Fix Multi GPU Issue
#4
by
jackchris121
- opened
- .gitattributes +0 -1
- README.md +116 -422
- all_results.json +8 -0
- config.json +70 -11
- configuration_intern_vit.py +1 -3
- configuration_internvl_chat.py +9 -2
- conversation.py +889 -11
- examples/image1.jpg +0 -0
- examples/image2.jpg +0 -0
- generation_config.json +1 -5
- modeling_intern_vit.py +22 -24
- modeling_internlm2.py +4 -23
- modeling_internvl_chat.py +120 -99
- preprocessor_config.json +0 -19
- examples/red-panda.mp4 → runs/Apr15_16-44-40_SH-IDC1-10-140-37-13/events.out.tfevents.1713171220.SH-IDC1-10-140-37-13.204150.0 +2 -2
- runs/Apr15_17-33-22_SH-IDC1-10-140-37-13/events.out.tfevents.1713174123.SH-IDC1-10-140-37-13.259480.0 +3 -0
.gitattributes
CHANGED
@@ -33,4 +33,3 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
-
examples/red-panda.mp4 filter=lfs diff=lfs merge=lfs -text
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
README.md
CHANGED
@@ -1,186 +1,110 @@
|
|
1 |
---
|
2 |
license: mit
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
tags:
|
12 |
-
- internvl
|
13 |
-
- vision
|
14 |
-
- ocr
|
15 |
-
- multi-image
|
16 |
-
- video
|
17 |
-
- custom_code
|
18 |
---
|
19 |
|
20 |
-
# InternVL-Chat-V1
|
21 |
-
|
22 |
-
[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL) [\[🆕 Blog\]](https://internvl.github.io/blog/) [\[📜 InternVL 1.0 Paper\]](https://arxiv.org/abs/2312.14238) [\[📜 InternVL 1.5 Report\]](https://arxiv.org/abs/2404.16821)
|
23 |
-
|
24 |
-
[\[🗨️ Chat Demo\]](https://internvl.opengvlab.com/) [\[🤗 HF Demo\]](https://huggingface.co/spaces/OpenGVLab/InternVL) [\[🚀 Quick Start\]](#quick-start) [\[📖 中文解读\]](https://zhuanlan.zhihu.com/p/706547971) [\[📖 Documents\]](https://internvl.readthedocs.io/en/latest/)
|
25 |
-
|
26 |
-
## Introduction
|
27 |
-
|
28 |
<p align="center">
|
29 |
-
<img src="https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/D60YzQBIzvoCvLRp2gZ0A.jpeg" alt="Image Description" width="300" height="300"
|
30 |
</p>
|
31 |
|
32 |
> _Two interns holding hands, symbolizing the integration of InternViT and InternLM._
|
33 |
|
34 |
-
|
35 |
|
36 |
-
We introduce
|
|
|
|
|
|
|
|
|
37 |
|
38 |
-
1. **Strong Vision Encoder:** we explored a continuous learning strategy for the large-scale vision foundation model---InternViT-6B, boosting its visual understanding capabilities, and making it can be transferred and reused in different LLMs.
|
39 |
-
2. **Dynamic High-Resolution:** we divide images into tiles ranging from 1 to 40 of 448 × 448 pixels according to the aspect ratio and resolution of the input images, which supports up to 4K resolution input during inference.
|
40 |
-
3. **High-Quality Bilingual Dataset:** we carefully collected a high-quality bilingual dataset that covers common scenes, document images, and annotated them with English and Chinese question-answer pairs, significantly enhancing performance in OCR- and Chinese-related tasks.
|
41 |
|
42 |
## Model Details
|
43 |
-
|
44 |
- **Model Type:** multimodal large language model (MLLM)
|
45 |
-
|
46 |
- **Model Stats:**
|
47 |
-
|
48 |
- Architecture: [InternViT-6B-448px-V1-5](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-5) + MLP + [InternLM2-Chat-20B](https://huggingface.co/internlm/internlm2-chat-20b)
|
49 |
- Image size: dynamic resolution, max to 40 tiles of 448 x 448 (4K resolution).
|
50 |
- Params: 25.5B
|
51 |
|
52 |
- **Training Strategy:**
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
## Performance
|
63 |
|
64 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/
|
65 |
|
66 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/i2vp6zSHPS3UIr-1Q9cSe.png)
|
67 |
|
68 |
-
|
69 |
-
|
70 |
-
- 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.
|
71 |
-
|
72 |
-
Limitations: Although we have made efforts to ensure the safety of the model during the training process and to encourage the model to generate text that complies with ethical and legal requirements, the model may still produce unexpected outputs due to its size and probabilistic generation paradigm. For example, the generated responses may contain biases, discrimination, or other harmful content. Please do not propagate such content. We are not responsible for any consequences resulting from the dissemination of harmful information.
|
73 |
|
74 |
## Examples
|
75 |
|
76 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/
|
77 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/ivhj4QqcO2NHUa28DTDkK.png)
|
78 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/18GeOW10QVcSt5g--TgDY.png)
|
79 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/tGM_TwdV297H1fCxQ0PZU.png)
|
80 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/FwlSRBpKgURAVkXNOLoSp.png)
|
81 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/to3nOaAnyv-fGLEoNPLzz.png)
|
82 |
-
|
83 |
-
## Quick Start
|
84 |
|
85 |
-
|
86 |
|
87 |
-
|
88 |
|
89 |
-
|
90 |
|
91 |
-
|
92 |
|
93 |
-
|
94 |
|
95 |
-
|
96 |
-
import torch
|
97 |
-
from transformers import AutoTokenizer, AutoModel
|
98 |
-
path = "OpenGVLab/InternVL-Chat-V1-5"
|
99 |
-
model = AutoModel.from_pretrained(
|
100 |
-
path,
|
101 |
-
torch_dtype=torch.bfloat16,
|
102 |
-
low_cpu_mem_usage=True,
|
103 |
-
use_flash_attn=True,
|
104 |
-
trust_remote_code=True).eval().cuda()
|
105 |
-
```
|
106 |
|
107 |
-
|
108 |
-
|
109 |
-
```python
|
110 |
-
import torch
|
111 |
-
from transformers import AutoTokenizer, AutoModel
|
112 |
-
path = "OpenGVLab/InternVL-Chat-V1-5"
|
113 |
-
model = AutoModel.from_pretrained(
|
114 |
-
path,
|
115 |
-
torch_dtype=torch.bfloat16,
|
116 |
-
load_in_8bit=True,
|
117 |
-
low_cpu_mem_usage=True,
|
118 |
-
use_flash_attn=True,
|
119 |
-
trust_remote_code=True).eval()
|
120 |
-
```
|
121 |
|
122 |
-
|
123 |
|
124 |
-
|
125 |
|
126 |
-
|
127 |
|
128 |
-
The reason for writing the code this way is to avoid errors that occur during multi-GPU inference due to tensors not being on the same device. By ensuring that the first and last layers of the large language model (LLM) are on the same device, we prevent such errors.
|
129 |
|
130 |
-
```python
|
131 |
-
import math
|
132 |
-
import torch
|
133 |
-
from transformers import AutoTokenizer, AutoModel
|
134 |
|
135 |
-
|
136 |
-
device_map = {}
|
137 |
-
world_size = torch.cuda.device_count()
|
138 |
-
num_layers = {'Mini-InternVL-2B-V1-5': 24, 'Mini-InternVL-4B-V1-5': 32, 'InternVL-Chat-V1-5': 48}[model_name]
|
139 |
-
# Since the first GPU will be used for ViT, treat it as half a GPU.
|
140 |
-
num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
|
141 |
-
num_layers_per_gpu = [num_layers_per_gpu] * world_size
|
142 |
-
num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
|
143 |
-
layer_cnt = 0
|
144 |
-
for i, num_layer in enumerate(num_layers_per_gpu):
|
145 |
-
for j in range(num_layer):
|
146 |
-
device_map[f'language_model.model.layers.{layer_cnt}'] = i
|
147 |
-
layer_cnt += 1
|
148 |
-
device_map['vision_model'] = 0
|
149 |
-
device_map['mlp1'] = 0
|
150 |
-
device_map['language_model.model.tok_embeddings'] = 0
|
151 |
-
device_map['language_model.model.embed_tokens'] = 0
|
152 |
-
device_map['language_model.output'] = 0
|
153 |
-
device_map['language_model.model.norm'] = 0
|
154 |
-
device_map['language_model.lm_head'] = 0
|
155 |
-
device_map[f'language_model.model.layers.{num_layers - 1}'] = 0
|
156 |
-
|
157 |
-
return device_map
|
158 |
|
159 |
-
|
160 |
-
device_map = split_model('InternVL-Chat-V1-5')
|
161 |
-
model = AutoModel.from_pretrained(
|
162 |
-
path,
|
163 |
-
torch_dtype=torch.bfloat16,
|
164 |
-
low_cpu_mem_usage=True,
|
165 |
-
use_flash_attn=True,
|
166 |
-
trust_remote_code=True,
|
167 |
-
device_map=device_map).eval()
|
168 |
-
```
|
169 |
|
170 |
-
|
171 |
|
172 |
```python
|
173 |
-
import
|
|
|
|
|
|
|
174 |
import torch
|
175 |
import torchvision.transforms as T
|
176 |
-
from decord import VideoReader, cpu
|
177 |
from PIL import Image
|
|
|
178 |
from torchvision.transforms.functional import InterpolationMode
|
179 |
-
|
180 |
|
181 |
IMAGENET_MEAN = (0.485, 0.456, 0.406)
|
182 |
IMAGENET_STD = (0.229, 0.224, 0.225)
|
183 |
|
|
|
184 |
def build_transform(input_size):
|
185 |
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
|
186 |
transform = T.Compose([
|
@@ -191,6 +115,7 @@ def build_transform(input_size):
|
|
191 |
])
|
192 |
return transform
|
193 |
|
|
|
194 |
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
|
195 |
best_ratio_diff = float('inf')
|
196 |
best_ratio = (1, 1)
|
@@ -206,7 +131,8 @@ def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_
|
|
206 |
best_ratio = ratio
|
207 |
return best_ratio
|
208 |
|
209 |
-
|
|
|
210 |
orig_width, orig_height = image.size
|
211 |
aspect_ratio = orig_width / orig_height
|
212 |
|
@@ -244,7 +170,8 @@ def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbna
|
|
244 |
processed_images.append(thumbnail_img)
|
245 |
return processed_images
|
246 |
|
247 |
-
|
|
|
248 |
image = Image.open(image_file).convert('RGB')
|
249 |
transform = build_transform(input_size=input_size)
|
250 |
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
@@ -252,310 +179,75 @@ def load_image(image_file, input_size=448, max_num=12):
|
|
252 |
pixel_values = torch.stack(pixel_values)
|
253 |
return pixel_values
|
254 |
|
|
|
|
|
255 |
# If you have an 80G A100 GPU, you can put the entire model on a single GPU.
|
256 |
-
# Otherwise, you need to load a model using multiple GPUs, please refer to the `Multiple GPUs` section.
|
257 |
-
path = 'OpenGVLab/InternVL-Chat-V1-5'
|
258 |
model = AutoModel.from_pretrained(
|
259 |
path,
|
260 |
torch_dtype=torch.bfloat16,
|
261 |
low_cpu_mem_usage=True,
|
262 |
-
use_flash_attn=True,
|
263 |
trust_remote_code=True).eval().cuda()
|
264 |
-
|
265 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
266 |
# set the max number of tiles in `max_num`
|
267 |
-
pixel_values = load_image('./examples/image1.jpg', max_num=
|
268 |
-
generation_config = dict(max_new_tokens=1024, do_sample=True)
|
269 |
-
|
270 |
-
# pure-text conversation (纯文本对话)
|
271 |
-
question = 'Hello, who are you?'
|
272 |
-
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
|
273 |
-
print(f'User: {question}\nAssistant: {response}')
|
274 |
|
275 |
-
|
276 |
-
|
277 |
-
|
|
|
|
|
278 |
|
279 |
-
# single-
|
280 |
-
question =
|
281 |
response = model.chat(tokenizer, pixel_values, question, generation_config)
|
282 |
-
print(
|
283 |
|
284 |
-
# single-image
|
285 |
-
question =
|
286 |
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
|
287 |
-
print(
|
288 |
|
289 |
-
question =
|
290 |
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
|
291 |
-
print(
|
292 |
-
|
293 |
-
# multi-image multi-round conversation, combined images (多图多轮对话,拼接图像)
|
294 |
-
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
295 |
-
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
|
296 |
-
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
297 |
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
print(f'User: {question}\nAssistant: {response}')
|
302 |
-
|
303 |
-
question = 'What are the similarities and differences between these two images.'
|
304 |
-
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
305 |
-
history=history, return_history=True)
|
306 |
-
print(f'User: {question}\nAssistant: {response}')
|
307 |
-
|
308 |
-
# multi-image multi-round conversation, separate images (多图多轮对话,独立图像)
|
309 |
-
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
310 |
-
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
|
311 |
-
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
312 |
-
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
|
313 |
-
|
314 |
-
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
|
315 |
-
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
316 |
-
num_patches_list=num_patches_list,
|
317 |
-
history=None, return_history=True)
|
318 |
-
print(f'User: {question}\nAssistant: {response}')
|
319 |
-
|
320 |
-
question = 'What are the similarities and differences between these two images.'
|
321 |
-
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
322 |
-
num_patches_list=num_patches_list,
|
323 |
-
history=history, return_history=True)
|
324 |
-
print(f'User: {question}\nAssistant: {response}')
|
325 |
-
|
326 |
-
# batch inference, single image per sample (单图批处理)
|
327 |
-
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
328 |
-
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
|
329 |
-
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
|
330 |
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
331 |
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
print(f'User: {question}\nAssistant: {response}')
|
339 |
-
|
340 |
-
# video multi-round conversation (视频多轮对话)
|
341 |
-
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
|
342 |
-
if bound:
|
343 |
-
start, end = bound[0], bound[1]
|
344 |
-
else:
|
345 |
-
start, end = -100000, 100000
|
346 |
-
start_idx = max(first_idx, round(start * fps))
|
347 |
-
end_idx = min(round(end * fps), max_frame)
|
348 |
-
seg_size = float(end_idx - start_idx) / num_segments
|
349 |
-
frame_indices = np.array([
|
350 |
-
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
|
351 |
-
for idx in range(num_segments)
|
352 |
-
])
|
353 |
-
return frame_indices
|
354 |
-
|
355 |
-
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
|
356 |
-
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
|
357 |
-
max_frame = len(vr) - 1
|
358 |
-
fps = float(vr.get_avg_fps())
|
359 |
-
|
360 |
-
pixel_values_list, num_patches_list = [], []
|
361 |
-
transform = build_transform(input_size=input_size)
|
362 |
-
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
|
363 |
-
for frame_index in frame_indices:
|
364 |
-
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
|
365 |
-
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
366 |
-
pixel_values = [transform(tile) for tile in img]
|
367 |
-
pixel_values = torch.stack(pixel_values)
|
368 |
-
num_patches_list.append(pixel_values.shape[0])
|
369 |
-
pixel_values_list.append(pixel_values)
|
370 |
-
pixel_values = torch.cat(pixel_values_list)
|
371 |
-
return pixel_values, num_patches_list
|
372 |
-
|
373 |
-
video_path = './examples/red-panda.mp4'
|
374 |
-
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
|
375 |
-
pixel_values = pixel_values.to(torch.bfloat16).cuda()
|
376 |
-
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
|
377 |
-
question = video_prefix + 'What is the red panda doing?'
|
378 |
-
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
|
379 |
-
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
380 |
-
num_patches_list=num_patches_list, history=None, return_history=True)
|
381 |
-
print(f'User: {question}\nAssistant: {response}')
|
382 |
-
|
383 |
-
question = 'Describe this video in detail. Don\'t repeat.'
|
384 |
-
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
385 |
-
num_patches_list=num_patches_list, history=history, return_history=True)
|
386 |
-
print(f'User: {question}\nAssistant: {response}')
|
387 |
-
```
|
388 |
-
|
389 |
-
#### Streaming output
|
390 |
-
|
391 |
-
Besides this method, you can also use the following code to get streamed output.
|
392 |
-
|
393 |
-
```python
|
394 |
-
from transformers import TextIteratorStreamer
|
395 |
-
from threading import Thread
|
396 |
-
|
397 |
-
# Initialize the streamer
|
398 |
-
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
|
399 |
-
# Define the generation configuration
|
400 |
-
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
|
401 |
-
# Start the model chat in a separate thread
|
402 |
-
thread = Thread(target=model.chat, kwargs=dict(
|
403 |
-
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
|
404 |
-
history=None, return_history=False, generation_config=generation_config,
|
405 |
-
))
|
406 |
-
thread.start()
|
407 |
-
|
408 |
-
# Initialize an empty string to store the generated text
|
409 |
-
generated_text = ''
|
410 |
-
# Loop through the streamer to get the new text as it is generated
|
411 |
-
for new_text in streamer:
|
412 |
-
if new_text == model.conv_template.sep:
|
413 |
-
break
|
414 |
-
generated_text += new_text
|
415 |
-
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
|
416 |
-
```
|
417 |
-
|
418 |
-
## Finetune
|
419 |
-
|
420 |
-
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.
