initial commit
Browse files- llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain.py +214 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/20240220_050613.log +0 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_10499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_10999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_11499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_11999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_12499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_12999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_13499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_1499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_1999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_2499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_2999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_3499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_3999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_4499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_4999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_5499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_5999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_6499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_6999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_7499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_7999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_8499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_8999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_9499.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_9999.txt +16 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/events.out.tfevents.1708405574.b447a125019d.77998.0 +3 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13500.pth/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13500.pth/mp_rank_00_model_states.pt +3 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13954.pth/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13954.pth/mp_rank_00_model_states.pt +3 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/last_checkpoint +1 -0
- work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/zero_to_fp32.py +604 -0
llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain.py
ADDED
@@ -0,0 +1,214 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) OpenMMLab. All rights reserved.
|
2 |
+
import torch
|
3 |
+
from mmengine.dataset import DefaultSampler
|
4 |
+
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
|
5 |
+
LoggerHook, ParamSchedulerHook)
|
6 |
+
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
|
7 |
+
from torch.optim import AdamW
|
8 |
+
from transformers import (AutoModelForCausalLM, AutoTokenizer,
|
9 |
+
BitsAndBytesConfig, CLIPImageProcessor,
|
10 |
+
CLIPVisionModel)
|
11 |
+
|
12 |
+
from xtuner.dataset import LLaVADataset
|
13 |
+
from xtuner.dataset.collate_fns import default_collate_fn
|
14 |
+
from xtuner.dataset.map_fns import llava_map_fn, template_map_fn_factory
|
15 |
+
from xtuner.engine.hooks import DatasetInfoHook, EvaluateChatHook
|
16 |
+
from xtuner.engine.runner import TrainLoop
|
17 |
+
from xtuner.model import LLaVAModel
|
18 |
+
from xtuner.utils import PROMPT_TEMPLATE
|
19 |
+
|
20 |
+
#######################################################################
|
21 |
+
# PART 1 Settings #
|
22 |
+
#######################################################################
|
23 |
+
# Model
|
24 |
+
llm_name_or_path = 'internlm/internlm2-chat-1_8b'
|
25 |
+
visual_encoder_name_or_path = 'openai/clip-vit-large-patch14-336'
|
26 |
+
|
27 |
+
# Data
|
28 |
+
data_root = './llava_data/'
|
29 |
+
data_path = data_root + 'LLaVA-Pretrain/blip_laion_cc_sbu_558k.json'
|
30 |
+
image_folder = data_root + 'LLaVA-Pretrain/images'
|
31 |
+
prompt_template = PROMPT_TEMPLATE.internlm2_chat
|
32 |
+
max_length = int(2048 - (336 / 14)**2)
|
33 |
+
|
34 |
+
# Scheduler & Optimizer
|
35 |
+
batch_size = 40 # per_device
|
36 |
+
accumulative_counts = 7
|
37 |
+
dataloader_num_workers = 2
|
38 |
+
max_epochs = 1
|
39 |
+
optim_type = AdamW
|
40 |
+
lr = 1e-3
|
41 |
+
betas = (0.9, 0.999)
|
42 |
+
weight_decay = 0
|
43 |
+
max_norm = 1 # grad clip
|
44 |
+
warmup_ratio = 0.03
|
45 |
+
|
46 |
+
# Save
|
47 |
+
save_steps = 500
|
48 |
+
save_total_limit = 2 # Maximum checkpoints to keep (-1 means unlimited)
|
49 |
+
|
50 |
+
# Evaluate the generation performance during the training
|
