding_window, self.config.sliding_window), + ) + + return attn_output + + # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input + def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): + batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape + + # On the first iteration we need to properly re-create the padding mask + # by slicing it on the proper place + if kv_seq_len != attention_mask.shape[-1]: + attention_mask_num_tokens = attention_mask.shape[-1] + attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :] + + indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) + + key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) + value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) + + if query_length == kv_seq_len: + query_layer = index_first_axis( + query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k + ) + cu_seqlens_q = cu_seqlens_k + max_seqlen_in_batch_q = max_seqlen_in_batch_k + indices_q = indices_k + elif query_length == 1: + max_seqlen_in_batch_q = 1 + cu_seqlens_q = torch.arange( + batch_size + 1, dtype=torch.int32, device=query_layer.device + ) # There is a memcpy here, that is very bad. + indices_q = cu_seqlens_q[:-1] + query_layer = query_layer.squeeze(1) + else: + # The -q_len: slice assumes left padding. + attention_mask = attention_mask[:, -query_length:] + query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) + + return ( + query_layer, + key_layer, + value_layer, + indices_q, + (cu_seqlens_q, cu_seqlens_k), + (max_seqlen_in_batch_q, max_seqlen_in_batch_k), + ) + + +# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3 +# TODO @Arthur no longer copied from LLama after static cache +class Phi3SdpaAttention(Phi3Attention): + """ + Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from + `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to + SDPA API. + """ + + # Adapted from Phi3Attention.forward + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + if output_attentions: + # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. + logger.warning_once( + 'Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, ' + 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + return super().forward( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + bsz, q_len, _ = hidden_states.size() + + qkv = self.qkv_proj(hidden_states) + query_pos = self.num_heads * self.head_dim + query_states = qkv[..., :query_pos] + key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim] + value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :] + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) + cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len) + + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + cache_kwargs = {'sin': sin, 'cos': cos} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): + raise ValueError( + f'Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}' + ) + + # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, + # Reference: https://github.com/pytorch/pytorch/issues/112577. + if query_states.device.type == 'cuda' and attention_mask is not None: + query_states = query_states.contiguous() + key_states = key_states.contiguous() + value_states = value_states.contiguous() + + attn_output = torch.nn.functional.scaled_dot_product_attention( + query_states, + key_states, + value_states, + attn_mask=attention_mask, + dropout_p=self.attention_dropout if self.training else 0.0, + # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. + is_causal=self.is_causal and attention_mask is None and q_len > 1, + ) + + attn_output = attn_output.transpose(1, 2).contiguous() + attn_output = attn_output.view(bsz, q_len, self.hidden_size) + + attn_output = self.o_proj(attn_output) + + return attn_output, None, past_key_value + + +PHI3_ATTENTION_CLASSES = { + 'eager': Phi3Attention, + 'flash_attention_2': Phi3FlashAttention2, + 'sdpa': Phi3SdpaAttention, +} + + +class Phi3DecoderLayer(nn.Module): + def __init__(self, config: Phi3Config, layer_idx: int): + super().__init__() + + self.config = config + self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx) + + self.mlp = Phi3MLP(config) + self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + self.resid_attn_dropout = nn.Dropout(config.resid_pdrop) + self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop) + self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + **kwargs, + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + if 'padding_mask' in kwargs: + warnings.warn( + 'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`' + ) + """ + Args: + hidden_states (`torch.FloatTensor`): + input to the layer of shape `(batch, seq_len, embed_dim)` + attention_mask (`torch.FloatTensor`, *optional*): attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + position_ids (`torch.LongTensor` of shape `({0})`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range + `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding + (see `past_key_values`). + past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + """ + + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + attn_outputs, self_attn_weights, present_key_value = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + hidden_states = residual + self.resid_attn_dropout(attn_outputs) + + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = residual + self.resid_mlp_dropout(hidden_states) + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights,) + + if use_cache: + outputs += (present_key_value,) + + return outputs + + +PHI3_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + + Parameters: + config ([`Phi3Config`]): + Model configuration class with all the parameters of the model. Initializing with a config file does not + load the weights associated with the model, only the configuration. Check out the + [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + + +@add_start_docstrings( + 'The bare Phi-3 model outputting raw hidden-states without any specific head on top.', + PHI3_START_DOCSTRING, +) +class Phi3PreTrainedModel(PreTrainedModel): + config_class = Phi3Config + base_model_prefix = 'model' + supports_gradient_checkpointing = True + _no_split_modules = ['Phi3DecoderLayer'] + _skip_keys_device_placement = 'past_key_values' + _supports_flash_attn_2 = True + _supports_sdpa = False + _supports_cache_class = True + + _version = '0.0.5' + + def __init__(self, config: Phi3Config): + if not has_flash_attn: + config._attn_implementation = 'eager' + print('Warning: Flash attention is not available, using eager attention instead.') + super().