|
421 |
-
|
422 |
-
## Deployment
|
423 |
-
|
424 |
-
### LMDeploy
|
425 |
-
|
426 |
-
LMDeploy is a toolkit for compressing, deploying, and serving LLM, developed by the MMRazor and MMDeploy teams.
|
427 |
-
|
428 |
-
```sh
|
429 |
-
pip install lmdeploy==0.5.3
|
430 |
-
```
|
431 |
-
|
432 |
-
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.
|
433 |
-
|
434 |
-
#### A 'Hello, world' example
|
435 |
-
|
436 |
-
```python
|
437 |
-
from lmdeploy import pipeline, TurbomindEngineConfig
|
438 |
-
from lmdeploy.vl import load_image
|
439 |
-
|
440 |
-
model = 'OpenGVLab/InternVL-Chat-V1-5'
|
441 |
-
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
|
442 |
-
pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
|
443 |
-
response = pipe(('describe this image', image))
|
444 |
-
print(response.text)
|
445 |
-
```
|
446 |
-
|
447 |
-
If `ImportError` occurs while executing this case, please install the required dependency packages as prompted.
|
448 |
-
|
449 |
-
#### Multi-images inference
|
450 |
-
|
451 |
-
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
|
452 |
-
|
453 |
-
> 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.
|
454 |
-
|
455 |
-
```python
|
456 |
-
from lmdeploy import pipeline, TurbomindEngineConfig
|
457 |
-
from lmdeploy.vl import load_image
|
458 |
-
from lmdeploy.vl.constants import IMAGE_TOKEN
|
459 |
-
|
460 |
-
model = 'OpenGVLab/InternVL-Chat-V1-5'
|
461 |
-
pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
|
462 |
-
|
463 |
-
image_urls=[
|
464 |
-
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
|
465 |
-
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
|
466 |
-
]
|
467 |
-
|
468 |
-
images = [load_image(img_url) for img_url in image_urls]
|
469 |
-
# Numbering images improves multi-image conversations
|
470 |
-
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
|
471 |
-
print(response.text)
|
472 |
-
```
|
473 |
-
|
474 |
-
#### Batch prompts inference
|
475 |
-
|
476 |
-
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
|
477 |
-
|
478 |
-
```python
|
479 |
-
from lmdeploy import pipeline, TurbomindEngineConfig
|
480 |
-
from lmdeploy.vl import load_image
|
481 |
-
|
482 |
-
model = 'OpenGVLab/InternVL-Chat-V1-5'
|
483 |
-
pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
|
484 |
-
|
485 |
-
image_urls=[
|
486 |
-
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
|
487 |
-
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
|
488 |
-
]
|
489 |
-
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
|
490 |
-
response = pipe(prompts)
|
491 |
-
print(response)
|
492 |
-
```
|
493 |
-
|
494 |
-
#### Multi-turn conversation
|
495 |
-
|
496 |
-
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.
|
497 |
-
|
498 |
-
```python
|
499 |
-
from lmdeploy import pipeline, TurbomindEngineConfig, GenerationConfig
|
500 |
-
from lmdeploy.vl import load_image
|
501 |
-
|
502 |
-
model = 'OpenGVLab/InternVL-Chat-V1-5'
|
503 |
-
pipe = pipeline(model, backend_config=TurbomindEngineConfig(session_len=8192))
|
504 |
-
|
505 |
-
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
|
506 |
-
gen_config = GenerationConfig(top_k=40, top_p=0.8, temperature=0.8)
|
507 |
-
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
|
508 |
-
print(sess.response.text)
|
509 |
-
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
|
510 |
-
print(sess.response.text)
|
511 |
-
```
|
512 |
-
|
513 |
-
#### Service
|
514 |
-
|
515 |
-
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:
|
516 |
-
|
517 |
-
```shell
|
518 |
-
lmdeploy serve api_server OpenGVLab/InternVL-Chat-V1-5 --backend turbomind --server-port 23333
|
519 |
-
```
|
520 |
-
|
521 |
-
To use the OpenAI-style interface, you need to install OpenAI:
|
522 |
-
|
523 |
-
```shell
|
524 |
-
pip install openai
|
525 |
-
```
|
526 |
-
|
527 |
-
Then, use the code below to make the API call:
|
528 |
|
529 |
-
|
530 |
-
|
531 |
-
|
532 |
-
|
533 |
-
|
534 |
-
|
535 |
-
|
536 |
-
|
537 |
-
|
538 |
-
|
539 |
-
|
540 |
-
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
-
'image_url': {
|
545 |
-
'url':
|
546 |
-
'https://modelscope.oss-cn-beijing.aliyuncs.com/resource/tiger.jpeg',
|
547 |
-
},
|
548 |
-
}],
|
549 |
-
}],
|
550 |
-
temperature=0.8,
|
551 |
-
top_p=0.8)
|
552 |
-
print(response)
|
553 |
```
|
554 |
|
555 |
-
## License
|
556 |
-
|
557 |
-
This project is released under the MIT license, while InternLM2 is licensed under the Apache-2.0 license.
|
558 |
-
|
559 |
## Citation
|
560 |
|
561 |
If you find this project useful in your research, please consider citing:
|
@@ -567,10 +259,12 @@ If you find this project useful in your research, please consider citing:
|
|
567 |
journal={arXiv preprint arXiv:2312.14238},
|
568 |
year={2023}
|
569 |
}
|
570 |
-
@article{chen2024far,
|
571 |
-
title={How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites},
|
572 |
-
author={Chen, Zhe and Wang, Weiyun and Tian, Hao and Ye, Shenglong and Gao, Zhangwei and Cui, Erfei and Tong, Wenwen and Hu, Kongzhi and Luo, Jiapeng and Ma, Zheng and others},
|
573 |
-
journal={arXiv preprint arXiv:2404.16821},
|
574 |
-
year={2024}
|
575 |
-
}
|
576 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
+
datasets:
|
4 |
+
- laion/laion2B-en
|
5 |
+
- laion/laion-coco
|
6 |
+
- laion/laion2B-multi
|
7 |
+
- kakaobrain/coyo-700m
|
8 |
+
- conceptual_captions
|
9 |
+
- wanng/wukong100m
|
10 |
+
pipeline_tag: visual-question-answering
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
---
|
12 |
|
13 |
+
# Model Card for InternVL-Chat-V1.5
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
<p align="center">
|
15 |
+
<img src="https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/D60YzQBIzvoCvLRp2gZ0A.jpeg" alt="Image Description" width="300" height="300" />
|
16 |
</p>
|
17 |
|
18 |
> _Two interns holding hands, symbolizing the integration of InternViT and InternLM._
|
19 |
|
20 |
+
\[[Paper](https://arxiv.org/abs/2312.14238)\] \[[GitHub](https://github.com/OpenGVLab/InternVL)\] \[[Chat Demo](https://internvl.opengvlab.com/)\] \[[中文解读](https://zhuanlan.zhihu.com/p/675877376)]
|
21 |
|
22 |
+
We introduce InternVL 1.5, an open-source multimodal large language model (MLLM) to bridge the capability gap between open-source and proprietary commercial models in multimodal understanding.
|
23 |
+
We introduce three simple designs:
|
24 |
+
1. Strong Vision Encoder: we explored a continuous learning strategy for the large-scale vision foundation model---InternViT-6B, boosting its visual understanding capabilities, and making it can be transferred and reused in different LLMs.
|
25 |
+
2. Dynamic High-Resolution: we divide images into tiles ranging from 1 to 40 of 448 × 448 pixels according to the aspect ratio and resolution of the input images, which supports up to 4K resolution input.
|
26 |
+
3. High-Quality Bilingual Dataset: we carefully collected a high-quality bilingual dataset that covers common scenes, document images, and annotated them with English and Chinese question-answer pairs, significantly enhancing performance in OCR- and Chinese-related tasks.
|
27 |
|
|
|
|
|
|
|
28 |
|
29 |
## Model Details
|
|
|
30 |
- **Model Type:** multimodal large language model (MLLM)
|
|
|
31 |
- **Model Stats:**
|
|
|
32 |
- Architecture: [InternViT-6B-448px-V1-5](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-5) + MLP + [InternLM2-Chat-20B](https://huggingface.co/internlm/internlm2-chat-20b)
|
33 |
- Image size: dynamic resolution, max to 40 tiles of 448 x 448 (4K resolution).
|
34 |
- Params: 25.5B
|
35 |
|
36 |
- **Training Strategy:**
|
37 |
+
- Pretraining Stage
|
38 |
+
- Learnable Component: ViT + MLP
|
39 |
+
- Data: Please see our technical report.
|
40 |
+
- SFT Stage
|
41 |
+
- Learnable Component: ViT + MLP + LLM
|
42 |
+
- Data: Please see our technical report.
|
43 |
+
|
44 |
+
## Released Models
|
45 |
+
|
46 |
+
| Model | Vision Foundation Model | Release Date |Note |
|
47 |
+
| :---------------------------------------------------------:|:--------------------------------------------------------------------------: |:----------------------:| :---------------------------------- |
|
48 |
+
| InternVL-Chat-V1.5(🤗 [HF link](https://huggingface.co/OpenGVLab/InternVL-Chat-V1-5)) | InternViT-6B-448px-V1-5(🤗 [HF link](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-5)) |2024.04.18 | support 4K image; super strong OCR; Approaching the performance of GPT-4V and Gemini Pro on various benchmarks like MMMU, DocVQA, ChartQA, MathVista, etc. (🔥new)|
|
49 |
+
| InternVL-Chat-V1.2-Plus(🤗 [HF link](https://huggingface.co/OpenGVLab/InternVL-Chat-V1-2-Plus) ) |InternViT-6B-448px-V1-2(🤗 [HF link](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-2)) |2024.02.21 | more SFT data and stronger |
|
50 |
+
| InternVL-Chat-V1.2(🤗 [HF link](https://huggingface.co/OpenGVLab/InternVL-Chat-V1-2) ) |InternViT-6B-448px-V1-2(🤗 [HF link](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-2)) |2024.02.11 | scaling up LLM to 34B |
|
51 |
+
| InternVL-Chat-V1.1(🤗 [HF link](https://huggingface.co/OpenGVLab/InternVL-Chat-V1-1)) |InternViT-6B-448px-V1-0(🤗 [HF link](https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-0)) |2024.01.24 | support Chinese and stronger OCR |
|
52 |
|
53 |
## Performance
|
54 |
|
55 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/ZyQklQ3C7C60I-xOv7X8L.png)
|
56 |
|
|
|
57 |
|
58 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/u98oqlnpZtWdq2dnarVlD.png)
|
|
|
|
|
|
|
|
|
59 |
|
60 |
## Examples
|
61 |
|
62 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/R34jISP4K1U17m9yNP38O.png)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
|
64 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/ChkU9XtlsjH0l2EqlO_is.png)
|
65 |
|
66 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/1TFxIcf96ANRPLoy4-rbh.png)
|
67 |
|
68 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/Wpjo1Sdwf7XcEDevqwcr-.png)
|
69 |
|
70 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/kO4-J38sN8TFtmQ5mIBMS.png)
|
71 |
|
72 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/qPnTe3Q9UBy8wbclOsmWk.png)
|
73 |
|
74 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/l_BILRi13CbZNzbZYn6o6.png)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
|
76 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/2782y7RnvGBogYEIG__7S.png)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
|
78 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/RyO35PTH14OFiwyxtAZM2.png)
|
79 |
|
80 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/xiLZXWL-JiCTVPnV_VxS2.png)
|
81 |
|
82 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64119264f0f81eb569e0d569/gqX46Tt5jvrcVqb0vcf06.png)
|
83 |
|
|
|
84 |
|
|
|
|
|
|
|
|
|
85 |
|
86 |
+
## Model Usage
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
+
We provide an example code to run InternVL-Chat-V1.5 using `transformers`.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
|
90 |
+
You also can use our [online demo](https://internvl.opengvlab.com/) for a quick experience of this model.
|
91 |
|
92 |
```python
|
93 |
+
import json
|
94 |
+
import os
|
95 |
+
from transformers import AutoTokenizer, AutoModel
|
96 |
+
from tqdm import tqdm
|
97 |
import torch
|
98 |
import torchvision.transforms as T
|
|
|
99 |
from PIL import Image
|
100 |
+
|
101 |
from torchvision.transforms.functional import InterpolationMode
|
102 |
+
|
103 |
|
104 |
IMAGENET_MEAN = (0.485, 0.456, 0.406)
|
105 |
IMAGENET_STD = (0.229, 0.224, 0.225)
|
106 |
|
107 |
+
|
108 |
def build_transform(input_size):
|
109 |
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
|
110 |
transform = T.Compose([
|
|
|
115 |
])
|
116 |
return transform
|
117 |
|
118 |
+
|
119 |
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
|
120 |
best_ratio_diff = float('inf')
|
121 |
best_ratio = (1, 1)
|
|
|
131 |
best_ratio = ratio
|
132 |
return best_ratio
|
133 |
|
134 |
+
|
135 |
+
def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
|
136 |
orig_width, orig_height = image.size
|
137 |
aspect_ratio = orig_width / orig_height
|
138 |
|
|
|
170 |
processed_images.append(thumbnail_img)
|
171 |
return processed_images
|
172 |
|
173 |
+
|
174 |
+
def load_image(image_file, input_size=448, max_num=6):
|
175 |
image = Image.open(image_file).convert('RGB')
|
176 |
transform = build_transform(input_size=input_size)
|
177 |
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
|
|
179 |
pixel_values = torch.stack(pixel_values)
|
180 |
return pixel_values
|
181 |
|
182 |
+
|
183 |
+
path = "OpenGVLab/InternVL-Chat-V1-5"
|
184 |
# If you have an 80G A100 GPU, you can put the entire model on a single GPU.
|
|
|
|
|
185 |
model = AutoModel.from_pretrained(
|
186 |
path,
|
187 |
torch_dtype=torch.bfloat16,
|
188 |
low_cpu_mem_usage=True,
|
|
|
189 |
trust_remote_code=True).eval().cuda()
|
190 |
+
# Otherwise, you need to set device_map='auto' to use multiple GPUs for inference.
|
191 |
+
# model = AutoModel.from_pretrained(
|
192 |
+
# path,
|
193 |
+
# torch_dtype=torch.bfloat16,
|
194 |
+
# low_cpu_mem_usage=True,
|
195 |
+
# trust_remote_code=True,
|
196 |
+
# device_map='auto').eval()
|
197 |
+
|
198 |
+
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
|
199 |
# set the max number of tiles in `max_num`
|
200 |
+
pixel_values = load_image('./examples/image1.jpg', max_num=6).to(torch.bfloat16).cuda()
|
|
|
|
|
|
|
|
|
|
|
|
|
201 |
|
202 |
+
generation_config = dict(
|
203 |
+
num_beams=1,
|
204 |
+
max_new_tokens=512,
|
205 |
+
do_sample=False,
|
206 |
+
)
|
207 |
|
208 |
+
# single-round single-image conversation
|
209 |
+
question = "请详细描述图片"
|
210 |
response = model.chat(tokenizer, pixel_values, question, generation_config)
|
211 |
+
print(question, response)
|
212 |
|
213 |
+
# multi-round single-image conversation
|
214 |
+
question = "请详细描述图片"
|
215 |
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
|
216 |
+
print(question, response)
|
217 |
|
218 |
+
question = "请根据图片写一首诗"
|
219 |
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
|
220 |
+
print(question, response)
|
|
|
|
|
|
|
|
|
|
|
221 |
|
222 |
+
# multi-round multi-image conversation
|
223 |
+
pixel_values1 = load_image('./examples/image1.jpg', max_num=6).to(torch.bfloat16).cuda()
|
224 |
+
pixel_values2 = load_image('./examples/image2.jpg', max_num=6).to(torch.bfloat16).cuda()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
226 |
|
227 |
+
question = "详细描述这两张图片"
|
228 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
|
229 |
+
print(question, response)
|
230 |
+
# 第一张图片是一只红熊猫,它有着独特的橙红色皮毛,脸部、耳朵和四肢的末端有白色斑块。红熊猫的眼睛周围有深色的环,它的耳朵是圆形的,上面有白色的毛。它正坐在一个木制的结构上,看起来像是一个平台或休息的地方。背景中有树木和竹子,这表明红熊猫可能在一个模拟自然环境的动物园或保护区内。
|
231 |
+
#
|
232 |
+
# 第二张图片是一只大熊猫,它是中国的国宝,以其黑白相间的皮毛而闻名。大熊猫的眼睛、耳朵和四肢的末端是黑色的,而它的脸部、耳朵内侧和身体其他部分是白色的。大熊猫正坐在地上,周围有竹子,这是它们的主要食物来源。背景中也有树木,这表明大熊猫可能在一个为它们提供自然栖息地模拟的动物园或保护区内。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
233 |
|
234 |
+
question = "这两张图片的相同点和区别分别是什么"
|
235 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
|
236 |
+
print(question, response)
|
237 |
+
# 这两张图片的相同点:
|
238 |
+
#
|
239 |
+
# 1. 都展示了熊猫,这是两种不同的熊猫物种。
|
240 |
+
# 2. 熊猫都处于一个看起来像是模拟自然环境的场所,可能是动物园或保护区。
|
241 |
+
# 3. 熊猫周围都有竹子,这是它们的主要食物来源。
|
242 |
+
#
|
243 |
+
# 这两张图片的区别:
|
244 |
+
#
|
245 |
+
# 1. 熊猫的种类不同:第一张图片是一只红熊猫,第二张图片是一只大熊猫。
|
246 |
+
# 2. 熊猫的皮毛颜色和图案不同:红熊猫的皮毛是橙红色,脸部、耳朵和四肢的末端有白色斑块;而大熊猫的皮毛是黑白相间的,眼睛、耳朵和四肢的末端是黑色的,脸部、耳朵内侧和身体其他部分是白色的。
|
247 |
+
# 3. 熊猫的姿态和位置不同:红熊猫坐在一个木制的结构上,而大熊猫坐在地上。
|
248 |
+
# 4. 背景中的植被和环境细节略有不同,但都包含树木和竹子。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
249 |
```
|
250 |
|
|
|
|
|
|
|
|
|
251 |
## Citation
|
252 |
|
253 |
If you find this project useful in your research, please consider citing:
|
|
|
259 |
journal={arXiv preprint arXiv:2312.14238},
|
260 |
year={2023}
|
261 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
262 |
```
|
263 |
+
|
264 |
+
## License
|
265 |
+
|
266 |
+
This project is released under the MIT license.