51 |
+
evaluation_freq = 500
|
52 |
+
SYSTEM = ''
|
53 |
+
evaluation_images = 'https://llava-vl.github.io/static/images/view.jpg'
|
54 |
+
evaluation_inputs = ['请描述一下这张照片', 'Please describe this picture']
|
55 |
+
|
56 |
+
#######################################################################
|
57 |
+
# PART 2 Model & Tokenizer & Image Processor #
|
58 |
+
#######################################################################
|
59 |
+
tokenizer = dict(
|
60 |
+
type=AutoTokenizer.from_pretrained,
|
61 |
+
pretrained_model_name_or_path=llm_name_or_path,
|
62 |
+
trust_remote_code=True,
|
63 |
+
padding_side='right')
|
64 |
+
|
65 |
+
image_processor = dict(
|
66 |
+
type=CLIPImageProcessor.from_pretrained,
|
67 |
+
pretrained_model_name_or_path=visual_encoder_name_or_path,
|
68 |
+
trust_remote_code=True)
|
69 |
+
|
70 |
+
model = dict(
|
71 |
+
type=LLaVAModel,
|
72 |
+
freeze_llm=True,
|
73 |
+
freeze_visual_encoder=True,
|
74 |
+
llm=dict(
|
75 |
+
type=AutoModelForCausalLM.from_pretrained,
|
76 |
+
pretrained_model_name_or_path=llm_name_or_path,
|
77 |
+
trust_remote_code=True,
|
78 |
+
torch_dtype=torch.float16,
|
79 |
+
quantization_config=dict(
|
80 |
+
type=BitsAndBytesConfig,
|
81 |
+
load_in_4bit=True,
|
82 |
+
load_in_8bit=False,
|
83 |
+
llm_int8_threshold=6.0,
|
84 |
+
llm_int8_has_fp16_weight=False,
|
85 |
+
bnb_4bit_compute_dtype=torch.float16,
|
86 |
+
bnb_4bit_use_double_quant=True,
|
87 |
+
bnb_4bit_quant_type='nf4')),
|
88 |
+
visual_encoder=dict(
|
89 |
+
type=CLIPVisionModel.from_pretrained,
|
90 |
+
pretrained_model_name_or_path=visual_encoder_name_or_path))
|
91 |
+
|
92 |
+
#######################################################################
|
93 |
+
# PART 3 Dataset & Dataloader #
|
94 |
+
#######################################################################
|
95 |
+
llava_dataset = dict(
|
96 |
+
type=LLaVADataset,
|
97 |
+
data_path=data_path,
|
98 |
+
image_folder=image_folder,
|
99 |
+
tokenizer=tokenizer,
|
100 |
+
image_processor=image_processor,
|
101 |
+
dataset_map_fn=llava_map_fn,
|
102 |
+
template_map_fn=dict(
|
103 |
+
type=template_map_fn_factory, template=prompt_template),
|
104 |
+
max_length=max_length,
|
105 |
+
pad_image_to_square=False)
|
106 |
+
|
107 |
+
train_dataloader = dict(
|
108 |
+
batch_size=batch_size,
|
109 |
+
num_workers=dataloader_num_workers,
|
110 |
+
dataset=llava_dataset,
|
111 |
+
sampler=dict(type=DefaultSampler, shuffle=True),
|
112 |
+
collate_fn=dict(type=default_collate_fn))
|
113 |
+
|
114 |
+
#######################################################################
|
115 |
+
# PART 4 Scheduler & Optimizer #
|
116 |
+
#######################################################################
|
117 |
+
# optimizer
|
118 |
+
optim_wrapper = dict(
|
119 |
+
type=AmpOptimWrapper,
|
120 |
+
optimizer=dict(
|
121 |
+
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
|
122 |
+
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
|
123 |
+
accumulative_counts=accumulative_counts,
|
124 |
+
loss_scale='dynamic',
|
125 |
+
dtype='float16')
|
126 |
+
|
127 |
+
# learning policy
|
128 |
+
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
|
129 |
+
param_scheduler = [
|
130 |
+
dict(
|
131 |
+
type=LinearLR,
|
132 |
+
start_factor=1e-5,
|
133 |
+
by_epoch=True,
|
134 |
+
begin=0,
|
135 |
+
end=warmup_ratio * max_epochs,
|
136 |
+
convert_to_iter_based=True),
|
137 |
+
dict(
|
138 |
+
type=CosineAnnealingLR,
|
139 |
+
eta_min=0.0,
|
140 |
+
by_epoch=True,
|
141 |
+
begin=warmup_ratio * max_epochs,
|
142 |
+
end=max_epochs,
|
143 |
+
convert_to_iter_based=True)
|
144 |
+
]
|
145 |
+
|
146 |
+
# train, val, test setting
|
147 |
+
train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
|
148 |
+
|
149 |
+
#######################################################################
|
150 |
+
# PART 5 Runtime #
|
151 |
+
#######################################################################
|
152 |
+
# Log the dialogue periodically during the training process, optional
|
153 |
+
custom_hooks = [
|
154 |
+
dict(type=DatasetInfoHook, tokenizer=tokenizer),
|
155 |
+
dict(
|
156 |
+
type=EvaluateChatHook,
|
157 |
+
tokenizer=tokenizer,
|
158 |
+
image_processor=image_processor,
|
159 |
+
every_n_iters=evaluation_freq,
|
160 |
+
evaluation_inputs=evaluation_inputs,
|
161 |
+
evaluation_images=evaluation_images,
|
162 |
+
system=SYSTEM,
|
163 |
+
prompt_template=prompt_template)
|
164 |
+
]