__init__(config) + + def _init_weights(self, module): + std = self.config.initializer_range + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + +PHI3_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + If `past_key_values` is used, optionally only the last `input_ids` have to be input (see + `past_key_values`). + + If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, + config.n_positions - 1]`. + + [What are position IDs?](../glossary#position-ids) + past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): + Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` + returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. + + Two formats are allowed: + - a [`~cache_utils.Cache`] instance; + - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of + shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy + cache format. + + The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the + legacy cache format will be returned. + + If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't + have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` + of shape `(batch_size, sequence_length)`. + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This + is useful if you want more control over how to convert `input_ids` indices into associated vectors than the + model's internal embedding lookup matrix. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. +""" + + +@add_start_docstrings( + 'The bare Phi-3 model outputting raw hidden-states without any specific head on top.', + PHI3_START_DOCSTRING, +) +class Phi3Model(Phi3PreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`] + + Args: + config: Phi3Config + """ + + def __init__(self, config: Phi3Config): + super().__init__(config) + self.padding_idx = config.pad_token_id + self.vocab_size = config.vocab_size + + self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) + self.embed_dropout = nn.Dropout(config.embd_pdrop) + self.layers = nn.ModuleList( + [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self._attn_implementation = config._attn_implementation + + self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + self.gradient_checkpointing = False + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError('You cannot specify both input_ids and inputs_embeds at the same time') + elif input_ids is not None: + batch_size, seq_length = input_ids.shape[:2] + elif inputs_embeds is not None: + batch_size, seq_length = inputs_embeds.shape[:2] + else: + raise ValueError('You have to specify either input_ids or inputs_embeds') + + past_key_values_length = 0 + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + '`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...' + ) + use_cache = False + + if use_cache: + use_legacy_cache = not isinstance(past_key_values, Cache) + if use_legacy_cache: + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + past_key_values_length = past_key_values.get_usable_length(seq_length) + + if position_ids is None: + device = input_ids.device if input_ids is not None else inputs_embeds.device + position_ids = torch.arange( + past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + ) + position_ids = position_ids.unsqueeze(0).view(-1, seq_length) + else: + position_ids = position_ids.view(-1, seq_length).long() + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if attention_mask is not None and self._attn_implementation == 'flash_attention_2' and use_cache: + is_padding_right = attention_mask[:, -1].sum().item() != batch_size + if is_padding_right: + raise ValueError( + "You are attempting to perform batched generation with padding_side='right'" + ' this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to ' + " call `tokenizer.padding_side = 'left'` before tokenizing the input. " + ) + + if self._attn_implementation == 'flash_attention_2': + # 2d mask is passed through the layers + attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None + else: + # 4d mask is passed through the layers + attention_mask = _prepare_4d_causal_attention_mask( + attention_mask, + (batch_size, seq_length), + inputs_embeds, + past_key_values_length, + sliding_window=self.config.sliding_window, + ) + + hidden_states = inputs_embeds + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = None + + for decoder_layer in self.layers: + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.gradient_checkpointing and self.training: + layer_outputs = self._gradient_checkpointing_func( + decoder_layer.__call__, + hidden_states, + attention_mask, + position_ids, + past_key_values, + output_attentions, + use_cache, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_values, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache = layer_outputs[2 if output_attentions else 1] + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + hidden_states = self.norm(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = None + if use_cache: + next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache + if not return_dict: + return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + +class Phi3ForCausalLM(Phi3PreTrainedModel): + _tied_weights_keys = ['lm_head.weight'] + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3 + def __init__(self, config): + super().