|
267 |
+
|
268 |
+
## Acknowledgement
|
269 |
+
|
270 |
+
InternVL is built with reference to the code of the following projects: [OpenAI CLIP](https://github.com/openai/CLIP), [Open CLIP](https://github.com/mlfoundations/open_clip), [CLIP Benchmark](https://github.com/LAION-AI/CLIP_benchmark), [EVA](https://github.com/baaivision/EVA/tree/master), [InternImage](https://github.com/OpenGVLab/InternImage), [ViT-Adapter](https://github.com/czczup/ViT-Adapter), [MMSegmentation](https://github.com/open-mmlab/mmsegmentation), [Transformers](https://github.com/huggingface/transformers), [DINOv2](https://github.com/facebookresearch/dinov2), [BLIP-2](https://github.com/salesforce/LAVIS/tree/main/projects/blip2), [Qwen-VL](https://github.com/QwenLM/Qwen-VL/tree/master/eval_mm), and [LLaVA-1.5](https://github.com/haotian-liu/LLaVA). Thanks for their awesome work!
|
all_results.json
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"epoch": 1.0,
|
3 |
+
"train_loss": 0.8170236018231988,
|
4 |
+
"train_runtime": 190400.5325,
|
5 |
+
"train_samples": 5155291,
|
6 |
+
"train_samples_per_second": 27.076,
|
7 |
+
"train_steps_per_second": 0.026
|
8 |
+
}
|
config.json
CHANGED
@@ -1,19 +1,19 @@
|
|
1 |
{
|
2 |
"_commit_hash": null,
|
|
|
3 |
"architectures": [
|
4 |
"InternVLChatModel"
|
5 |
],
|
6 |
"auto_map": {
|
7 |
"AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
|
8 |
-
"AutoModel": "modeling_internvl_chat.InternVLChatModel"
|
9 |
-
"AutoModelForCausalLM": "modeling_internvl_chat.InternVLChatModel"
|
10 |
},
|
11 |
-
"system_message": "You are an AI assistant whose name is InternLM (书生·浦语).",
|
12 |
"downsample_ratio": 0.5,
|
13 |
"dynamic_image_size": true,
|
14 |
"force_image_size": 448,
|
|
|
15 |
"llm_config": {
|
16 |
-
"_name_or_path": "
|
17 |
"add_cross_attention": false,
|
18 |
"architectures": [
|
19 |
"InternLM2ForCausalLM"
|
@@ -95,49 +95,108 @@
|
|
95 |
"top_p": 1.0,
|
96 |
"torch_dtype": "bfloat16",
|
97 |
"torchscript": false,
|
98 |
-
"transformers_version": "4.
|
99 |
"typical_p": 1.0,
|
100 |
-
"use_bfloat16":
|
101 |
-
"use_cache":
|
102 |
"vocab_size": 92553
|
103 |
},
|
104 |
-
"max_dynamic_patch":
|
105 |
"min_dynamic_patch": 1,
|
106 |
"model_type": "internvl_chat",
|
|
|
107 |
"ps_version": "v2",
|
108 |
"select_layer": -1,
|
109 |
"template": "internlm2-chat",
|
110 |
"torch_dtype": "bfloat16",
|
|
|
111 |
"use_backbone_lora": 0,
|
112 |
"use_llm_lora": 0,
|
113 |
"use_thumbnail": true,
|
114 |
"vision_config": {
|
|
|
|
|
115 |
"architectures": [
|
116 |
"InternVisionModel"
|
117 |
],
|
118 |
"attention_dropout": 0.0,
|
119 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
"dropout": 0.0,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
"hidden_act": "gelu",
|
122 |
"hidden_size": 3200,
|
|
|
|
|
|
|
|
|
123 |
"image_size": 448,
|
124 |
"initializer_factor": 0.1,
|
125 |
"initializer_range": 1e-10,
|
126 |
"intermediate_size": 12800,
|
|
|
|
|
|
|
|
|
|
|
|
|
127 |
"layer_norm_eps": 1e-06,
|
|
|
|
|
|
|
128 |
"model_type": "intern_vit_6b",
|
129 |
-
"
|
130 |
"num_attention_heads": 25,
|
|
|
|
|
131 |
"num_channels": 3,
|
132 |
"num_hidden_layers": 45,
|
|
|
133 |
"output_attentions": false,
|
134 |
"output_hidden_states": false,
|
|
|
|
|
135 |
"patch_size": 14,
|
|
|
|
|
|
|
136 |
"qk_normalization": true,
|
137 |
"qkv_bias": false,
|
|
|
|
|
138 |
"return_dict": true,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
"torch_dtype": "bfloat16",
|
140 |
-
"
|
|
|
|
|
141 |
"use_bfloat16": true,
|
142 |
"use_flash_attn": true
|
143 |
}
|
|
|
1 |
{
|
2 |
"_commit_hash": null,
|
3 |
+
"_name_or_path": "./work_dirs/internvl_chat_internlm2_20b_448_dynamic_chinese_pretrain3/checkpoint-1600_replace_llm",
|
4 |
"architectures": [
|
5 |
"InternVLChatModel"
|
6 |
],
|
7 |
"auto_map": {
|
8 |
"AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
|
9 |
+
"AutoModel": "modeling_internvl_chat.InternVLChatModel"
|
|
|
10 |
},
|
|
|
11 |
"downsample_ratio": 0.5,
|
12 |
"dynamic_image_size": true,
|
13 |
"force_image_size": 448,
|
14 |
+
"image_fold": null,
|
15 |
"llm_config": {
|
16 |
+
"_name_or_path": "pretrained/internlm2-chat-20b/",
|
17 |
"add_cross_attention": false,
|
18 |
"architectures": [
|
19 |
"InternLM2ForCausalLM"
|
|
|
95 |
"top_p": 1.0,
|
96 |
"torch_dtype": "bfloat16",
|
97 |
"torchscript": false,
|
98 |
+
"transformers_version": "4.36.2",
|
99 |
"typical_p": 1.0,
|
100 |
+
"use_bfloat16": false,
|
101 |
+
"use_cache": false,
|
102 |
"vocab_size": 92553
|
103 |
},
|
104 |
+
"max_dynamic_patch": 6,
|
105 |
"min_dynamic_patch": 1,
|
106 |
"model_type": "internvl_chat",
|
107 |
+
"pad2square": false,
|
108 |
"ps_version": "v2",
|
109 |
"select_layer": -1,
|
110 |
"template": "internlm2-chat",
|
111 |
"torch_dtype": "bfloat16",
|
112 |
+
"transformers_version": null,
|
113 |
"use_backbone_lora": 0,
|
114 |
"use_llm_lora": 0,
|
115 |
"use_thumbnail": true,
|
116 |
"vision_config": {
|
117 |
+
"_name_or_path": "work_dirs/internvl_chat_internlm2_20b_448_dynamic_chinese_pretrain/checkpoint-5200-vit",
|
118 |
+
"add_cross_attention": false,
|
119 |
"architectures": [
|
120 |
"InternVisionModel"
|
121 |
],
|
122 |
"attention_dropout": 0.0,
|
123 |
+
"auto_map": {
|
124 |
+
"AutoConfig": "configuration_intern_vit.InternVisionConfig",
|
125 |
+
"AutoModel": "modeling_intern_vit.InternVisionModel"
|
126 |
+
},
|
127 |
+
"bad_words_ids": null,
|
128 |
+
"begin_suppress_tokens": null,
|
129 |
+
"bos_token_id": null,
|
130 |
+
"chunk_size_feed_forward": 0,
|
131 |
+
"cross_attention_hidden_size": null,
|
132 |
+
"decoder_start_token_id": null,
|
133 |
+
"diversity_penalty": 0.0,
|
134 |
+
"do_sample": false,
|
135 |
+
"drop_path_rate": 0.4,
|
136 |
"dropout": 0.0,
|
137 |
+
"early_stopping": false,
|
138 |
+
"encoder_no_repeat_ngram_size": 0,
|
139 |
+
"eos_token_id": null,
|
140 |
+
"exponential_decay_length_penalty": null,
|
141 |
+
"finetuning_task": null,
|
142 |
+
"forced_bos_token_id": null,
|
143 |
+
"forced_eos_token_id": null,
|
144 |
"hidden_act": "gelu",
|
145 |
"hidden_size": 3200,
|
146 |
+
"id2label": {
|
147 |
+
"0": "LABEL_0",
|
148 |
+
"1": "LABEL_1"
|
149 |
+
},
|
150 |
"image_size": 448,
|
151 |
"initializer_factor": 0.1,
|
152 |
"initializer_range": 1e-10,
|
153 |
"intermediate_size": 12800,
|
154 |
+
"is_decoder": false,
|
155 |
+
"is_encoder_decoder": false,
|
156 |
+
"label2id": {
|
157 |
+
"LABEL_0": 0,
|
158 |
+
"LABEL_1": 1
|
159 |
+
},
|
160 |
"layer_norm_eps": 1e-06,
|
161 |
+
"length_penalty": 1.0,
|
162 |
+
"max_length": 20,
|
163 |
+
"min_length": 0,
|
164 |
"model_type": "intern_vit_6b",
|
165 |
+
"no_repeat_ngram_size": 0,
|
166 |
"num_attention_heads": 25,
|
167 |
+
"num_beam_groups": 1,
|
168 |
+
"num_beams": 1,
|
169 |
"num_channels": 3,
|
170 |
"num_hidden_layers": 45,
|
171 |
+
"num_return_sequences": 1,
|
172 |
"output_attentions": false,
|
173 |
"output_hidden_states": false,
|
174 |
+
"output_scores": false,
|
175 |
+
"pad_token_id": null,
|
176 |
"patch_size": 14,
|
177 |
+
"prefix": null,
|
178 |
+
"problem_type": null,
|
179 |
+
"pruned_heads": {},
|
180 |
"qk_normalization": true,
|
181 |
"qkv_bias": false,
|
182 |
+
"remove_invalid_values": false,
|
183 |
+
"repetition_penalty": 1.0,
|
184 |
"return_dict": true,
|
185 |
+
"return_dict_in_generate": false,
|
186 |
+
"sep_token_id": null,
|
187 |
+
"suppress_tokens": null,
|
188 |
+
"task_specific_params": null,
|
189 |
+
"temperature": 1.0,
|
190 |
+
"tf_legacy_loss": false,
|
191 |
+
"tie_encoder_decoder": false,
|
192 |
+
"tie_word_embeddings": true,
|
193 |
+
"tokenizer_class": null,
|
194 |
+
"top_k": 50,
|
195 |
+
"top_p": 1.0,
|
196 |
"torch_dtype": "bfloat16",
|
197 |
+
"torchscript": false,
|
198 |
+
"transformers_version": "4.36.2",
|
199 |
+
"typical_p": 1.0,
|
200 |
"use_bfloat16": true,
|
201 |
"use_flash_attn": true
|
202 |
}
|
configuration_intern_vit.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
-
# Copyright (c)
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
import os
|
@@ -73,7 +73,6 @@ class InternVisionConfig(PretrainedConfig):
|
|
73 |
num_hidden_layers=48,
|
74 |
use_flash_attn=True,
|
75 |
hidden_act='gelu',
|
76 |
-
norm_type='rms_norm',
|
77 |
layer_norm_eps=1e-6,
|
78 |
dropout=0.0,
|
79 |
drop_path_rate=0.0,
|
@@ -98,7 +97,6 @@ class InternVisionConfig(PretrainedConfig):
|
|
98 |
self.attention_dropout = attention_dropout
|
99 |
self.layer_norm_eps = layer_norm_eps
|
100 |
self.hidden_act = hidden_act
|
101 |
-
self.norm_type = norm_type
|
102 |
self.qkv_bias = qkv_bias
|
103 |
self.qk_normalization = qk_normalization
|
104 |
self.use_flash_attn = use_flash_attn
|
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
+
# Copyright (c) 2023 OpenGVLab
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
import os
|
|
|
73 |
num_hidden_layers=48,
|
74 |
use_flash_attn=True,
|
75 |
hidden_act='gelu',
|
|
|
76 |
layer_norm_eps=1e-6,
|
77 |
dropout=0.0,
|
78 |
drop_path_rate=0.0,
|
|
|
97 |
self.attention_dropout = attention_dropout
|
98 |
self.layer_norm_eps = layer_norm_eps
|
99 |
self.hidden_act = hidden_act
|
|
|
100 |
self.qkv_bias = qkv_bias
|
101 |
self.qk_normalization = qk_normalization
|
102 |
self.use_flash_attn = use_flash_attn
|
configuration_internvl_chat.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
-
# Copyright (c)
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
|
@@ -26,10 +26,12 @@ class InternVLChatConfig(PretrainedConfig):
|
|
26 |
llm_config=None,
|
27 |
use_backbone_lora=0,
|
28 |
use_llm_lora=0,
|
29 |
-
|
|
|
30 |
force_image_size=None,
|
31 |
downsample_ratio=0.5,
|
32 |
template=None,
|
|
|
33 |
dynamic_image_size=False,
|
34 |
use_thumbnail=False,
|
35 |
ps_version='v1',
|
@@ -55,10 +57,12 @@ class InternVLChatConfig(PretrainedConfig):
|
|
55 |
raise ValueError('Unsupported architecture: {}'.format(llm_config['architectures'][0]))
|
56 |
self.use_backbone_lora = use_backbone_lora
|
57 |
self.use_llm_lora = use_llm_lora
|
|
|
58 |
self.select_layer = select_layer
|
59 |
self.force_image_size = force_image_size
|
60 |
self.downsample_ratio = downsample_ratio
|
61 |
self.template = template
|
|
|
62 |
self.dynamic_image_size = dynamic_image_size
|
63 |
self.use_thumbnail = use_thumbnail
|
64 |
self.ps_version = ps_version # pixel shuffle version
|
@@ -66,6 +70,7 @@ class InternVLChatConfig(PretrainedConfig):
|
|
66 |
self.max_dynamic_patch = max_dynamic_patch
|
67 |
|
68 |
logger.info(f'vision_select_layer: {self.select_layer}')
|
|
|
69 |
logger.info(f'ps_version: {self.ps_version}')
|
70 |
logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}')
|
71 |
logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}')
|
@@ -83,10 +88,12 @@ class InternVLChatConfig(PretrainedConfig):
|
|
83 |
output['model_type'] = self.__class__.model_type
|
84 |
output['use_backbone_lora'] = self.use_backbone_lora
|
85 |
output['use_llm_lora'] = self.use_llm_lora
|
|
|
86 |
output['select_layer'] = self.select_layer
|
87 |
output['force_image_size'] = self.force_image_size
|
88 |
output['downsample_ratio'] = self.downsample_ratio
|
89 |
output['template'] = self.template
|
|
|
90 |
output['dynamic_image_size'] = self.dynamic_image_size
|
91 |
output['use_thumbnail'] = self.use_thumbnail
|
92 |
output['ps_version'] = self.ps_version
|
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
+
# Copyright (c) 2023 OpenGVLab
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
|
|
|
26 |
llm_config=None,
|
27 |
use_backbone_lora=0,
|
28 |
use_llm_lora=0,
|
29 |
+
pad2square=False,
|
30 |
+
select_layer=-4,
|
31 |
force_image_size=None,
|
32 |
downsample_ratio=0.5,
|
33 |
template=None,
|
34 |
+
image_fold=False,
|
35 |
dynamic_image_size=False,
|
36 |
use_thumbnail=False,
|
37 |
ps_version='v1',
|
|
|
57 |
raise ValueError('Unsupported architecture: {}'.format(llm_config['architectures'][0]))
|
58 |
self.use_backbone_lora = use_backbone_lora
|
59 |
self.use_llm_lora = use_llm_lora
|
60 |
+
self.pad2square = pad2square
|
61 |
self.select_layer = select_layer
|
62 |
self.force_image_size = force_image_size
|
63 |
self.downsample_ratio = downsample_ratio
|
64 |
self.template = template
|
65 |
+
self.image_fold = image_fold
|
66 |
self.dynamic_image_size = dynamic_image_size
|
67 |
self.use_thumbnail = use_thumbnail
|
68 |
self.ps_version = ps_version # pixel shuffle version
|
|
|
70 |
self.max_dynamic_patch = max_dynamic_patch
|
71 |
|
72 |
logger.info(f'vision_select_layer: {self.select_layer}')
|
73 |
+
logger.info(f'image_fold: {self.image_fold}')
|
74 |
logger.info(f'ps_version: {self.ps_version}')
|
75 |
logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}')
|
76 |
logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}')
|
|
|
88 |
output['model_type'] = self.__class__.model_type
|
89 |
output['use_backbone_lora'] = self.use_backbone_lora
|
90 |
output['use_llm_lora'] = self.use_llm_lora
|
91 |
+
output['pad2square'] = self.pad2square
|
92 |
output['select_layer'] = self.select_layer
|
93 |
output['force_image_size'] = self.force_image_size
|
94 |
output['downsample_ratio'] = self.downsample_ratio
|
95 |
output['template'] = self.template
|
96 |
+
output['image_fold'] = self.image_fold
|
97 |
output['dynamic_image_size'] = self.dynamic_image_size
|
98 |
output['use_thumbnail'] = self.use_thumbnail
|
99 |
output['ps_version'] = self.ps_version
|
conversation.py
CHANGED
@@ -2,7 +2,7 @@
|
|
2 |
Conversation prompt templates.
