From 1d0ac1e1c25762f3f7eaef197ec00ceb2153cd68 Mon Sep 17 00:00:00 2001 From: gitee_code_template Date: Thu, 28 Mar 2024 14:07:33 +0800 Subject: [PATCH 1/3] =?UTF-8?q?=E6=96=B0=E5=A2=9Ecogvlm=E6=A8=A1=E5=9E=8B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- PyTorch/built-in/foundation/CogVLM/README.md | 190 +++++ .../foundation/CogVLM/cogvlm_utils/dataset.py | 62 ++ .../foundation/CogVLM/cogvlm_utils/env_npu.sh | 55 ++ .../CogVLM/cogvlm_utils/eva_clip_model.py | 175 ++++ .../cogvlm_utils/eval_cogvlm_base_224.sh | 53 ++ .../cogvlm_utils/evaluate_cogvlm_demo.py | 223 +++++ .../cogvlm_utils/finetune_cogvlm_base_224.sh | 57 ++ .../cogvlm_utils/finetune_cogvlm_demo.py | 281 ++++++ .../CogVLM/cogvlm_utils/inference.py | 33 + .../CogVLM/cogvlm_utils/layernorm.py | 40 + .../foundation/CogVLM/cogvlm_utils/mixin.py | 275 ++++++ .../CogVLM/cogvlm_utils/modeling_cogvlm.py | 806 ++++++++++++++++++ .../CogVLM/cogvlm_utils/rotary_embeddings.py | 136 +++ .../cogvlm_utils/transformer_defaults.py | 270 ++++++ .../cogvlm_utils/triton_rotary_embeddings.py | 15 + .../foundation/CogVLM/cogvlm_utils/visual.py | 169 ++++ .../CogVLM/fintune_cogvlm_change.sh | 133 +++ .../CogVLM/inference_cogvlm_change.sh | 108 +++ .../foundation/CogVLM/requirements.txt | 24 + 19 files changed, 3105 insertions(+) create mode 100644 PyTorch/built-in/foundation/CogVLM/README.md create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/dataset.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/env_npu.sh create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eva_clip_model.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eval_cogvlm_base_224.sh create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/evaluate_cogvlm_demo.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_demo.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/inference.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/layernorm.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/mixin.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/modeling_cogvlm.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/rotary_embeddings.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/transformer_defaults.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/triton_rotary_embeddings.py create mode 100644 PyTorch/built-in/foundation/CogVLM/cogvlm_utils/visual.py create mode 100644 PyTorch/built-in/foundation/CogVLM/fintune_cogvlm_change.sh create mode 100644 PyTorch/built-in/foundation/CogVLM/inference_cogvlm_change.sh create mode 100644 PyTorch/built-in/foundation/CogVLM/requirements.txt diff --git a/PyTorch/built-in/foundation/CogVLM/README.md b/PyTorch/built-in/foundation/CogVLM/README.md new file mode 100644 index 0000000000..e570216f16 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/README.md @@ -0,0 +1,190 @@ +# CogVLM for Pytorch +# 目录 + +- [简介](#简介) + - [模型介绍](#模型介绍) + - [支持任务列表](#支持任务列表) + - [代码实现](#代码实现) +- [CogVLM](#CogVLM) + - [准备训练环境](#准备训练环境) + - [准备数据集](#准备数据集) + - [快速开始](#快速开始) + - [微调任务](#微调任务) + - [推理任务](#推理任务) +- [公网地址变更说明](#公网地址变更说明) +- [变更说明](#变更说明) +- [FAQ](#FAQ) + + + +# 简介 +## 模型介绍 +CogVLM is VISUAL EXPERT FOR LARGE LANGUAGE 是一个多模态视觉-文本模型，它强调“视觉优先”，使用11B参数建模图像特征，多于文本的7B参数量。该模型包含ViT编码器、MLP适配器、预训练大语言模型和视觉专家模块，通过深度整合语言和视觉信息，提升了跨模态任务的性能。在多个基准测试中，CogVLM展现出领先或次领先的性能，显示出其在视觉理解研究和工业应用中的巨大潜力。 + +官方仓：https://github.com/THUDM/CogVLM + +## 支持任务列表 +本仓已支持以下模型任务类型。 + +| 模型 | 模型大小 | 任务类型 | 是否支持 | +|------------|------|------------| ------------ | +| CogVLM | base-224 | 微调 | ✅ | + +## 代码实现 +- 参考实现 + ``` + CogVLM仓: https://github.com/THUDM/CogVLM + commit id: eb2367f54b95da2ee64f996305ab1baa45df7479 + ``` +- 适配昇腾 AI 处理器的实现： + ``` + url=https://gitee.com/ascend/ModelZoo-PyTorch.git + code_path=PyTorch/built-in/foundation/CogVLM + ``` + +# CogVLM + +## 准备训练环境 +### 安装昇腾环境 + +请参考昇腾社区中《[Pytorch框架训练环境准备](https://www.hiascend.com/document/detail/zh/ModelZoo/pytorchframework/ptes)》文档搭建昇腾环境，本仓已支持表1中软件版本。 + +**表 1** 昇腾软件版本支持表 + +| 软件类型 | 支持版本 | +|:------------------:|:--------:| +| FrameworkPTAdapter | 6.0.RC1 | +| CANN | 8.0.RC1 | +| 昇腾NPU固件 | 24.1.RC1 | +| 昇腾NPU驱动 | 24.1.RC1 | + +### 安装模型环境 + +**表 2** 三方库版本支持表 + +| 三方库 | 支持版本 | +|:---------:|:------:| +| PyTorch | 2.1.0 | + +在模型根目录下执行以下命令，安装模型对应PyTorch版本需要的依赖, 需要先安装PTA包。 +```shell +pip install -r requirements.txt +``` +### 准备数据集 + +1) 微调数据集: +训练与评估所使用的数据集为Captcha Images dataset(验证码数据集)[下载](https://www.kaggle.com/datasets/aadhavvignesh/captcha-images)，该数据集是官网提供的一个预训练数据集。 + +2) label数据: +数据的label信息为图像的文件名。 + +下载完后文件夹结构如下所示： + +```text +archive +├── 004rVO6G09.jpg +├── 00949IT0LT.jpg +├── 00bAQwhAZU.jpg +├── 01S19jY65H.jpg +... +``` +#### 微调数据预处理 +数据下载完成后，需要对其进行数据集划分，train/validation/test的划分比例为80/5/15，在官网utils/split_dataset.py中指定源文件路径，如下面代码中的"archive"路径： +```python +all_files = find_all_files('archive') +``` +执行数据划分操作，如下命令： +```shell[dataset.py](cogvlm_utils%2Fdataset.py) +python utils/split_dataset.py +``` +划分后会生成train/valid/test文件，文件中分别包含划分后的图像。 +```text +archive_split +├── test +├── train +├── valid +``` + +### 获取预训练权重 + +1) 下载并安装en_core_web_sm-any-py3-none-any.whl[下载](https://huggingface.co/spacy/en_core_web_sm/tree/main)， +en_core_web_sm是spaCy 自然语言处理（NLP）工具库中的一种语言模型，专为英语设计。 + +2) 官方提供微调权重cogvlm-base-224[下载](https://huggingface.co/THUDM/CogVLM/tree/main)。 + +3) 分词器权重[下载](https://huggingface.co/lmsys/vicuna-7b-v1.5/tree/main)。 + +## 快速开始 + +### 微调任务 +主要提供基于Captcha Images数据集微调的8卡训练脚本。 +#### 模型适配 + +1) 模型文件替换: + +替换fintune_cogvlm_change.sh脚本中开头三个路径，model_zoo路径、Cogvlm路径和sat路径为实际路径，执行脚本进行替换，如下命令： +```shell +sh fintune_cogvlm_change.sh +``` +#### 执行微调 +1) finetune_demo/finetune_cogvlm_base_224.sh文件夹下，修改文件中微调权重路径、分词器权重路径和数据集路径（train_data和valid_data）为实际路径。 + +2) 执行训练，如下命令： +``` +cd finetune_demo +bash finetune_cogvlm_base_224.sh +``` + +#### 训练结果 + +##### 精度 + +基于Captcha Images数据集训练800步、1600步和2000步验证下游任务，由于模型本身有确定性问题，因此下游任务在评估数据上略有波动： + +| 芯片 | 800 | 1600 | 2000 | +|:--------:|:-----:|--------|-------| +| GPU | 95% | 96.25% | 97.5% | +| Atlas A2 | 95% | 95% | 97.5% | + +##### 性能 + + +| 芯片 | 卡数 | samples/s | batch_size | AMP_Type | Torch_Version | +|:--------:| :----: |:---------:|:----------:|:--------:| :-----------: | +| GPU | 8p | 340 | 4 | bfp16 | 2.1 | +| Atlas A2 | 8p | 280 | 4 | bfp16 | 2.1 | + +#### 微调后推理 +1) finetune_demo/eval_cogvlm_base_224.sh文件夹下，修改文件中微调后权重路径、分词器路径和数据集路径(test_data)为实际路径。 +2) 执行推理，如下命令： +```shell +bash eval_cogvlm_base_224.sh +``` + +### 推理任务 +该处的推理为官方hf权重的推理方式，与微调后推理方式略有区别，用户可根据实际情况判断是否执行。 +#### 推理前准备 +1) 预训练权重cogvlm-base-224-hf[下载](https://huggingface.co/THUDM/cogvlm-base-224-hf)。 + +2) 替换inference_cogvlm_change.sh脚本中model_zoo路径、Cogvlm路径和HF权重路径为实际路径,执行脚本进行替换。 +```shell +sh inference_cogvlm_change.sh +``` +3) finetune_demo/inference.py文件夹下，并根据实际路径修改推理权重路径、分词器权重路径和图片路径 +#### 启动推理 +```shell +cd finetune_demo +source ./env_npu.sh +python inference.py +``` + +# 公网地址变更说明 +暂无。 + +# 变更说明 +暂无。 + + +# FAQ + +暂无。 \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/dataset.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/dataset.py new file mode 100644 index 0000000000..d3e3a9107e --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/dataset.py @@ -0,0 +1,62 @@ +import os +import logging +import random +import logging +import jsonlines +from io import BytesIO +from PIL import Image +from torch.utils.data import Dataset +from sat.helpers import print_rank0 + +def find_all_files(path, suffix=".jpg"): + target_files = [] + for cur_dir, _, files in os.walk(path, followlinks=True): + for f in files: + if f.endswith(suffix): + target_files.append(os.path.join(cur_dir, f)) + print_rank0(f'find {len(target_files)} files...') + return target_files + +class ItemDataset(Dataset): + def __init__(self, image_processor, text_processor, args, data_dirs, cross_image_processor=None, **kwargs): + super().__init__() + self.data = self.load_data(data_dirs) + self.image_processor, self.text_processor, self.cross_image_processor = image_processor, text_processor, cross_image_processor + + def process_img(self, img): + img_dict = {'vision': self.image_processor(img)} + if self.cross_image_processor: + img_dict.update({'cross': self.cross_image_processor(img)}) + return img_dict + + def process_text(self, answer, prompt): + return self.text_processor(answer, prompt) + + def load_data(self, data_dir): + all_files = find_all_files(data_dir, suffix=".jpg") + all_files.sort() + print_rank0(f"find {len(all_files)} samples in all...") + return all_files + + def __len__(self): + return len(self.data) + + def __getitem__(self, index): + data = self.data[index] + # img + try: + img = Image.open(data).convert('RGB') + except Exception as e: + print_rank0(e, level=logging.WARNING) + return {} + img_dict = self.process_img(img) + # text + label = data.split('/')[-1].split('.')[0] + uni_key = label + text_dict = self.process_text(label, "CAPTCHA:") + if text_dict is None: + print_rank0(f"Process text failed. Please check the max_target_length & max_source_length.\n The data is {data}", level=logging.WARNING) + return {} + # other attr + ret = {**img_dict, **text_dict, "question_id": uni_key} + return ret \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/env_npu.sh b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/env_npu.sh new file mode 100644 index 0000000000..11c1a42d04 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/env_npu.sh @@ -0,0 +1,55 @@ +#!/bin/bash +CANN_INSTALL_PATH_CONF='/etc/Ascend/ascend_cann_install.info' + +if [ -f $CANN_INSTALL_PATH_CONF ]; then + CANN_INSTALL_PATH=$(cat $CANN_INSTALL_PATH_CONF | grep Install_Path | cut -d "=" -f 2) +else + CANN_INSTALL_PATH="/usr/local/Ascend" +fi + +if [ -d ${CANN_INSTALL_PATH}/ascend-toolkit/latest ]; then + source ${CANN_INSTALL_PATH}/ascend-toolkit/set_env.sh +else + source ${CANN_INSTALL_PATH}/nnae/set_env.sh +fi + +count=$(npu-smi info -l | grep -c "NPU ID") + +for ((i=0; i<${count}; i=i+1)) +do + msnpureport -g error -d ${i} +done + +#将Host日志输出到串口,0-关闭/1-开启 +export ASCEND_SLOG_PRINT_TO_STDOUT=0 +#设置默认日志级别,0-debug/1-info/2-warning/3-error +export ASCEND_GLOBAL_LOG_LEVEL=3 +#设置Event日志开启标志,0-关闭/1-开启 +export ASCEND_GLOBAL_EVENT_ENABLE=0 +#设置是否开启taskque,0-关闭/1-开启 +export TASK_QUEUE_ENABLE=1 +#设置是否开启combined标志,0-关闭/1-开启 +export COMBINED_ENABLE=1 +#HCCL白名单开关,1-关闭/0-开启 +export HCCL_WHITELIST_DISABLE=1 +export HCCL_IF_IP=$(hostname -I |awk '{print $1}') +export HCCL_CONNECT_TIMEOUT=1200 + +path_lib=$(python3 -c """ +import sys +import re +result='' +for index in range(len(sys.path)): + match_sit = re.search('-packages', sys.path[index]) + if match_sit is not None: + match_lib = re.search('lib', sys.path[index]) + + if match_lib is not None: + end=match_lib.span()[1] + result += sys.path[index][0:end] + ':' + + result+=sys.path[index] + '/torch/lib:' +print(result)""" +) + +echo ${path_lib} \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eva_clip_model.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eva_clip_model.py new file mode 100644 index 0000000000..e6f5eabee7 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eva_clip_model.py @@ -0,0 +1,175 @@ +import torch +import torch_npu +import torch.nn.functional as F +from sat.model.base_model import BaseModel +from sat.model.mixins import BaseMixin +from sat.model.official.vit_model import ViTProperty, ImagePatchEmbeddingMixin, InterpolatedPositionEmbeddingMixin, gelu +from sat import mpu + + +class FlashSelfAttention(torch.nn.Module): + """Implement the scaled dot product attention with softmax. + Arguments + --------- + softmax_scale: The temperature to use for the softmax attention. + (default: 1/sqrt(d_keys) where d_keys is computed at + runtime) + attention_dropout: The dropout rate to apply to the attention + (default: 0.0) + """ + + def __init__(self, causal=False, softmax_scale=1., attention_dropout=0.): + super().__init__() + self.causal = causal + self.softmax_scale = softmax_scale + self.dropout_p = attention_dropout + + def forward(self, q, k, v, n, attention_mask, pse): + + if self.causal: + output = torch_npu.npu_fusion_attention( + q, k, v, n, "BSND",# SBH + pse=pse, + padding_mask=None, + atten_mask=attention_mask, + scale=self.softmax_scale, + pre_tockens=65536, # seq_len + next_tockens=65536, # 0 + keep_prob=1 - self.dropout_p, + )[0] + return output + + +class IdentityMixin(BaseMixin): + def __init__(self): + super().__init__() + + def final_forward(self, logits, **kwargs): + return logits[:, 1:] + + +def memory_efficient_attention_min(query, key, value, dropout): + """ + xformers.ops.memory_efficient_attention 小算子实现 + """ + scale = 1.0 / query.shape[-1] ** 0.5 + query = query * scale + query = query.transpose(1, 2) + key = key.transpose(1, 2) + value = value.transpose(1, 2) + attn = query @ key.transpose(-2, -1) + attn = attn.softmax(-1) + attn = F.dropout(attn, dropout) + attn = attn @ value + return attn.transpose(1, 2) + + +class XAttn(BaseMixin): + def __init__(self, head_dim): + super().__init__() + self.scale = head_dim ** -0.5 + self.core_attention_flash = FlashSelfAttention(causal=True, softmax_scale=self.scale, attention_dropout=0.) + + def attention_fn(self, query_layer, key_layer, value_layer, attention_mask, + attention_dropout=None, log_attention_weights=None, scaling_attention_score=True, **kwargs): + dropout_p = 0. # xformers does not support dropout for eva hidden size + + query_layer = query_layer.permute(0, 2, 1, 3) # B, num_heads, N, C -> B, N, num_heads, C + key_layer = key_layer.permute(0, 2, 1, 3) + value_layer = value_layer.permute(0, 2, 1, 3) + out = self.core_attention_flash(query_layer, key_layer, value_layer, query_layer.shape[2], None, None) + return out + + def attention_forward(self, hidden_states, mask, **kw_args): + self = self.transformer.layers[kw_args['layer_id']].attention + attention_fn = self.hooks['attention_fn'] + + mixed_raw_layer = self.query_key_value(hidden_states) + + B, N, C = hidden_states.shape + mixed_raw_layer = mixed_raw_layer.reshape(B, N, 3, self.num_attention_heads_per_partition, -1).permute(2, 0, 3, 1, 4) # 3, B, num_heads, N, C + query_layer, key_layer, value_layer = mixed_raw_layer[0], mixed_raw_layer[1], mixed_raw_layer[2] + + dropout_fn = self.attention_dropout if self.training else None + + context_layer = attention_fn(query_layer, key_layer, value_layer, mask, dropout_fn, **kw_args) + + context_layer = context_layer.contiguous().view(B, N, -1) + output = self.dense(context_layer) + + if self.training: + output = self.output_dropout(output) + return output + + +class NewLayerForward(BaseMixin): + def __init__(self): + super().