|
165 |
+
|
166 |
+
# configure default hooks
|
167 |
+
default_hooks = dict(
|
168 |
+
# record the time of every iteration.
|
169 |
+
timer=dict(type=IterTimerHook),
|
170 |
+
# print log every 10 iterations.
|
171 |
+
logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
|
172 |
+
# enable the parameter scheduler.
|
173 |
+
param_scheduler=dict(type=ParamSchedulerHook),
|
174 |
+
# save checkpoint per `save_steps`.
|
175 |
+
checkpoint=dict(
|
176 |
+
type=CheckpointHook,
|
177 |
+
by_epoch=False,
|
178 |
+
interval=save_steps,
|
179 |
+
max_keep_ckpts=save_total_limit),
|
180 |
+
# set sampler seed in distributed evrionment.
|
181 |
+
sampler_seed=dict(type=DistSamplerSeedHook),
|
182 |
+
)
|
183 |
+
|
184 |
+
# configure environment
|
185 |
+
env_cfg = dict(
|
186 |
+
# whether to enable cudnn benchmark
|
187 |
+
cudnn_benchmark=False,
|
188 |
+
# set multi process parameters
|
189 |
+
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
|
190 |
+
# set distributed parameters
|
191 |
+
dist_cfg=dict(backend='nccl'),
|
192 |
+
)
|
193 |
+
|
194 |
+
# set visualizer
|
195 |
+
from mmengine.visualization import Visualizer, TensorboardVisBackend
|
196 |
+
visualizer = dict(
|
197 |
+
type=Visualizer,
|
198 |
+
vis_backends=[dict(type=TensorboardVisBackend)]
|
199 |
+
)
|
200 |
+
|
201 |
+
# set log level
|
202 |
+
log_level = 'INFO'
|
203 |
+
|
204 |
+
# load from which checkpoint
|
205 |
+
load_from = None
|
206 |
+
|
207 |
+
# whether to resume training from the loaded checkpoint
|
208 |
+
resume = False
|
209 |
+
|
210 |
+
# Defaults to use random seed and disable `deterministic`
|
211 |
+
randomness = dict(seed=None, deterministic=False)
|
212 |
+
|
213 |
+
# set log processor
|
214 |
+
log_processor = dict(by_epoch=False)
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/20240220_050613.log
ADDED
The diff for this file is too large to render.
See raw diff
|
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_10499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake tahoe, california<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake lakelake in the mountains<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_10999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at the lake with a boat on it<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at a lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_11499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at the lake with a boat on the shore<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_11999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_12499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at the lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_12999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_13499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_1499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a bridge over a lake in the mountains with a view of the water<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a bridge over a lake in the mountains with a view of the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_1999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a wooden pier on a lake in the mountains<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a wooden pier with a boat on the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_2499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock on a lake with a boat on the water<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock in the lake with a boat on the shore<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_2999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock on a lake with a boat on the water<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock on a lake with a boat on the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_3499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock on a lake with a boat and a dock in the background<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock on a lake with a boat and a boat dock in the background<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_3999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock with a boat and a lake in the background<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock with a boat on the water and a bridge over it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_4499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a pier on a lake with a boat on the pier<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a bridge over a lake with a boat on the shore<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a wooden bridge in the mountains with a view of the