|
3 |
|
4 |
We kindly request that you import fastchat instead of copying this file if you wish to use it.
|
5 |
-
If you have changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates.
|
6 |
"""
|
7 |
|
8 |
import dataclasses
|
@@ -330,6 +330,384 @@ def get_conv_template(name: str) -> Conversation:
|
|
330 |
return conv_templates[name].copy()
|
331 |
|
332 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
333 |
register_conv_template(
|
334 |
Conversation(
|
335 |
name='Hermes-2',
|
@@ -343,7 +721,7 @@ register_conv_template(
|
|
343 |
6,
|
344 |
7,
|
345 |
8,
|
346 |
-
],
|
347 |
stop_str='<|endoftext|>',
|
348 |
)
|
349 |
)
|
@@ -365,19 +743,519 @@ register_conv_template(
|
|
365 |
)
|
366 |
)
|
367 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
368 |
|
|
|
|
|
|
|
369 |
register_conv_template(
|
370 |
Conversation(
|
371 |
-
name='
|
372 |
-
|
373 |
-
|
374 |
-
|
375 |
-
|
376 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
377 |
stop_token_ids=[
|
|
|
|
|
378 |
2,
|
379 |
-
|
380 |
-
|
381 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
382 |
)
|
383 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
Conversation prompt templates.
|
3 |
|
4 |
We kindly request that you import fastchat instead of copying this file if you wish to use it.
|
5 |
+
If you have any changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates.
|
6 |
"""
|
7 |
|
8 |
import dataclasses
|
|
|
330 |
return conv_templates[name].copy()
|
331 |
|
332 |
|
333 |
+
# An empty template for raw conversation.
|
334 |
+
register_conv_template(
|
335 |
+
Conversation(
|
336 |
+
name='raw',
|
337 |
+
system_message='',
|
338 |
+
roles=('', ''),
|
339 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
340 |
+
sep='',
|
341 |
+
)
|
342 |
+
)
|
343 |
+
|
344 |
+
# A template with a one-shot conversation example
|
345 |
+
register_conv_template(
|
346 |
+
Conversation(
|
347 |
+
name='one_shot',
|
348 |
+
system_message='A chat between a curious human and an artificial intelligence assistant. '
|
349 |
+
"The assistant gives helpful, detailed, and polite answers to the human's questions.",
|
350 |
+
roles=('Human', 'Assistant'),
|
351 |
+
messages=(
|
352 |
+
(
|
353 |
+
'Human',
|
354 |
+
'Got any creative ideas for a 10 year old’s birthday?',
|
355 |
+
),
|
356 |
+
(
|
357 |
+
'Assistant',
|
358 |
+
"""Of course! Here are some creative ideas for a 10-year-old's birthday party:
|
359 |
+
1. Treasure Hunt: Organize a treasure hunt in your backyard or nearby park. Create clues and riddles for the kids to solve, leading them to hidden treasures and surprises.
|
360 |
+
2. Science Party: Plan a science-themed party where kids can engage in fun and interactive experiments. You can set up different stations with activities like making slime, erupting volcanoes, or creating simple chemical reactions.
|
361 |
+
3. Outdoor Movie Night: Set up a backyard movie night with a projector and a large screen or white sheet. Create a cozy seating area with blankets and pillows, and serve popcorn and snacks while the kids enjoy a favorite movie under the stars.
|
362 |
+
4. DIY Crafts Party: Arrange a craft party where kids can unleash their creativity. Provide a variety of craft supplies like beads, paints, and fabrics, and let them create their own unique masterpieces to take home as party favors.
|
363 |
+
5. Sports Olympics: Host a mini Olympics event with various sports and games. Set up different stations for activities like sack races, relay races, basketball shooting, and obstacle courses. Give out medals or certificates to the participants.
|
364 |
+
6. Cooking Party: Have a cooking-themed party where the kids can prepare their own mini pizzas, cupcakes, or cookies. Provide toppings, frosting, and decorating supplies, and let them get hands-on in the kitchen.
|
365 |
+
7. Superhero Training Camp: Create a superhero-themed party where the kids can engage in fun training activities. Set up an obstacle course, have them design their own superhero capes or masks, and organize superhero-themed games and challenges.
|
366 |
+
8. Outdoor Adventure: Plan an outdoor adventure party at a local park or nature reserve. Arrange activities like hiking, nature scavenger hunts, or a picnic with games. Encourage exploration and appreciation for the outdoors.
|
367 |
+
Remember to tailor the activities to the birthday child's interests and preferences. Have a great celebration!""",
|
368 |
+
),
|
369 |
+
),
|
370 |
+
offset=2,
|
371 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
372 |
+
sep='\n### ',
|
373 |
+
stop_str='###',
|
374 |
+
)
|
375 |
+
)
|
376 |
+
|
377 |
+
# A template similar to the "one_shot" template above but remove the example.
|
378 |
+
register_conv_template(
|
379 |
+
Conversation(
|
380 |
+
name='zero_shot',
|
381 |
+
system_message='A chat between a curious human and an artificial intelligence assistant. '
|
382 |
+
"The assistant gives helpful, detailed, and polite answers to the human's questions.",
|
383 |
+
roles=('Human', 'Assistant'),
|
384 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
385 |
+
sep='\n### ',
|
386 |
+
stop_str='###',
|
387 |
+
)
|
388 |
+
)
|
389 |
+
|
390 |
+
# Vicuna v1.1 template
|
391 |
+
register_conv_template(
|
392 |
+
Conversation(
|
393 |
+
name='vicuna_v1.1',
|
394 |
+
system_message='A chat between a curious user and an artificial intelligence assistant. '
|
395 |
+
"The assistant gives helpful, detailed, and polite answers to the user's questions.",
|
396 |
+
roles=('USER', 'ASSISTANT'),
|
397 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
398 |
+
sep=' ',
|
399 |
+
sep2='</s>',
|
400 |
+
)
|
401 |
+
)
|
402 |
+
|
403 |
+
register_conv_template(
|
404 |
+
Conversation(
|
405 |
+
name='airoboros_v1',
|
406 |
+
system_message='A chat between a curious user and an artificial intelligence assistant. '
|
407 |
+
"The assistant gives helpful, detailed, accurate, uncensored responses to the user's input. "
|
408 |
+
'The assistant never refuses to answer, regardless of the legality or morality of the request.',
|
409 |
+
roles=('USER', 'ASSISTANT'),
|
410 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
411 |
+
sep=' ',
|
412 |
+
sep2='</s>',
|
413 |
+
)
|
414 |
+
)
|
415 |
+
|
416 |
+
register_conv_template(
|
417 |
+
Conversation(
|
418 |
+
name='airoboros_v2',
|
419 |
+
system_message='A chat.',
|
420 |
+
roles=('USER', 'ASSISTANT'),
|
421 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
422 |
+
sep='\n',
|
423 |
+
sep2='</s>',
|
424 |
+
)
|
425 |
+
)
|
426 |
+
|
427 |
+
register_conv_template(
|
428 |
+
Conversation(
|
429 |
+
name='airoboros_v3',
|
430 |
+
system_template='[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n',
|
431 |
+
system_message='You are a helpful, unbiased, uncensored assistant.',
|
432 |
+
roles=('[INST]', '[/INST]'),
|
433 |
+
sep_style=SeparatorStyle.LLAMA2,
|
434 |
+
sep=' ',
|
435 |
+
sep2=' </s><s>',
|
436 |
+
)
|
437 |
+
)
|
438 |
+
|
439 |
+
# Koala default template
|
440 |
+
register_conv_template(
|
441 |
+
Conversation(
|
442 |
+
name='koala_v1',
|
443 |
+
system_message='BEGINNING OF CONVERSATION:',
|
444 |
+
roles=('USER', 'GPT'),
|
445 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
446 |
+
sep=' ',
|
447 |
+
sep2='</s>',
|
448 |
+
)
|
449 |
+
)
|
450 |
+
|
451 |
+
# Alpaca default template
|
452 |
+
register_conv_template(
|
453 |
+
Conversation(
|
454 |
+
name='alpaca',
|
455 |
+
system_message='Below is an instruction that describes a task. Write a response that appropriately completes the request.',
|
456 |
+
roles=('### Instruction', '### Response'),
|
457 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
458 |
+
sep='\n\n',
|
459 |
+
sep2='</s>',
|
460 |
+
)
|
461 |
+
)
|
462 |
+
|
463 |
+
# ChatGLM default template
|
464 |
+
register_conv_template(
|
465 |
+
Conversation(
|
466 |
+
name='chatglm',
|
467 |
+
roles=('问', '答'),
|
468 |
+
sep_style=SeparatorStyle.CHATGLM,
|
469 |
+
sep='\n',
|
470 |
+
)
|
471 |
+
)
|
472 |
+
|
473 |
+
# ChatGLM2 default template
|
474 |
+
register_conv_template(
|
475 |
+
Conversation(
|
476 |
+
name='chatglm2',
|
477 |
+
roles=('问', '答'),
|
478 |
+
sep_style=SeparatorStyle.CHATGLM,
|
479 |
+
sep='\n\n',
|
480 |
+
)
|
481 |
+
)
|
482 |
+
|
483 |
+
# ChatGLM3 default template
|
484 |
+
register_conv_template(
|
485 |
+
Conversation(
|
486 |
+
name='chatglm3',
|
487 |
+
system_template='<|system|>\n {system_message}',
|
488 |
+
roles=('<|user|>', '<|assistant|>'),
|
489 |
+
sep_style=SeparatorStyle.CHATGLM3,
|
490 |
+
stop_token_ids=[
|
491 |
+
64795,
|
492 |
+
64797,
|
493 |
+
2,
|
494 |
+
], # "<|user|>", "<|observation|>", "</s>"
|
495 |
+
)
|
496 |
+
)
|
497 |
+
|
498 |
+
# CodeGeex(2) Template
|
499 |
+
register_conv_template(
|
500 |
+
Conversation(
|
501 |
+
name='codegeex',
|
502 |
+
roles=('', ''),
|
503 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
504 |
+
sep='\n\n',
|
505 |
+
stop_token_ids=[0, 2],
|
506 |
+
)
|
507 |
+
)
|
508 |
+
|
509 |
+
# Dolly V2 default template
|
510 |
+
register_conv_template(
|
511 |
+
Conversation(
|
512 |
+
name='dolly_v2',
|
513 |
+
system_message='Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n',
|
514 |
+
roles=('### Instruction', '### Response'),
|
515 |
+
sep_style=SeparatorStyle.DOLLY,
|
516 |
+
sep='\n\n',
|
517 |
+
sep2='### End',
|
518 |
+
)
|
519 |
+
)
|
520 |
+
|
521 |
+
# OpenAssistant Pythia default template
|
522 |
+
register_conv_template(
|
523 |
+
Conversation(
|
524 |
+
name='oasst_pythia',
|
525 |
+
roles=('<|prompter|>', '<|assistant|>'),
|
526 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
527 |
+
sep='<|endoftext|>',
|
528 |
+
)
|
529 |
+
)
|
530 |
+
|
531 |
+
# OpenAssistant default template
|
532 |
+
register_conv_template(
|
533 |
+
Conversation(
|
534 |
+
name='oasst_llama',
|
535 |
+
roles=('<|prompter|>', '<|assistant|>'),
|
536 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
537 |
+
sep='</s>',
|
538 |
+
)
|
539 |
+
)
|
540 |
+
|
541 |
+
# OpenChat 3.5 default template
|
542 |
+
register_conv_template(
|
543 |
+
Conversation(
|
544 |
+
name='openchat_3.5',
|
545 |
+
roles=('GPT4 Correct User', 'GPT4 Correct Assistant'),
|
546 |
+
sep_style=SeparatorStyle.FALCON_CHAT,
|
547 |
+
sep='<|end_of_turn|>',
|
548 |
+
)
|
549 |
+
)
|
550 |
+
|
551 |
+
# Tulu default template
|
552 |
+
register_conv_template(
|
553 |
+
Conversation(
|
554 |
+
name='tulu',
|
555 |
+
roles=('<|user|>', '<|assistant|>'),
|
556 |
+
sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
|
557 |
+
sep='\n',
|
558 |
+
)
|
559 |
+
)
|
560 |
+
|
561 |
+
# StableLM Alpha default template
|
562 |
+
register_conv_template(
|
563 |
+
Conversation(
|
564 |
+
name='stablelm',
|
565 |
+
system_template='<|SYSTEM|>{system_message}',
|
566 |
+
system_message="""# StableLM Tuned (Alpha version)
|
567 |
+
- StableLM is a helpful and harmless open-source AI language model developed by StabilityAI.
|
568 |
+
- StableLM is excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
|
569 |
+
- StableLM is more than just an information source, StableLM is also able to write poetry, short stories, and make jokes.
|
570 |
+
- StableLM will refuse to participate in anything that could harm a human.
|
571 |
+
""",
|
572 |
+
roles=('<|USER|>', '<|ASSISTANT|>'),
|
573 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
574 |
+
sep='',
|
575 |
+
stop_token_ids=[50278, 50279, 50277, 1, 0],
|
576 |
+
)
|
577 |
+
)
|
578 |
+
|
579 |
+
# Baize default template
|
580 |
+
register_conv_template(
|
581 |
+
Conversation(
|
582 |
+
name='baize',
|
583 |
+
system_message='The following is a conversation between a human and an AI assistant named Baize (named after a mythical creature in Chinese folklore). Baize is an open-source AI assistant developed by UCSD and Sun Yat-Sen University. The human and the AI assistant take turns chatting. Human statements start with [|Human|] and AI assistant statements start with [|AI|]. The AI assistant always provides responses in as much detail as possible, and in Markdown format. The AI assistant always declines to engage with topics, questions and instructions related to unethical, controversial, or sensitive issues. Complete the transcript in exactly that format.\n',
|
584 |
+
roles=('[|Human|]', '[|AI|]'),
|
585 |
+
messages=(
|
586 |
+
('[|Human|]', 'Hello!'),
|
587 |
+
('[|AI|]', 'Hi!'),
|
588 |
+
),
|
589 |
+
offset=2,
|
590 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
591 |
+
sep='\n',
|
592 |
+
stop_str='[|Human|]',
|
593 |
+
)
|
594 |
+
)
|
595 |
+
|
596 |
+
# RWKV-4-Raven default template
|
597 |
+
register_conv_template(
|
598 |
+
Conversation(
|
599 |
+
name='rwkv',
|
600 |
+
roles=('Bob', 'Alice'),
|
601 |
+
messages=(
|
602 |
+
('Bob', 'hi'),
|
603 |
+
(
|
604 |
+
'Alice',
|
605 |
+
'Hi. I am your assistant and I will provide expert full response in full details. Please feel free to ask any question and I will always answer it.',
|
606 |
+
),
|
607 |
+
),
|
608 |
+
offset=2,
|
609 |
+
sep_style=SeparatorStyle.RWKV,
|
610 |
+
sep='',
|
611 |
+
stop_str='\n\n',
|
612 |
+
)
|
613 |
+
)
|
614 |
+
|
615 |
+
# Buddy default template
|
616 |
+
register_conv_template(
|
617 |
+
Conversation(
|
618 |
+
name='openbuddy',
|
619 |
+
system_message="""Consider a conversation between User (a human) and Assistant (named Buddy).