__init__() + + def layer_forward(self, hidden_states, mask, *args, **kw_args): + ''' + hidden_states: [batch, seq_len, hidden_size] + mask: [(1, 1), seq_len, seq_len] + ''' + self = self.transformer.layers[kw_args['layer_id']] + + attention_input = hidden_states + + # Self attention. + attention_output = self.input_layernorm(self.attention(attention_input, mask, **kw_args)) + + # DropPath for attention + if self.training and self.drop_path > 0.: + if mpu.get_cuda_rng_tracker is not None: + # drop_path must use model parallel rng tracker + # the tracker is initialized as seed of `seed + model_parallel_rank` + # deepspeed act-ckpt record the model parallel tracker states + with mpu.get_cuda_rng_tracker().fork(): + # drop_path percentage 0, others 1/(1-p) + random_tensor = (1-self.drop_path + + torch.rand((attention_output.shape[0],), dtype=attention_output.dtype, device=attention_output.device)).floor_() / (1-self.drop_path) + attention_output = random_tensor.view(-1, 1, 1) * attention_output + + # Residual connection. + hidden_states = attention_input + attention_output + mlp_input = hidden_states + + # MLP. + mlp_output = self.post_attention_layernorm(self.mlp(mlp_input, **kw_args)) + + # DropPath for mlp + if self.training and self.drop_path > 0.: + if mpu.get_cuda_rng_tracker is not None: + with mpu.get_cuda_rng_tracker().fork(): + random_tensor = (1-self.drop_path + + torch.rand((mlp_output.shape[0],), dtype=mlp_output.dtype, device=mlp_output.device)).floor_() / (1-self.drop_path) + mlp_output = random_tensor.view(-1, 1, 1) * mlp_output + + # Second residual connection. + output = mlp_input + mlp_output + + return output + +class EVA2CLIPModel(BaseModel): + def __init__(self, args, transformer=None, parallel_output=True, **kwargs): + property = ViTProperty(args.image_size, args.patch_size, args.pre_len, args.post_len) + args.max_sequence_length = property.pre_len + property.num_patches + property.post_len + if 'activation_func' not in kwargs: + kwargs['activation_func'] = gelu + super().__init__(args, transformer=transformer, parallel_output=parallel_output, **kwargs) + self.transformer.property = property + self.add_mixin("patch_embedding", ImagePatchEmbeddingMixin(args.in_channels, args.hidden_size, property)) + self.add_mixin("pos_embedding", InterpolatedPositionEmbeddingMixin()) + self.add_mixin("final", IdentityMixin()) + self.add_mixin("newpost", NewLayerForward()) + self.add_mixin("xattn", XAttn(args.hidden_size // args.num_attention_heads)) + + @classmethod + def add_model_specific_args(cls, parser): + group = parser.add_argument_group('EVA2CLIP', 'EVA2CLIP Configurations') + group.add_argument('--image-size', nargs='+', type=int, default=[224, 224]) + group.add_argument('--pre-len', type=int, default=1) # [cls] by default + group.add_argument('--post-len', type=int, default=0) # empty by default, but sometimes with special tokens, such as [det] in yolos. + group.add_argument('--in-channels', type=int, default=3) + group.add_argument('--patch-size', type=int, default=16) + return parser diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eval_cogvlm_base_224.sh b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eval_cogvlm_base_224.sh new file mode 100644 index 0000000000..ec07f64abd --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/eval_cogvlm_base_224.sh @@ -0,0 +1,53 @@ +#! /bin/bash +# export PATH=/usr/local/cuda/bin:$PATH +# export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH + +NUM_GPUS_PER_WORKER=8 +MP_SIZE=1 + +script_path=$(realpath $0) +script_dir=$(dirname $script_path) +main_dir=$(dirname $script_dir) +MODEL_TYPE="微调后权重路径" +VERSION="base" +MODEL_ARGS="--from_pretrained $MODEL_TYPE \ + --max_length 490 \ + --local_tokenizer 分词器权重路径 \ + --version $VERSION" +# Tips: If training models of resolution 244, you can set --max_length smaller + + +OPTIONS_SAT="SAT_HOME=~/.sat_models" +OPTIONS_NCCL="NCCL_DEBUG=info NCCL_IB_DISABLE=0 NCCL_NET_GDR_LEVEL=2 LOCAL_WORLD_SIZE=$NUM_GPUS_PER_WORKER" +HOST_FILE_PATH="hostfile" + +test_data="./archive_split/test" + +gpt_options=" \ + --experiment-name finetune-$MODEL_TYPE \ + --model-parallel-size ${MP_SIZE} \ + --mode finetune \ + --train-iters 0 \ + --resume-dataloader \ + $MODEL_ARGS \ + --train-data ${train_data} \ + --test-data ${test_data} \ + --distributed-backend nccl \ + --lr-decay-style cosine \ + --warmup .02 \ + --checkpoint-activations \ + --strict-eval \ + --eval-batch-size 1 \ + --split 1. \ + --deepspeed_config test_config_bf16.json \ + --skip-init \ + --seed 1234 +" + + + +run_cmd="${OPTIONS_NCCL} ${OPTIONS_SAT} deepspeed --master_port 18888 --hostfile ${HOST_FILE_PATH} evaluate_cogvlm_demo.py ${gpt_options}" +echo ${run_cmd} +eval ${run_cmd} + +set +x \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/evaluate_cogvlm_demo.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/evaluate_cogvlm_demo.py new file mode 100644 index 0000000000..0966c6e149 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/evaluate_cogvlm_demo.py @@ -0,0 +1,223 @@ +import os +import torch +import torch_npu +from torch_npu.contrib import transfer_to_npu +import argparse +from functools import partial +import sys +sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) + +from sat import mpu, get_args, get_tokenizer +from sat.training.deepspeed_training import training_main +from sat.helpers import print_rank0 +from utils.models import FineTuneTestCogVLMModel +from utils.utils import llama2_text_processor, llama2_text_processor_inference, get_image_processor + + +def data_collator(examples): + examples = [ex for ex in examples if len(ex) > 0] # drop {} + for example in examples: + for k in example: + if isinstance(example[k], list): + example[k] = torch.tensor(example[k]) + elif isinstance(example[k], np.ndarray): + example[k] = torch.from_numpy(example[k]) + img_args = {} + tmp_example = examples[0] + for k in tmp_example['vision']: + if type(tmp_example['vision'][k]) is torch.Tensor: + img_args['vision_'+k] = torch.cat([example['vision'][k] for example in examples]) + else: + img_args['vision_'+k] = example['vision'][k] + for example in examples: + example.pop('vision') + if 'cross' in example: + example.pop('cross') + + model_args = {} + tmp_example = examples[0] + for k in tmp_example: + if type(tmp_example[k]) is torch.Tensor: + model_args[k] = torch.cat([example[k] for example in examples]) + else: + model_args[k] = tmp_example[k] + model_args.update(img_args) + return model_args + +from collections import defaultdict + +def broadcast_auto(data_dict): + type2list = defaultdict(list) + other = [] + for k in data_dict: + if type(data_dict[k]) is torch.Tensor: + type2list[data_dict[k].dtype].append(k) + else: + other.append(k) + new_data = {} + for k in type2list: + new_data.update(mpu.broadcast_data(type2list[k], data_dict, k)) + for k in other: + new_data[k] = data_dict[k] + return new_data + +def get_batch(data_iterator, args, timers): + # Broadcast data. + timers('data loader').start() + if data_iterator is not None: + data = next(data_iterator) + else: + data = None + timers('data loader').stop() + data_b = broadcast_auto(data) + for k in data_b: + if type(data_b[k]) is torch.Tensor and data_b[k].dtype is not torch.int32 and data_b[k].dtype is not torch.long: + if args.fp16: + data_b[k] = data_b[k].half() + elif args.bf16: + data_b[k] = data_b[k].bfloat16() + return data_b + +from torch.nn import CrossEntropyLoss +import numpy as np + +from sat.model.mixins import CachedAutoregressiveMixin +from sat.generation.autoregressive_sampling import filling_sequence +from sat.generation.sampling_strategies import BaseStrategy, BeamSearchStrategy + + +def chat(model, tokenizer, tokens, + max_length: int = 1800, num_beams=5, top_p=0.95, top_k=0, temperature=0.8, **kwargs): + inputs = tokens.to(model.parameters().__next__().device)[0] + seq = torch.cat( + [inputs, torch.tensor([-1] * (max_length - len(inputs)), device=inputs.device)], dim=0 + ) + strategy = BaseStrategy(temperature=temperature, top_p=0.4, top_k=1, end_tokens=[tokenizer.eos_token_id]) + # strategy = BeamSearchStrategy(temperature=temperature, top_p=top_p, top_k=top_k, end_tokens=[tokenizer.eos_token_id], + # num_beams=num_beams, consider_end=True) + get_func = llama2_text_processor_inference.get_func(None, None, image_rope_mask=kwargs['image_rope_mask']) + output = filling_sequence( + model, seq, + batch_size=1, + strategy=strategy, + get_masks_and_position_ids=get_func, + **kwargs + )[0] # drop memory + + return output + + +def forward_step_eval(data_iterator, model, args, timers): + def compute_metrics(eval_preds): + preds, labels, device = eval_preds + preds = preds.unsqueeze(0) + if isinstance(preds, tuple): + preds = preds[0] + decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True) + if args.ignore_pad_token_for_loss: + # Replace -100 in the labels as we can't decode them. + labels = np.where(labels != -100, labels, tokenizer.pad_token_id) + decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True) + + score_dict = { + "acc": [], + "acc_w/o_case": [], + } + for pred, label in zip(decoded_preds, decoded_labels): + if args.rank == 0: + print('pred', pred, 'label', label, flush=True) + if pred == label: + score_dict['acc'].append(1.) + else: + score_dict['acc'].append(0.) + if pred.lower() == label.lower(): + score_dict['acc_w/o_case'].append(1.) + else: + score_dict['acc_w/o_case'].append(0.) + + + for k, v in score_dict.items(): + score_dict[k] = float(np.mean(v)) + return score_dict + + # Get the batch. + timers('batch generator').start() + data_b = get_batch( + data_iterator, args, timers) + timers('batch generator').stop() + + context_len = int(data_b['context_length'][0]) + tokens = data_b['input_ids'][:, :context_len] + data_b['vision_expert_mask'] = data_b['vision_expert_mask'][:, :context_len] + data_b['image_embed_mask'] = data_b['image_embed_mask'][:, :context_len] + data_b['image_rope_mask'] = data_b['image_rope_mask'][:, :context_len] + + data_b.pop('input_ids') + data_b.pop('attention_mask') + data_b.pop('position_ids') + labels = data_b.pop('labels') + qid = data_b.pop('question_id') + + model.add_mixin('auto-regressive', CachedAutoregressiveMixin()) + outputs = chat(model, tokenizer, tokens, **data_b)[0][context_len:] + # print(outputs) + model.del_mixin('auto-regressive') + + return torch.tensor(0, device=outputs.device), {k: torch.tensor(v, device=outputs.device) for k, v in + compute_metrics( + (outputs.cpu(), labels.cpu(), outputs.device)).items()} + + +from torch.nn import CrossEntropyLoss +def forward_step(data_iterator, model, args, timers): + """Forward step.""" + + # Get the batch. + timers('batch generator').start() + data_b = get_batch( + data_iterator, args, timers) + labels = data_b.pop('labels') + timers('batch generator').stop() + logits = model(**data_b)[0] + lm_logits = logits.to(torch.float32) + # Shift so that tokens < n predict n + shift_labels = labels[..., 1:].contiguous() + shift_logits = lm_logits[..., -1-shift_labels.size(-1):-1, :].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss(ignore_index=-100) + loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) + loss = loss.to(torch.float32) + + return loss, {'loss': loss} + +from utils.utils import ItemDataset +def create_dataset_function(image_processor, text_processor, path, args): + dataset = ItemDataset(image_processor, text_processor, args, path) + return dataset + +if __name__ == '__main__': + torch.npu.set_compile_mode(jit_compile=False) + torch.npu.config.allow_internal_format = False + py_parser = argparse.ArgumentParser(add_help=False) + py_parser.add_argument('--max_length', type=int) + py_parser.add_argument('--ignore_pad_token_for_loss', action='store_false') + py_parser.add_argument("--version", type=str, default="chat", help='version to interact with') + py_parser.add_argument("--from_pretrained", type=str, default="cogvlm-chat", help='pretrained ckpt') + py_parser.add_argument("--local_tokenizer", type=str, default="lmsys/vicuna-7b-v1.5", help='tokenizer path') + py_parser.add_argument("--vit_checkpoint_activations", action='store_true') + py_parser = FineTuneTestCogVLMModel.add_model_specific_args(py_parser) + known, args_list = py_parser.parse_known_args() + args = get_args(args_list) + args = argparse.Namespace(**vars(args), **vars(known)) + if args.use_qlora: + args.device = 'cpu' + + model, args = FineTuneTestCogVLMModel.from_pretrained(args.from_pretrained, args, overwrite_args={'model_parallel_size': args.model_parallel_size} if args.model_parallel_size != 1 else {}) + if args.use_qlora and torch.cuda.is_available(): + model = model.to('cuda') + from utils.utils import llama2_tokenizer + tokenizer = llama2_tokenizer(args.local_tokenizer, signal_type=args.version) + image_processor = get_image_processor(args.eva_args["image_size"][0]) + text_processor = llama2_text_processor(tokenizer, args.max_length, args.image_length) + + training_main(args, model_cls=model, forward_step_function=forward_step, create_dataset_function=partial(create_dataset_function, image_processor, text_processor), collate_fn=data_collator, forward_step_eval=forward_step_eval) \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh new file mode 100644 index 0000000000..22b3e9911f --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh @@ -0,0 +1,57 @@ +#! /bin/bash +source ./env_npu.sh +# export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH + +NUM_GPUS_PER_WORKER=8 +MP_SIZE=1 + +script_path=$(realpath $0) +script_dir=$(dirname $script_path) +main_dir=$(dirname $script_dir) +MODEL_TYPE="预训练权重路径" +VERSION="base" +MODEL_ARGS="--from_pretrained $MODEL_TYPE \ + --max_length 512 \ + --local_tokenizer 分词器权重路径 \ + --version $VERSION" +# Tips: If training models of resolution 244, you can set --max_length smaller + +OPTIONS_SAT="SAT_HOME=~/.sat_models" +OPTIONS_NCCL="NCCL_DEBUG=info NCCL_IB_DISABLE=0 NCCL_NET_GDR_LEVEL=2 LOCAL_WORLD_SIZE=$NUM_GPUS_PER_WORKER" +HOST_FILE_PATH="hostfile" + +# 用户需要指定train_data和valid_data为实际路径 +train_data="../archive_split/train" +valid_data="../archive_split/valid" + +gpt_options=" \ + --experiment-name finetune-$MODEL_TYPE \ + --model-parallel-size ${MP_SIZE} \ + --mode finetune \ + --train-iters 1000 \ + --resume-dataloader \ + $MODEL_ARGS \ + --train-data ${train_data} \ + --valid-data ${valid_data} \ + --distributed-backend nccl \ + --lr-decay-style cosine \ + --warmup .02 \ + --log-interval 1 \ + --save-interval 2000 \ + --eval-interval 200 \ + --save "./checkpoints" \ + --eval-iters 10 \ + --eval-batch-size 1 \ + --split 1. \ + --deepspeed_config test_config_bf16.json \ + --skip-init \ + --seed 1234 +" + + + +run_cmd="${OPTIONS_NCCL} ${OPTIONS_SAT} deepspeed --master_port 16666 --hostfile ${HOST_FILE_PATH} finetune_cogvlm_demo.py ${gpt_options}" +echo ${run_cmd} +eval ${run_cmd} + +set +x diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_demo.