skypilot

1
# Usage:
2
#
3
#  Unmanaged spot (no auto-recovery; for debugging):
4
#   HF_TOKEN=abc BUCKET=<unique-name> sky launch -c axolotl-spot axolotl-spot.yaml --env HF_TOKEN --env BUCKET -i30 --down
5
#
6
#  Managed spot (auto-recovery; for full runs):
7
#   HF_TOKEN=abc BUCKET=<unique-name> sky spot launch -n axolotl-spot axolotl-spot.yaml --env HF_TOKEN --env BUCKET
8

9
name: axolotl
10

11
resources:
12
  accelerators: A100:1
13
  cloud: gcp # optional
14
  use_spot: True
15

16
workdir: mistral
17

18
file_mounts:
19
  /sky-notebook:
20
    name: ${BUCKET}
21
    mode: MOUNT
22

23
setup: |
24
  docker pull winglian/axolotl:main-py3.10-cu118-2.0.1
25

26
run: |
27
  docker run --gpus all \
28
    -v ~/sky_workdir:/sky_workdir \
29
    -v /root/.cache:/root/.cache \
30
    winglian/axolotl:main-py3.10-cu118-2.0.1 \
31
    huggingface-cli login --token ${HF_TOKEN}
32
  
33
  docker run --gpus all \
34
    -v ~/sky_workdir:/sky_workdir \
35
    -v /root/.cache:/root/.cache \
36
    -v /sky-notebook:/sky-notebook \
37
    winglian/axolotl:main-py3.10-cu118-2.0.1 \
38
    accelerate launch -m axolotl.cli.train /sky_workdir/qlora-checkpoint.yaml
39

40
envs:
41
  HF_TOKEN: <your-huggingface-token> # TODO: Replace with huggingface token
42
  BUCKET: <a-unique-bucket-name-to-use>
43
  
44
  
45

46

47
  
48

49