vllm support

README.md

@@ -1,20 +1,6 @@
 ---
 license: mit
 pipeline_tag: image-text-to-text
-library_name: transformers
-base_model:
-  - OpenGVLab/InternViT-300M-448px
-  - internlm/internlm2_5-7b-chat
-base_model_relation: merge
-language:
-  - multilingual
-tags:
-  - internvl
-  - vision
-  - ocr
-  - multi-image
-  - video
-  - custom_code
 ---
 
 # InternVL2-8B
@@ -76,13 +62,11 @@ InternVL 2.0 is a multimodal large language model series, featuring models of va
 | MathVista<sub>testmini</sub> |         54.3         |        53.5        |     58.3     |
 |  OpenCompass<sub>avg</sub>   |         58.8         |        61.7        |     64.1     |
 
-- For more details and evaluation reproduction, please refer to our [Evaluation Guide](https://internvl.readthedocs.io/en/latest/internvl2.0/evaluation.html).
-
-- We simultaneously use [InternVL](https://github.com/OpenGVLab/InternVL) and [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) repositories for model evaluation. Specifically, the results reported for DocVQA, ChartQA, InfoVQA, TextVQA, MME, AI2D, MMBench, CCBench, MMVet, and SEED-Image were tested using the InternVL repository. OCRBench, RealWorldQA, HallBench, and MathVista were evaluated using the VLMEvalKit.
+- We simultaneously use InternVL and VLMEvalKit repositories for model evaluation. Specifically, the results reported for DocVQA, ChartQA, InfoVQA, TextVQA, MME, AI2D, MMBench, CCBench, MMVet, and SEED-Image were tested using the InternVL repository. OCRBench, RealWorldQA, HallBench, and MathVista were evaluated using the VLMEvalKit.
 
 - For MMMU, we report both the original scores (left side: evaluated using the InternVL codebase for InternVL series models, and sourced from technical reports or webpages for other models) and the VLMEvalKit scores (right side: collected from the OpenCompass leaderboard).
 
-- Please note that evaluating the same model using different testing toolkits like [InternVL](https://github.com/OpenGVLab/InternVL) and [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) can result in slight differences, which is normal. Updates to code versions and variations in environment and hardware can also cause minor discrepancies in results.
+- Please note that evaluating the same model using different testing toolkits like InternVL and VLMEvalKit can result in slight differences, which is normal. Updates to code versions and variations in environment and hardware can also cause minor discrepancies in results.
 
 ### Video Benchmarks
 
@@ -144,7 +128,6 @@ model = AutoModel.from_pretrained(
     path,
     torch_dtype=torch.bfloat16,
     low_cpu_mem_usage=True,
-    use_flash_attn=True,
     trust_remote_code=True).eval().cuda()
 ```
 
@@ -159,7 +142,6 @@ model = AutoModel.from_pretrained(
     torch_dtype=torch.bfloat16,
     load_in_8bit=True,
     low_cpu_mem_usage=True,
-    use_flash_attn=True,
     trust_remote_code=True).eval()
 ```
 
@@ -174,7 +156,6 @@ model = AutoModel.from_pretrained(
     torch_dtype=torch.bfloat16,
     load_in_4bit=True,
     low_cpu_mem_usage=True,
-    use_flash_attn=True,
     trust_remote_code=True).eval()
 ```
 
@@ -219,7 +200,6 @@ model = AutoModel.from_pretrained(
     path,
     torch_dtype=torch.bfloat16,
     low_cpu_mem_usage=True,
-    use_flash_attn=True,
     trust_remote_code=True,
     device_map=device_map).eval()
 ```
@@ -315,13 +295,12 @@ model = AutoModel.from_pretrained(
     path,
     torch_dtype=torch.bfloat16,
     low_cpu_mem_usage=True,
-    use_flash_attn=True,
     trust_remote_code=True).eval().cuda()
 tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
 
 # set the max number of tiles in `max_num`
 pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
-generation_config = dict(max_new_tokens=1024, do_sample=True)
+generation_config = dict(max_new_tokens=1024, do_sample=False)
 