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_demo.py new file mode 100644 index 0000000000..a7cc6f2fba --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_demo.py @@ -0,0 +1,281 @@ +import os +import torch +import torch_npu +from torch_npu.contrib import transfer_to_npu +import argparse +from functools import partial +import sys +sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) + +from sat import mpu, get_args, get_tokenizer +from sat.training.deepspeed_training import training_main +from sat.helpers import print_rank0 +from utils.models import FineTuneTrainCogVLMModel +from utils.utils import llama2_text_processor, llama2_text_processor_inference, get_image_processor + +import random + + +def seed_all(seed=1234, mode=False): + random.seed(seed) + os.environ['PYTHONHASHSEED'] = str(seed) + np.random.seed(seed) + torch.manual_seed(seed) + torch.use_deterministic_algorithms(mode) + torch_npu.npu.manual_seed_all(seed) + torch_npu.npu.manual_seed(seed) + + +def disable_untrainable_params(self): + total_trainable = 0 + enable = [('mlp', 'vit')] + if self.args.use_ptuning: + enable.extend(['ptuning']) + if self.args.use_lora or self.args.use_qlora: + enable.extend(['matrix_A', 'matrix_B']) + for n, p in self.named_parameters(): + flag = False + for e in enable: + if type(e) is tuple: + if e[0].lower() in n.lower() and e[1].lower() in n.lower() and 55 > int(n[:n.find('.mlp')].split('.')[-1]) > 45: + flag = True + break + else: + if e.lower() in n.lower(): + flag = True + break + if not flag: + p.requires_grad_(False) + else: + total_trainable += p.numel() + print_rank0(n) + print_rank0("***** Total trainable parameters: "+str(total_trainable)+" *****") + +FineTuneTrainCogVLMModel.disable_untrainable_params = disable_untrainable_params + +def data_collator(examples): + examples = [ex for ex in examples if len(ex) > 0] # drop {} + for example in examples: + for k in example: + if isinstance(example[k], list): + example[k] = torch.tensor(example[k]) + elif isinstance(example[k], np.ndarray): + example[k] = torch.from_numpy(example[k]) + img_args = {} + tmp_example = examples[0] + for k in tmp_example['vision']: + if type(tmp_example['vision'][k]) is torch.Tensor: + img_args['vision_'+k] = torch.cat([example['vision'][k] for example in examples]) + else: + img_args['vision_'+k] = example['vision'][k] + for example in examples: + example.pop('vision') + if 'cross' in example: + example.pop('cross') + + model_args = {} + tmp_example = examples[0] + for k in tmp_example: + if type(tmp_example[k]) is torch.Tensor: + model_args[k] = torch.cat([example[k] for example in examples]) + else: + model_args[k] = tmp_example[k] + model_args.update(img_args) + return model_args + +from collections import defaultdict + +def broadcast_auto(data_dict): + type2list = defaultdict(list) + other = [] + for k in data_dict: + if type(data_dict[k]) is torch.Tensor: + type2list[data_dict[k].dtype].append(k) + else: + other.append(k) + new_data = {} + for k in type2list: + new_data.update(mpu.broadcast_data(type2list[k], data_dict, k)) + for k in other: + new_data[k] = data_dict[k] + return new_data + +def get_batch(data_iterator, args, timers): + # Broadcast data. + timers('data loader').start() + if data_iterator is not None: + data = next(data_iterator) + else: + data = None + timers('data loader').stop() + for k in data: + if type(data[k]) is torch.Tensor: + data[k] = data[k].to(torch.cuda.current_device()) + if type(data[k]) is torch.Tensor and data[k].dtype is not torch.int32 and data[k].dtype is not torch.long: + if args.fp16: + data[k] = data[k].half() + elif args.bf16: + data[k] = data[k].bfloat16() + return data + +from torch.nn import CrossEntropyLoss +import numpy as np + +from sat.model.mixins import CachedAutoregressiveMixin +from sat.generation.autoregressive_sampling import filling_sequence +from sat.generation.sampling_strategies import BaseStrategy, BeamSearchStrategy + + +def chat(model, tokenizer, tokens, + max_length: int = 1800, num_beams=5, top_p=0.95, top_k=0, temperature=0.8, **kwargs): + inputs = tokens.to(model.parameters().__next__().device)[0] + seq = torch.cat( + [inputs, torch.tensor([-1] * (max_length - len(inputs)), device=inputs.device)], dim=0 + ) + strategy = BaseStrategy(temperature=temperature, top_p=0.4, top_k=1, end_tokens=[tokenizer.eos_token_id]) + # strategy = BeamSearchStrategy(temperature=temperature, top_p=top_p, top_k=top_k, end_tokens=[tokenizer.eos_token_id], + # num_beams=num_beams, consider_end=True) + get_func = llama2_text_processor_inference.get_func(None, None, image_rope_mask=kwargs['image_rope_mask']) + output = filling_sequence( + model, seq, + batch_size=1, + strategy=strategy, + get_masks_and_position_ids=get_func, + **kwargs + )[0] # drop memory + + return output + + +def forward_step_eval(data_iterator, model, args, timers): + def compute_metrics(eval_preds): + preds, labels, device = eval_preds + preds = preds.unsqueeze(0) + if isinstance(preds, tuple): + preds = preds[0] + decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True) + if args.ignore_pad_token_for_loss: + # Replace -100 in the labels as we can't decode them. + labels = np.where(labels != -100, labels, tokenizer.pad_token_id) + decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True) + + score_dict = { + "acc": [], + "acc_w/o_case": [], + } + for pred, label in zip(decoded_preds, decoded_labels): + if args.rank == 0: + print('pred', pred, 'label', label, flush=True) + if pred == label: + score_dict['acc'].append(1.) + else: + score_dict['acc'].append(0.) + if pred.lower() == label.lower(): + score_dict['acc_w/o_case'].append(1.) + else: + score_dict['acc_w/o_case'].append(0.) + + + for k, v in score_dict.items(): + score_dict[k] = float(np.mean(v)) + return score_dict + + # Get the batch. + timers('batch generator').start() + data_b = get_batch( + data_iterator, args, timers) + timers('batch generator').stop() + + context_len = int(data_b['context_length'][0]) + tokens = data_b['input_ids'][:, :context_len] + data_b['vision_expert_mask'] = data_b['vision_expert_mask'][:, :context_len] + data_b['image_embed_mask'] = data_b['image_embed_mask'][:, :context_len] + data_b['image_rope_mask'] = data_b['image_rope_mask'][:, :context_len] + + data_b.pop('input_ids') + data_b.pop('attention_mask') + data_b.pop('position_ids') + labels = data_b.pop('labels') + qid = data_b.pop('question_id') + + model.add_mixin('auto-regressive', CachedAutoregressiveMixin()) + outputs = chat(model, tokenizer, tokens, **data_b)[0][context_len:] + model.del_mixin('auto-regressive') + + return torch.tensor(0, device=outputs.device), {k: torch.tensor(v, device=outputs.device) for k, v in + compute_metrics( + (outputs.cpu(), labels.cpu(), outputs.device)).items()} + + +from torch.nn import CrossEntropyLoss +def forward_step(data_iterator, model, args, timers): + """Forward step.""" + + # Get the batch. + timers('batch generator').start() + data_b = get_batch( + data_iterator, args, timers) + labels = data_b.pop('labels') + timers('batch generator').stop() + logits = model(**data_b)[0] + lm_logits = logits.to(torch.float32) + # Shift so that tokens < n predict n + shift_labels = labels[..., 1:].contiguous() + shift_logits = lm_logits[..., -1-shift_labels.size(-1):-1, :].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss(ignore_index=-100) + loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) + loss = loss.to(torch.float32) + + return loss, {'loss': loss} + +from utils.utils import ItemDataset +def create_dataset_function(image_processor, text_processor, path, args): + dataset = ItemDataset(image_processor, text_processor, args, path) + return dataset + +from sat.model.finetune.lora2 import LoraMixin +from sat.model.finetune.prompt_tuning import PTuningV2Mixin + +if __name__ == '__main__': + torch.npu.set_compile_mode(jit_compile=False) + torch.npu.config.allow_internal_format = False + # seed_all(seed=1234, mode=True) # 根据实际情况开启确定性计算 + py_parser = argparse.ArgumentParser(add_help=False) + py_parser.add_argument('--max_length', type=int) + py_parser.add_argument('--ignore_pad_token_for_loss', action='store_false') + py_parser.add_argument("--version", type=str, default="chat_old", help='version to interact with') + py_parser.add_argument("--from_pretrained", type=str, default="cogvlm-chat", help='pretrained ckpt') + py_parser.add_argument("--local_tokenizer", type=str, default="lmsys/vicuna-7b-v1.5", help='tokenizer path') + py_parser.add_argument("--vit_checkpoint_activations", action='store_true') + py_parser = FineTuneTrainCogVLMModel.add_model_specific_args(py_parser) + known, args_list = py_parser.parse_known_args() + args = get_args(args_list) + args = argparse.Namespace(**vars(args), **vars(known)) + if args.use_qlora: + args.device = 'cpu' + + model, args = FineTuneTrainCogVLMModel.from_pretrained(args.from_pretrained, args, overwrite_args={'model_parallel_size': args.model_parallel_size} if args.model_parallel_size != 1 else {}) + if args.use_ptuning: + model.add_mixin("ptuning", PTuningV2Mixin(args.num_layers, args.hidden_size // args.num_attention_heads, args.num_attention_heads, args.pre_seq_len)) + if args.use_lora: + model.add_mixin("lora", LoraMixin(args.num_layers, args.lora_rank, layer_range=args.layer_range), reinit=True) + model.get_mixin("eva").vit_model.add_mixin("lora", LoraMixin(args.eva_args['num_layers'], args.lora_rank, layer_range=args.layer_range), reinit=True) + elif args.use_qlora: + model.add_mixin("lora", LoraMixin(args.num_layers, args.lora_rank, layer_range=args.layer_range, qlora=True), reinit=True) + + if args.use_qlora and torch.cuda.is_available(): + model = model.to('cuda') + from utils.utils import llama2_tokenizer + tokenizer = llama2_tokenizer(args.local_tokenizer, signal_type=args.version) + image_processor = get_image_processor(args.eva_args["image_size"][0]) + text_processor = llama2_text_processor(tokenizer, args.max_length, args.image_length) + + model = training_main(args, model_cls=model, forward_step_function=forward_step, create_dataset_function=partial(create_dataset_function, image_processor, text_processor), collate_fn=data_collator, forward_step_eval=forward_step_eval) + if args.use_lora: + model.get_mixin("lora").merge_lora() + model.get_mixin("eva").vit_model.get_mixin("lora").merge_lora() + args.use_lora = False + args.save = "checkpoints/merged_lora_cogvlm{}".format(args.eva_args["image_size"][0]) + from sat.training.model_io import save_checkpoint + save_checkpoint(1, model, None, None, args) diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/inference.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/inference.py new file mode 100644 index 0000000000..2814f37dae --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/inference.py @@ -0,0 +1,33 @@ +import torch +import torch_npu +from torch_npu.contrib import transfer_to_npu +import requests +from PIL import Image +from transformers import AutoModelForCausalLM, LlamaTokenizer + +tokenizer = LlamaTokenizer.from_pretrained('分词器权重路径') +model = AutoModelForCausalLM.from_pretrained( + '推理权重路径', + torch_dtype=torch.bfloat16, + low_cpu_mem_usage=True, + trust_remote_code=True +).to('cuda').eval() + +image = Image.open("图片路径").convert('RGB') +inputs = model.build_conversation_input_ids(tokenizer, query='', images=[image]) +inputs = { + 'input_ids': inputs['input_ids'].unsqueeze(0).to('cuda'), + 'token_type_ids': inputs['token_type_ids'].unsqueeze(0).to('cuda'), + 'attention_mask': inputs['attention_mask'].unsqueeze(0).to('cuda'), + 'images': [[inputs['images'][0].to('cuda').to(torch.bfloat16)]], +} + +gen_kwargs = {"max_length": 2048, "do_sample": False} + +with torch.no_grad(): + print("Begin inference") + outputs = model.generate(**inputs, **gen_kwargs) + print("Inference End") + outputs = outputs[:, inputs['input_ids'].shape[1]:] + response = tokenizer.decode(outputs[0]) + print("\nCog:", response) diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/layernorm.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/layernorm.py new file mode 100644 index 0000000000..e5e773a23e --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/layernorm.py @@ -0,0 +1,40 @@ +try: + from apex.normalization.fused_layer_norm import FusedLayerNorm + class LayerNorm(FusedLayerNorm): + def __init__(self, *args, pb_relax=False, **kwargs): + super().