Add ZeRO-2 resume metadata
Browse files- zero2_resume_info.json +17 -0
zero2_resume_info.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 62000,
|
| 3 |
+
"single_file_checkpoint": "0062000.pt",
|
| 4 |
+
"zero2_checkpoint_dir": "checkpoints/0062000",
|
| 5 |
+
"optimizer_resume_note": "The single .pt has optimizer=None under DeepSpeed ZeRO-2. Use --ckpt 0062000.pt with sibling checkpoints/0062000/ present so train_robot.py can call DeepSpeed load_checkpoint(tag=0062000).",
|
| 6 |
+
"required_zero2_files": [
|
| 7 |
+
"bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt",
|
| 8 |
+
"bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt",
|
| 9 |
+
"bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt",
|
| 10 |
+
"bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt",
|
| 11 |
+
"bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt",
|
| 12 |
+
"bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt",
|
| 13 |
+
"bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt",
|
| 14 |
+
"bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt",
|
| 15 |
+
"mp_rank_00_model_states.pt"
|
| 16 |
+
]
|
| 17 |
+
}
|