 # pure-text conversation (纯文本对话)
 question = 'Hello, who are you?'
@@ -473,7 +452,7 @@ for new_text in streamer:
 
 ## Finetune
 
-Many repositories now support fine-tuning of the InternVL series models, including [InternVL](https://github.com/OpenGVLab/InternVL), [SWIFT](https://github.com/modelscope/ms-swift), [XTurner](https://github.com/InternLM/xtuner), and others. Please refer to their documentation for more details on fine-tuning.
+SWIFT from ModelScope community has supported the fine-tuning (Image/Video) of InternVL, please check [this link](https://github.com/modelscope/swift/blob/main/docs/source_en/Multi-Modal/internvl-best-practice.md) for more details.
 
 ## Deployment
 
@@ -482,7 +461,7 @@ Many repositories now support fine-tuning of the InternVL series models, includi
 LMDeploy is a toolkit for compressing, deploying, and serving LLM, developed by the MMRazor and MMDeploy teams.
 
 ```sh
-pip install lmdeploy==0.5.3
+pip install lmdeploy
 ```
 
 LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
@@ -490,12 +469,16 @@ LMDeploy abstracts the complex inference process of multi-modal Vision-Language
 #### A 'Hello, world' example
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
 from lmdeploy.vl import load_image
 
 model = 'OpenGVLab/InternVL2-8B'
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
 image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 response = pipe(('describe this image', image))
 print(response.text)
 ```
@@ -509,12 +492,16 @@ When dealing with multiple images, you can put them all in one list. Keep in min
 > Warning: Due to the scarcity of multi-image conversation data, the performance on multi-image tasks may be unstable, and it may require multiple attempts to achieve satisfactory results.
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
 from lmdeploy.vl import load_image
 from lmdeploy.vl.constants import IMAGE_TOKEN
 
 model = 'OpenGVLab/InternVL2-8B'
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 
 image_urls=[
     'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
@@ -532,11 +519,15 @@ print(response.text)
 Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
 from lmdeploy.vl import load_image
 
 model = 'OpenGVLab/InternVL2-8B'
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 
 image_urls=[
     "https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
@@ -552,11 +543,15 @@ print(response)
 There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig, GenerationConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig, GenerationConfig
 from lmdeploy.vl import load_image
 
 model = 'OpenGVLab/InternVL2-8B'
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 
 image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
 gen_config = GenerationConfig(top_k=40, top_p=0.8, temperature=0.8)
@@ -568,10 +563,20 @@ print(sess.response.text)
 
 #### Service
 
+To deploy InternVL2 as an API, please configure the chat template config first. Create the following JSON file `chat_template.json`.
+
+```json
+{
+    "model_name":"internvl-internlm2",
+    "meta_instruction":"我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。",
+    "stop_words":["<|im_start|>", "<|im_end|>"]
+}
+```
+
 LMDeploy's `api_server` enables models to be easily packed into services with a single command. The provided RESTful APIs are compatible with OpenAI's interfaces. Below are an example of service startup:
 
 ```shell
-lmdeploy serve api_server OpenGVLab/InternVL2-8B --backend turbomind --server-port 23333
+lmdeploy serve api_server OpenGVLab/InternVL2-8B --backend turbomind --server-port 23333 --chat-template chat_template.json
 ```
 
 To use the OpenAI-style interface, you need to install OpenAI:
@@ -608,6 +613,14 @@ response = client.chat.completions.create(
 print(response)
 ```
 
+### vLLM
+
+TODO
+
+### Ollama
+
+TODO
+
 ## License
 
 This project is released under the MIT license, while InternLM2 is licensed under the Apache-2.0 license.
@@ -680,8 +693,6 @@ InternVL 2.0 是一个多模态大语言模型系列，包含各种规模的模
 | MathVista<sub>testmini</sub> |         54.3         |        53.5        |     58.3     |
 |  OpenCompass<sub>avg</sub>   |         58.8         |        61.7        |     64.1     |
 
