File size: 2,655 Bytes
c4b1fea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
98
#!/bin/bash
#SBATCH -o ../watch_folder/%x_%j.out     # output file (%j expands to jobID)
#SBATCH -N 1                          # Total number of nodes requested
#SBATCH --get-user-env                # retrieve the users login environment
#SBATCH --mem=64000                   # server memory requested (per node)
#SBATCH -t 960:00:00                  # Time limit (hh:mm:ss)
#SBATCH --constraint="[a100|a6000|a5000|3090]"
#SBATCH --ntasks-per-node=4
#SBATCH --gres=gpu:4                  # Type/number of GPUs needed
#SBATCH --open-mode=append            # Do not overwrite logs
#SBATCH --requeue                     # Requeue upon preemption

<<comment
#  Usage:
cd scripts/
MODEL=<ar|mdlm|udlm>
sbatch \
  --export=ALL,MODEL=${MODEL} \
  --job-name=train_text8_${MODEL} \
  train_text8.sh
comment

# Setup environment
cd ../ || exit  # Go to the root directory of the repo
source setup_env.sh
export NCCL_P2P_LEVEL=NVL
export HYDRA_FULL_ERROR=1

# Expecting:
#  - MODEL (ar, mdlm, udlm)
if [ -z "${MODEL}" ]; then
  echo "MODEL is not set"
  exit 1
fi
RUN_NAME="${MODEL}"

if [ "${MODEL}" = "ar" ]; then
  # AR
  DIFFUSION="absorbing_state"
  PARAMETERIZATION="ar"
  T=0
  TIME_COND=False
  ZERO_RECON_LOSS=False
  sampling_use_cache=False
elif [ "${MODEL}" = "mdlm" ]; then
  # MDLM
  DIFFUSION="absorbing_state"
  PARAMETERIZATION="subs"
  T=0
  TIME_COND=False
  ZERO_RECON_LOSS=False
  sampling_use_cache=True
elif [ "${MODEL}" = "udlm" ]; then
  # UDLM
  DIFFUSION="uniform"
  PARAMETERIZATION="d3pm"
  T=0
  TIME_COND=True
  ZERO_RECON_LOSS=True
  sampling_use_cache=False
else
  echo "MODEL must be one of ar, mdlm, udlm"
  exit 1
fi


# To enable preemption re-loading, set `hydra.run.dir` or
# `checkpointing.save_dir` explicitly.
srun python -u -m main \
  diffusion="${DIFFUSION}" \
  parameterization="${PARAMETERIZATION}" \
  T=${T} \
  time_conditioning=${TIME_COND} \
  zero_recon_loss=${ZERO_RECON_LOSS} \
  data="text8" \
  data.wrap=True \
  data.tokenizer_name_or_path=text8 \
  loader.global_batch_size=512 \
  loader.eval_global_batch_size=1024 \
  backbone="dit" \
  model=small \
  model.length=256 \
  optim.lr=3e-4 \
  training.guidance=null \
  callbacks.checkpoint_every_n_steps.every_n_train_steps=100_000 \
  trainer.log_every_n_steps=100 \
  trainer.max_steps=1_000_000 \
  trainer.precision=bf16 \
  trainer.val_check_interval=5_000 \
  +trainer.check_val_every_n_epoch=null \
  eval.generate_samples=True \
  sampling.num_sample_batches=1 \
  sampling.batch_size=2 \
  sampling.use_cache=${sampling_use_cache} \
  sampling.steps=256 \
  wandb.name="text8_${RUN_NAME}" \
  hydra.run.dir="${PWD}/outputs/text8/${RUN_NAME}"