skypilot

1
name: squad_v2
2

3
resources:
4
  candidates:
5
  - {accelerators: T4:8}
6
  - {accelerators: V100:8}
7

8
workdir: ./examples/benchmark/transformers_qa
9

10
setup: |
11
  conda create -n hf python=3.8 -y
12
  conda activate hf
13

14
  # Install SkyCallback
15
  pip install "git+https://github.com/skypilot-org/skypilot.git#egg=sky-callback&subdirectory=sky/callbacks/"
16

17
  # User setup
18
  pip install transformers
19
  git clone https://github.com/huggingface/transformers.git
20
  cd transformers
21
  git checkout v4.20.0
22
  pip install -r examples/pytorch/question-answering/requirements.txt
23

24
  # Apply the patch to enable SkyCallback
25
  git apply ../callback.patch
26

27
run: |
28
  conda activate hf
29
  cd transformers/examples/pytorch/question-answering/
30
  python run_qa.py \
31
    --model_name_or_path bert-base-uncased \
32
    --dataset_name squad_v2 \
33
    --do_train \
34
    --do_eval \
35
    --per_device_train_batch_size 12 \
36
    --learning_rate 3e-5 \
37
    --num_train_epochs 2 \
38
    --max_seq_length 384 \
39
    --doc_stride 128 \
40
    --version_2_with_negative \
41
    --output_dir outputs/
42