-- 关于更多的细节以及评测复现，请看我们的[评测指南](https://internvl.readthedocs.io/en/latest/internvl2.0/evaluation.html)。
-
 - 我们同时使用 InternVL 和 VLMEvalKit 仓库进行模型评估。具体来说，DocVQA、ChartQA、InfoVQA、TextVQA、MME、AI2D、MMBench、CCBench、MMVet 和 SEED-Image 的结果是使用 InternVL 仓库测试的。OCRBench、RealWorldQA、HallBench 和 MathVista 是使用 VLMEvalKit 进行评估的。
 
 - 对于MMMU，我们报告了原始分数（左侧：InternVL系列模型使用InternVL代码库评测，其他模型的分数来自其技术报告或网页）和VLMEvalKit分数（右侧：从OpenCompass排行榜收集）。
@@ -740,7 +751,7 @@ InternVL 2.0 是一个多模态大语言模型系列，包含各种规模的模
 
 ## 微调
 
-许多仓库现在都支持 InternVL 系列模型的微调，包括 [InternVL](https://github.com/OpenGVLab/InternVL)、[SWIFT](https://github.com/modelscope/ms-swift)、[XTurner](https://github.com/InternLM/xtuner) 等。请参阅它们的文档以获取更多微调细节。
+来自ModelScope社区的SWIFT已经支持对InternVL进行微调（图像/视频），详情请查看[此链接](https://github.com/modelscope/swift/blob/main/docs/source_en/Multi-Modal/internvl-best-practice.md)。
 
 ## 部署
 
@@ -749,7 +760,7 @@ InternVL 2.0 是一个多模态大语言模型系列，包含各种规模的模
 LMDeploy 是由 MMRazor 和 MMDeploy 团队开发的用于压缩、部署和服务大语言模型（LLM）的工具包。
 
 ```sh
-pip install lmdeploy==0.5.3
+pip install lmdeploy
 ```
 
 LMDeploy 将多模态视觉-语言模型（VLM）的复杂推理过程抽象为一个易于使用的管道，类似于大语言模型（LLM）的推理管道。
@@ -757,12 +768,16 @@ LMDeploy 将多模态视觉-语言模型（VLM）的复杂推理过程抽象为
 #### 一个“你好，世界”示例
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
 from lmdeploy.vl import load_image
 
 model = 'OpenGVLab/InternVL2-8B'
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
 image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 response = pipe(('describe this image', image))
 print(response.text)
 ```
@@ -774,12 +789,16 @@ print(response.text)
 在处理多张图像时，可以将它们全部放入一个列表中。请注意，多张图像会导致输入 token 数量增加，因此通常需要增加上下文窗口的大小。
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
 from lmdeploy.vl import load_image
 from lmdeploy.vl.constants import IMAGE_TOKEN
 
 model = 'OpenGVLab/InternVL2-8B'
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 
 image_urls=[
     'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
@@ -787,7 +806,6 @@ image_urls=[
 ]
 
 images = [load_image(img_url) for img_url in image_urls]
-# Numbering images improves multi-image conversations
 response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
 print(response.text)
 ```
@@ -797,11 +815,15 @@ print(response.text)
 使用批量Prompt进行推理非常简单；��需将它们放在一个列表结构中：
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
 from lmdeploy.vl import load_image
 