lake<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a wooden bridge in the mountains with a view of the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_4999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at the lake with a boat and a boat house<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at the lake with a boat on the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_5499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock and pier on a lake with a boat<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock with a boat on the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_5999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at a lake with a boat on the water<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at a lake with a boat on the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_6499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a bridge over a lake with a boat on the shore<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a bridge over a lake with a boat on the shore<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_6999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock on a lake with a boat and a dock<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock on a lake with a boat and a dock<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_7499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock on a lake with a boat on the shore<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock on a lake with a boat and a dock<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_7999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat on the shore<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat on the shore<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_8499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock on a lake with a boat on the shore<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock on a lake with a boat on the shore<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_8999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat on the shore photo<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat on the shore photo<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_9499.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at the lake with a boat on the water<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at a lake with a boat on it<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a bridge over a lake in the mountains with a mountain range in the background<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a bridge over a lake in the mountains with the sun setting over the water<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/eval_outputs_iter_9999.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Eval output 1:
|
2 |
+
<|im_start|>user
|
3 |
+
<image>
|
4 |
+
请描述一下这张照片<|im_end|>
|
5 |
+
<|im_start|>assistant
|
6 |
+
<s>a dock at lake with a boat and a mountain in the background<|im_end|>
|
7 |
+
|
8 |
+
|
9 |
+
Eval output 2:
|
10 |
+
<|im_start|>user
|
11 |
+
<image>
|
12 |
+
Please describe this picture<|im_end|>
|
13 |
+
<|im_start|>assistant
|
14 |
+
<s>a dock at lake with a boat and a mountain in the background<|im_end|>
|
15 |
+
|
16 |
+
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/20240220_050613/vis_data/events.out.tfevents.1708405574.b447a125019d.77998.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:863398f4d30a127fd2f17932e0255bfa5bfa92e3d4c6d0ff923551593d2ceb9e
|
3 |
+
size 374431
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13500.pth/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a0232b6c3371d3eea84f49847dbe95ea31d06a597ac58be51fb665d91b4af35b
|
3 |
+
size 75549943
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13500.pth/mp_rank_00_model_states.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:49a35652d7c91ea73e7e30eb753801f822f4e6f59b12406dbbcbd2dd6fa419a4
|
3 |
+
size 13577563
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13954.pth/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:853fcd7c7865d95e9016b23b703eb1ff1a80448f63342b28c2163b55b243d8f4
|
3 |
+
size 75549943
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/iter_13954.pth/mp_rank_00_model_states.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:680d780055ad0f155550e80fc19719369865df98d9faa5ab5c4badbf205a902d
|
3 |
+
size 13610395
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/last_checkpoint
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
/workspace/work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain_copy/iter_13954.pth
|
work_dirs/llava_internlm2_chat_1_8b_clip_vit_large_p14_336_e1_gpu1_pretrain/zero_to_fp32.py
