forked from pytorch/torchtune
-
Notifications
You must be signed in to change notification settings - Fork 0
/
knowledge_distillation_single_device.yaml
97 lines (83 loc) · 2.64 KB
/
knowledge_distillation_single_device.yaml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
# Config for single device knowledge distillation in kd_single_device.py
# using a teacher and student model
#
# This config assumes that you've ran the following commands before launching KD:
# First download the student and teacher models
# tune download Qwen/Qwen2-0.5B-Instruct --output-dir /tmp/Qwen2-0.5B-Instruct --ignore-patterns None
# tune download Qwen/Qwen2-1.5B-Instruct --output-dir /tmp/Qwen2-1.5B-Instruct --ignore-patterns None
#
# You get better results using KD if the teacher model has already been fine-tuned on the target dataset:
# tune run lora_finetune_single_device --config qwen2/1.5B_lora_single_device
#
# To launch on a single device, run the following command from root:
# tune run knowledge_distillation_single_device --config qwen2/knowledge_distillation_single_device
#
# This config works only for distilling on a single device.
# Model Arguments
model:
_component_: torchtune.models.qwen2.lora_qwen2_0_5b
lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
apply_lora_to_mlp: False
lora_rank: 32
lora_alpha: 64
teacher_model:
_component_: torchtune.models.qwen2.qwen2_1_5b
tokenizer:
_component_: torchtune.models.qwen2.qwen2_tokenizer
path: /tmp/Qwen2-0.5B-Instruct/vocab.json
merges_file: /tmp/Qwen2-0.5B-Instruct/merges.txt
max_seq_len: null
checkpointer:
_component_: torchtune.training.FullModelHFCheckpointer
checkpoint_dir: /tmp/Qwen2-0.5B-Instruct
checkpoint_files: [
model.safetensors
]
recipe_checkpoint: null
output_dir: /tmp/Qwen2-0.5B-Instruct
model_type: QWEN2
teacher_checkpointer:
_component_: torchtune.training.FullModelHFCheckpointer
checkpoint_dir: /tmp/Qwen2-1.5B-Instruct
checkpoint_files: [
model.safetensors
]
recipe_checkpoint: null
output_dir: /tmp/Qwen2-1.5B-Instruct
model_type: QWEN2
resume_from_checkpoint: False
# Dataset and Sampler
dataset:
_component_: torchtune.datasets.alpaca_cleaned_dataset
seed: null
shuffle: True
batch_size: 8
# Optimizer and Scheduler
optimizer:
_component_: torch.optim.AdamW
weight_decay: 0.01
lr: 3e-4
lr_scheduler:
_component_: torchtune.modules.get_cosine_schedule_with_warmup
num_warmup_steps: 100
loss:
_component_: torchtune.modules.loss.CEWithChunkedOutputLoss
kd_loss:
_component_: torchtune.modules.loss.ForwardKLWithChunkedOutputLoss
kd_ratio: 0.5
# Training
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 2
compile: False
# Logging
output_dir: /tmp/qwen_kd
metric_logger:
_component_: torchtune.training.metric_logging.DiskLogger
log_dir: ${output_dir}
log_every_n_steps: 1
log_peak_memory_stats: False
# Environment
device: cuda
dtype: bf16
enable_activation_checkpointing: False