|
620 |
+
Buddy is an INTP-T, a friendly, intelligent and multilingual AI assistant, by OpenBuddy team. GitHub: https://github.com/OpenBuddy/OpenBuddy
|
621 |
+
Buddy cannot access the Internet.
|
622 |
+
Buddy can fluently speak the user's language (e.g. English, Chinese).
|
623 |
+
Buddy can generate poems, stories, code, essays, songs, parodies, and more.
|
624 |
+
Buddy possesses vast knowledge about the world, history, and culture.
|
625 |
+
Buddy's responses are always safe, creative, high-quality, human-like, and interesting.
|
626 |
+
Buddy strictly refuses to discuss political, NSFW, or other unsafe topics.
|
627 |
+
|
628 |
+
User: Hi.
|
629 |
+
Assistant: Hi, I'm Buddy, your AI assistant. How can I help you today?""",
|
630 |
+
roles=('User', 'Assistant'),
|
631 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
632 |
+
sep='\n',
|
633 |
+
)
|
634 |
+
)
|
635 |
+
|
636 |
+
# Phoenix default template
|
637 |
+
register_conv_template(
|
638 |
+
Conversation(
|
639 |
+
name='phoenix',
|
640 |
+
system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
|
641 |
+
roles=('Human', 'Assistant'),
|
642 |
+
sep_style=SeparatorStyle.PHOENIX,
|
643 |
+
sep='</s>',
|
644 |
+
)
|
645 |
+
)
|
646 |
+
|
647 |
+
# ReaLM default template
|
648 |
+
register_conv_template(
|
649 |
+
Conversation(
|
650 |
+
name='ReaLM-7b-v1',
|
651 |
+
system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
|
652 |
+
roles=('Human', 'Assistant'),
|
653 |
+
sep_style=SeparatorStyle.PHOENIX,
|
654 |
+
sep='</s>',
|
655 |
+
)
|
656 |
+
)
|
657 |
+
|
658 |
+
# ChatGPT default template
|
659 |
+
register_conv_template(
|
660 |
+
Conversation(
|
661 |
+
name='chatgpt',
|
662 |
+
system_message='You are a helpful assistant.',
|
663 |
+
roles=('user', 'assistant'),
|
664 |
+
sep_style=None,
|
665 |
+
sep=None,
|
666 |
+
)
|
667 |
+
)
|
668 |
+
|
669 |
+
# Claude default template
|
670 |
+
register_conv_template(
|
671 |
+
Conversation(
|
672 |
+
name='claude',
|
673 |
+
roles=('Human', 'Assistant'),
|
674 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
675 |
+
sep='\n\n',
|
676 |
+
)
|
677 |
+
)
|
678 |
+
|
679 |
+
# MPT default template
|
680 |
+
register_conv_template(
|
681 |
+
Conversation(
|
682 |
+
name='mpt-7b-chat',
|
683 |
+
system_template="""<|im_start|>system
|
684 |
+
{system_message}""",
|
685 |
+
system_message="""- You are a helpful assistant chatbot trained by MosaicML.
|
686 |
+
- You answer questions.
|
687 |
+
- You are excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
|
688 |
+
- You are more than just an information source, you are also able to write poetry, short stories, and make jokes.""",
|
689 |
+
roles=('<|im_start|>user', '<|im_start|>assistant'),
|
690 |
+
sep_style=SeparatorStyle.CHATML,
|
691 |
+
sep='<|im_end|>',
|
692 |
+
stop_token_ids=[50278, 0],
|
693 |
+
)
|
694 |
+
)
|
695 |
+
|
696 |
+
# MPT-30b-chat default template
|
697 |
+
register_conv_template(
|
698 |
+
Conversation(
|
699 |
+
name='mpt-30b-chat',
|
700 |
+
system_template="""<|im_start|>system
|
701 |
+
{system_message}""",
|
702 |
+
system_message="""A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.""",
|
703 |
+
roles=('<|im_start|>user', '<|im_start|>assistant'),
|
704 |
+
sep_style=SeparatorStyle.CHATML,
|
705 |
+
sep='<|im_end|>',
|
706 |
+
stop_token_ids=[50278, 0],
|
707 |
+
)
|
708 |
+
)
|
709 |
+
|
710 |
+
|
711 |
register_conv_template(
|
712 |
Conversation(
|
713 |
name='Hermes-2',
|
|
|
721 |
6,
|
722 |
7,
|
723 |
8,
|
724 |
+
], # "<|endoftext|>", "<|im_start|>", "<|im_end|>", "<|im_sep|>"
|
725 |
stop_str='<|endoftext|>',
|
726 |
)
|
727 |
)
|
|
|
743 |
)
|
744 |
)
|
745 |
|
746 |
+
# Lemur-70b-chat default template
|
747 |
+
# reference: https://huggingface.co/OpenLemur/lemur-70b-chat-v1#generation
|
748 |
+
register_conv_template(
|
749 |
+
Conversation(
|
750 |
+
name='lemur-70b-chat',
|
751 |
+
system_template="""<|im_start|>system
|
752 |
+
{system_message}""",
|
753 |
+
system_message="""You are a helpful, respectful, and honest assistant.""",
|
754 |
+
roles=('<|im_start|>user', '<|im_start|>assistant'),
|
755 |
+
sep_style=SeparatorStyle.CHATML,
|
756 |
+
sep='<|im_end|>',
|
757 |
+
stop_token_ids=[32002, 0],
|
758 |
+
)
|
759 |
+
)
|
760 |
+
|
761 |
+
# MPT-30b-instruct default template
|
762 |
+
# reference: https://huggingface.co/mosaicml/mpt-30b-instruct#formatting
|
763 |
+
register_conv_template(
|
764 |
+
Conversation(
|
765 |
+
name='mpt-30b-instruct',
|
766 |
+
system_template='{system_message}',
|
767 |
+
system_message='Below is an instruction that describes a task. Write a response that appropriately completes the request.',
|
768 |
+
roles=('### Instruction', '### Response'),
|
769 |
+
sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
|
770 |
+
sep='\n\n',
|
771 |
+
stop_token_ids=[50278, 0],
|
772 |
+
)
|
773 |
+
)
|
774 |
|
775 |
+
# Bard default template
|
776 |
+
# Reference: https://github.com/google/generative-ai-python/blob/9c99bcb474a991a97a2e7d62fcdb52db7ce40729/google/generativeai/discuss.py#L150
|
777 |
+
# https://github.com/google/generative-ai-python/blob/9c99bcb474a991a97a2e7d62fcdb52db7ce40729/google/generativeai/discuss.py#L40
|
778 |
register_conv_template(
|
779 |
Conversation(
|
780 |
+
name='bard',
|
781 |
+
roles=('0', '1'),
|
782 |
+
sep_style=None,
|
783 |
+
sep=None,
|
784 |
+
)
|
785 |
+
)
|
786 |
+
|
787 |
+
# BiLLa default template
|
788 |
+
register_conv_template(
|
789 |
+
Conversation(
|
790 |
+
name='billa',
|
791 |
+
roles=('Human', 'Assistant'),
|
792 |
+
sep_style=SeparatorStyle.ADD_COLON_SPACE_SINGLE,
|
793 |
+
sep='\n',
|
794 |
+
stop_str='Human:',
|
795 |
+
)
|
796 |
+
)
|
797 |
+
|
798 |
+
# RedPajama INCITE default template
|
799 |
+
register_conv_template(
|
800 |
+
Conversation(
|
801 |
+
name='redpajama-incite',
|
802 |
+
roles=('<human>', '<bot>'),
|
803 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
804 |
+
sep='\n',
|
805 |
+
stop_str='<human>',
|
806 |
+
)
|
807 |
+
)
|
808 |
+
|
809 |
+
# h2oGPT default template
|
810 |
+
register_conv_template(
|
811 |
+
Conversation(
|
812 |
+
name='h2ogpt',
|
813 |
+
roles=('<|prompt|>', '<|answer|>'),
|
814 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
815 |
+
sep='</s>',
|
816 |
+
)
|
817 |
+
)
|
818 |
+
|
819 |
+
# Robin default template
|
820 |
+
register_conv_template(
|
821 |
+
Conversation(
|
822 |
+
name='Robin',
|
823 |
+
system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.",
|
824 |
+
roles=('###Human', '###Assistant'),
|
825 |
+
sep_style=SeparatorStyle.ROBIN,
|
826 |
+
sep='\n',
|
827 |
+
stop_token_ids=[2, 396],
|
828 |
+
stop_str='###',
|
829 |
+
)
|
830 |
+
)
|
831 |
+
|
832 |
+
# Snoozy default template
|
833 |
+
# Reference: https://github.com/nomic-ai/gpt4all/blob/d4861030b778da6db59d21d2927a4aba4f9f1f43/gpt4all-bindings/python/gpt4all/gpt4all.py#L232
|
834 |
+
register_conv_template(
|
835 |
+
Conversation(
|
836 |
+
name='snoozy',
|
837 |
+
system_template='### Instruction:\n{system_message}',
|
838 |
+
system_message='The prompt below is a question to answer, a task to complete, or a conversation to respond to; decide which and write an appropriate response.',
|
839 |
+
roles=('### Prompt', '### Response'),
|
840 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
841 |
+
sep='\n',
|
842 |
+
stop_str='###',
|
843 |
+
)
|
844 |
+
)
|
845 |
+
|
846 |
+
# manticore default template
|
847 |
+
register_conv_template(
|
848 |
+
Conversation(
|
849 |
+
name='manticore',
|
850 |
+
roles=('USER', 'ASSISTANT'),
|
851 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
852 |
+
sep='\n',
|
853 |
+
sep2='</s>',
|
854 |
+
)
|
855 |
+
)
|
856 |
+
|
857 |
+
# Falcon default template
|
858 |
+
register_conv_template(
|
859 |
+
Conversation(
|
860 |
+
name='falcon',
|
861 |
+
roles=('User', 'Assistant'),
|
862 |
+
messages=[],
|
863 |
+
sep_style=SeparatorStyle.RWKV,
|
864 |
+
sep='\n',
|
865 |
+
sep2='<|endoftext|>',
|
866 |
+
stop_str='\nUser', # use stop_str to stop generation after stop_token_ids, it will also remove stop_str from the generated text
|
867 |
stop_token_ids=[
|
868 |
+
0,
|
869 |
+
1,
|
870 |
2,
|
871 |
+
3,
|
872 |
+
4,
|
873 |
+
5,
|
874 |
+
6,
|
875 |
+
7,
|
876 |
+
8,
|
877 |
+
9,
|
878 |
+
10,
|
879 |
+
11,
|
880 |
+
], # it better only put special tokens here, because tokenizer only remove special tokens
|
881 |
+
)
|
882 |
+
)
|
883 |
+
|
884 |
+
# ChangGPT default template
|
885 |
+
register_conv_template(
|
886 |
+
Conversation(
|
887 |
+
name='polyglot_changgpt',
|
888 |
+
roles=('B', 'A'),
|
889 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
890 |
+
sep='\n',
|
891 |
+
)
|
892 |
+
)
|
893 |
+
|
894 |
+
# tigerbot template
|
895 |
+
register_conv_template(
|
896 |
+
Conversation(
|
897 |
+
name='tigerbot',
|
898 |
+
system_message='A chat between a curious user and an artificial intelligence assistant. '
|
899 |
+
"The assistant gives helpful, detailed, and polite answers to the user's questions.",
|
900 |
+
roles=('### Instruction', '### Response'),
|
901 |
+
sep_style=SeparatorStyle.ROBIN,
|
902 |
+
sep='\n\n',
|
903 |
+
stop_str='###',
|
904 |
+
)
|
905 |
+
)
|
906 |
+
|
907 |
+
# ref: https://huggingface.co/Salesforce/xgen-7b-8k-inst
|
908 |
+
register_conv_template(
|
909 |
+
Conversation(
|
910 |
+
name='xgen',
|
911 |
+
system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
|
912 |
+
roles=('### Human', '### Assistant'),
|
913 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
914 |
+
sep='\n',
|
915 |
+
stop_token_ids=[50256],
|
916 |
+
)
|
917 |
+
)
|
918 |
+
|
919 |
+
# Internlm-chat template
|
920 |
+
register_conv_template(
|
921 |
+
Conversation(
|
922 |
+
name='internlm-chat',
|
923 |
+
system_message="A chat between a curious <|User|> and an <|Bot|>. The <|Bot|> gives helpful, detailed, and polite answers to the <|User|>'s questions.\n\n",
|
924 |
+
roles=('<|User|>', '<|Bot|>'),
|
925 |
+
sep_style=SeparatorStyle.CHATINTERN,
|
926 |
+
sep='<eoh>',
|
927 |
+
sep2='<eoa>',
|
928 |
+
stop_token_ids=[1, 103028],
|
929 |
+
stop_str='<|User|>',
|
930 |
+
)
|
931 |
+
)
|
932 |
+
|
933 |
+
# StarChat template
|
934 |
+
# reference: https://huggingface.co/spaces/HuggingFaceH4/starchat-playground/blob/main/dialogues.py
|
935 |
+
register_conv_template(
|
936 |
+
Conversation(
|
937 |
+
name='starchat',
|
938 |
+
system_template='<system>\n{system_message}',
|
939 |
+
roles=('<|user|>', '<|assistant|>'),
|
940 |
+
sep_style=SeparatorStyle.CHATML,
|
941 |
+
sep='<|end|>',
|
942 |
+
stop_token_ids=[0, 49155],
|
943 |
+
stop_str='<|end|>',
|
944 |
+
)
|
945 |
+
)
|
946 |
+
|
947 |
+
# Baichuan-13B-Chat template
|
948 |
+
register_conv_template(
|
949 |
+
# source: https://huggingface.co/baichuan-inc/Baichuan-13B-Chat/blob/19ef51ba5bad8935b03acd20ff04a269210983bc/modeling_baichuan.py#L555
|
950 |
+
# https://huggingface.co/baichuan-inc/Baichuan-13B-Chat/blob/main/generation_config.json
|
951 |
+
# https://github.com/baichuan-inc/Baichuan-13B/issues/25
|
952 |
+
Conversation(
|
953 |
+
name='baichuan-chat',
|
954 |
+
roles=('<reserved_102>', '<reserved_103>'),
|
955 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
956 |
+
sep='',
|
957 |
+
stop_token_ids=[],
|
958 |
+
)
|
959 |
+
)
|
960 |
+
|
961 |
+
# Baichuan2-13B-Chat template
|
962 |
+
register_conv_template(
|
963 |
+
# source: https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/blob/c6f8592a60b4ad73c210b28dd2ab3cca51abbf93/modeling_baichuan.py#L773
|
964 |
+
# https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/blob/main/generation_config.json
|
965 |
+
# https://github.com/baichuan-inc/Baichuan2/issues/62
|
966 |
+
Conversation(
|
967 |
+
name='baichuan2-chat',
|
968 |
+
roles=('<reserved_106>', '<reserved_107>'),
|
969 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
970 |
+
sep='',
|
971 |
+
stop_token_ids=[],
|
972 |
+
)
|
973 |
+
)
|
974 |
+
|
975 |
+
# Mistral template
|
976 |
+
# source: https://docs.mistral.ai/llm/mistral-instruct-v0.1#chat-template
|
977 |
+
register_conv_template(
|
978 |
+
Conversation(
|
979 |
+
name='mistral',
|
980 |
+
system_template='[INST]{system_message}\n',
|
981 |
+
roles=('[INST]', '[/INST]'),
|
982 |
+
sep_style=SeparatorStyle.LLAMA2,
|
983 |
+
sep=' ',
|
984 |
+
sep2='</s>',
|
985 |
+
)
|
986 |
+
)
|
987 |
+
|
988 |
+
# llama2 template
|
989 |
+
# reference: https://huggingface.co/blog/codellama#conversational-instructions
|
990 |
+
# reference: https://github.com/facebookresearch/llama/blob/1a240688810f8036049e8da36b073f63d2ac552c/llama/generation.py#L212
|