 model = 'OpenGVLab/InternVL2-8B'
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 
 image_urls=[
     "https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
@@ -817,11 +839,15 @@ print(response)
 使用管道进行多轮对话有两种方法。一种是根据 OpenAI 的格式构建消息并使用上述方法，另一种是使用 `pipeline.chat` 接口。
 
 ```python
-from lmdeploy import pipeline, TurbomindEngineConfig, GenerationConfig
+from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig, GenerationConfig
 from lmdeploy.vl import load_image
 
 model = 'OpenGVLab/InternVL2-8B'
-pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
+system_prompt = '我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。'
+chat_template_config = ChatTemplateConfig('internvl-internlm2')
+chat_template_config.meta_instruction = system_prompt
+pipe = pipeline(model, chat_template_config=chat_template_config,
+                backend_config=TurbomindEngineConfig(session_len=8192))
 
 image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
 gen_config = GenerationConfig(top_k=40, top_p=0.8, temperature=0.8)
@@ -833,10 +859,20 @@ print(sess.response.text)
 
 #### API部署
 
+为了将InternVL2部署成API，请先配置聊天模板配置文件。创建如下的 JSON 文件 `chat_template.json`。
+
+```json
+{
+    "model_name":"internvl-internlm2",
+    "meta_instruction":"我是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。",
+    "stop_words":["<|im_start|>", "<|im_end|>"]
+}
+```
+
 LMDeploy 的 `api_server` 使模型能够通过一个命令轻松打包成服务。提供的 RESTful API 与 OpenAI 的接口兼容。以下是服务启动的示例：
 
 ```shell
-lmdeploy serve api_server OpenGVLab/InternVL2-8B --backend turbomind --server-port 23333
+lmdeploy serve api_server OpenGVLab/InternVL2-8B --backend turbomind --server-port 23333 --chat-template chat_template.json
 ```
 
 为了使用OpenAI风格的API接口，您需要安装OpenAI:
@@ -873,6 +909,14 @@ response = client.chat.completions.create(
 print(response)
 ```
 
+### vLLM
+
+TODO
+
+### Ollama
+
+TODO
+
 ## 开源许可证
 
 该项目采用 MIT 许可证发布，而 InternLM2 则采用 Apache-2.0 许可证。

configuration_intern_vit.py

@@ -3,7 +3,6 @@
 # Copyright (c) 2024 OpenGVLab
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
-
 import os
 from typing import Union
 

configuration_internvl_chat.py

@@ -47,12 +47,12 @@ class InternVLChatConfig(PretrainedConfig):
             logger.info('llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).')
 
         self.vision_config = InternVisionConfig(**vision_config)
-        if llm_config.get('architectures')[0] == 'LlamaForCausalLM':
+        if llm_config['architectures'][0] == 'LlamaForCausalLM':
             self.llm_config = LlamaConfig(**llm_config)
-        elif llm_config.get('architectures')[0] == 'InternLM2ForCausalLM':
+        elif llm_config['architectures'][0] == 'InternLM2ForCausalLM':
             self.llm_config = InternLM2Config(**llm_config)
         else:
-            raise ValueError('Unsupported architecture: {}'.format(llm_config.get('architectures')[0]))
+            raise ValueError('Unsupported architecture: {}'.format(llm_config['architectures'][0]))
         self.use_backbone_lora = use_backbone_lora
         self.use_llm_lora = use_llm_lora
         self.select_layer = select_layer

conversation.py

@@ -3,13 +3,11 @@ Conversation prompt templates.
 
 We kindly request that you import fastchat instead of copying this file if you wish to use it.
 If you have changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates.
-
-Modified from https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
 """
 
 import dataclasses
 from enum import IntEnum, auto
-from typing import Dict, List, Tuple, Union
+from typing import Any, Dict, List, Tuple, Union
 
 
 class SeparatorStyle(IntEnum):
@@ -346,6 +344,12 @@ register_conv_template(
         roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
         sep_style=SeparatorStyle.MPT,
         sep='<|im_end|>',
+        stop_token_ids=[
+            2,
+            6,
+            7,
+            8,
+        ],
         stop_str='<|endoftext|>',
     )
 )
@@ -361,6 +365,11 @@ register_conv_template(
         roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
         sep_style=SeparatorStyle.MPT,
         sep='<|im_end|>',
+        stop_token_ids=[
+            2,
+            92543,
+            92542
+        ]
     )
 )
 