ADDED
@@ -0,0 +1,604 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
# Copyright (c) Microsoft Corporation.
|
4 |
+
# SPDX-License-Identifier: Apache-2.0
|
5 |
+
|
6 |
+
# DeepSpeed Team
|
7 |
+
|
8 |
+
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
9 |
+
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
10 |
+
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
11 |
+
# application.
|
12 |
+
#
|
13 |
+
# example: python zero_to_fp32.py . pytorch_model.bin
|
14 |
+
|
15 |
+
import argparse
|
16 |
+
import torch
|
17 |
+
import glob
|
18 |
+
import math
|
19 |
+
import os
|
20 |
+
import re
|
21 |
+
from collections import OrderedDict
|
22 |
+
from dataclasses import dataclass
|
23 |
+
|
24 |
+
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
25 |
+
# DeepSpeed data structures it has to be available in the current python environment.
|
26 |
+
from deepspeed.utils import logger
|
27 |
+
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
28 |
+
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
29 |
+
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
30 |
+
|
31 |
+
|
32 |
+
@dataclass
|
33 |
+
class zero_model_state:
|
34 |
+
buffers: dict()
|
35 |
+
param_shapes: dict()
|
36 |
+
shared_params: list
|
37 |
+
ds_version: int
|
38 |
+
frozen_param_shapes: dict()
|
39 |
+
frozen_param_fragments: dict()
|
40 |
+
|
41 |
+
|
42 |
+
debug = 0
|
43 |
+
|
44 |
+
# load to cpu
|
45 |
+
device = torch.device('cpu')
|
46 |
+
|
47 |
+
|
48 |
+
def atoi(text):
|
49 |
+
return int(text) if text.isdigit() else text
|
50 |
+
|
51 |
+
|
52 |
+
def natural_keys(text):
|
53 |
+
'''
|
54 |
+
alist.sort(key=natural_keys) sorts in human order
|
55 |
+
http://nedbatchelder.com/blog/200712/human_sorting.html
|
56 |
+
(See Toothy's implementation in the comments)
|
57 |
+
'''
|
58 |
+
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
59 |
+
|
60 |
+
|
61 |
+
def get_model_state_file(checkpoint_dir, zero_stage):
|
62 |
+
if not os.path.isdir(checkpoint_dir):
|
63 |
+
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
64 |
+
|
65 |
+
# there should be only one file
|
66 |
+
if zero_stage <= 2:
|
67 |
+
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
68 |
+
elif zero_stage == 3:
|
69 |
+
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
70 |
+
|
71 |
+
if not os.path.exists(file):
|
72 |
+
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
73 |
+
|
74 |
+
return file
|
75 |
+
|
76 |
+
|
77 |
+
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
78 |
+
# XXX: need to test that this simple glob rule works for multi-node setup too
|
79 |
+
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
80 |
+
|
81 |
+
if len(ckpt_files) == 0:
|
82 |
+
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
83 |
+
|
84 |
+
return ckpt_files
|
85 |
+
|
86 |
+
|
87 |
+
def get_optim_files(checkpoint_dir):
|
88 |
+
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
89 |
+
|
90 |
+
|
91 |
+
def get_model_state_files(checkpoint_dir):
|
92 |
+
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
93 |
+
|
94 |
+
|
95 |
+
def parse_model_states(files):
|
96 |
+
zero_model_states = []
|
97 |
+
for file in files:
|
98 |
+
state_dict = torch.load(file, map_location=device)
|
99 |
+
|
100 |
+
if BUFFER_NAMES not in state_dict:
|
101 |
+
raise ValueError(f"{file} is not a model state checkpoint")
|
102 |
+
buffer_names = state_dict[BUFFER_NAMES]
|
103 |
+
if debug:
|
104 |
+
print("Found buffers:", buffer_names)
|
105 |
+
|
106 |
+
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
107 |
+
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
108 |
+
param_shapes = state_dict[PARAM_SHAPES]
|
109 |
+
|
110 |
+
# collect parameters that are included in param_shapes
|
111 |
+
param_names = []
|
112 |
+
for s in param_shapes:
|
113 |
+
for name in s.keys():
|
114 |
+
param_names.append(name)
|
115 |
+
|
116 |
+
# update with frozen parameters
|
117 |
+
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
118 |
+
if frozen_param_shapes is not None:
|
119 |
+
if debug:
|
120 |
+
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
121 |
+
param_names += list(frozen_param_shapes.keys())
|
122 |
+
|
123 |
+
# handle shared params
|
124 |
+
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
125 |
+
|
126 |
+
ds_version = state_dict.get(DS_VERSION, None)
|
127 |
+
|
128 |
+
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
129 |
+
|
130 |
+
z_model_state = zero_model_state(buffers=buffers,
|
131 |
+
param_shapes=param_shapes,
|
132 |
+
shared_params=shared_params,
|
133 |
+
ds_version=ds_version,
|
134 |
+
frozen_param_shapes=frozen_param_shapes,
|
135 |
+
frozen_param_fragments=frozen_param_fragments)
|
136 |
+
zero_model_states.append(z_model_state)
|
137 |
+
|
138 |
+
return zero_model_states
|
139 |
+
|
140 |
+
|
141 |
+
def parse_optim_states(files, ds_checkpoint_dir):
|
142 |
+
|
143 |
+
total_files = len(files)
|
144 |
+
state_dicts = []
|
145 |
+
for f in files:
|
146 |
+
state_dict = torch.load(f, map_location=device)
|
147 |
+
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
148 |
+
# and also handle the case where it was already removed by another helper script
|
149 |
+
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
150 |
+
state_dicts.append(state_dict)