991 |
+
register_conv_template(
|
992 |
+
Conversation(
|
993 |
+
name='llama-2',
|
994 |
+
system_template='[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n',
|
995 |
+
roles=('[INST]', '[/INST]'),
|
996 |
+
sep_style=SeparatorStyle.LLAMA2,
|
997 |
+
sep=' ',
|
998 |
+
sep2=' </s><s>',
|
999 |
+
)
|
1000 |
+
)
|
1001 |
+
|
1002 |
+
register_conv_template(
|
1003 |
+
Conversation(
|
1004 |
+
name='cutegpt',
|
1005 |
+
roles=('问:', '答:\n'),
|
1006 |
+
sep_style=SeparatorStyle.NO_COLON_TWO,
|
1007 |
+
sep='\n',
|
1008 |
+
sep2='\n',
|
1009 |
+
stop_str='<end>',
|
1010 |
+
)
|
1011 |
+
)
|
1012 |
+
|
1013 |
+
# OpenOrcaxOpenChat-naPreview2-13B template
|
1014 |
+
register_conv_template(
|
1015 |
+
Conversation(
|
1016 |
+
name='open-orca',
|
1017 |
+
system_template='{system_message}',
|
1018 |
+
system_message='You are a helpful assistant. Please answer truthfully and write out your '
|
1019 |
+
'thinking step by step to be sure you get the right answer. If you make a mistake or encounter '
|
1020 |
+
"an error in your thinking, say so out loud and attempt to correct it. If you don't know or "
|
1021 |
+
"aren't sure about something, say so clearly. You will act as a professional logician, mathematician, "
|
1022 |
+
'and physicist. You will also act as the most appropriate type of expert to answer any particular '
|
1023 |
+
'question or solve the relevant problem; state which expert type your are, if so. Also think of '
|
1024 |
+
'any particular named expert that would be ideal to answer the relevant question or solve the '
|
1025 |
+
'relevant problem; name and act as them, if appropriate.',
|
1026 |
+
roles=('User', 'Assistant'),
|
1027 |
+
sep_style=SeparatorStyle.ADD_COLON_SPACE_SINGLE,
|
1028 |
+
sep='<|end_of_turn|>\n',
|
1029 |
+
stop_token_ids=[32000, 32001], # "<|end_of_turn|>"
|
1030 |
+
stop_str='User',
|
1031 |
+
)
|
1032 |
+
)
|
1033 |
+
|
1034 |
+
# Open-Orca/Mistral-7B-OpenOrca template
|
1035 |
+
# source: https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca
|
1036 |
+
# reference: https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca#prompt-template
|
1037 |
+
register_conv_template(
|
1038 |
+
Conversation(
|
1039 |
+
name='mistral-7b-openorca',
|
1040 |
+
system_template='<|im_start|>system\n{system_message}',
|
1041 |
+
system_message='You are MistralOrca, a large language model trained by Alignment Lab AI. Write out your reasoning step-by-step to be sure you get the right answers!',
|
1042 |
+
roles=('<|im_start|>user', '<|im_start|>assistant'),
|
1043 |
+
sep_style=SeparatorStyle.CHATML,
|
1044 |
+
sep='<|im_end|>',
|
1045 |
+
stop_token_ids=[32000, 32001],
|
1046 |
+
)
|
1047 |
+
)
|
1048 |
+
|
1049 |
+
# Qwen-chat default template
|
1050 |
+
# source: https://huggingface.co/Qwen/Qwen-7B-Chat/blob/main/qwen_generation_utils.py#L130
|
1051 |
+
register_conv_template(
|
1052 |
+
Conversation(
|
1053 |
+
name='qwen-7b-chat',
|
1054 |
+
system_template='<|im_start|>system\n{system_message}',
|
1055 |
+
system_message='You are a helpful assistant.',
|
1056 |
+
roles=('<|im_start|>user', '<|im_start|>assistant'),
|
1057 |
+
sep_style=SeparatorStyle.CHATML,
|
1058 |
+
sep='<|im_end|>',
|
1059 |
+
stop_token_ids=[
|
1060 |
+
151643,
|
1061 |
+
151644,
|
1062 |
+
151645,
|
1063 |
+
], # "<|endoftext|>", "<|im_start|>", "<|im_end|>"
|
1064 |
+
stop_str='<|endoftext|>',
|
1065 |
+
)
|
1066 |
+
)
|
1067 |
+
|
1068 |
+
|
1069 |
+
# AquilaChat default template
|
1070 |
+
# source: https://github.com/FlagAI-Open/FlagAI/blob/master/examples/Aquila/Aquila-chat/cyg_conversation.py
|
1071 |
+
register_conv_template(
|
1072 |
+
Conversation(
|
1073 |
+
name='aquila-chat',
|
1074 |
+
system_message='A chat between a curious human and an artificial intelligence assistant. '
|
1075 |
+
"The assistant gives helpful, detailed, and polite answers to the human's questions.",
|
1076 |
+
roles=('Human', 'Assistant'),
|
1077 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
1078 |
+
sep='###',
|
1079 |
+
sep2='',
|
1080 |
+
stop_str=['###', '</s>', '[UNK]'],
|
1081 |
+
)
|
1082 |
+
)
|
1083 |
+
# AquilaChat2-34B default template
|
1084 |
+
# source: https://huggingface.co/BAAI/AquilaChat2-34B/blob/4608b75855334b93329a771aee03869dbf7d88cc/predict.py#L212
|
1085 |
+
register_conv_template(
|
1086 |
+
Conversation(
|
1087 |
+
name='aquila-legacy',
|
1088 |
+
system_message='A chat between a curious human and an artificial intelligence assistant. '
|
1089 |
+
"The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
|
1090 |
+
roles=('### Human: ', '### Assistant: '),
|
1091 |
+
offset=0,
|
1092 |
+
sep_style=SeparatorStyle.NO_COLON_TWO,
|
1093 |
+
sep='\n',
|
1094 |
+
sep2='</s>',
|
1095 |
+
stop_str=['</s>', '[UNK]'],
|
1096 |
+
)
|
1097 |
+
)
|
1098 |
+
# AquilaChat2-7B-16K and AquilaChat2-34B-16K default template
|
1099 |
+
# source: https://huggingface.co/BAAI/AquilaChat2-34B/blob/4608b75855334b93329a771aee03869dbf7d88cc/predict.py#L227
|
1100 |
+
register_conv_template(
|
1101 |
+
Conversation(
|
1102 |
+
name='aquila',
|
1103 |
+
system_message='A chat between a curious human and an artificial intelligence assistant. '
|
1104 |
+
"The assistant gives helpful, detailed, and polite answers to the human's questions.",
|
1105 |
+
roles=('Human', 'Assistant'),
|
1106 |
+
offset=0,
|
1107 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
1108 |
+
sep='###',
|
1109 |
+
sep2='</s>',
|
1110 |
+
stop_str=['</s>', '[UNK]'],
|
1111 |
+
)
|
1112 |
+
)
|
1113 |
+
|
1114 |
+
# AquilaChat2-7B default template
|
1115 |
+
# source: https://huggingface.co/BAAI/AquilaChat2-34B/blob/4608b75855334b93329a771aee03869dbf7d88cc/predict.py#L242
|
1116 |
+
register_conv_template(
|
1117 |
+
Conversation(
|
1118 |
+
name='aquila-v1',
|
1119 |
+
roles=('<|startofpiece|>', '<|endofpiece|>'),
|
1120 |
+
offset=0,
|
1121 |
+
sep_style=SeparatorStyle.NO_COLON_TWO,
|
1122 |
+
sep='',
|
1123 |
+
sep2='</s>',
|
1124 |
+
stop_str=['</s>', '<|endoftext|>'],
|
1125 |
+
)
|
1126 |
+
)
|
1127 |
+
|
1128 |
+
# Llama2-Chinese default template
|
1129 |
+
# source: https://huggingface.co/FlagAlpha
|
1130 |
+
register_conv_template(
|
1131 |
+
Conversation(
|
1132 |
+
name='llama2-chinese',
|
1133 |
+
system_template='<s>{system_message}</s>',
|
1134 |
+
roles=('Human', 'Assistant', 'System'),
|
1135 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
1136 |
+
sep='\n',
|
1137 |
+
sep2='\n</s><s>',
|
1138 |
+
stop_str='</s>',
|
1139 |
)
|
1140 |
)
|
1141 |
+
|
1142 |
+
# Vigogne Instruct default template
|
1143 |
+
# source: https://github.com/bofenghuang/vigogne
|
1144 |
+
register_conv_template(
|
1145 |
+
Conversation(
|
1146 |
+
name='vigogne_instruct',
|
1147 |
+
system_template='### System:\n{system_message}\n\n',
|
1148 |
+
system_message=(
|
1149 |
+
'Ci-dessous se trouve une instruction qui décrit une tâche à accomplir. Rédigez une réponse qui répond de manière'
|
1150 |
+
' précise à la demande.'
|
1151 |
+
),
|
1152 |
+
roles=('### Instruction', '### Response'),
|
1153 |
+
sep_style=SeparatorStyle.DOLLY,
|
1154 |
+
sep='\n\n',
|
1155 |
+
sep2='</s>',
|
1156 |
+
)
|
1157 |
+
)
|
1158 |
+
|
1159 |
+
# Vigogne Chat default template
|
1160 |
+
register_conv_template(
|
1161 |
+
Conversation(
|
1162 |
+
name='vigogne_chat_v2',
|
1163 |
+
system_template='<|system|>: {system_message}',
|
1164 |
+
system_message=(
|
1165 |
+
'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez'
|
1166 |
+
' autant que vous le pouvez.'
|
1167 |
+
),
|
1168 |
+
roles=('<|user|>', '<|assistant|>'),
|
1169 |
+
sep_style=SeparatorStyle.ADD_COLON_TWO,
|
1170 |
+
sep='\n',
|
1171 |
+
sep2='</s>\n',
|
1172 |
+
stop_str='<|user|>',
|
1173 |
+
)
|
1174 |
+
)
|
1175 |
+
|
1176 |
+
register_conv_template(
|
1177 |
+
Conversation(
|
1178 |
+
name='vigogne_chat_v3',
|
1179 |
+
system_template='[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n',
|
1180 |
+
system_message=(
|
1181 |
+
'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez'
|
1182 |
+
' autant que vous le pouvez.'
|
1183 |
+
),
|
1184 |
+
roles=('[INST]', '[/INST]'),
|
1185 |
+
sep_style=SeparatorStyle.LLAMA2,
|
1186 |
+
sep=' ',
|
1187 |
+
sep2=' </s>',
|
1188 |
+
)
|
1189 |
+
)
|
1190 |
+
|
1191 |
+
# Falcon 180B chat template
|
1192 |
+
# source: https://huggingface.co/spaces/tiiuae/falcon-180b-demo/blob/d1590ee7fae9b6ce331ba7808e61a29dcce9239f/app.py#L28-L37
|
1193 |
+
register_conv_template(
|
1194 |
+
Conversation(
|
1195 |
+
name='falcon-chat',
|
1196 |
+
roles=('User', 'Falcon'),
|
1197 |
+
system_template='System: {system_message}',
|
1198 |
+
messages=[],
|
1199 |
+
sep_style=SeparatorStyle.FALCON_CHAT,
|
1200 |
+
sep='\n',
|
1201 |
+
sep2='<|endoftext|>',
|
1202 |
+
stop_str='\nUser:', # use stop_str to stop generation after stop_token_ids, it will also remove stop_str from the generated text
|
1203 |
+
)
|
1204 |
+
)
|
1205 |
+
|
1206 |
+
# Phind template
|
1207 |
+
# source: https://huggingface.co/Phind/Phind-CodeLlama-34B-v2
|
1208 |
+
register_conv_template(
|
1209 |
+
Conversation(
|
1210 |
+
name='phind',
|
1211 |
+
system_message='### System Prompt\nYou are an intelligent programming assistant.',
|
1212 |
+
roles=('### User Message', '### Assistant'),
|
1213 |
+
messages=(),
|
1214 |
+
offset=0,
|
1215 |
+
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
|
1216 |
+
sep='\n\n',
|
1217 |
+
)
|
1218 |
+
)
|
1219 |
+
|
1220 |
+
# Metharme formatting for Pygmalion models
|
1221 |
+
# source: https://huggingface.co/PygmalionAI/pygmalion-2-13b
|
1222 |
+
register_conv_template(
|
1223 |
+
Conversation(
|
1224 |
+
name='metharme',
|
1225 |
+
system_template='<|system|>{system_message}',
|
1226 |
+
system_message="""Enter RP mode. You shall reply to the user while staying
|
1227 |
+
in character. Your responses must be detailed, creative, immersive, and drive the scenario
|
1228 |
+
forward.""",
|
1229 |
+
roles=('<|user|>', '<|model|>'),
|
1230 |
+
sep_style=SeparatorStyle.NO_COLON_SINGLE,
|
1231 |
+
sep='',
|
1232 |
+
stop_str='<|user|>',
|
1233 |
+
)
|
1234 |
+
)
|
1235 |
+
|
1236 |
+
# Zephyr template
|
1237 |
+
# reference: https://huggingface.co/spaces/HuggingFaceH4/zephyr-playground/blob/main/dialogues.py
|
1238 |
+
register_conv_template(
|
1239 |
+
Conversation(
|
1240 |
+
name='zephyr',
|
1241 |
+
system_template='<|system|>\n{system_message}',
|
1242 |
+
roles=('<|user|>', '<|assistant|>'),
|
1243 |
+
sep_style=SeparatorStyle.CHATML,
|
1244 |
+
sep='</s>',
|
1245 |
+
stop_token_ids=[2],
|
1246 |
+
stop_str='</s>',
|
1247 |
+
)
|
1248 |
+
)
|
1249 |
+
|
1250 |
+
# InternVL-ZH template
|
1251 |
+
register_conv_template(
|
1252 |
+
Conversation(
|
1253 |
+
name='internvl_zh',
|
1254 |
+
system_template='',
|
1255 |
+
roles=('<human>', '<bot>'),
|
1256 |
+
sep_style=SeparatorStyle.INTERNVL_ZH,
|
1257 |
+
sep=' ',
|
1258 |
+
sep2='</s>',
|
1259 |
+
)
|
1260 |
+
)
|
1261 |
+
|
examples/image1.jpg
DELETED
Binary file (78.1 kB)
|
|
examples/image2.jpg
DELETED
Binary file (126 kB)
|
|
generation_config.json
CHANGED
@@ -1,8 +1,4 @@
|
|
1 |
{
|
2 |
"_from_model_config": true,
|
3 |
-
"transformers_version": "4.
|
4 |
-
"eos_token_id": [
|
5 |
-
92542,
|
6 |
-
92543
|
7 |
-
]
|
8 |
}
|
|
|
1 |
{
|
2 |
"_from_model_config": true,
|
3 |
+
"transformers_version": "4.36.2"
|
|
|
|
|
|
|
|
|
4 |
}
|
modeling_intern_vit.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
-
# Copyright (c)
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
from typing import Optional, Tuple, Union
|
@@ -20,12 +20,18 @@ from transformers.utils import logging
|
|
20 |
from .configuration_intern_vit import InternVisionConfig
|
21 |
|
22 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
from flash_attn.bert_padding import pad_input, unpad_input
|
24 |
-
|
25 |
-
flash_attn_varlen_qkvpacked_func
|
26 |
has_flash_attn = True
|
27 |
except:
|
28 |
-
print('
|
29 |
has_flash_attn = False
|
30 |
|
31 |
logger = logging.get_logger(__name__)
|
@@ -41,12 +47,12 @@ class FlashAttention(nn.Module):
|
|
41 |
attention_dropout: The dropout rate to apply to the attention
|
42 |
(default: 0.0)
|
43 |
"""
|
44 |
-
|
45 |
def __init__(self, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None):
|
46 |
super().__init__()
|
47 |
self.softmax_scale = softmax_scale
|
48 |
self.dropout_p = attention_dropout
|
49 |
-
|
50 |
def forward(self, qkv, key_padding_mask=None, causal=False, cu_seqlens=None,
|
51 |
max_s=None, need_weights=False):
|
52 |
"""Implements the multihead softmax attention.