__init__(*args, **kwargs) + self.pb_relax = pb_relax + + def forward(self, x): + if not self.pb_relax: + return super().forward(x) + return super().forward(x / (x.abs().max().detach() / 8)) +except ModuleNotFoundError: + from sat.helpers import print_rank0 + print_rank0('Please install apex to use fused_layer_norm, fall back to torch.nn.LayerNorm', level='DEBUG') + import torch + class LayerNorm(torch.nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + def forward(self, x): + # if cpu and float16, calculate in float32 for both x and weight, bias. + if str(x.device) == 'cpu' and x.dtype in[torch.float16, torch.bfloat16]: + return torch.nn.functional.layer_norm(x.float(), self.normalized_shape, self.weight.float(), self.bias.float()).to(x.dtype) + else: + return super().forward(x) + +import torch +import torch_npu +import torch.nn as nn +class RMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + RMSNorm 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 + return torch_npu.npu_rms_norm(hidden_states.to(torch.float32), self.weight.to(torch.float32), epsilon=self.variance_epsilon)[0].to(input_dtype) diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/mixin.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/mixin.py new file mode 100644 index 0000000000..1b39d7cf36 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/mixin.py @@ -0,0 +1,275 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from sat.transformer_defaults import attention_fn_default +from sat.model.base_model import BaseMixin, non_conflict +from sat.mpu.layers import ColumnParallelLinear, RowParallelLinear +from sat.mpu.utils import split_tensor_along_last_dim +from sat import mpu +from sat.model.position_embedding.triton_rotary_embeddings import apply_rotary_pos_emb_index_bhs + + +class LlamaVisionExpertFCMixin(BaseMixin): + def __init__(self, in_features, hidden_features, num_layers=32, num_vision_layers=0, vision_layer_range=None, + params_dtype=torch.float, device=torch.device('cpu')): + super().__init__() + + self.num_layers = num_layers + self.num_vision_layers = num_vision_layers + if vision_layer_range is None: + vision_layer_range = [i for i in range(min(num_vision_layers, num_layers))] + self.vision_layer_range = vision_layer_range + self.gate_proj = nn.ModuleList([ColumnParallelLinear( + in_features, + hidden_features, + gather_output=False, + init_method=None, + bias=False, + params_dtype=params_dtype, + module=self, + name="dense_h_to_4h_gate", + skip_init=True, + device=device + ) for i in range(num_layers)]) + # Trainable vision expert parameters + vision_dense_h_to_4h_list = [] + vision_dense_4h_to_h_list = [] + gate_proj_list = [] + + + for i in vision_layer_range: + vision_dense_h_to_4h = ColumnParallelLinear( + in_features, + hidden_features, + gather_output=False, + init_method=None, + bias=False, + params_dtype=params_dtype, + module=self, + name="vision_dense_h_to_4h", + skip_init=True, + device=device + ) + + # Project back to h. + vision_dense_4h_to_h = RowParallelLinear( + hidden_features, + in_features, + input_is_parallel=True, + init_method=None, + bias=False, + params_dtype=params_dtype, + module=self, + name="vision_dense_4h_to_h", + skip_init=True, + device=device + ) + + gate_proj = ColumnParallelLinear( + in_features, + hidden_features, + gather_output=False, + init_method=None, + bias=False, + params_dtype=params_dtype, + module=self, + name="vision_gate_proj", + skip_init=True, + device=device + ) + + vision_dense_h_to_4h_list.append(vision_dense_h_to_4h) + vision_dense_4h_to_h_list.append(vision_dense_4h_to_h) + gate_proj_list.append(gate_proj) + + self.vision_dense_h_to_4h_list = nn.ModuleDict([ + (str(layer_id), vision_dense_h_to_4h) + for layer_id, vision_dense_h_to_4h in zip(vision_layer_range, vision_dense_h_to_4h_list) + ]) + self.vision_dense_4h_to_h_list = nn.ModuleDict([ + (str(layer_id), vision_dense_4h_to_h) + for layer_id, vision_dense_4h_to_h in zip(vision_layer_range, vision_dense_4h_to_h_list) + ]) + self.vision_gate_proj = nn.ModuleDict([ + (str(layer_id), gate_proj) + for layer_id, gate_proj in zip(vision_layer_range, gate_proj_list) + ]) + + def mlp_forward(self, hidden_states, **kw_args): + mixin_self = self + self = self.transformer.layers[kw_args['layer_id']].mlp + if "vision_expert_mask" in kw_args: + vision_expert_mask = kw_args['vision_expert_mask'] + else: + vision_expert_mask = None + + layer_id_key = str(int(kw_args['layer_id'])) + + if kw_args['layer_id'] in mixin_self.vision_layer_range and (vision_expert_mask is not None) and vision_expert_mask.any(): + vision_dense_h_to_4h = mixin_self.vision_dense_h_to_4h_list[layer_id_key] + vision_dense_4h_to_h = mixin_self.vision_dense_4h_to_h_list[layer_id_key] + vision_gate_proj = mixin_self.vision_gate_proj[layer_id_key] + output = torch.empty(hidden_states.shape, dtype=hidden_states.dtype, device=hidden_states.device) + + language_hidden_state = hidden_states[~vision_expert_mask.bool()] + language_intermediate_parallel = self.activation_func(mixin_self.gate_proj[kw_args['layer_id']](language_hidden_state)) * self.dense_h_to_4h(language_hidden_state) + output[~vision_expert_mask.bool()] = self.dense_4h_to_h(language_intermediate_parallel) # language_output + + vision_hidden_state = hidden_states[vision_expert_mask.bool()] + vision_intermediate_parallel = vision_dense_h_to_4h(vision_hidden_state) + gate_output = vision_gate_proj(vision_hidden_state) + + vision_intermediate_parallel *= self.activation_func(gate_output) + output[vision_expert_mask.bool()] = vision_dense_4h_to_h(vision_intermediate_parallel) # vision_output + else: + intermediate_parallel = self.activation_func(mixin_self.gate_proj[kw_args['layer_id']](hidden_states)) * self.dense_h_to_4h(hidden_states) + output = self.dense_4h_to_h(intermediate_parallel) + + return output.contiguous() + + def copy_param(self): + with torch.no_grad(): + for i in self.vision_layer_range: + self.vision_gate_proj[str(i)].weight.data.copy_(self.gate_proj[i].weight.data) + self.vision_dense_4h_to_h_list[str(i)].weight.data.copy_(self.transformer.layers[i].mlp.dense_4h_to_h.weight.data) + self.vision_dense_h_to_4h_list[str(i)].weight.data.copy_(self.transformer.layers[i].mlp.dense_h_to_4h.weight.data) + +from sat.mpu import get_model_parallel_world_size +from sat.mpu.utils import divide +from sat.model.position_embedding.triton_rotary_embeddings import FastRotaryEmbedding + +class LlamaVisionExpertAttnMixin(BaseMixin): + def __init__(self, hidden_size, num_heads, num_layers=28, num_vision_layers=0, use_vision_expert=True, vision_layer_range=None, + params_dtype=torch.float, device=torch.device('cpu')): + super().__init__() + + world_size = get_model_parallel_world_size() + self.hidden_size = hidden_size + self.num_attention_heads = num_heads + self.hidden_size_per_attention_head = divide(hidden_size, num_heads) + self.num_attention_heads_per_partition = divide(num_heads, world_size) + self.inner_hidden_size = num_heads * self.hidden_size_per_attention_head + + self.rotary_emb = FastRotaryEmbedding( + hidden_size // num_heads + ) + + self.num_vision_layers = num_vision_layers + self.num_layers = num_layers + if vision_layer_range is None: + vision_layer_range = [i for i in range(min(num_vision_layers, num_layers))] + self.vision_layer_range = vision_layer_range + + self.use_vision_expert = use_vision_expert + # Trainable vision expert parameters + + if self.use_vision_expert: + vision_query_key_value_list = [] + vision_dense_list = [] + for i in vision_layer_range: + vision_query_key_value = ColumnParallelLinear( + hidden_size, + 3 * hidden_size, + stride=3, + gather_output=False, + init_method=None, + bias=False, + params_dtype=params_dtype, + module=self, + name="vision_query_key_value", + skip_init=True, + device=device + ) + + vision_dense = RowParallelLinear( + self.inner_hidden_size, + hidden_size, + input_is_parallel=True, + init_method=None, + bias=False, + params_dtype=params_dtype, + module=self, + name="vision_dense", + skip_init=True, + device=device, + final_bias=False + ) + + vision_query_key_value_list.append(vision_query_key_value) + vision_dense_list.append(vision_dense) + + self.vision_query_key_value_list = nn.ModuleDict([ + (str(layer_id), vision_query_key_value) + for layer_id, vision_query_key_value in zip(vision_layer_range, vision_query_key_value_list) + ]) + self.vision_dense_list = nn.ModuleDict([ + (str(layer_id), vision_dense) + for layer_id, vision_dense in zip(vision_layer_range, vision_dense_list) + ]) + + def attention_forward(self, hidden_states, mask, **kw_args): + mixin_self = self + self = self.transformer.layers[kw_args['layer_id']].attention + attention_fn = attention_fn_default + if 'attention_fn' in self.hooks: + attention_fn = self.hooks['attention_fn'] + if "vision_expert_mask" in kw_args: + vision_expert_mask = kw_args['vision_expert_mask'] + else: + vision_expert_mask = None + + layer_id_key = str(int(kw_args['layer_id'])) + if mixin_self.use_vision_expert and kw_args['layer_id'] in mixin_self.vision_layer_range and ( + vision_expert_mask is not None) and vision_expert_mask.any(): + shape = list(hidden_states.shape) + parallel_size = mpu.get_model_parallel_world_size() + shape[-1] = shape[-1] * 3 // parallel_size + vision_query_key_value = mixin_self.vision_query_key_value_list[layer_id_key] + mixed_raw_layer = torch.empty(shape, dtype=hidden_states.dtype, device=hidden_states.device) + language_hidden_states = hidden_states[~vision_expert_mask.bool()] + vision_hidden_states = hidden_states[vision_expert_mask.bool()] + mixed_raw_layer[~vision_expert_mask.bool()] = self.query_key_value( + language_hidden_states) # language_mixed_raw_layer + mixed_raw_layer[vision_expert_mask.bool()] = vision_query_key_value( + vision_hidden_states) # vision_mixed_raw_layer + else: + mixed_raw_layer = self.query_key_value(hidden_states) + + (mixed_query_layer, + mixed_key_layer, + mixed_value_layer) = split_tensor_along_last_dim(mixed_raw_layer, 3) + + dropout_fn = self.attention_dropout if self.training else None + + query_layer = self._transpose_for_scores(mixed_query_layer) + key_layer = self._transpose_for_scores(mixed_key_layer) + value_layer = self._transpose_for_scores(mixed_value_layer) + cos, sin = mixin_self.rotary_emb(value_layer, seq_len=kw_args['position_ids'].max()+1) + query_layer, key_layer = apply_rotary_pos_emb_index_bhs(query_layer, key_layer, cos, sin, kw_args['position_ids']) + + context_layer = attention_fn(query_layer, key_layer, value_layer, mask, dropout_fn, **kw_args) + + context_layer = context_layer.permute(0, 2, 1, 3).contiguous() + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.view(*new_context_layer_shape) + + if mixin_self.use_vision_expert and kw_args['layer_id'] in mixin_self.vision_layer_range and ( + vision_expert_mask is not None) and vision_expert_mask.any(): + vision_dense = mixin_self.vision_dense_list[layer_id_key] + parallel_size = mpu.get_model_parallel_world_size() + target_shape = context_layer.shape[:-1] + (context_layer.shape[-1] * parallel_size,) + output = torch.empty(target_shape, dtype=hidden_states.dtype, device=hidden_states.device) + output[~vision_expert_mask.bool()] = self.dense(context_layer[~vision_expert_mask.bool()]) # language + output[vision_expert_mask.bool()] = vision_dense(context_layer[vision_expert_mask.bool()]) # vision + else: + output = self.dense(context_layer) + + if self.training: + output = self.output_dropout(output) + return output.contiguous() + + def copy_param(self): + with torch.no_grad(): + for i in self.vision_layer_range: + self.vision_query_key_value_list[str(i)].weight.data.copy_(self.transformer.layers[i].attention.query_key_value.weight.data) + self.vision_dense_list[str(i)].weight.data.copy_(self.transformer.layers[i].attention.dense.weight.data) diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/modeling_cogvlm.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/modeling_cogvlm.py new file mode 100644 index 0000000000..4eaa77b6c3 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/modeling_cogvlm.py @@ -0,0 +1,806 @@ +"""largely copy from llama and adapt for cogvlm""" +import warnings +from typing import TYPE_CHECKING, Optional, Tuple, List, Union, Literal, Dict, Any + +import math +import torch +from torch import nn +from torch.nn import CrossEntropyLoss +from torchvision import transforms +from einops import rearrange + +from transformers import PreTrainedModel, PreTrainedTokenizer +from transformers.utils.logging import get_logger +from transformers.activations import ACT2FN +from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast + +from .configuration_cogvlm import CogVLMConfig +from .rotary_embeddings import RotaryEmbedding as FastRotaryEmbedding, apply_rotary_pos_emb_index_bhs +from .visual import EVA2CLIPModel + +if TYPE_CHECKING: + from transformers.utils import ModelOutput + +logger = get_logger(__name__) + +LANGUAGE_TOKEN_TYPE = 0 +VISION_TOKEN_TYPE = 1 + + +# Copied from transformers.models.bart.modeling_bart._make_causal_mask +def _make_causal_mask( + input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0 +): + """ + Make causal mask used for bi-directional self-attention. + """ + bsz, tgt_len = input_ids_shape + mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device) + mask_cond = torch.arange(mask.size(-1), device=device) + mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0) + mask = mask.to(dtype) + + if past_key_values_length > 0: + mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1) + return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length) + + +# Copied from transformers.models.bart.modeling_bart._expand_mask +def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None): + """ + Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. + """ + bsz, src_len = mask.size() + tgt_len = tgt_len if tgt_len is not None else src_len + + expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype) + + inverted_mask = 1.0 - expanded_mask + + return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min) + + +class RMSNorm(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) + + +class MLP(nn.Module): + def __init__(self, config): + super().