blob: 51076b1ae9464bb703f324e3d6d0e61fafc6cf8f (
plain)
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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
|
#!/usr/bin/env bash
set -eo pipefail
FLOSS_DIR="/home/yurenh2/rrm/research/flossing"
CONDA_SH="/home/yurenh2/miniconda3/etc/profile.d/conda.sh"
HRM_ROOT="/home/yurenh2/rrm/hrm/checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 righteous-python"
TRM_ROOT="/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_singleGPU"
wait_for_pid() {
local pid="$1"
if [[ "${pid}" == "0" ]]; then
return 0
fi
while kill -0 "${pid}" 2>/dev/null; do
sleep 60
done
}
activate_env() {
source "${CONDA_SH}"
conda activate rrm
cd "${FLOSS_DIR}"
}
run_hrm_queue() {
wait_for_pid "${1:-0}"
activate_env
CUDA_VISIBLE_DEVICES=0 python step7_interfloss.py \
--model hrm \
--ckpt-root "${HRM_ROOT}" \
--ckpt-name step_26040 \
--train-steps 10000 \
--batch-size 8 \
--train-lr 1e-5 \
--floss-lr 1e-4 \
--floss-steps 500 \
--interfloss-at 0,500 \
--floss-mode engelken_l2 \
--lambda-star 0 \
--k-lyap 8 \
--lyap-act-steps 4 \
--lyap-start-act 12 \
--kl-beta 10 \
--kl-replay-size 64 \
--kl-batch-size 4 \
--kl-temperature 1 \
--seed 42 \
--eval-every 1000 \
--eval-n 512 \
--eval-batch-size 32 \
--floss-log-every 10 \
--out step7_E_hrm_late_engelken_interfloss_kl10_start12_26040_k8_10k.json \
> step7_E_hrm_late_engelken_interfloss_kl10_start12_26040_k8_10k.log 2>&1
CUDA_VISIBLE_DEVICES=0 python step7_interfloss.py \
--model hrm \
--ckpt-root "${HRM_ROOT}" \
--ckpt-name step_26040 \
--train-steps 10000 \
--batch-size 8 \
--train-lr 1e-5 \
--floss-lr 1e-4 \
--floss-steps 500 \
--interfloss-at 0,500 \
--floss-mode volume_cf \
--lambda-star -0.15 \
--k-lyap 8 \
--lyap-act-steps 4 \
--kl-beta 10 \
--kl-replay-size 64 \
--kl-batch-size 4 \
--kl-temperature 1 \
--seed 42 \
--eval-every 1000 \
--eval-n 512 \
--eval-batch-size 32 \
--floss-log-every 10 \
--out step7_G_hrm_volume_envelope_interfloss_kl10_lstar_neg015_26040_k8_10k.json \
> step7_G_hrm_volume_envelope_interfloss_kl10_lstar_neg015_26040_k8_10k.log 2>&1
CUDA_VISIBLE_DEVICES=0 python step8_basin_consistency.py \
--model hrm \
--ckpt-root "${HRM_ROOT}" \
--ckpt-name step_26040 \
--train-steps 10000 \
--batch-size 8 \
--lr 1e-5 \
--consistency-beta 1 \
--consistency-every 1 \
--perturb-after 8 \
--noise-std 0.02 \
--kl-temperature 1 \
--seed 42 \
--eval-every 1000 \
--eval-n 512 \
--eval-batch-size 32 \
--out step8_A_hrm_basin_consistency_beta1_noise002_after8_26040_10k.json \
> step8_A_hrm_basin_consistency_beta1_noise002_after8_26040_10k.log 2>&1
}
run_trm_queue() {
wait_for_pid "${1:-0}"
activate_env
CUDA_VISIBLE_DEVICES=2 python step7_interfloss.py \
--model trm \
--ckpt-root "${TRM_ROOT}" \
--ckpt-name step_26041 \
--train-steps 10000 \
--batch-size 4 \
--train-lr 1e-5 \
--floss-lr 1e-4 \
--floss-steps 500 \
--interfloss-at 0,500 \
--floss-mode engelken_l2 \
--lambda-star 0 \
--k-lyap 4 \
--lyap-act-steps 4 \
--lyap-start-act 12 \
--kl-beta 10 \
--kl-replay-size 64 \
--kl-batch-size 4 \
--kl-temperature 1 \
--seed 42 \
--eval-every 1000 \
--eval-n 512 \
--eval-batch-size 32 \
--floss-log-every 10 \
--out step7_F_trm_late_engelken_interfloss_kl10_start12_26041_k4_batch4_10k.json \
> step7_F_trm_late_engelken_interfloss_kl10_start12_26041_k4_batch4_10k.log 2>&1
CUDA_VISIBLE_DEVICES=2 python step7_interfloss.py \
--model trm \
--ckpt-root "${TRM_ROOT}" \
--ckpt-name step_26041 \
--train-steps 10000 \
--batch-size 4 \
--train-lr 1e-5 \
--floss-lr 1e-4 \
--floss-steps 500 \
--interfloss-at 0,500 \
--floss-mode volume_cf \
--lambda-star 0.02 \
--k-lyap 4 \
--lyap-act-steps 4 \
--kl-beta 10 \
--kl-replay-size 64 \
--kl-batch-size 4 \
--kl-temperature 1 \
--seed 42 \
--eval-every 1000 \
--eval-n 512 \
--eval-batch-size 32 \
--floss-log-every 10 \
--out step7_H_trm_volume_envelope_interfloss_kl10_lstar002_26041_k4_batch4_10k.json \
> step7_H_trm_volume_envelope_interfloss_kl10_lstar002_26041_k4_batch4_10k.log 2>&1
CUDA_VISIBLE_DEVICES=2 python step8_basin_consistency.py \
--model trm \
--ckpt-root "${TRM_ROOT}" \
--ckpt-name step_26041 \
--train-steps 10000 \
--batch-size 4 \
--lr 1e-5 \
--consistency-beta 1 \
--consistency-every 1 \
--perturb-after 8 \
--noise-std 0.02 \
--kl-temperature 1 \
--seed 42 \
--eval-every 1000 \
--eval-n 512 \
--eval-batch-size 32 \
--out step8_B_trm_basin_consistency_beta1_noise002_after8_26041_batch4_10k.json \
> step8_B_trm_basin_consistency_beta1_noise002_after8_26041_batch4_10k.log 2>&1
}
cmd="${1:?usage: launch_dynamics_variants_queue.sh MODE [wait_pid]}"
wait_pid="${2:-0}"
case "${cmd}" in
hrm) run_hrm_queue "${wait_pid}" ;;
trm) run_trm_queue "${wait_pid}" ;;
*) echo "unknown command: ${cmd}" >&2; exit 2 ;;
esac
|