|
151 |
+
|
152 |
+
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
153 |
+
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
154 |
+
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
155 |
+
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
156 |
+
|
157 |
+
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
158 |
+
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
159 |
+
# use the max of the partition_count to get the dp world_size.
|
160 |
+
|
161 |
+
if type(world_size) is list:
|
162 |
+
world_size = max(world_size)
|
163 |
+
|
164 |
+
if world_size != total_files:
|
165 |
+
raise ValueError(
|
166 |
+
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
167 |
+
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
168 |
+
)
|
169 |
+
|
170 |
+
# the groups are named differently in each stage
|
171 |
+
if zero_stage <= 2:
|
172 |
+
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
173 |
+
elif zero_stage == 3:
|
174 |
+
fp32_groups_key = FP32_FLAT_GROUPS
|
175 |
+
else:
|
176 |
+
raise ValueError(f"unknown zero stage {zero_stage}")
|
177 |
+
|
178 |
+
if zero_stage <= 2:
|
179 |
+
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
180 |
+
elif zero_stage == 3:
|
181 |
+
# if there is more than one param group, there will be multiple flattened tensors - one
|
182 |
+
# flattened tensor per group - for simplicity merge them into a single tensor
|
183 |
+
#
|
184 |
+
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
185 |
+
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
186 |
+
|
187 |
+
fp32_flat_groups = [
|
188 |
+
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
189 |
+
]
|
190 |
+
|
191 |
+
return zero_stage, world_size, fp32_flat_groups
|
192 |
+
|
193 |
+
|
194 |
+
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
|
195 |
+
"""
|
196 |
+
Returns fp32 state_dict reconstructed from ds checkpoint
|
197 |
+
|
198 |
+
Args:
|
199 |
+
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
200 |
+
|
201 |
+
"""
|
202 |
+
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
203 |
+
|
204 |
+
optim_files = get_optim_files(ds_checkpoint_dir)
|
205 |
+
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
206 |
+
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
207 |
+
|
208 |
+
model_files = get_model_state_files(ds_checkpoint_dir)
|
209 |
+
|
210 |
+
zero_model_states = parse_model_states(model_files)
|
211 |
+
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
212 |
+
|
213 |
+
if zero_stage <= 2:
|
214 |
+
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
215 |
+
exclude_frozen_parameters)
|
216 |
+
elif zero_stage == 3:
|
217 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
218 |
+
exclude_frozen_parameters)
|
219 |
+
|
220 |
+
|
221 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
222 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
223 |
+
return
|
224 |
+
|
225 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
226 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
227 |
+
|
228 |
+
if debug:
|
229 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
230 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
231 |
+
|
232 |
+
wanted_params = len(frozen_param_shapes)
|
233 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
234 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
235 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
236 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
237 |
+
|
238 |
+
total_params = 0
|
239 |
+
total_numel = 0
|
240 |
+
for name, shape in frozen_param_shapes.items():
|
241 |
+
total_params += 1
|
242 |
+
unpartitioned_numel = shape.numel()
|
243 |
+
total_numel += unpartitioned_numel
|
244 |
+
|
245 |
+
state_dict[name] = frozen_param_fragments[name]
|
246 |
+
|
247 |
+
if debug:
|
248 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
249 |
+
|
250 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
251 |
+
|
252 |
+
|
253 |
+
def _has_callable(obj, fn):
|
254 |
+
attr = getattr(obj, fn, None)
|
255 |
+
return callable(attr)
|
256 |
+
|
257 |
+
|
258 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
259 |
+
param_shapes = zero_model_states[0].param_shapes
|
260 |
+
|
261 |
+
# Reconstruction protocol:
|
262 |
+
#
|
263 |
+
# XXX: document this
|
264 |
+
|
265 |
+
if debug:
|
266 |
+
for i in range(world_size):
|
267 |
+
for j in range(len(fp32_flat_groups[0])):
|
268 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
269 |
+
|
270 |
+
# XXX: memory usage doubles here (zero2)
|
271 |
+
num_param_groups = len(fp32_flat_groups[0])
|
272 |
+
merged_single_partition_of_fp32_groups = []
|
273 |
+
for i in range(num_param_groups):
|
274 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
275 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
276 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
277 |
+
avail_numel = sum(
|
278 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
279 |
+
|
280 |
+
if debug:
|
281 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
282 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
283 |
+
# not asserting if there is a mismatch due to possible padding
|
284 |
+
print(f"Have {avail_numel} numels to process.")