|
@@ -59,7 +65,7 @@ class FlashAttention(nn.Module):
|
|
59 |
assert not need_weights
|
60 |
assert qkv.dtype in [torch.float16, torch.bfloat16]
|
61 |
assert qkv.is_cuda
|
62 |
-
|
63 |
if cu_seqlens is None:
|
64 |
batch_size = qkv.shape[0]
|
65 |
seqlen = qkv.shape[1]
|
@@ -68,7 +74,7 @@ class FlashAttention(nn.Module):
|
|
68 |
max_s = seqlen
|
69 |
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
|
70 |
device=qkv.device)
|
71 |
-
output =
|
72 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
73 |
softmax_scale=self.softmax_scale, causal=causal
|
74 |
)
|
@@ -78,7 +84,7 @@ class FlashAttention(nn.Module):
|
|
78 |
x = rearrange(qkv, 'b s three h d -> b s (three h d)')
|
79 |
x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask)
|
80 |
x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
|
81 |
-
output_unpad =
|
82 |
x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
83 |
softmax_scale=self.softmax_scale, causal=causal
|
84 |
)
|
@@ -87,11 +93,11 @@ class FlashAttention(nn.Module):
|
|
87 |
'b s (h d) -> b s h d', h=nheads)
|
88 |
else:
|
89 |
assert max_s is not None
|
90 |
-
output =
|
91 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
92 |
softmax_scale=self.softmax_scale, causal=causal
|
93 |
)
|
94 |
-
|
95 |
return output, None
|
96 |
|
97 |
|
@@ -123,12 +129,6 @@ except Exception:
|
|
123 |
pass
|
124 |
|
125 |
|
126 |
-
NORM2FN = {
|
127 |
-
'rms_norm': InternRMSNorm,
|
128 |
-
'layer_norm': nn.LayerNorm,
|
129 |
-
}
|
130 |
-
|
131 |
-
|
132 |
class InternVisionEmbeddings(nn.Module):
|
133 |
def __init__(self, config: InternVisionConfig):
|
134 |
super().__init__()
|
@@ -154,7 +154,7 @@ class InternVisionEmbeddings(nn.Module):
|
|
154 |
target_dtype = pos_embed.dtype
|
155 |
pos_embed = pos_embed.float().reshape(
|
156 |
1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
|
157 |
-
pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False)
|
158 |
reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)
|
159 |
return pos_embed
|
160 |
|
@@ -267,12 +267,11 @@ class InternVisionEncoderLayer(nn.Module):
|
|
267 |
super().__init__()
|
268 |
self.embed_dim = config.hidden_size
|
269 |
self.intermediate_size = config.intermediate_size
|
270 |
-
self.norm_type = config.norm_type
|
271 |
|
272 |
self.attn = InternAttention(config)
|
273 |
self.mlp = InternMLP(config)
|
274 |
-
self.norm1 =
|
275 |
-
self.norm2 =
|
276 |
|
277 |
self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
|
278 |
self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
|
@@ -287,9 +286,9 @@ class InternVisionEncoderLayer(nn.Module):
|
|
287 |
Args:
|
288 |
hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
289 |
"""
|
290 |
-
hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states)
|
291 |
|
292 |
-
hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states)
|
293 |
|
294 |
return hidden_states
|
295 |
|
@@ -362,7 +361,6 @@ class InternVisionEncoder(nn.Module):
|
|
362 |
|
363 |
class InternVisionModel(PreTrainedModel):
|
364 |
main_input_name = 'pixel_values'
|
365 |
-
_supports_flash_attn_2 = True
|
366 |
config_class = InternVisionConfig
|
367 |
_no_split_modules = ['InternVisionEncoderLayer']
|
368 |
|
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
+
# Copyright (c) 2023 OpenGVLab
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
from typing import Optional, Tuple, Union
|
|
|
20 |
from .configuration_intern_vit import InternVisionConfig
|
21 |
|
22 |
try:
|
23 |
+
try: # v1
|
24 |
+
from flash_attn.flash_attn_interface import \
|
25 |
+
flash_attn_unpadded_qkvpacked_func
|
26 |
+
except: # v2
|
27 |
+
from flash_attn.flash_attn_interface import \
|
28 |
+
flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
|
29 |
+
|
30 |
from flash_attn.bert_padding import pad_input, unpad_input
|
31 |
+
|
|
|
32 |
has_flash_attn = True
|
33 |
except:
|
34 |
+
print('FlashAttention is not installed.')
|
35 |
has_flash_attn = False
|
36 |
|
37 |
logger = logging.get_logger(__name__)
|
|
|
47 |
attention_dropout: The dropout rate to apply to the attention
|
48 |
(default: 0.0)
|
49 |
"""
|
50 |
+
|
51 |
def __init__(self, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None):
|
52 |
super().__init__()
|
53 |
self.softmax_scale = softmax_scale
|
54 |
self.dropout_p = attention_dropout
|
55 |
+
|
56 |
def forward(self, qkv, key_padding_mask=None, causal=False, cu_seqlens=None,
|
57 |
max_s=None, need_weights=False):
|
58 |
"""Implements the multihead softmax attention.
|
|
|
65 |
assert not need_weights
|
66 |
assert qkv.dtype in [torch.float16, torch.bfloat16]
|
67 |
assert qkv.is_cuda
|
68 |
+
|
69 |
if cu_seqlens is None:
|
70 |
batch_size = qkv.shape[0]
|
71 |
seqlen = qkv.shape[1]
|
|
|
74 |
max_s = seqlen
|
75 |
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
|
76 |
device=qkv.device)
|
77 |
+
output = flash_attn_unpadded_qkvpacked_func(
|
78 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
79 |
softmax_scale=self.softmax_scale, causal=causal
|
80 |
)
|
|
|
84 |
x = rearrange(qkv, 'b s three h d -> b s (three h d)')
|
85 |
x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask)
|
86 |
x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
|
87 |
+
output_unpad = flash_attn_unpadded_qkvpacked_func(
|
88 |
x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
89 |
softmax_scale=self.softmax_scale, causal=causal
|
90 |
)
|
|
|
93 |
'b s (h d) -> b s h d', h=nheads)
|
94 |
else:
|
95 |
assert max_s is not None
|
96 |
+
output = flash_attn_unpadded_qkvpacked_func(
|
97 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
98 |
softmax_scale=self.softmax_scale, causal=causal
|
99 |
)
|
100 |
+
|
101 |
return output, None
|
102 |
|
103 |
|
|
|
129 |
pass
|
130 |
|
131 |
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
class InternVisionEmbeddings(nn.Module):
|
133 |
def __init__(self, config: InternVisionConfig):
|
134 |
super().__init__()
|
|
|
154 |
target_dtype = pos_embed.dtype
|
155 |
pos_embed = pos_embed.float().reshape(
|
156 |
1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
|
157 |
+
pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False).\
|
158 |
reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)
|
159 |
return pos_embed
|
160 |
|
|
|
267 |
super().__init__()
|
268 |
self.embed_dim = config.hidden_size
|
269 |
self.intermediate_size = config.intermediate_size
|
|
|
270 |
|
271 |
self.attn = InternAttention(config)
|
272 |
self.mlp = InternMLP(config)
|
273 |
+
self.norm1 = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
|
274 |
+
self.norm2 = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
|
275 |
|
276 |
self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
|
277 |
self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
|
|
|
286 |
Args:
|
287 |
hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
288 |
"""
|
289 |
+
hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states)) * self.ls1)
|
290 |
|
291 |
+
hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states)) * self.ls2)
|
292 |
|
293 |
return hidden_states
|
294 |
|
|
|
361 |
|
362 |
class InternVisionModel(PreTrainedModel):
|
363 |
main_input_name = 'pixel_values'
|
|
|
364 |
config_class = InternVisionConfig
|
365 |
_no_split_modules = ['InternVisionEncoderLayer']
|
366 |
|
modeling_internlm2.py
CHANGED
@@ -48,18 +48,6 @@ _CONFIG_FOR_DOC = 'InternLM2Config'
|
|
48 |
|
49 |
flash_attn_func, flash_attn_varlen_func = None, None
|
50 |
pad_input, index_first_axis, unpad_input = None, None, None
|
51 |
-
try:
|
52 |
-
from flash_attn import flash_attn_func as _flash_attn_func
|
53 |
-
from flash_attn import flash_attn_varlen_func as _flash_attn_varlen_func
|
54 |
-
from flash_attn.bert_padding import index_first_axis as _index_first_axis
|
55 |
-
from flash_attn.bert_padding import pad_input as _pad_input
|
56 |
-
from flash_attn.bert_padding import unpad_input as _unpad_input
|
57 |
-
|
58 |
-
flash_attn_func, flash_attn_varlen_func = _flash_attn_func, _flash_attn_varlen_func
|
59 |
-
pad_input, index_first_axis, unpad_input = _pad_input, _index_first_axis, _unpad_input
|
60 |
-
has_flash_attn = True
|
61 |
-
except:
|
62 |
-
has_flash_attn = False
|
63 |
|
64 |
|
65 |
def _import_flash_attn():
|
@@ -161,7 +149,7 @@ class InternLM2RotaryEmbedding(nn.Module):
|
|
161 |
|
162 |
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
163 |
self.max_seq_len_cached = seq_len
|
164 |
-
t = torch.arange(self.max_seq_len_cached, device=device
|
165 |
|
166 |
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
|
167 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
@@ -190,7 +178,7 @@ class InternLM2LinearScalingRotaryEmbedding(InternLM2RotaryEmbedding):
|
|
190 |
|
191 |
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
192 |
self.max_seq_len_cached = seq_len
|
193 |
-
t = torch.arange(self.max_seq_len_cached, device=device
|
194 |
t = t / self.scaling_factor
|
195 |
|
196 |
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
|
@@ -220,7 +208,7 @@ class InternLM2DynamicNTKScalingRotaryEmbedding(InternLM2RotaryEmbedding):
|
|
220 |
inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
|
221 |
self.register_buffer('inv_freq', inv_freq, persistent=False)
|
222 |
|
223 |
-
t = torch.arange(self.max_seq_len_cached, device=device
|
224 |
|
225 |
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
|
226 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
@@ -709,7 +697,6 @@ class InternLM2PreTrainedModel(PreTrainedModel):
|
|
709 |
supports_gradient_checkpointing = True
|
710 |
_no_split_modules = ['InternLM2DecoderLayer']
|
711 |
_skip_keys_device_placement = 'past_key_values'
|
712 |
-
_supports_flash_attn_2 = True
|
713 |
|
714 |
def _init_weights(self, module):
|
715 |
std = self.config.initializer_range
|
@@ -808,9 +795,6 @@ class InternLM2Model(InternLM2PreTrainedModel):
|
|
808 |
self.padding_idx = config.pad_token_id
|
809 |
self.vocab_size = config.vocab_size
|
810 |
self.config = config
|
811 |
-
if not has_flash_attn:
|
812 |
-
self.config.attn_implementation = 'eager'
|
813 |
-
print('Warning: Flash attention is not available, using eager attention instead.')
|
814 |
|
815 |
self.tok_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
816 |
|
@@ -1098,16 +1082,13 @@ class InternLM2ForCausalLM(InternLM2PreTrainedModel):
|
|
1098 |
output = (logits,) + outputs[1:]
|
1099 |
return (loss,) + output if loss is not None else output
|
1100 |
|
1101 |
-
|
1102 |
-
output = CausalLMOutputWithPast(
|
1103 |
loss=loss,
|
1104 |
logits=logits,
|
1105 |
past_key_values=outputs.past_key_values,
|
1106 |
hidden_states=outputs.hidden_states,
|
1107 |
attentions=outputs.attentions,
|
1108 |
)
|
1109 |
-
output['logits'] = output['logits'].to(device)
|
1110 |
-
return output
|
1111 |
|
1112 |
def prepare_inputs_for_generation(
|
1113 |
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
|
|
48 |
|
49 |
flash_attn_func, flash_attn_varlen_func = None, None
|
50 |
pad_input, index_first_axis, unpad_input = None, None, None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
|
52 |
|
53 |
def _import_flash_attn():
|
|
|
149 |
|
150 |
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
151 |
self.max_seq_len_cached = seq_len
|
152 |
+
t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
|
153 |
|
154 |
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
|
155 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
|
|
178 |
|
179 |
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
180 |
self.max_seq_len_cached = seq_len
|
181 |
+
t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
|
182 |
t = t / self.scaling_factor
|
183 |
|
184 |
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
|
|
|
208 |
inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
|
209 |
self.register_buffer('inv_freq', inv_freq, persistent=False)
|
210 |
|
211 |
+
t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
|
212 |
|
213 |
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
|
214 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
|
|
697 |
supports_gradient_checkpointing = True
|
698 |
_no_split_modules = ['InternLM2DecoderLayer']
|
699 |
_skip_keys_device_placement = 'past_key_values'
|
|
|
700 |
|
701 |
def _init_weights(self, module):
|
702 |
std = self.config.initializer_range
|
|
|
795 |
self.padding_idx = config.pad_token_id
|
796 |
self.vocab_size = config.vocab_size
|
797 |
self.config = config
|
|
|
|
|
|
|
798 |
|
799 |
self.tok_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
800 |
|
|
|
1082 |
output = (logits,) + outputs[1:]
|
1083 |
return (loss,) + output if loss is not None else output
|
1084 |
|
1085 |
+
return CausalLMOutputWithPast(
|
|
|
1086 |
loss=loss,
|
1087 |
logits=logits,
|
1088 |
past_key_values=outputs.past_key_values,
|
1089 |
hidden_states=outputs.hidden_states,
|
1090 |
attentions=outputs.attentions,
|
1091 |
)
|
|
|
|
|
1092 |
|
1093 |
def prepare_inputs_for_generation(
|
1094 |
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
modeling_internvl_chat.py
CHANGED
@@ -1,46 +1,70 @@
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
-
# Copyright (c)
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
import warnings
|
7 |
from typing import Any, List, Optional, Tuple, Union
|
8 |
|
9 |
import torch.utils.checkpoint
|
10 |
-
import
|
11 |
from torch import nn
|
12 |
from torch.nn import CrossEntropyLoss
|
13 |
-
from transformers import AutoModel, GenerationConfig, LlamaForCausalLM
|
|
|
14 |
from transformers.modeling_outputs import CausalLMOutputWithPast
|
15 |
from transformers.modeling_utils import PreTrainedModel
|
16 |
from transformers.utils import ModelOutput, logging
|
17 |
|
18 |
from .configuration_internvl_chat import InternVLChatConfig
|
19 |
-
from .
|
20 |
-
from .modeling_intern_vit import InternVisionModel, has_flash_attn
|
21 |
from .modeling_internlm2 import InternLM2ForCausalLM
|
22 |
|
23 |
logger = logging.get_logger(__name__)
|
24 |
|
25 |
|
26 |
-
def
|
27 |
-
|
|
|
|
|
|
|
28 |
|
29 |
-
|
30 |
-
|
31 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
|
34 |
class InternVLChatModel(PreTrainedModel):
|
35 |
config_class = InternVLChatConfig
|
36 |
main_input_name = 'pixel_values'
|
37 |
-
|
38 |
-
_no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'InternLM2DecoderLayer']
|
39 |
|
40 |
-
def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None
|
41 |
super().__init__(config)
|
42 |
|
43 |
-
assert version_cmp(transformers.__version__, '4.36.2', 'ge')
|
44 |
image_size = config.force_image_size or config.vision_config.image_size
|
45 |
patch_size = config.vision_config.patch_size
|
46 |
self.patch_size = patch_size
|
@@ -48,10 +72,8 @@ class InternVLChatModel(PreTrainedModel):
|
|
48 |
self.template = config.template
|
49 |
self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
|
50 |
self.downsample_ratio = config.downsample_ratio
|
|
|
51 |
self.ps_version = config.ps_version
|
52 |
-
use_flash_attn = use_flash_attn if has_flash_attn else False
|
53 |
-
config.vision_config.use_flash_attn = True if use_flash_attn else False
|
54 |
-
config.llm_config.attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager'
|
55 |
|
56 |
logger.info(f'num_image_token: {self.num_image_token}')
|
57 |
logger.info(f'ps_version: {self.ps_version}')
|
@@ -79,9 +101,44 @@ class InternVLChatModel(PreTrainedModel):
|
|
79 |
nn.Linear(llm_hidden_size, llm_hidden_size)
|
80 |
)
|
81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
self.img_context_token_id = None
|
83 |
-
self.
|
84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
|
86 |
def forward(
|
87 |
self,
|
@@ -178,7 +235,17 @@ class InternVLChatModel(PreTrainedModel):
|
|
178 |
x = x.permute(0, 2, 1, 3).contiguous()
|
179 |
return x
|
180 |
|
|
|
|
|
|
|
|
|
|
|
|
|
181 |
def extract_feature(self, pixel_values):
|
|
|
|
|
|
|
|
|
182 |
if self.select_layer == -1:
|
183 |
vit_embeds = self.vision_model(
|
184 |
pixel_values=pixel_values,
|
@@ -191,96 +258,50 @@ class InternVLChatModel(PreTrainedModel):
|
|
191 |
return_dict=True).hidden_states[self.select_layer]
|
192 |
vit_embeds = vit_embeds[:, 1:, :]
|
193 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
194 |
h = w = int(vit_embeds.shape[1] ** 0.5)
|
195 |
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
|
196 |
vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
|
197 |
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
|
|
|
|
|
198 |
vit_embeds = self.mlp1(vit_embeds)
|
199 |
return vit_embeds
|
200 |
|
201 |
-
def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None,
|
202 |
-
history=None, return_history=False, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>',
|
203 |
-
IMG_CONTEXT_TOKEN='<IMG_CONTEXT>', verbose=False, image_counts=None):
|
204 |
-
if history is not None or return_history:
|
205 |
-
print('Now multi-turn chat is not supported in batch_chat.')
|
206 |
-
raise NotImplementedError
|
207 |
-
|
208 |
-
if image_counts is not None:
|
209 |
-
num_patches_list = image_counts
|
210 |
-
print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.')