@@ -375,17 +384,10 @@ register_conv_template(
         roles=('<|user|>\n', '<|assistant|>\n'),
         sep_style=SeparatorStyle.MPT,
         sep='<|end|>',
-    )
-)
-
-
-register_conv_template(
-    Conversation(
-        name='internvl2_5',
-        system_template='<|im_start|>system\n{system_message}',
-        system_message='你是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
-        roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
-        sep_style=SeparatorStyle.MPT,
-        sep='<|im_end|>\n',
+        stop_token_ids=[
+            2,
+            32000,
+            32007
+        ]
     )
 )

eval_llm_benchmark.log

@@ -1,53 +0,0 @@
-/mnt/petrelfs/wangweiyun/miniconda3/envs/internvl_eval/lib/python3.10/site-packages/bitsandbytes/cextension.py:34: UserWarning: The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable.
-  warn("The installed version of bitsandbytes was compiled without GPU support. "
-/mnt/petrelfs/wangweiyun/miniconda3/envs/internvl_eval/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cpu.so: undefined symbol: cadam32bit_grad_fp32
-model path is /mnt/petrelfs/wangweiyun/workspace_cz/InternVL/internvl_chat_dev/share_internvl/InternVL2-8B
-09/30 19:08:03 - OpenCompass - WARNING - No previous results to reuse!
-09/30 19:08:03 - OpenCompass - INFO - Reusing experiements from 20240930_190803
-09/30 19:08:03 - OpenCompass - INFO - Current exp folder: /mnt/petrelfs/wangweiyun/workspace_cz/InternVL/internvl_chat_dev/share_internvl/InternVL2-8B/20240930_190803
-09/30 19:08:06 - OpenCompass - INFO - Partitioned into 64 tasks.
-[                                                  ] 0/64, elapsed: 0s, ETA:
-09/30 19:52:33 - OpenCompass - INFO - Partitioned into 287 tasks.
-[                                                  ] 0/287, elapsed: 0s, ETA:
-dataset                       version    metric                        mode    internvl-chat-20b
-----------------------------  ---------  ----------------------------  ------  -------------------
-mmlu                          -          naive_average                 gen     73.17
-cmmlu                         -          naive_average                 gen     79.21
-ceval                         -          naive_average                 gen     80.14
-agieval                       -          -                             -       -
-GaokaoBench                   -          weighted_average              gen     74.99
-triviaqa                      2121ce     score                         gen     62.03
-triviaqa_wiki_1shot           -          -                             -       -
-nq                            3dcea1     score                         gen     28.12
-C3                            8c358f     accuracy                      gen     94.19
-race-high                     9a54b6     accuracy                      gen     90.82
-flores_100                    -          -                             -       -
-winogrande                    b36770     accuracy                      gen     85.87
-hellaswag                     e42710     accuracy                      gen     94.91
-bbh                           -          naive_average                 gen     72.67
-gsm8k                         1d7fe4     accuracy                      gen     75.59
-math                          393424     accuracy                      gen     39.50
-TheoremQA                     6f0af8     score                         gen     15.62
-MathBench                     -          -                             -       -
-openai_humaneval              8e312c     humaneval_pass@1              gen     69.51
-humanevalx                    -          -                             -       -
-sanitized_mbpp                a447ff     score                         gen     58.75
-mbpp_cn                       6fb572     score                         gen     48.20
-leval                         -          -                             -       -
-leval_closed                  -          -                             -       -
-leval_open                    -          -                             -       -
-longbench                     -          -                             -       -
-longbench_single-document-qa  -          -                             -       -
-longbench_multi-document-qa   -          -                             -       -
-longbench_summarization       -          -                             -       -
-longbench_few-shot-learning   -          -                             -       -
-longbench_synthetic-tasks     -          -                             -       -
-longbench_code-completion     -          -                             -       -
-teval                         -          -                             -       -
-teval_zh                      -          -                             -       -
-IFEval                        3321a3     Prompt-level-strict-accuracy  gen     52.31
-IFEval                        3321a3     Inst-level-strict-accuracy    gen     62.71
-IFEval                        3321a3     Prompt-level-loose-accuracy   gen     54.90
-IFEval                        3321a3     Inst-level-loose-accuracy     gen     64.87
-09/30 19:55:16 - OpenCompass - INFO - write summary to /mnt/petrelfs/wangweiyun/workspace_cz/InternVL/internvl_chat_dev/share_internvl/InternVL2-8B/20240930_190803/summary/summary_20240930_190803.txt
-09/30 19:55:16 - OpenCompass - INFO - write csv to /mnt/petrelfs/wangweiyun/workspace_cz/InternVL/internvl_chat_dev/share_internvl/InternVL2-8B/20240930_190803/summary/summary_20240930_190803.csv