|
285 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
286 |
+
|
287 |
+
# params
|
288 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
289 |
+
# out-of-core computing solution
|
290 |
+
total_numel = 0
|
291 |
+
total_params = 0
|
292 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
293 |
+
offset = 0
|
294 |
+
avail_numel = full_single_fp32_vector.numel()
|
295 |
+
for name, shape in shapes.items():
|
296 |
+
|
297 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
298 |
+
total_numel += unpartitioned_numel
|
299 |
+
total_params += 1
|
300 |
+
|
301 |
+
if debug:
|
302 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
303 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
304 |
+
offset += unpartitioned_numel
|
305 |
+
|
306 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
307 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
308 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
309 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
310 |
+
align_to = 2 * world_size
|
311 |
+
|
312 |
+
def zero2_align(x):
|
313 |
+
return align_to * math.ceil(x / align_to)
|
314 |
+
|
315 |
+
if debug:
|
316 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
317 |
+
|
318 |
+
offset = zero2_align(offset)
|
319 |
+
avail_numel = zero2_align(avail_numel)
|
320 |
+
|
321 |
+
if debug:
|
322 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
323 |
+
|
324 |
+
# Sanity check
|
325 |
+
if offset != avail_numel:
|
326 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
327 |
+
|
328 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
329 |
+
|
330 |
+
|
331 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
332 |
+
exclude_frozen_parameters):
|
333 |
+
state_dict = OrderedDict()
|
334 |
+
|
335 |
+
# buffers
|
336 |
+
buffers = zero_model_states[0].buffers
|
337 |
+
state_dict.update(buffers)
|
338 |
+
if debug:
|
339 |
+
print(f"added {len(buffers)} buffers")
|
340 |
+
|
341 |
+
if not exclude_frozen_parameters:
|
342 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
343 |
+
|
344 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
345 |
+
|
346 |
+
# recover shared parameters
|
347 |
+
for pair in zero_model_states[0].shared_params:
|
348 |
+
if pair[1] in state_dict:
|
349 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
350 |
+
|
351 |
+
return state_dict
|
352 |
+
|
353 |
+
|
354 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
355 |
+
remainder = unpartitioned_numel % world_size
|
356 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
357 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
358 |
+
return partitioned_numel, padding_numel
|
359 |
+
|
360 |
+
|
361 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
362 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
363 |
+
return
|
364 |
+
|
365 |
+
if debug:
|
366 |
+
for i in range(world_size):
|
367 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
368 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
369 |
+
|
370 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
371 |
+
wanted_params = len(frozen_param_shapes)
|
372 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
373 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
374 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
375 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
376 |
+
|
377 |
+
total_params = 0
|
378 |
+
total_numel = 0
|
379 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
380 |
+
total_params += 1
|
381 |
+
unpartitioned_numel = shape.numel()
|
382 |
+
total_numel += unpartitioned_numel
|
383 |
+
|
384 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
385 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
386 |
+
|
387 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
388 |
+
|
389 |
+
if debug:
|
390 |
+
print(
|
391 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
392 |
+
)
|
393 |
+
|
394 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
395 |
+
|
396 |
+
|
397 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
398 |
+
param_shapes = zero_model_states[0].param_shapes
|
399 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
400 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
401 |
+
# param, re-consolidating each param, while dealing with padding if any
|
402 |
+
|
403 |
+
# merge list of dicts, preserving order
|
404 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
405 |
+
|
406 |
+
if debug:
|
407 |
+
for i in range(world_size):
|
408 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
409 |
+
|
410 |
+
wanted_params = len(param_shapes)
|
411 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
412 |
+
# not asserting if there is a mismatch due to possible padding
|
413 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
414 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
415 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
416 |
+
|
417 |
+
# params
|
418 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
419 |
+
# out-of-core computing solution
|
420 |
+
offset = 0
|
421 |
+
total_numel = 0
|
422 |
+
total_params = 0
|
423 |
+
for name, shape in param_shapes.items():
|
424 |
+
|
425 |
+
unpartitioned_numel = shape.numel()
|
426 |
+
total_numel += unpartitioned_numel
|
427 |
+
total_params += 1
|
428 |
+
|
429 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
430 |
+
|
431 |
+
if debug:
|
432 |
+
print(
|
433 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
434 |
+
)
|
435 |
+
|
436 |
+
# XXX: memory usage doubles here
|
437 |
+
state_dict[name] = torch.cat(
|
438 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
439 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
440 |
+
offset += partitioned_numel
|
441 |
+
|
442 |
+
offset *= world_size
|
443 |
+
|
444 |
+
# Sanity check
|
445 |
+
if offset != avail_numel:
|
446 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
447 |
+
|
448 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
449 |
+
|
450 |
+
|
451 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
452 |
+
exclude_frozen_parameters):
|
453 |
+
state_dict = OrderedDict()
|
454 |
+
|
455 |
+
# buffers
|
456 |
+
buffers = zero_model_states[0].buffers
|
457 |
+
state_dict.update(buffers)
|
458 |
+
if debug:
|
459 |
+
print(f"added {len(buffers)} buffers")
|
460 |
+
|
461 |
+
if not exclude_frozen_parameters:
|
462 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
463 |
+
|
464 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
465 |
+
|
466 |
+
# recover shared parameters
|
467 |
+
for pair in zero_model_states[0].shared_params:
|
468 |
+
if pair[1] in state_dict:
|
469 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
470 |
+
|
471 |
+
return state_dict
|
472 |
+
|
473 |
+
|
474 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
|
475 |
+
"""
|
476 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
477 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
478 |
+
via a model hub.