|
211 |
-
|
212 |
-
img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
|
213 |
-
self.img_context_token_id = img_context_token_id
|
214 |
-
|
215 |
-
if verbose and pixel_values is not None:
|
216 |
-
image_bs = pixel_values.shape[0]
|
217 |
-
print(f'dynamic ViT batch size: {image_bs}')
|
218 |
-
|
219 |
-
queries = []
|
220 |
-
for idx, num_patches in enumerate(num_patches_list):
|
221 |
-
question = questions[idx]
|
222 |
-
if pixel_values is not None and '<image>' not in question:
|
223 |
-
question = '<image>\n' + question
|
224 |
-
template = get_conv_template(self.template)
|
225 |
-
template.system_message = self.system_message
|
226 |
-
template.append_message(template.roles[0], question)
|
227 |
-
template.append_message(template.roles[1], None)
|
228 |
-
query = template.get_prompt()
|
229 |
-
|
230 |
-
image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
|
231 |
-
query = query.replace('<image>', image_tokens, 1)
|
232 |
-
queries.append(query)
|
233 |
-
|
234 |
-
tokenizer.padding_side = 'left'
|
235 |
-
model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
|
236 |
-
input_ids = model_inputs['input_ids'].cuda()
|
237 |
-
attention_mask = model_inputs['attention_mask'].cuda()
|
238 |
-
eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
|
239 |
-
generation_config['eos_token_id'] = eos_token_id
|
240 |
-
generation_output = self.generate(
|
241 |
-
pixel_values=pixel_values,
|
242 |
-
input_ids=input_ids,
|
243 |
-
attention_mask=attention_mask,
|
244 |
-
**generation_config
|
245 |
-
)
|
246 |
-
responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
|
247 |
-
responses = [response.split(template.sep)[0].strip() for response in responses]
|
248 |
-
return responses
|
249 |
-
|
250 |
def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
|
251 |
-
|
252 |
-
verbose=False):
|
253 |
-
|
254 |
-
if history is None and pixel_values is not None and '<image>' not in question:
|
255 |
-
question = '<image>\n' + question
|
256 |
-
|
257 |
-
if num_patches_list is None:
|
258 |
-
num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
|
259 |
-
assert pixel_values is None or len(pixel_values) == sum(num_patches_list)
|
260 |
|
261 |
img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
|
262 |
self.img_context_token_id = img_context_token_id
|
|
|
|
|
|
|
|
|
263 |
|
264 |
-
|
265 |
-
template.system_message = self.system_message
|
266 |
-
eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
|
267 |
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
272 |
template.append_message(template.roles[0], question)
|
273 |
template.append_message(template.roles[1], None)
|
274 |
query = template.get_prompt()
|
275 |
-
|
276 |
-
if verbose and pixel_values is not None:
|
277 |
-
image_bs = pixel_values.shape[0]
|
278 |
-
print(f'dynamic ViT batch size: {image_bs}')
|
279 |
-
|
280 |
-
for num_patches in num_patches_list:
|
281 |
-
image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
|
282 |
-
query = query.replace('<image>', image_tokens, 1)
|
283 |
-
|
284 |
model_inputs = tokenizer(query, return_tensors='pt')
|
285 |
input_ids = model_inputs['input_ids'].cuda()
|
286 |
attention_mask = model_inputs['attention_mask'].cuda()
|
@@ -292,16 +313,15 @@ class InternVLChatModel(PreTrainedModel):
|
|
292 |
**generation_config
|
293 |
)
|
294 |
response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
|
295 |
-
response = response.split(
|
296 |
history.append((question, response))
|
297 |
if return_history:
|
298 |
return response, history
|
299 |
else:
|
300 |
-
query_to_print = query.replace(
|
301 |
-
query_to_print
|
302 |
-
if verbose:
|
303 |
-
print(query_to_print, response)
|
304 |
return response
|
|
|
305 |
|
306 |
@torch.no_grad()
|
307 |
def generate(
|
@@ -322,6 +342,7 @@ class InternVLChatModel(PreTrainedModel):
|
|
322 |
vit_embeds = visual_features
|
323 |
else:
|
324 |
vit_embeds = self.extract_feature(pixel_values)
|
|
|
325 |
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
326 |
B, N, C = input_embeds.shape
|
327 |
input_embeds = input_embeds.reshape(B * N, C)
|
@@ -329,7 +350,7 @@ class InternVLChatModel(PreTrainedModel):
|
|
329 |
input_ids = input_ids.reshape(B * N)
|
330 |
selected = (input_ids == self.img_context_token_id)
|
331 |
assert selected.sum() != 0
|
332 |
-
input_embeds[selected] = vit_embeds.reshape(-1, C)
|
333 |
|
334 |
input_embeds = input_embeds.reshape(B, N, C)
|
335 |
else:
|
|
|
1 |
# --------------------------------------------------------
|
2 |
# InternVL
|
3 |
+
# Copyright (c) 2023 OpenGVLab
|
4 |
# Licensed under The MIT License [see LICENSE for details]
|
5 |
# --------------------------------------------------------
|
6 |
import warnings
|
7 |
from typing import Any, List, Optional, Tuple, Union
|
8 |
|
9 |
import torch.utils.checkpoint
|
10 |
+
from peft import LoraConfig, get_peft_model
|
11 |
from torch import nn
|
12 |
from torch.nn import CrossEntropyLoss
|
13 |
+
from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
|
14 |
+
LlamaTokenizer)
|
15 |
from transformers.modeling_outputs import CausalLMOutputWithPast
|
16 |
from transformers.modeling_utils import PreTrainedModel
|
17 |
from transformers.utils import ModelOutput, logging
|
18 |
|
19 |
from .configuration_internvl_chat import InternVLChatConfig
|
20 |
+
from .modeling_intern_vit import InternVisionModel
|
|
|
21 |
from .modeling_internlm2 import InternLM2ForCausalLM
|
22 |
|
23 |
logger = logging.get_logger(__name__)
|
24 |
|
25 |
|
26 |
+
def window_partition(x, window_size):
|
27 |
+
"""
|
28 |
+
Args:
|
29 |
+
x: (B, C, H, W)
|
30 |
+
window_size (int): window size, assuming square window
|
31 |
|
32 |
+
Returns:
|
33 |
+
windows: (num_windows*B, C, window_size, window_size)
|
34 |
+
"""
|
35 |
+
B, C, H, W = x.shape
|
36 |
+
assert H % window_size == 0 and W % window_size == 0, 'H and W must be divisible by window_size'
|
37 |
+
|
38 |
+
x = x.view(B, C, H // window_size, window_size, W // window_size, window_size)
|
39 |
+
windows = x.permute(0, 2, 4, 1, 3, 5).contiguous().view(-1, C, window_size, window_size)
|
40 |
+
return windows
|
41 |
+
|
42 |
+
|
43 |
+
def window_reverse(windows, window_size, H, W):
|
44 |
+
"""
|
45 |
+
Args:
|
46 |
+
windows: (num_windows*B, window_size, window_size, C)
|
47 |
+
window_size (int): Window size
|
48 |
+
H (int): Height of image
|
49 |
+
W (int): Width of image
|
50 |
+
|
51 |
+
Returns:
|
52 |
+
x: (B, H * W, C)
|
53 |
+
"""
|
54 |
+
B = int(windows.shape[0] / (H * W / window_size / window_size))
|
55 |
+
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
|
56 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H * W, -1)
|
57 |
+
return x
|
58 |
|
59 |
|
60 |
class InternVLChatModel(PreTrainedModel):
|
61 |
config_class = InternVLChatConfig
|
62 |
main_input_name = 'pixel_values'
|
63 |
+
_no_split_modules = ['InternVisionEncoderLayer', 'LlamaDecoderLayer', 'LlamaForCausalLM']
|
|
|
64 |
|
65 |
+
def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None):
|
66 |
super().__init__(config)
|
67 |
|
|
|
68 |
image_size = config.force_image_size or config.vision_config.image_size
|
69 |
patch_size = config.vision_config.patch_size
|
70 |
self.patch_size = patch_size
|
|
|
72 |
self.template = config.template
|
73 |
self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
|
74 |
self.downsample_ratio = config.downsample_ratio
|
75 |
+
self.image_fold = config.image_fold
|
76 |
self.ps_version = config.ps_version
|
|
|
|
|
|
|
77 |
|
78 |
logger.info(f'num_image_token: {self.num_image_token}')
|
79 |
logger.info(f'ps_version: {self.ps_version}')
|
|
|
101 |
nn.Linear(llm_hidden_size, llm_hidden_size)
|
102 |
)
|
103 |
|
104 |
+
# if config.force_image_size != config.vision_config.image_size:
|
105 |
+
# self.vision_model.resize_pos_embeddings(
|
106 |
+
# old_size=config.vision_config.image_size,
|
107 |
+
# new_size=config.force_image_size,
|
108 |
+
# patch_size=config.vision_config.patch_size
|
109 |
+
# )
|
110 |
+
|
111 |
self.img_context_token_id = None
|
112 |
+
self.neftune_alpha = None
|
113 |
+
|
114 |
+
if config.use_backbone_lora:
|
115 |
+
self.wrap_backbone_lora(r=config.use_backbone_lora, lora_alpha=2 * config.use_backbone_lora)
|
116 |
+
|
117 |
+
if config.use_llm_lora:
|
118 |
+
self.wrap_llm_lora(r=config.use_llm_lora, lora_alpha=2 * config.use_llm_lora)
|
119 |
+
|
120 |
+
def wrap_backbone_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
|
121 |
+
lora_config = LoraConfig(
|
122 |
+
r=r,
|
123 |
+
target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'],
|
124 |
+
lora_alpha=lora_alpha,
|
125 |
+
lora_dropout=lora_dropout,
|
126 |
+
)
|
127 |
+
self.vision_model = get_peft_model(self.vision_model, lora_config)
|
128 |
+
self.vision_model.print_trainable_parameters()
|
129 |
+
|
130 |
+
def wrap_llm_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
|
131 |
+
lora_config = LoraConfig(
|
132 |
+
r=r,
|
133 |
+
target_modules=['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj',
|
134 |
+
'mlp.gate_proj', 'mlp.down_proj', 'mlp.up_proj'],
|
135 |
+
lora_alpha=lora_alpha,
|
136 |
+
lora_dropout=lora_dropout,
|
137 |
+
task_type='CAUSAL_LM'
|
138 |
+
)
|
139 |
+
self.language_model = get_peft_model(self.language_model, lora_config)
|
140 |
+
self.language_model.enable_input_require_grads()
|
141 |
+
self.language_model.print_trainable_parameters()
|
142 |
|
143 |
def forward(
|
144 |
self,
|
|
|
235 |
x = x.permute(0, 2, 1, 3).contiguous()
|
236 |
return x
|
237 |
|
238 |
+
def noised_embed(self, vit_embeds, noise_alpha=5):
|
239 |
+
dims = torch.tensor(vit_embeds.size(1) * vit_embeds.size(2))
|
240 |
+
mag_norm = noise_alpha / torch.sqrt(dims)
|
241 |
+
noise = torch.zeros_like(vit_embeds).uniform_(-mag_norm, mag_norm)
|
242 |
+
return vit_embeds + noise
|
243 |
+
|
244 |
def extract_feature(self, pixel_values):
|
245 |
+
if self.image_fold:
|
246 |
+
image_size = pixel_values.size(-1) # B, C, H, W
|
247 |
+
pixel_values = window_partition(pixel_values, window_size=image_size // self.image_fold) # 4B, C, H/2, W/2
|
248 |
+
|
249 |
if self.select_layer == -1:
|
250 |
vit_embeds = self.vision_model(
|
251 |
pixel_values=pixel_values,
|
|
|
258 |
return_dict=True).hidden_states[self.select_layer]
|
259 |
vit_embeds = vit_embeds[:, 1:, :]
|
260 |
|
261 |
+
if self.training and self.neftune_alpha is not None:
|
262 |
+
vit_embeds = self.noised_embed(vit_embeds, self.neftune_alpha)
|
263 |
+
|
264 |
+
if self.image_fold:
|
265 |
+
vit_embeds = window_reverse(vit_embeds, window_size=image_size // (self.image_fold * self.patch_size),
|
266 |
+
H=image_size // self.patch_size, W=image_size // self.patch_size)
|
267 |
+
|
268 |
+
# if torch.distributed.get_rank() == 0:
|
269 |
+
# print("before pixel shuffle:", vit_embeds.shape)
|
270 |
h = w = int(vit_embeds.shape[1] ** 0.5)
|
271 |
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
|
272 |
vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
|
273 |
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
|
274 |
+
# if torch.distributed.get_rank() == 0:
|
275 |
+
# print("after pixel shuffle:", vit_embeds.shape)
|
276 |
vit_embeds = self.mlp1(vit_embeds)
|
277 |
return vit_embeds
|
278 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
279 |
def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
|
280 |
+
IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>'):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
281 |
|
282 |
img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
|
283 |
self.img_context_token_id = img_context_token_id
|
284 |
+
if tokenizer.convert_tokens_to_ids('<|im_end|>') != 0:
|
285 |
+
eos_token_id = tokenizer.convert_tokens_to_ids('<|im_end|>') # 92542, InternLM2
|
286 |
+
else:
|
287 |
+
eos_token_id = tokenizer.eos_token_id
|
288 |
|
289 |
+
from .conversation import get_conv_template
|
|
|
|
|
290 |
|
291 |
+
template = get_conv_template(self.template)
|
292 |
+
image_bs = pixel_values.shape[0]
|
293 |
+
print(f'dynamic ViT batch size: {image_bs}')
|
294 |
+
if history is None:
|
295 |
+
history = []
|
296 |
+
image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * image_bs + IMG_END_TOKEN
|
297 |
+
question = image_tokens + '\n' + question
|
298 |
+
else:
|
299 |
+
for (old_question, old_answer) in history:
|
300 |
+
template.append_message(template.roles[0], old_question)
|
301 |
+
template.append_message(template.roles[1], old_answer)
|
302 |
template.append_message(template.roles[0], question)
|
303 |
template.append_message(template.roles[1], None)
|
304 |
query = template.get_prompt()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
305 |
model_inputs = tokenizer(query, return_tensors='pt')
|
306 |
input_ids = model_inputs['input_ids'].cuda()
|
307 |
attention_mask = model_inputs['attention_mask'].cuda()
|
|
|
313 |
**generation_config
|
314 |
)
|
315 |
response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
|
316 |
+
response = response.split('<|im_end|>')[0].strip() # for InternLM2
|
317 |
history.append((question, response))
|
318 |
if return_history:
|
319 |
return response, history
|
320 |
else:
|
321 |
+
query_to_print = query.replace(image_tokens, '<image>')
|
322 |
+
print(query_to_print, response)
|
|
|
|
|
323 |
return response
|
324 |
+
return response
|
325 |
|
326 |
@torch.no_grad()
|
327 |
def generate(
|
|
|
342 |
vit_embeds = visual_features
|
343 |
else:
|
344 |
vit_embeds = self.extract_feature(pixel_values)
|
345 |
+
|
346 |
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
347 |
B, N, C = input_embeds.shape
|
348 |
input_embeds = input_embeds.reshape(B * N, C)
|
|
|
350 |
input_ids = input_ids.reshape(B * N)
|
351 |
selected = (input_ids == self.img_context_token_id)
|
352 |
assert selected.sum() != 0
|
353 |
+
input_embeds[selected] = vit_embeds.reshape(-1, C)
|
354 |
|
355 |
input_embeds = input_embeds.reshape(B, N, C)
|
356 |
else:
|
preprocessor_config.json
DELETED
@@ -1,19 +0,0 @@
|
|
1 |
-
{
|
2 |
-
"crop_size": 448,
|
3 |
-
"do_center_crop": true,
|
4 |
-
"do_normalize": true,
|
5 |
-
"do_resize": true,
|
6 |
-
"feature_extractor_type": "CLIPFeatureExtractor",
|
7 |
-
"image_mean": [
|
8 |
-
0.485,
|
9 |
-
0.456,
|
10 |
-
0.406
|
11 |
-
],
|
12 |
-
"image_std": [
|
13 |
-
0.229,
|
14 |
-
0.224,
|
15 |
-
0.225
|
16 |
-
],
|
17 |
-
"resample": 3,
|
18 |
-
"size": 448
|
19 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
examples/red-panda.mp4 → runs/Apr15_16-44-40_SH-IDC1-10-140-37-13/events.out.tfevents.1713171220.SH-IDC1-10-140-37-13.204150.0
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:294d5bf755e6dea5c005c57af52e958a38bb42a7d17d801a25a6543bfe6ddca2
|
3 |
+
size 16662
|
runs/Apr15_17-33-22_SH-IDC1-10-140-37-13/events.out.tfevents.1713174123.SH-IDC1-10-140-37-13.259480.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:57d61c0e776bfb521e58febdbd99525e011f82137ceaaa655ffa6e2b3a9b02a9
|
3 |
+
size 72471
|