__init__() + self.hidden_size = config.hidden_size + self.intermediate_size = config.intermediate_size + self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) + self.act_fn = ACT2FN[config.hidden_act] + + def forward(self, x): + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + + +def get_expert_mask(token_type_ids: "torch.LongTensor(B, L)") -> "[torch.BoolTensor(B, L), torch.BoolTensor(B, L)]": + vision_token_mask = torch.zeros_like(token_type_ids, dtype=torch.bool) + vision_token_mask[:, :-1] = (token_type_ids[:, :-1] == VISION_TOKEN_TYPE) & (token_type_ids[:, 1:] == VISION_TOKEN_TYPE) + language_token_mask = ~vision_token_mask + return vision_token_mask, language_token_mask + + +def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None) -> torch.Tensor: + # Efficient implementation equivalent to the following: + L, S = query.size(-2), key.size(-2) + scale_factor = 1 / math.sqrt(query.size(-1)) if scale is None else scale + attn_bias = torch.zeros(L, S, dtype=query.dtype).to(query.device) + if is_causal: + assert attn_mask is None + temp_mask = torch.ones(L, S, dtype=torch.bool).tril(diagonal=0).to(query.device) + attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf")) + attn_bias.to(query.dtype) + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_mask.masked_fill_(attn_mask.logical_not(), float("-inf")) + else: + attn_bias += attn_mask + attn_weight = query @ key.transpose(-2, -1) * scale_factor + attn_weight += attn_bias + attn_weight = torch.softmax(attn_weight, dim=-1) + attn_weight = torch.dropout(attn_weight, dropout_p, train=True) + return attn_weight @ value + + +class VisionExpertMLP(nn.Module): + def __init__(self, config): + super().__init__() + self.language_mlp = MLP(config) + self.vision_mlp = MLP(config) + + def forward(self, hidden_states: "torch.Tensor(B, L, D)", token_type_ids: "torch.LongTensor(B, L)"): + output = torch.empty(hidden_states.shape, dtype=hidden_states.dtype, device=hidden_states.device) + vision_token_mask, language_token_mask = get_expert_mask(token_type_ids) + output[vision_token_mask] = self.vision_mlp(hidden_states[vision_token_mask]) + output[language_token_mask] = self.language_mlp(hidden_states[language_token_mask]) + return output + + +def attention_fn( + query_layer: "torch.tensor(B, H, L, HD)", + key_layer: "torch.tensor(B, H, L, HD)", + value_layer: "torch.tensor(B, H, L, HD)", + attention_mask: "torch.tensor(B, H, L, HD)", + *, + scaling_attention_score: bool = True, + attention_dropout: nn.Module = None +): + attention_mask_bool = (attention_mask == 0) + is_low_triangle = (attention_mask_bool == torch.ones_like(attention_mask_bool, dtype=torch.float).tril()).all() + is_full = (attention_mask_bool > 0).all() + if not (int(torch.__version__.split('.')[0]) >= 2): + warnings.warn("It's recommended to use torch2.0 or higher.") + if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score and (is_full or is_low_triangle): + dropout_p = 0. if attention_dropout is None or not attention_dropout.training else attention_dropout.p + scaled_dot_product_attention(query_layer, key_layer, value_layer, dropout_p=dropout_p, is_causal=not is_full) + else: + if scaling_attention_score: + query_layer = query_layer / math.sqrt(query_layer.shape[-1]) + attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2)) + attention_scores = attention_scores + attention_mask + attention_scores = nn.functional.softmax(attention_scores, dim=-1, dtype=torch.float32).to(query_layer.dtype) + if attention_dropout is not None: + attention_scores = attention_dropout(attention_scores) + context_layer = torch.matmul(attention_scores, value_layer) + return context_layer + +class VisionExpertAttention(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = self.hidden_size // self.num_heads + self.max_position_embeddings = config.max_position_embeddings + + self.rotary_emb = FastRotaryEmbedding(self.hidden_size // self.num_heads) + self.vision_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False) + self.vision_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False) + self.language_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False) + self.language_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False) + + def _transpose_for_scores(self, tensor): + """Transpose a 3D tensor [B, L, H*HD] into a 4D tensor with size [B H L HD].""" + new_tensor_shape = tensor.size()[:-1] + (self.num_heads, self.head_dim) + tensor = tensor.view(*new_tensor_shape) + return tensor.permute(0, 2, 1, 3) + + def forward( + self, + hidden_states: torch.Tensor, + token_type_ids: torch.LongTensor, + position_ids: torch.LongTensor, + attention_mask: Optional[torch.Tensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: bool = False, + use_cache: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + bsz, q_len, _ = hidden_states.size() + vision_token_mask, language_token_mask = get_expert_mask(token_type_ids) + + shape = list(hidden_states.shape) + shape[-1] = shape[-1] * 3 + mixed_raw_layer = torch.empty(shape, dtype=hidden_states.dtype, device=hidden_states.device) + mixed_raw_layer[vision_token_mask] = self.vision_expert_query_key_value(hidden_states[vision_token_mask]) + mixed_raw_layer[language_token_mask] = self.language_expert_query_key_value(hidden_states[language_token_mask]) + + query_states, key_states, value_states = torch.split(mixed_raw_layer, self.hidden_size, dim=-1) + query_states = self._transpose_for_scores(query_states) # B, H, L, HD + key_states = self._transpose_for_scores(key_states) # B, H, L, HD + value_states = self._transpose_for_scores(value_states) # B, H, L, HD + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + kv_seq_len += past_key_value[0].shape[-2] + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + kv_seq_len += past_key_value[0].shape[-2] + + cos, sin = self.rotary_emb(value_states, seq_len=position_ids.max()+1) + query_states, key_states = apply_rotary_pos_emb_index_bhs(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) + + past_key_value = (key_states, value_states) if use_cache else None + + context_layer = attention_fn( + query_layer=query_states, key_layer=key_states, value_layer=value_states, attention_mask=attention_mask, + scaling_attention_score=True, attention_dropout=None) + if context_layer.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" {context_layer.size()}" + ) + context_layer = context_layer.transpose(1, 2).contiguous().reshape(bsz, q_len, self.hidden_size) + + attn_output = torch.empty(context_layer.shape, dtype=hidden_states.dtype, device=hidden_states.device) + attn_output[vision_token_mask] = self.vision_expert_dense(context_layer[vision_token_mask]) + attn_output[language_token_mask] = self.language_expert_dense(context_layer[language_token_mask]) + + if output_attentions: + warnings.warn("output_attentions is not implemented.") + + return attn_output, None, past_key_value + + +class CogVLMDecoderLayer(nn.Module): + def __init__(self, config): + super().__init__() + self.hidden_size = config.hidden_size + self.self_attn = VisionExpertAttention(config=config) + self.mlp = VisionExpertMLP(config) + self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + token_type_ids: torch.LongTensor, + position_ids: torch.LongTensor, + attention_mask: Optional[torch.Tensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights, present_key_value = self.self_attn( + hidden_states=hidden_states, + token_type_ids=token_type_ids, + position_ids=position_ids, + attention_mask=attention_mask, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states, token_type_ids=token_type_ids) + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights,) + + if use_cache: + outputs += (present_key_value,) + + return outputs # type: ignore + + +class CogVLMPreTrainedModel(PreTrainedModel): + config_class = CogVLMConfig + base_model_prefix = "model" + supports_gradient_checkpointing = False + _no_split_modules = ["CogVLMDecoderLayer"] + _skip_keys_device_placement = "past_key_values" + + 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_() + + +def is_empty(images_list: Optional[List[List[torch.Tensor]]]): + if images_list is None or len(images_list) == 0: + return True + for image_list in images_list: + if len(image_list): + return False + return True + + +def build_position_ids(x: "torch.BoolTensor(B, L)", attention_mask: Optional["torch.BoolTensor(B, L)"] = None) -> "torch.LongTensor(B, L)": + if attention_mask is not None: + tmp = x.clone() + tmp[~(attention_mask.bool())] = -1 + else: + tmp = x.clone() + # image boi eoi token as LANGUAGE_TOKEN_TYPE + is_boi_eoi = torch.zeros_like(x, dtype=torch.bool) + is_boi_eoi[:, 1:] |= (tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE) + is_boi_eoi[:, 0] |= (tmp[:, 0] == VISION_TOKEN_TYPE) + is_boi_eoi[:, :-1] |= (tmp[:, :-1] == VISION_TOKEN_TYPE) & (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE) + is_boi_eoi[:, -1] |= (tmp[:, -1] == VISION_TOKEN_TYPE) + tmp[is_boi_eoi] = LANGUAGE_TOKEN_TYPE + # final position ids + y = torch.zeros_like(x, dtype=torch.long) + y[:, 1:] = (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE) | ((tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE)) + y = y.cumsum(dim=-1) + return y + + +class CogVLMModel(CogVLMPreTrainedModel): + def __init__(self, config): + 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.layers = nn.ModuleList([CogVLMDecoderLayer(config) for _ in range(config.num_hidden_layers)]) + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + self.vision = EVA2CLIPModel(config) + + self.gradient_checkpointing = False + # Initialize weights and apply final processing + self.post_init() + + def encode_images(self, images: List[List[torch.Tensor]]) -> torch.Tensor: + images_list, images = images, [] + + images = [] + for image_list in images_list: + for image in image_list: + images.append(image) + + images = torch.stack(images) + images_features = self.vision(images) + return images_features + + def forward( + self, + input_ids: torch.LongTensor = None, + images: List[List[torch.Tensor]] = None, + token_type_ids: Optional[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]: + """take care of image_encode, token_type_ids, position_ids and (attention_mask = None is fine)""" + + if past_key_values is not None: + pass # generate mode with past_key_values. the image features are already mapped + else: + # not allow for inputs_embeds, because we want to process image feature + assert input_ids is not None and inputs_embeds is None, f"{input_ids} {inputs_embeds}" + if not is_empty(images): # multi-modality + assert token_type_ids is not None, f"multi-modality requires `token_type_ids`!" + assert len(input_ids) == len(images), f"{len(input_ids)} {len(images)}" + inputs_embeds = self.embed_tokens(input_ids) + images_features = self.encode_images(images) + images_features = rearrange(images_features, 'b n d -> (b n) d') + images_features = images_features.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device) + inputs_embeds = inputs_embeds.index_put([token_type_ids == VISION_TOKEN_TYPE], images_features) + else: # single-modality + if token_type_ids is None: + token_type_ids = torch.ones_like(input_ids, dtype=torch.long, device=input_ids.device) * LANGUAGE_TOKEN_TYPE + assert not (token_type_ids == VISION_TOKEN_TYPE).any(), f"{(token_type_ids == VISION_TOKEN_TYPE).sum()}" + inputs_embeds = self.embed_tokens(input_ids) + + if position_ids is None: + position_ids = build_position_ids(token_type_ids, attention_mask) + input_ids = None + + return self.llm_forward( + input_ids=input_ids, + token_type_ids=token_type_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, + ) + + def llm_forward( + self, + input_ids: torch.LongTensor = None, + token_type_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]: + """largely copy from llama forward and adapt for cogvlm with `token_type_ids`""" + 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 decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + batch_size, seq_length = input_ids.shape + elif inputs_embeds is not None: + batch_size, seq_length, _ = inputs_embeds.shape + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + + seq_length_with_past = seq_length + past_key_values_length = 0 + + if past_key_values is not None: + past_key_values_length = past_key_values[0][0].shape[2] + seq_length_with_past = seq_length_with_past + past_key_values_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) + # embed positions + if attention_mask is None: + attention_mask = torch.ones( + (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device + ) + attention_mask = self._prepare_decoder_attention_mask( + attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length + ) + + 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 = () if use_cache else None + + for idx, decoder_layer in enumerate(self.layers): + if output_hidden_states: + all_hidden_states += (hidden_states,) + + past_key_value = past_key_values[idx] if past_key_values is not None else None + layer_outputs = decoder_layer( + hidden_states, + token_type_ids=token_type_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + 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 = next_decoder_cache if use_cache else None + 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, + ) + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + # noinspection PyMethodMayBeStatic + # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask + def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length): + # create causal mask + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + combined_attention_mask = None + if input_shape[-1] > 1: + combined_attention_mask = _make_causal_mask( + input_shape, + inputs_embeds.dtype, + device=inputs_embeds.device, + past_key_values_length=past_key_values_length, + ) + + if attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to( + inputs_embeds.device + ) + combined_attention_mask = ( + expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask + ) + + return combined_attention_mask + + +def chat_history_to_prompt(history, query): + prompt = " [INST] " + for i, (old_query, response) in enumerate(history): + prompt += old_query + " [/INST] " + response + " [INST] " + prompt += query + " [/INST] " + return prompt + + +def base_history_to_prompt(history, query): + prompt = query + return prompt + + +_history_to_prompt = { + "base": base_history_to_prompt, + "chat": chat_history_to_prompt +} + + +class CogVLMForCausalLM(CogVLMPreTrainedModel): + _auto_class = "AutoModelForCausalLM" + + def __init__(self, config): + super().