generation_config.json

@@ -1,8 +1,4 @@
 {
   "_from_model_config": true,
-  "transformers_version": "4.37.2",
-  "eos_token_id": [
-    92542,
-    92543
-  ]
+  "transformers_version": "4.37.2"
 }

modeling_intern_vit.py

@@ -3,7 +3,6 @@
 # Copyright (c) 2024 OpenGVLab
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
-
 from typing import Optional, Tuple, Union
 
 import torch
@@ -21,12 +20,18 @@ from transformers.utils import logging
 from .configuration_intern_vit import InternVisionConfig
 
 try:
+    try:  # v1
+        from flash_attn.flash_attn_interface import \
+            flash_attn_unpadded_qkvpacked_func
+    except:  # v2
+        from flash_attn.flash_attn_interface import \
+            flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
+
     from flash_attn.bert_padding import pad_input, unpad_input
-    from flash_attn.flash_attn_interface import \
-        flash_attn_varlen_qkvpacked_func
+
     has_flash_attn = True
 except:
-    print('FlashAttention2 is not installed.')
+    print('FlashAttention is not installed.')
     has_flash_attn = False
 
 logger = logging.get_logger(__name__)
@@ -69,7 +74,7 @@ class FlashAttention(nn.Module):
                 max_s = seqlen
                 cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
                                           device=qkv.device)
-                output = flash_attn_varlen_qkvpacked_func(
+                output = flash_attn_unpadded_qkvpacked_func(
                     qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
                     softmax_scale=self.softmax_scale, causal=causal
                 )
@@ -79,7 +84,7 @@ class FlashAttention(nn.Module):
                 x = rearrange(qkv, 'b s three h d -> b s (three h d)')
                 x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask)
                 x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
-                output_unpad = flash_attn_varlen_qkvpacked_func(
+                output_unpad = flash_attn_unpadded_qkvpacked_func(
                     x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
                     softmax_scale=self.softmax_scale, causal=causal
                 )
@@ -88,7 +93,7 @@ class FlashAttention(nn.Module):
                                    'b s (h d) -> b s h d', h=nheads)
         else:
             assert max_s is not None
-            output = flash_attn_varlen_qkvpacked_func(
+            output = flash_attn_unpadded_qkvpacked_func(
                 qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
                 softmax_scale=self.softmax_scale, causal=causal
             )
@@ -288,9 +293,9 @@ class InternVisionEncoderLayer(nn.Module):
         Args:
             hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
         """
-        hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states).to(hidden_states.dtype)) * self.ls1)
+        hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states)) * self.ls1)
 
-        hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states).to(hidden_states.dtype)) * self.ls2)
+        hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states)) * self.ls2)
 
         return hidden_states
 

modeling_internvl_chat.py

@@ -3,9 +3,8 @@
 # Copyright (c) 2024 OpenGVLab
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
-
 import warnings
-from typing import List, Optional, Tuple, Union
+from typing import Any, List, Optional, Tuple, Union
 
 import torch.utils.checkpoint
 import transformers
@@ -19,7 +18,7 @@ from transformers.utils import ModelOutput, logging
 
 from .configuration_internvl_chat import InternVLChatConfig
 from .conversation import get_conv_template
-from .modeling_intern_vit import InternVisionModel, has_flash_attn
+from .modeling_intern_vit import InternVisionModel
 from .modeling_internlm2 import InternLM2ForCausalLM
 
 logger = logging.get_logger(__name__)
@@ -36,11 +35,10 @@ def version_cmp(v1, v2, op='eq'):
 class InternVLChatModel(PreTrainedModel):
     config_class = InternVLChatConfig
     main_input_name = 'pixel_values'
-    base_model_prefix = 'language_model'
     _supports_flash_attn_2 = True
     _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'InternLM2DecoderLayer']
 