|
479 |
+
|
480 |
+
Args:
|
481 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
482 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
|
483 |
+
- ``exclude_frozen_parameters``: exclude frozen parameters
|
484 |
+
|
485 |
+
Returns:
|
486 |
+
- pytorch ``state_dict``
|
487 |
+
|
488 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
489 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
490 |
+
the checkpoint.
|
491 |
+
|
492 |
+
A typical usage might be ::
|
493 |
+
|
494 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
495 |
+
# do the training and checkpoint saving
|
496 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
497 |
+
model = model.cpu() # move to cpu
|
498 |
+
model.load_state_dict(state_dict)
|
499 |
+
# submit to model hub or save the model to share with others
|
500 |
+
|
501 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
502 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
503 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
504 |
+
|
505 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
506 |
+
|
507 |
+
"""
|
508 |
+
if tag is None:
|
509 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
510 |
+
if os.path.isfile(latest_path):
|
511 |
+
with open(latest_path, 'r') as fd:
|
512 |
+
tag = fd.read().strip()
|
513 |
+
else:
|
514 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
515 |
+
|
516 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
517 |
+
|
518 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
519 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
520 |
+
|
521 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
|
522 |
+
|
523 |
+
|
524 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
|
525 |
+
"""
|
526 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
527 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
528 |
+
|
529 |
+
Args:
|
530 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
531 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
532 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
533 |
+
- ``exclude_frozen_parameters``: exclude frozen parameters
|
534 |
+
"""
|
535 |
+
|
536 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
|
537 |
+
print(f"Saving fp32 state dict to {output_file}")
|
538 |
+
torch.save(state_dict, output_file)
|
539 |
+
|
540 |
+
|
541 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
542 |
+
"""
|
543 |
+
1. Put the provided model to cpu
|
544 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
545 |
+
3. Load it into the provided model
|
546 |
+
|
547 |
+
Args:
|
548 |
+
- ``model``: the model object to update
|
549 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
550 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
551 |
+
|
552 |
+
Returns:
|
553 |
+
- ``model`: modified model
|
554 |
+
|
555 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
556 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
557 |
+
conveniently placed for you in the checkpoint folder.
|
558 |
+
|
559 |
+
A typical usage might be ::
|
560 |
+
|
561 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
562 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
563 |
+
# submit to model hub or save the model to share with others
|
564 |
+
|
565 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
566 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
567 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
568 |
+
|
569 |
+
"""
|
570 |
+
logger.info(f"Extracting fp32 weights")
|
571 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
572 |
+
|
573 |
+
logger.info(f"Overwriting model with fp32 weights")
|
574 |
+
model = model.cpu()
|
575 |
+
model.load_state_dict(state_dict, strict=False)
|
576 |
+
|
577 |
+
return model
|
578 |
+
|
579 |
+
|
580 |
+
if __name__ == "__main__":
|
581 |
+
|
582 |
+
parser = argparse.ArgumentParser()
|
583 |
+
parser.add_argument("checkpoint_dir",
|
584 |
+
type=str,
|
585 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
586 |
+
parser.add_argument(
|
587 |
+
"output_file",
|
588 |
+
type=str,
|
589 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
590 |
+
parser.add_argument("-t",
|
591 |
+
"--tag",
|
592 |
+
type=str,
|
593 |
+
default=None,
|
594 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
595 |
+
parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
|
596 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
597 |
+
args = parser.parse_args()
|
598 |
+
|
599 |
+
debug = args.debug
|
600 |
+
|
601 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
|
602 |
+
args.output_file,
|
603 |
+
tag=args.tag,
|
604 |
+
exclude_frozen_parameters=args.exclude_frozen_parameters)
|