__init__(config) + self.model = CogVLMModel(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() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + def forward( + self, + input_ids: torch.LongTensor = None, + images: List[List[torch.Tensor]] = None, + token_type_ids: Optional[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, + labels: Optional[torch.LongTensor] = None, + ) -> Union[Tuple, CausalLMOutputWithPast]: + 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, + images=images, + token_type_ids=token_type_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, + ) + + def _prepare_attention_mask_for_generation( + self, + inputs: torch.Tensor, + pad_token_id: Optional[int], + eos_token_id: Optional[Union[int, List[int]]], + ) -> torch.LongTensor: + return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device) # type: ignore + + def prepare_inputs_for_generation( + self, input_ids, token_type_ids, images=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + ): + # build position_ids if needed + position_ids = kwargs.get("position_ids", None) + if position_ids is None: + position_ids = build_position_ids(token_type_ids, attention_mask) + + if past_key_values: + input_ids = input_ids[:, -1:] + token_type_ids = token_type_ids[:, -1:] + position_ids = position_ids[:, -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: + model_inputs = {"inputs_embeds": inputs_embeds} + else: + model_inputs = {"input_ids": input_ids} + + model_inputs.update( + { + "token_type_ids": token_type_ids, + "images": images, + "position_ids": position_ids, + "past_key_values": past_key_values, + "use_cache": kwargs.get("use_cache"), + "attention_mask": attention_mask, + } + ) + return model_inputs + + def _update_model_kwargs_for_generation( + self, + outputs: "ModelOutput", + model_kwargs: Dict[str, Any], + is_encoder_decoder: bool = False, + standardize_cache_format: bool = False, + ) -> Dict[str, Any]: + # update past_key_values + model_kwargs["past_key_values"] = self._extract_past_from_model_output( + outputs, standardize_cache_format=standardize_cache_format + ) + if getattr(outputs, "state", None) is not None: + model_kwargs["state"] = outputs.state + + # update token_type_ids with last value + if "token_type_ids" in model_kwargs: + token_type_ids = model_kwargs["token_type_ids"] + new_token_type_ids = torch.ones(size=(token_type_ids.shape[0], 1), dtype=token_type_ids.dtype, device=token_type_ids.device) * LANGUAGE_TOKEN_TYPE + model_kwargs["token_type_ids"] = torch.cat([token_type_ids, new_token_type_ids], dim=-1) + + if not is_encoder_decoder: + # update attention mask + if "attention_mask" in model_kwargs: + attention_mask = model_kwargs["attention_mask"] + model_kwargs["attention_mask"] = torch.cat( + [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1 + ) + else: + # update decoder attention mask + if "decoder_attention_mask" in model_kwargs: + decoder_attention_mask = model_kwargs["decoder_attention_mask"] + model_kwargs["decoder_attention_mask"] = torch.cat( + [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))], + dim=-1, + ) + + return model_kwargs + + def _reorder_cache(self, 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 + + def build_conversation_input_ids( + self, + tokenizer: "PreTrainedTokenizer", + *, + query: str, + history: Optional[List[Tuple[str, str]]] = None, + images: Optional[List["PIL.Image"]] = None, + template_version: Optional[Literal["base", "chat"]] = None, + ): + image_size: int = self.config.vision_config['image_size'] + patch_size: int = self.config.vision_config['patch_size'] + template_version = template_version or self.config.template_version + assert images is None or len(images) <= 1, f"not support multi images by now." + history = history or [] + text = _history_to_prompt[template_version](history, query) + + input_ids = [tokenizer.bos_token_id] + token_type_ids = [LANGUAGE_TOKEN_TYPE] + if images is not None and len(images) == 1: + # vision + transform = transforms.Compose( + [ + transforms.Resize( + (image_size, image_size), interpolation=transforms.InterpolationMode.BICUBIC + ), + transforms.ToTensor(), + transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)), + ] + ) + images = [transform(images[0])] + # language + vision_token_num = (image_size // patch_size) * (image_size // patch_size) + 2 + input_ids += [tokenizer.pad_token_id] * vision_token_num + token_type_ids += [VISION_TOKEN_TYPE] * vision_token_num + text_ids = tokenizer.encode(text, add_special_tokens=False) + + input_ids += text_ids + token_type_ids += [LANGUAGE_TOKEN_TYPE] * len(text_ids) + attention_mask = [1] * len(input_ids) + + return { + 'input_ids': torch.tensor(input_ids, dtype=torch.long), + 'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long), + 'attention_mask': torch.tensor(attention_mask, dtype=torch.long), + 'images': images, + } diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/rotary_embeddings.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/rotary_embeddings.py new file mode 100644 index 0000000000..a9e17df12e --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/rotary_embeddings.py @@ -0,0 +1,136 @@ +import torch +import torch.nn.functional as F + + +class RotaryEmbedding(torch.nn.Module): + + def __init__(self, dim, base=10000, precision=torch.half, learnable=False, device=torch.device('cpu')): + super().__init__() + inv_freq = 1. / (base ** (torch.arange(0, dim, 2, device=device).float() / dim)) + # inv_freq = inv_freq.half() + self.learnable = learnable + if learnable: + self.inv_freq = torch.nn.Parameter(inv_freq) + self.max_seq_len_cached = None + else: + self.register_buffer('inv_freq', inv_freq) + self.max_seq_len_cached = None + self.cos_cached = None + self.sin_cached = None + self.precision = precision + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): + pass + + def forward(self, x, seq_dim=1, seq_len=None): + if seq_len is None: + seq_len = x.shape[seq_dim] + if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached): + self.max_seq_len_cached = None if self.learnable else seq_len + t = torch.arange(seq_len, device=x.device, dtype=torch.float32) + freqs = torch.einsum('i,j->ij', t, self.inv_freq) + # Different from paper, but it uses a different permutation in order to obtain the same calculation + emb = torch.cat((freqs, freqs), dim=-1).to(x.device) + if self.precision == torch.bfloat16: + emb = emb.float() + + # [sx, 1 (b * np), hn] + cos_cached = emb.cos()[:, None, :] + sin_cached = emb.sin()[:, None, :] + cos_cached = cos_cached.to(x.dtype) + sin_cached = sin_cached.to(x.dtype) + if self.learnable: + return cos_cached, sin_cached + self.cos_cached, self.sin_cached = cos_cached, sin_cached + return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...] + + +class RotaryPositionalEmbeddingFunction(torch.autograd.Function): + + @staticmethod + def forward(ctx, q, cos, sin): + import rotary_positional_embedding_cuda + + q_ = q.contiguous() + cos_ = cos.contiguous() + sin_ = sin.contiguous() + output = rotary_positional_embedding_cuda.forward(*q.shape, q_, cos_, sin_) + ctx.save_for_backward(cos_, sin_) + + return output + + @staticmethod + def backward(ctx, grad_output): + import rotary_positional_embedding_cuda + + cos_, sin_ = ctx.saved_tensors + grad_q = rotary_positional_embedding_cuda.backward(*grad_output.shape, grad_output, cos_, sin_) + + return grad_q, None, None + +# rotary pos emb helpers: + +def rotate_half(x): + x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:] + return torch.cat((-x2, x1), dim=x1.ndim - 1) # dim=-1 triggers a bug in earlier torch versions + + +def apply_rotary_pos_emb_index_bhs(q, k, cos, sin, position_id): + # batch_size, num_head, seq_len, hidden_size + cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(1), \ + F.embedding(position_id, sin.squeeze(1)).unsqueeze(1) + q = (q * cos) + (rotate_half(q) * sin) + k = (k * cos) + (rotate_half(k) * sin) + return q, k + + +@torch.jit.script +def apply_rotary_pos_emb(q, k, cos, sin, offset: int = 0): + cos, sin = cos[offset:q.shape[0] + offset, ...], sin[offset:q.shape[0] + offset, ...] + return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin) + + +def apply_rotary_pos_emb_torch(q, k, cos, sin, offset: int = 0): # jitting fails with bf16 + cos, sin = cos[offset:q.shape[0] + offset, ...], sin[offset:q.shape[0] + offset, ...] + return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin) + + +def apply_rotary_pos_emb_fused(q, k, cos, sin, offset: int = 0): + cos, sin = cos[offset:q.shape[0] + offset, ...], sin[offset:q.shape[0] + offset, ...] + q = RotaryPositionalEmbeddingFunction.apply(q, cos, sin) + k = RotaryPositionalEmbeddingFunction.apply(k, cos, sin) + return q, k + + +@torch.jit.script +def apply_rotary_pos_emb_index_single(q, cos, sin, position_id): + # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn] + cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \ + F.embedding(position_id, sin.squeeze(1)).unsqueeze(2) + return (q * cos) + (rotate_half(q) * sin) + + +@torch.jit.script +def apply_rotary_pos_emb_index(q, k, cos, sin, position_id): + # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn] + cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \ + F.embedding(position_id, sin.squeeze(1)).unsqueeze(2) + q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin) + return q, k + + +def apply_rotary_pos_emb_index_torch(q, k, cos, sin, position_id): # jitting fails with bf16 + # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn] + cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \ + F.embedding(position_id, sin.squeeze(1)).unsqueeze(2) + q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin) + return q, k + + +def apply_rotary_pos_emb_index_fused(q, k, cos, sin, position_id): + # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn] + cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \ + F.embedding(position_id, sin.squeeze(1)).unsqueeze(2) + q = RotaryPositionalEmbeddingFunction.apply(q, cos, sin) + k = RotaryPositionalEmbeddingFunction.apply(k, cos, sin) + return q, k diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/transformer_defaults.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/transformer_defaults.py new file mode 100644 index 0000000000..da4a2da6ad --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/transformer_defaults.py @@ -0,0 +1,270 @@ +# coding=utf-8 +# -*- encoding: utf-8 -*- +''' +@File : transformer_defaults.py +@Time : 2022/06/01 21:44:17 +@Author : Ming Ding +@Contact : dm18@mails.tsinghua.edu.cn +''' + +import math +import torch +import torch.nn.functional as F + +from sat import mpu + +from sat.mpu.utils import split_tensor_along_last_dim + +def standard_attention(query_layer, key_layer, value_layer, attention_mask, + attention_dropout=None, log_attention_weights=None, scaling_attention_score=True, **kwargs): + # We disable the PB-relax-Attention and only changes the order of computation, because it is enough for most of training. + # The implementation in the paper can be done very easily, if you really need it to train very deep transformers. + + if scaling_attention_score: + query_layer = query_layer / math.sqrt(query_layer.shape[-1]) + attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2)) + if log_attention_weights is not None: + attention_scores += log_attention_weights + + if not (attention_mask.shape[-2] == 1 and (attention_mask > 0).all()): + # if auto-regressive, skip + attention_scores = torch.mul(attention_scores, attention_mask) - \ + 10000.0 * (1.0 - attention_mask) + + attention_probs = F.softmax(attention_scores, dim=-1) + + if attention_dropout is not None: + if mpu.get_cuda_rng_tracker is not None: + with mpu.get_cuda_rng_tracker().fork(): + attention_probs = attention_dropout(attention_probs) + else: + attention_probs = attention_dropout(attention_probs) + + context_layer = torch.matmul(attention_probs, value_layer) + return context_layer + + +def attention_fn_default(query_layer, key_layer, value_layer, attention_mask, + attention_dropout=None, log_attention_weights=None, scaling_attention_score=True, **kwargs): + # expand head dim to query dim, if necessary + # only useful for multi-query attention + batch_size, num_query_heads = query_layer.shape[:2] # [b, np, s, hn] + num_kv_heads = key_layer.shape[1] # [b, np, s, hn] + key_layer = key_layer.unsqueeze(2).expand(-1, -1, num_query_heads//num_kv_heads, -1, -1).contiguous().view(batch_size, num_query_heads, *key_layer.shape[2:]) + value_layer = value_layer.unsqueeze(2).expand(-1, -1, num_query_heads//num_kv_heads, -1, -1).contiguous().view(batch_size, num_query_heads, *value_layer.shape[2:]) + + is_low_triangle = (attention_mask == torch.ones_like(attention_mask, dtype=torch.float).tril()).all() + is_full = (attention_mask is None) or (attention_mask > 0).all() + + if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score and (is_full or is_low_triangle): + # Pytorch 2.0 attention uses very much memory if attention_mask is float, and has NaN bug if attention_mask is None. + dropout_p = 0. if attention_dropout is None or not attention_dropout.training else attention_dropout.p + return torch.nn.functional.scaled_dot_product_attention( + query_layer, key_layer, value_layer, + attn_mask=None, + dropout_p=dropout_p, + is_causal=not is_full + ) + else: + return standard_attention( + query_layer, key_layer, value_layer, attention_mask, + attention_dropout=attention_dropout, log_attention_weights=log_attention_weights, + scaling_attention_score=scaling_attention_score, **kwargs + ) + +def attention_forward_default(self, hidden_states, mask, **kw_args): + self = self.transformer.layers[kw_args['layer_id']].attention + attention_fn = attention_fn_default + if 'attention_fn' in self.hooks: + attention_fn = self.hooks['attention_fn'] + + mixed_raw_layer = self.query_key_value(hidden_states) + (mixed_query_layer, + mixed_key_layer, + mixed_value_layer) = split_tensor_along_last_dim(mixed_raw_layer, self.stride) + + dropout_fn = self.attention_dropout if self.training else None + + query_layer = self._transpose_for_scores(mixed_query_layer) + key_layer = self._transpose_for_scores(mixed_key_layer) + value_layer = self._transpose_for_scores(mixed_value_layer) + + context_layer = attention_fn(query_layer, key_layer, value_layer, mask, dropout_fn, **kw_args) + + context_layer = context_layer.permute(0, 2, 1, 3).contiguous() + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.view(*new_context_layer_shape) + output = self.dense(context_layer) + + if self.training: + output = self.output_dropout(output) + return output + +def cross_attention_forward_default(self, hidden_states, cross_attention_mask, encoder_outputs, **kw_args): + self = self.transformer.layers[kw_args['layer_id']].cross_attention + attention_fn = attention_fn_default + if 'attention_fn' in self.hooks: + attention_fn = self.hooks['attention_fn'] + + mixed_query_layer = self.query(hidden_states) + query_layer = self._transpose_for_scores(mixed_query_layer) + dropout_fn = self.attention_dropout if self.training else None + if isinstance(encoder_outputs, torch.Tensor): + mixed_x_layer = self.key_value(encoder_outputs) + (mixed_key_layer, mixed_value_layer) = split_tensor_along_last_dim(mixed_x_layer, 2) + # Reshape and transpose [b, np, s, hn] + key_layer = self._transpose_for_scores(mixed_key_layer) + value_layer = self._transpose_for_scores(mixed_value_layer) + mem_cross = (key_layer, value_layer) + else: + key_layer, value_layer = encoder_outputs[kw_args['layer_id']] + mem_cross = (key_layer, value_layer) + + context_layer = attention_fn(query_layer, key_layer, value_layer, cross_attention_mask, dropout_fn, cross_attention=True, mem_cross=mem_cross, **kw_args) + context_layer = context_layer.permute(0, 2, 1, 3).contiguous() + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + # [b, s, hp] + context_layer = context_layer.view(*new_context_layer_shape) + + # Output. [b, s, h] + output = self.dense(context_layer) + if self.training: + output = self.output_dropout(output) + return output + +def mlp_forward_default(self, hidden_states, **kw_args): + self = self.transformer.layers[kw_args['layer_id']].mlp + intermediate_parallel = self.dense_h_to_4h(hidden_states) + intermediate_parallel = self.activation_func(intermediate_parallel) + output = self.dense_4h_to_h(intermediate_parallel) + return output + +def gated_mlp_forward_default(self, hidden_states, **kw_args): + self = self.transformer.layers[kw_args['layer_id']].mlp + intermediate_parallel = self.dense_h_to_4h(hidden_states) + gated_intermediate_parallel = self.dense_h_to_4h_gate(hidden_states) + intermediate_parallel = self.activation_func(gated_intermediate_parallel) * intermediate_parallel + output = self.dense_4h_to_h(intermediate_parallel) + return output + +def word_embedding_forward_default(self, input_ids, output_cross_layer, **kw_args): + return self.transformer.word_embeddings(input_ids) + +def position_embedding_forward_default(self, position_ids, output_cross_layer, **kw_args): + return self.transformer.position_embeddings(position_ids) + +from sat.mpu import gather_from_model_parallel_region +def final_forward_default(self, logits, **kw_args): + logits_parallel = F.linear(logits, self.transformer.word_embeddings.weight) + if not kw_args['parallel_output']: + logits_parallel = gather_from_model_parallel_region(logits_parallel) + return logits_parallel + +def layer_forward_default(self, hidden_states, mask, *args, **kw_args): + ''' + hidden_states: [batch, seq_len, hidden_size] + mask: [(1, 1), seq_len, seq_len] + ''' + self = self.transformer.layers[kw_args['layer_id']] + + # Layer norm at the begining of the transformer layer. + attention_input = self.input_layernorm(hidden_states) + # Self attention. + attention_output = self.attention(attention_input, mask, **kw_args) + + # Third LayerNorm + if self.layernorm_order == 'sandwich': + attention_output = self.third_layernorm(attention_output) + + # DropPath for attention + if self.training and self.drop_path > 0.: + # drop_path percentage 0, others 1/(1-p) + random_tensor = (1-self.drop_path + + torch.rand((attention_output.shape[0],), dtype=attention_output.dtype, device=attention_output.device)).floor_() / (1-self.drop_path) + attention_output = random_tensor.view(-1, 1, 1) * attention_output + + # Residual connection. + if self.layernorm_order == 'post': + hidden_states = attention_input + attention_output + mlp_input = self.post_attention_layernorm(hidden_states) + else: + hidden_states = hidden_states + attention_output + + if self.is_decoder: + encoder_outputs = kw_args['encoder_outputs'] + if encoder_outputs is not None: + assert 'cross_attention_mask' in kw_args + # Cross attention + if self.layernorm_order == 'post': + attention_output = self.cross_attention(mlp_input, **kw_args) + # Residual connection. + hidden_states = mlp_input + attention_output + # Layer norm post the cross attention + mlp_input = self.post_cross_attention_layernorm(hidden_states) + else: + cross_input = self.post_cross_attention_layernorm(hidden_states) + attention_output = self.cross_attention(cross_input, **kw_args) + hidden_states = hidden_states + attention_output + + if self.layernorm_order != 'post': + mlp_input = self.post_attention_layernorm(hidden_states) + + # MLP. + mlp_output = self.mlp(mlp_input, **kw_args) + + # Fourth LayerNorm + if self.layernorm_order == 'sandwich': + mlp_output = self.fourth_layernorm(mlp_output) + + # DropPath for mlp + if self.training and self.drop_path > 0.: + random_tensor = (1-self.drop_path + + torch.rand((mlp_output.shape[0],), dtype=mlp_output.dtype, device=mlp_output.device)).floor_() / (1-self.drop_path) + mlp_output = random_tensor.view(-1, 1, 1) * mlp_output + + # Second residual connection. + if self.layernorm_order == 'post': + output = mlp_input + mlp_output + else: + output = hidden_states + mlp_output + + return output + +HOOKS_DEFAULT = { + 'attention_fn': attention_fn_default, + 'attention_forward': attention_forward_default, + 'cross_attention_forward': cross_attention_forward_default, + 'mlp_forward': mlp_forward_default, + 'gated_mlp_forward': gated_mlp_forward_default, + 'word_embedding_forward': word_embedding_forward_default, + 'position_embedding_forward': position_embedding_forward_default, + 'final_forward': final_forward_default, + 'layer_forward': layer_forward_default +} + +ARGS_DEFAULT = { + 'embedding_dropout_prob': ('hidden_dropout', 0), + 'attention_dropout_prob': ('attention_dropout', 0), + 'output_dropout_prob': ('hidden_dropout', 0), + 'inner_hidden_size': ('inner_hidden_size', None), + 'hidden_size_per_attention_head': ('hidden_size_per_attention_head', None), + 'cross_hidden_size_per_attention_head': ('cross_hidden_size_per_attention_head', None), + 'checkpoint_activations': ('checkpoint_activations', False), + 'checkpoint_num_layers': ('checkpoint_num_layers', 1), + 'checkpoint_skip_layers': ('checkpoint_skip_layers', 0), + 'is_decoder': ('is_decoder', False), + 'cross_attn_hidden_size': ('cross_attn_hidden_size', None), + 'use_final_layernorm': ('use_final_layernorm', True), + 'layernorm_epsilon': ('layernorm_epsilon', 1e-5), + 'use_bias': ('use_bias', True), + 'use_qkv_bias': ('use_qkv_bias', False), + 'num_multi_query_heads': ('num_multi_query_heads', 0), + 'cross_num_multi_query_heads': ('cross_num_multi_query_heads', 0), + 'drop_path': ('drop_path', 0.), + 'row_parallel_linear_final_bias': ('row_parallel_linear_final_bias', True), + 'is_gated_mlp': ('is_gated_mlp', False) +} + +from sat.ops.layernorm import LayerNorm, RMSNorm + +NO_WD_MODULES = [LayerNorm, torch.nn.LayerNorm, RMSNorm] diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/triton_rotary_embeddings.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/triton_rotary_embeddings.py new file mode 100644 index 0000000000..d11b0ac794 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/triton_rotary_embeddings.py @@ -0,0 +1,15 @@ +from .rotary_embeddings import RotaryEmbedding as FastRotaryEmbedding, rotate_half +import torch.nn.functional as F + + +def apply_rotary_pos_emb_index_bhs(q, k, cos, sin, position_id): + """ + 位置编码计算 + """ + # batch_size, num_head, seq_len, hidden_size + cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(1), \ + F.embedding(position_id, sin.squeeze(1)).unsqueeze(1) + q = (q * cos) + (rotate_half(q) * sin) + k = (k * cos) + (rotate_half(k) * sin) + return q, k + diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/visual.py b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/visual.py new file mode 100644 index 0000000000..434e603eaf --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/visual.py @@ -0,0 +1,169 @@ +import torch +import torch_npu +from torch import nn +from argparse import Namespace +from transformers.activations import ACT2FN + + +class FlashSelfAttention(torch.nn.Module): + """Implement the scaled dot product attention with softmax. + Arguments + --------- + softmax_scale: The temperature to use for the softmax attention. + (default: 1/sqrt(d_keys) where d_keys is computed at + runtime) + attention_dropout: The dropout rate to apply to the attention + (default: 0.0) + """ + + def __init__(self, causal=False, softmax_scale=1., attention_dropout=0.): + super().__init__() + self.causal = causal + self.softmax_scale = softmax_scale + self.dropout_p = attention_dropout + + def forward(self, q, k, v, n, attention_mask, pse): + + if self.causal: + output = torch_npu.npu_fusion_attention( + q, k, v, n, "BSND",# SBH + pse=pse, + padding_mask=None, + atten_mask=attention_mask, + scale=self.softmax_scale, + pre_tockens=k.shape[1], # seq_len + next_tockens=0, # 0 + keep_prob=1 - self.dropout_p, + )[0] + return output + raise Exception("the attention type {} is not support!".format(self.attention_type)) + + +class PatchEmbedding(nn.Module): + def __init__(self, config): + super().__init__() + self.proj = nn.Conv2d(config.in_channels, config.hidden_size, kernel_size=config.patch_size, stride=config.patch_size) + self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size)) + self.position_embedding = nn.Embedding(config.num_positions, config.hidden_size) + + def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)": + x = self.proj(images) + x = x.flatten(2).transpose(1, 2) + cls_token = self.cls_embedding.expand(x.shape[0], -1, -1) + x = torch.cat((cls_token, x), dim=1) + x += self.position_embedding.weight.unsqueeze(0) + return x + + +class Attention(nn.Module): + def __init__(self, config): + super().__init__() + self.num_heads = config.num_heads + head_dim = config.hidden_size // config.num_heads + self.scale = head_dim ** -0.5 + self.query_key_value = nn.Linear(config.hidden_size, config.hidden_size * 3) + self.dense = nn.Linear(config.hidden_size, config.hidden_size) + self.output_dropout = torch.nn.Dropout(config.dropout_prob) + self.core_attention_flash = FlashSelfAttention( + causal=True, softmax_scale=self.scale, attention_dropout=config.dropout_prob + ) + + def forward(self, x: "tensor(B, L, D)") -> "tensor(B, L, D)": + B, L, _ = x.shape + qkv = self.query_key_value(x) + qkv = qkv.reshape(B, L, 3, self.num_heads, -1).permute(2, 0, 1, 3, 4) # 3, B, L, H, D + q, k, v = qkv[0], qkv[1], qkv[2] + out = self.core_attention_flash(q, k, v, self.num_heads, None, None) + output = self.dense(out.view(B, L, -1)) + output = self.output_dropout(output) + return output + + def attention(self, q, k, v): + attn_weights = torch.matmul(q * self.scale, k.transpose(-2, -1)) + attn_weights = attn_weights.softmax(dim=-1) + output = torch.matmul(attn_weights, v) + return output + + +class MLP(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.activation_fn = 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, x: torch.Tensor) -> torch.Tensor: + x = self.fc1(x) + x = self.activation_fn(x) + x = self.fc2(x) + return x + + +class TransformerLayer(nn.Module): + def __init__(self, config): + super().__init__() + self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.attention = Attention(config) + self.mlp = MLP(config) + self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + def forward(self, hidden_states): + attention_input = hidden_states + attention_output = self.input_layernorm(self.attention(attention_input)) + hidden_states = attention_input + attention_output + mlp_input = hidden_states + mlp_output = self.post_attention_layernorm(self.mlp(mlp_input)) + output = mlp_input + mlp_output + return output + + +class Transformer(nn.Module): + def __init__(self, config): + super().__init__() + self.layers = nn.ModuleList([TransformerLayer(config) for _ in range(config.num_hidden_layers)]) + + def forward(self, hidden_states): + for layer_module in self.layers: + hidden_states = layer_module(hidden_states) + return hidden_states + + +class GLU(nn.Module): + def __init__(self, config, in_features): + super().__init__() + self.linear_proj = nn.Linear(in_features, config.hidden_size, bias=False) + self.norm1 = nn.LayerNorm(config.hidden_size) + self.act1 = nn.GELU() + self.act2 = nn.functional.silu + self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size, bias=False) + self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False) + self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) + + def forward(self, x): + x = self.linear_proj(x) + x = self.act1(self.norm1(x)) + x = self.act2(self.gate_proj(x)) * self.dense_h_to_4h(x) + x = self.dense_4h_to_h(x) + return x + + +class EVA2CLIPModel(nn.Module): + def __init__(self, config): + super().