-    def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None, use_flash_attn=True):
+    def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None):
         super().__init__(config)
 
         assert version_cmp(transformers.__version__, '4.36.2', 'ge')
@@ -52,9 +50,6 @@ class InternVLChatModel(PreTrainedModel):
         self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
         self.downsample_ratio = config.downsample_ratio
         self.ps_version = config.ps_version
-        use_flash_attn = use_flash_attn if has_flash_attn else False
-        config.vision_config.use_flash_attn = True if use_flash_attn else False
-        config.llm_config.attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager'
 
         logger.info(f'num_image_token: {self.num_image_token}')
         logger.info(f'ps_version: {self.ps_version}')
@@ -103,7 +98,7 @@ class InternVLChatModel(PreTrainedModel):
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         image_flags = image_flags.squeeze(-1)
-        input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
+        input_embeds = self.language_model.get_input_embeddings()(input_ids)
 
         vit_embeds = self.extract_feature(pixel_values)
         vit_embeds = vit_embeds[image_flags == 1]
@@ -236,9 +231,9 @@ class InternVLChatModel(PreTrainedModel):
 
         tokenizer.padding_side = 'left'
         model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
-        input_ids = model_inputs['input_ids'].to(self.device)
-        attention_mask = model_inputs['attention_mask'].to(self.device)
-        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
+        input_ids = model_inputs['input_ids'].cuda()
+        attention_mask = model_inputs['attention_mask'].cuda()
+        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
         generation_config['eos_token_id'] = eos_token_id
         generation_output = self.generate(
             pixel_values=pixel_values,
@@ -247,7 +242,7 @@ class InternVLChatModel(PreTrainedModel):
             **generation_config
         )
         responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
-        responses = [response.split(template.sep.strip())[0].strip() for response in responses]
+        responses = [response.split(template.sep)[0].strip() for response in responses]
         return responses
 
     def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
@@ -266,7 +261,7 @@ class InternVLChatModel(PreTrainedModel):
 
         template = get_conv_template(self.template)
         template.system_message = self.system_message
-        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
+        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
 
         history = [] if history is None else history
         for (old_question, old_answer) in history:
@@ -285,8 +280,8 @@ class InternVLChatModel(PreTrainedModel):
             query = query.replace('<image>', image_tokens, 1)
 
         model_inputs = tokenizer(query, return_tensors='pt')
-        input_ids = model_inputs['input_ids'].to(self.device)
-        attention_mask = model_inputs['attention_mask'].to(self.device)
+        input_ids = model_inputs['input_ids'].cuda()
+        attention_mask = model_inputs['attention_mask'].cuda()
         generation_config['eos_token_id'] = eos_token_id
         generation_output = self.generate(
             pixel_values=pixel_values,
@@ -295,7 +290,7 @@ class InternVLChatModel(PreTrainedModel):
             **generation_config
         )
         response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
-        response = response.split(template.sep.strip())[0].strip()
+        response = response.split(template.sep)[0].strip()
         history.append((question, response))
         if return_history:
             return response, history
@@ -315,6 +310,7 @@ class InternVLChatModel(PreTrainedModel):
             visual_features: Optional[torch.FloatTensor] = None,
             generation_config: Optional[GenerationConfig] = None,
             output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
             **generate_kwargs,
     ) -> torch.LongTensor:
 
@@ -342,6 +338,7 @@ class InternVLChatModel(PreTrainedModel):
             attention_mask=attention_mask,
             generation_config=generation_config,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             use_cache=True,
             **generate_kwargs,
         )