__init__(config) + self.model = Phi3Model(config) + self.vocab_size = config.vocab_size + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings + def get_input_embeddings(self): + return self.model.embed_tokens + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings + def get_output_embeddings(self): + return self.lm_head + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder + def set_decoder(self, decoder): + self.model = decoder + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder + def get_decoder(self): + return self.model + + # Ignore copy + @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + Args: + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, Phi3ForCausalLM + + >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct") + >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct") + + >>> prompt = "This is an example script ." + >>> inputs = tokenizer(prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(inputs.input_ids, max_length=30) + >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + 'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum' + ```""" + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + outputs = self.model( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + hidden_states = outputs[0] + logits = self.lm_head(hidden_states) + logits = logits.float() + + loss = None + if labels is not None: + # Shift so that tokens < n predict n + shift_logits = logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss() + shift_logits = shift_logits.view(-1, self.config.vocab_size) + shift_labels = shift_labels.view(-1) + # Enable model parallelism + shift_labels = shift_labels.to(shift_logits.device) + loss = loss_fct(shift_logits, shift_labels) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation + def prepare_inputs_for_generation( + self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + ): + if past_key_values is not None: + if isinstance(past_key_values, Cache): + cache_length = past_key_values.get_seq_length() + past_length = past_key_values.seen_tokens + max_cache_length = past_key_values.get_max_length() + else: + cache_length = past_length = past_key_values[0][0].shape[2] + max_cache_length = None + + # Keep only the unprocessed tokens: + # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where + # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as + # input) + if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: + input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] + # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard + # input_ids based on the past_length. + elif past_length < input_ids.shape[1]: + input_ids = input_ids[:, past_length:] + # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. + + # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. + if ( + max_cache_length is not None + and attention_mask is not None + and cache_length + input_ids.shape[1] > max_cache_length + ): + attention_mask = attention_mask[:, -max_cache_length:] + + position_ids = kwargs.get('position_ids', None) + if attention_mask is not None and position_ids is None: + # create position_ids on the fly for batch generation + position_ids = attention_mask.long().cumsum(-1) - 1 + position_ids.masked_fill_(attention_mask == 0, 1) + if past_key_values: + position_ids = position_ids[:, -input_ids.shape[1] :] + + # if `inputs_embeds` are passed, we only want to use them in the 1st generation step + if (inputs_embeds is not None and past_key_values is None) or (inputs_embeds is not None and len(past_key_values) == 0): + model_inputs = {'inputs_embeds': inputs_embeds} + else: + model_inputs = {'input_ids': input_ids} + + model_inputs.update( + { + 'position_ids': position_ids, + 'past_key_values': past_key_values, + 'use_cache': kwargs.get('use_cache'), + 'attention_mask': attention_mask, + } + ) + return model_inputs + + @staticmethod + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache + def _reorder_cache(past_key_values, beam_idx): + reordered_past = () + for layer_past in past_key_values: + reordered_past += ( + tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past), + ) + return reordered_past + + +@add_start_docstrings( + """ + The [`Phi3Model`] with a sequence classification head on top (linear layer). + + [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-2) do. + + Since it does classification on the last token, it requires to know the position of the last token. If a + `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If + no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the + padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in + each row of the batch). + """, + PHI3_START_DOCSTRING, +) +# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs +class Phi3ForSequenceClassification(Phi3PreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Phi3Model(config) + self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, SequenceClassifierOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + model_outputs = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = model_outputs[0] + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size = input_ids.shape[0] + else: + batch_size = inputs_embeds.shape[0] + + if self.config.pad_token_id is None and batch_size != 1: + raise ValueError('Cannot handle batch sizes > 1 if no padding token is defined.') + if self.config.pad_token_id is None: + sequence_lengths = -1 + else: + if input_ids is not None: + # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility + sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 + sequence_lengths = sequence_lengths % input_ids.shape[-1] + sequence_lengths = sequence_lengths.to(logits.device) + else: + sequence_lengths = -1 + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] + + loss = None + if labels is not None: + labels = labels.to(logits.device) + if self.config.problem_type is None: + if self.num_labels == 1: + self.config.problem_type = 'regression' + elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int): + self.config.problem_type = 'single_label_classification' + else: + self.config.problem_type = 'multi_label_classification' + + if self.config.problem_type == 'regression': + loss_fct = MSELoss() + if self.num_labels == 1: + loss = loss_fct(pooled_logits.squeeze(), labels.squeeze()) + else: + loss = loss_fct(pooled_logits, labels) + elif self.config.problem_type == 'single_label_classification': + loss_fct = CrossEntropyLoss() + loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1)) + elif self.config.problem_type == 'multi_label_classification': + loss_fct = BCEWithLogitsLoss() + loss = loss_fct(pooled_logits, labels) + if not return_dict: + output = (pooled_logits,) + model_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=model_outputs.past_key_values, + hidden_states=model_outputs.hidden_states, + attentions=model_outputs.attentions, + ) + + +@add_start_docstrings( + """ + [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for + Named-Entity-Recognition (NER) tasks. + """, + PHI3_START_DOCSTRING, +) +# Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs +class Phi3ForTokenClassification(Phi3PreTrainedModel): + def __init__(self, config: Phi3Config): + super().__init__(config) + self.num_labels = config.num_labels + + self.model = Phi3Model(config) + if hasattr(config, 'classifier_dropout') and config.classifier_dropout is not None: + classifier_dropout = config.classifier_dropout + elif hasattr(config, 'hidden_dropout') and config.hidden_dropout is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.classifier = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=TokenClassifierOutput, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, + attention_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + **deprecated_arguments, + ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + model_outputs = self.model( + input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + hidden_states = model_outputs[0] + hidden_states = self.dropout(hidden_states) + logits = self.classifier(hidden_states) + + loss = None + if labels is not None: + # move labels to correct device to enable model parallelism + labels = labels.to(logits.device) + batch_size, seq_length = labels.shape + loss_fct = CrossEntropyLoss() + loss = loss_fct( + logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length) + ) + + if not return_dict: + output = (logits,) + model_outputs[2:] + return ((loss,) + output) if loss is not None else output + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=model_outputs.hidden_states, + attentions=model_outputs.attentions, + ) diff --git a/preprocessor_config.json b/preprocessor_config.json new file mode 100644 index 0000000..dfd7e50 --- /dev/null +++ b/preprocessor_config.json @@ -0,0 +1,19 @@ +{ + "crop_size": 448, + "do_center_crop": true, + "do_normalize": true, + "do_resize": true, + "feature_extractor_type": "CLIPFeatureExtractor", + "image_mean": [ + 0.485, + 0.456, + 0.406 + ], + "image_std": [ + 0.229, + 0.224, + 0.225 + ], + "resample": 3, + "size": 448 +} diff --git a/special_tokens_map.json b/special_tokens_map.json new file mode 100644 index 0000000..7b24044 --- /dev/null +++ b/special_tokens_map.json @@ -0,0 +1,41 @@ +{ + "additional_special_tokens": [ + "", + "", + "", + "", + "", + "", + "", + "", + "" + ], + "bos_token": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false + }, + "eos_token": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false + }, + "pad_token": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false + }, + "unk_token": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false + } +} diff --git a/tokenizer.model b/tokenizer.model new file mode 100644 index 0000000..6c00c74 --- /dev/null +++ b/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347 +size 499723 diff --git a/tokenizer_config.json b/tokenizer_config.json new file mode 100644 index 0000000..152c3ac --- /dev/null +++ b/tokenizer_config.json @@ -0,0 +1,214 @@ +{ + "add_bos_token": true, + "add_eos_token": false, + "added_tokens_decoder": { + "0": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "1": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "2": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32000": { + "content": "<|endoftext|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32001": { + "content": "<|assistant|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32002": { + "content": "<|placeholder1|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32003": { + "content": "<|placeholder2|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32004": { + "content": "<|placeholder3|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32005": { + "content": "<|placeholder4|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32006": { + "content": "<|system|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32007": { + "content": "<|end|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32008": { + "content": "<|placeholder5|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32009": { + "content": "<|placeholder6|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32010": { + "content": "<|user|>", + "lstrip": false, + "normalized": false, + "rstrip": true, + "single_word": false, + "special": true + }, + "32011": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32012": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32013": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32014": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32015": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32016": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32017": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32018": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "32019": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + } + }, + "additional_special_tokens": [ + "", + "", + "", + "", + "", + "", + "", + "", + "" + ], + "bos_token": "", + "chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}", + "clean_up_tokenization_spaces": false, + "eos_token": "", + "legacy": false, + "model_max_length": 8192, + "pad_token": "", + "padding_side": "right", + "sp_model_kwargs": {}, + "spaces_between_special_tokens": false, + "tokenizer_class": "LlamaTokenizer", + "unk_token": "", + "use_default_system_prompt": false +}