__init__() + vision_config = Namespace(**config.vision_config) + self.patch_embedding = PatchEmbedding(vision_config) + self.transformer = Transformer(vision_config) + self.linear_proj = GLU(config, in_features=vision_config.hidden_size) + self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) + self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) + + def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)": + x = self.patch_embedding(images) + x = self.transformer(x) + x = x[:, 1:] + x = self.linear_proj(x) + boi = self.boi.expand(x.shape[0], -1, -1) + eoi = self.eoi.expand(x.shape[0], -1, -1) + x = torch.cat((boi, x, eoi), dim=1) + return x diff --git a/PyTorch/built-in/foundation/CogVLM/fintune_cogvlm_change.sh b/PyTorch/built-in/foundation/CogVLM/fintune_cogvlm_change.sh new file mode 100644 index 0000000000..67e5e93fdf --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/fintune_cogvlm_change.sh @@ -0,0 +1,133 @@ +#!/bin/bash + +# 设置modelzoo项目下cogvlm_utils路径，例如"xxx/cogvlm_utils"，用户需要自行设定 +MODEL_ZOO_SOURCE_DIR="model_zoo路径" +# 设置CogVLM项目路径，例如"xxx/CogVLM-main"，用户需要自行设定 +COGVLM_SOURCE_DIR="Cogvlm路径" +# 设置三方件sat路径，为conda环境下三方件安装路径，例如"xxx/python3.8/site-packages/sat"，用户需要自行设定 +SAT_SOURCE_DIR="sat路径" + +# 定义一个函数，用于给文件添加modelzoobackup后缀 +rename_file() { + local ORIGINAL_FILE_PATH=$1 + local BACKUP_FILE_PATH + + # 提取原始文件的基本名称和目录路径 + local FILE_NAME=$(basename "$ORIGINAL_FILE_PATH") + local DIRECTORY_PATH=$(dirname "$ORIGINAL_FILE_PATH") + + # 构建带有backup后缀的新文件名 + local BACKUP_FILE_NAME="${FILE_NAME}.modelzoobackup" + + # 构建新文件的完整路径 + BACKUP_FILE_PATH="$DIRECTORY_PATH/$BACKUP_FILE_NAME" + + # 检查原始文件是否存在 + if [ ! -f "$ORIGINAL_FILE_PATH" ]; then + echo "Error: Original file '$ORIGINAL_FILE_PATH' does not exist." + return 1 + fi + + # 检查备份文件是否已经存在 + if [ -e "$BACKUP_FILE_PATH" ]; then + echo "Warning: Backup file '$BACKUP_FILE_PATH' already exists." + # 如果需要覆盖备份文件，可以取消注释下面的行 + # rm "$BACKUP_FILE_PATH" + fi + + # 使用mv命令创建备份文件 + mv "$ORIGINAL_FILE_PATH" "$BACKUP_FILE_PATH" + echo "Backup created: $BACKUP_FILE_PATH (from $ORIGINAL_FILE_PATH)" +} + +# 定义拷贝文件的函数 +copy_file() { + local SOURCE_FILE=$1 + local DESTINATION_FILE=$2 + + # 检查源文件是否存在 + if [ ! -f "$SOURCE_FILE" ]; then + echo "错误：源文件 '$SOURCE_FILE' 不存在。" + return 1 + fi + + # 使用cp命令拷贝文件 + cp "$SOURCE_FILE" "$DESTINATION_FILE" + + # 检查拷贝是否成功 + if [ $? -eq 0 ]; then + echo "文件拷贝成功，从 '$SOURCE_FILE' 到 '$DESTINATION_FILE'。" + else + echo "错误：文件拷贝失败。" + fi +} + +# 检查源目录是否存在 +if [ ! -d "$MODEL_ZOO_SOURCE_DIR" ]; then + echo "Error: Source directory $MODEL_ZOO_SOURCE_DIR does not exist." + exit 1 +fi + +if [ ! -d "$COGVLM_SOURCE_DIR" ]; then + echo "Error: Source directory $COGVLM_SOURCE_DIR does not exist." + exit 1 +fi + +if [ ! -d "$SAT_SOURCE_DIR" ]; then + echo "Error: Source directory $SAT_SOURCE_DIR does not exist." + exit 1 +fi + +# model_zoo path +eva_clip_model_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/eva_clip_model.py" +mixin_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/mixin.py" +finetune_cogvlm_demo_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/finetune_cogvlm_demo.py" +triton_rotary_embeddings_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/triton_rotary_embeddings.py" +transformer_defaults_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/transformer_defaults.py" +finetune_cogvlm_base_224_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/finetune_cogvlm_base_224.sh" +layernorm_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/layernorm.py" +dataset_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/dataset.py" +evaluate_cogvlm_demo_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/evaluate_cogvlm_demo.py" +eval_cogvlm_base_224_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/eval_cogvlm_base_224.sh" +env_npu_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/env_npu.sh" + +# cogvlm path +eva_clip_model_cogvlm_path="$COGVLM_SOURCE_DIR/utils/models/eva_clip_model.py" +mixin_cogvlm_path="$COGVLM_SOURCE_DIR/utils/models/mixin.py" +dataset_cogvlm_path="$COGVLM_SOURCE_DIR/utils/utils/dataset.py" +finetune_cogvlm_demo_cogvlm_path="$COGVLM_SOURCE_DIR/finetune_demo/finetune_cogvlm_demo.py" +evaluate_cogvlm_demo_cogvlm_path="$COGVLM_SOURCE_DIR/finetune_demo/evaluate_cogvlm_demo.py" +finetune_demo_cogvlm_path="$COGVLM_SOURCE_DIR/finetune_demo/" + +# sat path +triton_rotary_embeddings_sat_path="$SAT_SOURCE_DIR/model/position_embedding/triton_rotary_embeddings.py" +transformer_defaults_sat_path="$SAT_SOURCE_DIR/transformer_defaults.py" +layernorm_sat_path="$SAT_SOURCE_DIR/ops/layernorm.py" + +# cogvlm文件改名 +rename_file $eva_clip_model_cogvlm_path +rename_file $mixin_cogvlm_path +rename_file $finetune_cogvlm_demo_cogvlm_path +rename_file $evaluate_cogvlm_demo_cogvlm_path +rename_file $dataset_cogvlm_path + +# sat文件改名 +rename_file $triton_rotary_embeddings_sat_path +rename_file $transformer_defaults_sat_path +rename_file $layernorm_sat_path + +# cogvlm文件替换 +copy_file $eva_clip_model_model_zoo_path $(dirname "$eva_clip_model_cogvlm_path") +copy_file $mixin_model_zoo_path $(dirname "$mixin_cogvlm_path") +copy_file $finetune_cogvlm_demo_model_zoo_path $(dirname "$finetune_cogvlm_demo_cogvlm_path") +copy_file $dataset_model_zoo_path $(dirname "$dataset_cogvlm_path") +copy_file $finetune_cogvlm_base_224_model_zoo_path $finetune_demo_cogvlm_path +copy_file $eval_cogvlm_base_224_model_zoo_path $finetune_demo_cogvlm_path +copy_file $env_npu_model_zoo_path $finetune_demo_cogvlm_path + +# sat文件替换 +copy_file $triton_rotary_embeddings_model_zoo_path $(dirname "$triton_rotary_embeddings_sat_path") +copy_file $transformer_defaults_model_zoo_path $(dirname "$transformer_defaults_sat_path") +copy_file $layernorm_model_zoo_path $(dirname "$layernorm_sat_path") + +echo "All files copied successfully." \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/inference_cogvlm_change.sh b/PyTorch/built-in/foundation/CogVLM/inference_cogvlm_change.sh new file mode 100644 index 0000000000..1758a25b70 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/inference_cogvlm_change.sh @@ -0,0 +1,108 @@ +#!/bin/bash + +# 设置modelzoo项目下cogvlm_utils路径，例如"xxx/cogvlm_utils"，用户需要自行设定 +MODEL_ZOO_SOURCE_DIR="model_zoo路径" +# 设置CogVLM项目路径，例如"xxx/CogVLM-main"，用户需要自行设定 +COGVLM_SOURCE_DIR="Cogvlm路径" +# 设置HF权重路径，例如"xxx/models_weight/"，用户需要自行设定 +WEIGHT_SOURCE_DIR="HF权重路径" + +# 定义一个函数，用于给文件添加modelzoobackup后缀 +rename_file() { + local ORIGINAL_FILE_PATH=$1 + local BACKUP_FILE_PATH + + # 提取原始文件的基本名称和目录路径 + local FILE_NAME=$(basename "$ORIGINAL_FILE_PATH") + local DIRECTORY_PATH=$(dirname "$ORIGINAL_FILE_PATH") + + # 构建带有backup后缀的新文件名 + local BACKUP_FILE_NAME="${FILE_NAME}.modelzoobackup" + + # 构建新文件的完整路径 + BACKUP_FILE_PATH="$DIRECTORY_PATH/$BACKUP_FILE_NAME" + + # 检查原始文件是否存在 + if [ ! -f "$ORIGINAL_FILE_PATH" ]; then + echo "Error: Original file '$ORIGINAL_FILE_PATH' does not exist." + return 1 + fi + + # 检查备份文件是否已经存在 + if [ -e "$BACKUP_FILE_PATH" ]; then + echo "Warning: Backup file '$BACKUP_FILE_PATH' already exists." + # 如果需要覆盖备份文件，可以取消注释下面的行 + # rm "$BACKUP_FILE_PATH" + fi + + # 使用mv命令创建备份文件 + mv "$ORIGINAL_FILE_PATH" "$BACKUP_FILE_PATH" + echo "Backup created: $BACKUP_FILE_PATH (from $ORIGINAL_FILE_PATH)" +} + +# 定义拷贝文件的函数 +copy_file() { + local SOURCE_FILE=$1 + local DESTINATION_FILE=$2 + + # 检查源文件是否存在 + if [ ! -f "$SOURCE_FILE" ]; then + echo "错误：源文件 '$SOURCE_FILE' 不存在。" + return 1 + fi + + # 使用cp命令拷贝文件 + cp "$SOURCE_FILE" "$DESTINATION_FILE" + + # 检查拷贝是否成功 + if [ $? -eq 0 ]; then + echo "文件拷贝成功，从 '$SOURCE_FILE' 到 '$DESTINATION_FILE'。" + else + echo "错误：文件拷贝失败。" + fi +} + +# 检查源目录是否存在 +if [ ! -d "$MODEL_ZOO_SOURCE_DIR" ]; then + echo "Error: Source directory $MODEL_ZOO_SOURCE_DIR does not exist." + exit 1 +fi + +if [ ! -d "$COGVLM_SOURCE_DIR" ]; then + echo "Error: Source directory $COGVLM_SOURCE_DIR does not exist." + exit 1 +fi + +if [ ! -d "$SAT_SOURCE_DIR" ]; then + echo "Error: Source directory $SAT_SOURCE_DIR does not exist." + exit 1 +fi + +# model_zoo path +modeling_cogvlm_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/modeling_cogvlm.py" +rotary_embeddings_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/rotary_embeddings.py" +visual_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/visual.py" +inference_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/inference.py" +env_npu_model_zoo_path="$MODEL_ZOO_SOURCE_DIR/env_npu.sh" + +# hf weight path +modeling_cogvlm_weight_path="$WEIGHT_SOURCE_DIR/modeling_cogvlm.py" +visual_weight_path="$WEIGHT_SOURCE_DIR/visual.py" + +# cogvlm path +finetune_demo_cogvlm_path="$COGVLM_SOURCE_DIR/finetune_demo/" + + +# hf 权重文件改名 +rename_file $modeling_cogvlm_weight_path +rename_file $visual_weight_path + +# cogvlm文件替换 +copy_file $modeling_cogvlm_model_zoo_path $(dirname "$modeling_cogvlm_weight_path") +copy_file $rotary_embeddings_model_zoo_path $(dirname "$modeling_cogvlm_weight_path") +copy_file $visual_model_zoo_path $(dirname "$modeling_cogvlm_weight_path") +copy_file $inference_model_zoo_path $finetune_demo_cogvlm_path +copy_file $env_npu_model_zoo_path $finetune_demo_cogvlm_path + + +echo "All files copied successfully." \ No newline at end of file diff --git a/PyTorch/built-in/foundation/CogVLM/requirements.txt b/PyTorch/built-in/foundation/CogVLM/requirements.txt new file mode 100644 index 0000000000..ef17523c83 --- /dev/null +++ b/PyTorch/built-in/foundation/CogVLM/requirements.txt @@ -0,0 +1,24 @@ +numpy +torchvision==0.16.0 +SwissArmyTransformer==0.4.11 +transformers>=4.36.2 +xformers==0.0.22.post7 +spacy>=3.6.0 +pillow>=10.2.0 +deepspeed>=0.13.1 +seaborn>=0.13.2 +loguru~=0.7.2 +streamlit>=1.31.0 +timm>=0.9.12 +accelerate==0.26.1 +pydantic>=2.6.0 +scipy +decorator +streamlit +jsonlines +# for openai demo +openai>=1.11.1 +sse-starlette>=1.8.2 +fastapi>=0.109.0 +httpx>=0.26.0 +uvicorn>=0.27.0 -- Gitee From b5a6fac5dc1806d86ed6ea9ac1c0db0cd78a5ce3 Mon Sep 17 00:00:00 2001 From: gitee_code_template Date: Thu, 28 Mar 2024 16:23:37 +0800 Subject: [PATCH 2/3] =?UTF-8?q?=E6=9B=B4=E6=96=B0readme?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- PyTorch/built-in/foundation/CogVLM/README.md | 18 +++++++++++++----- .../cogvlm_utils/finetune_cogvlm_base_224.sh | 6 +++--- .../foundation/CogVLM/requirements.txt | 2 ++ 3 files changed, 18 insertions(+), 8 deletions(-) diff --git a/PyTorch/built-in/foundation/CogVLM/README.md b/PyTorch/built-in/foundation/CogVLM/README.md index e570216f16..4cef1ba0ff 100644 --- a/PyTorch/built-in/foundation/CogVLM/README.md +++ b/PyTorch/built-in/foundation/CogVLM/README.md @@ -66,10 +66,13 @@ CogVLM is VISUAL EXPERT FOR LARGE LANGUAGE 是一个多模态视觉-文本模型 |:---------:|:------:| | PyTorch | 2.1.0 | -在模型根目录下执行以下命令，安装模型对应PyTorch版本需要的依赖, 需要先安装PTA包。 +1) 安装模型对应PyTorch版本需要的依赖, 需要先安装PTA包。 +2) 下载model_zoo下面的CogVLM相关文件，并依赖其下的requirements.txt进行三方件安装。 ```shell pip install -r requirements.txt ``` +3) 下载并安装en_core_web_sm-any-py3-none-any.whl[下载](https://huggingface.co/spacy/en_core_web_sm/tree/main)， +en_core_web_sm是spaCy 自然语言处理（NLP）工具库中的一种语言模型，专为英语设计。 ### 准备数据集 1) 微调数据集: @@ -107,9 +110,6 @@ archive_split ### 获取预训练权重 -1) 下载并安装en_core_web_sm-any-py3-none-any.whl[下载](https://huggingface.co/spacy/en_core_web_sm/tree/main)， -en_core_web_sm是spaCy 自然语言处理（NLP）工具库中的一种语言模型，专为英语设计。 - 2) 官方提供微调权重cogvlm-base-224[下载](https://huggingface.co/THUDM/CogVLM/tree/main)。 3) 分词器权重[下载](https://huggingface.co/lmsys/vicuna-7b-v1.5/tree/main)。 @@ -137,9 +137,17 @@ bash finetune_cogvlm_base_224.sh #### 训练结果 +#### 随机性说明 +模型中包含多种随机问题，会影响loss曲线和下游任务，用户可根据需要自行修改，本代码不做更换： +1) Cogvlm项目路径的utils/utils/dataset.py 中load_data中的all_files是无序的，可以通过以下方式固定： +```python +all_files.sort() +``` +2) 模型本身有确定性问题 +3) SwissArmyTransformer三方件的sat/model/transformer.py文件中embedding_dropout_prob、attention_dropout_prob和output_dropout_prob三个dropout不为0 ##### 精度 -基于Captcha Images数据集训练800步、1600步和2000步验证下游任务，由于模型本身有确定性问题，因此下游任务在评估数据上略有波动： +基于Captcha Images数据集训练800步、1600步和2000步验证下游任务，由于模型本身有随机性问题，因此下游任务在评估数据上略有波动： | 芯片 | 800 | 1600 | 2000 | |:--------:|:-----:|--------|-------| diff --git a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh index 22b3e9911f..c9f825c48c 100644 --- a/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh +++ b/PyTorch/built-in/foundation/CogVLM/cogvlm_utils/finetune_cogvlm_base_224.sh @@ -20,9 +20,9 @@ OPTIONS_SAT="SAT_HOME=~/.sat_models" OPTIONS_NCCL="NCCL_DEBUG=info NCCL_IB_DISABLE=0 NCCL_NET_GDR_LEVEL=2 LOCAL_WORLD_SIZE=$NUM_GPUS_PER_WORKER" HOST_FILE_PATH="hostfile" -# 用户需要指定train_data和valid_data为实际路径 -train_data="../archive_split/train" -valid_data="../archive_split/valid" +# 用户需要指定train_data和valid_data为实际路径，例如"../archive_split/train"和"../archive_split/valid" +train_data="训练数据路径" +valid_data="验证数据路径" gpt_options=" \ --experiment-name finetune-$MODEL_TYPE \ diff --git a/PyTorch/built-in/foundation/CogVLM/requirements.txt b/PyTorch/built-in/foundation/CogVLM/requirements.txt index ef17523c83..78af436304 100644 --- a/PyTorch/built-in/foundation/CogVLM/requirements.txt +++ b/PyTorch/built-in/foundation/CogVLM/requirements.txt @@ -16,6 +16,8 @@ scipy decorator streamlit jsonlines +absl-py +cloudpickle # for openai demo openai>=1.11.1 sse-starlette>=1.8.2 -- Gitee From 57b1251e4de1d6796d2343272cc66bffaecb4c2f Mon Sep 17 00:00:00 2001 From: gitee_code_template Date: Thu, 28 Mar 2024 16:38:59 +0800 Subject: [PATCH 3/3] =?UTF-8?q?=E4=BF=AE=E6=94=B9readme?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- PyTorch/built-in/foundation/CogVLM/README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/PyTorch/built-in/foundation/CogVLM/README.md b/PyTorch/built-in/foundation/CogVLM/README.md index 4cef1ba0ff..b7029a3e22 100644 --- a/PyTorch/built-in/foundation/CogVLM/README.md +++ b/PyTorch/built-in/foundation/CogVLM/README.md @@ -67,7 +67,7 @@ CogVLM is VISUAL EXPERT FOR LARGE LANGUAGE 是一个多模态视觉-文本模型 | PyTorch | 2.1.0 | 1) 安装模型对应PyTorch版本需要的依赖, 需要先安装PTA包。 -2) 下载model_zoo下面的CogVLM相关文件，并依赖其下的requirements.txt进行三方件安装。 +2) 下载model_zoo下面的CogVLM相关文件，并依赖该路径下的requirements.txt进行三方件安装。 ```shell pip install -r requirements.txt ``` @@ -144,7 +144,7 @@ bash finetune_cogvlm_base_224.sh all_files.sort() ``` 2) 模型本身有确定性问题 -3) SwissArmyTransformer三方件的sat/model/transformer.py文件中embedding_dropout_prob、attention_dropout_prob和output_dropout_prob三个dropout不为0 +3) SwissArmyTransformer三方件的sat/model/transformer.py文件中的BaseTransformer下embedding_dropout_prob、attention_dropout_prob和output_dropout_prob三个dropout不为0 ##### 精度 基于Captcha Images数据集训练800步、1600步和2000步验证下游任务，由于模型本身有随机性问题，因此下游任务在评估数据上略有波动： -- Gitee