Upload config.py with huggingface_hub
Browse files
config.py
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
DS_CONFIG_ZERO0 = {
|
| 2 |
+
"bf16": {
|
| 3 |
+
"enabled": True,
|
| 4 |
+
},
|
| 5 |
+
"optimizer": {
|
| 6 |
+
"type": "AdamW",
|
| 7 |
+
"params": {
|
| 8 |
+
"lr": 3e-5,
|
| 9 |
+
"betas": [0.98, 0.999],
|
| 10 |
+
"eps": 1e-9
|
| 11 |
+
}
|
| 12 |
+
},
|
| 13 |
+
"scheduler": {
|
| 14 |
+
"type": "WarmupLR",
|
| 15 |
+
"params": {
|
| 16 |
+
"warmup_min_lr": 2e-5,
|
| 17 |
+
"warmup_max_lr": 5e-5,
|
| 18 |
+
"warmup_num_steps": 300
|
| 19 |
+
}
|
| 20 |
+
},
|
| 21 |
+
"zero_optimization": {
|
| 22 |
+
"stage": 0
|
| 23 |
+
},
|
| 24 |
+
"gradient_accumulation_steps": 16,
|
| 25 |
+
"train_micro_batch_size_per_gpu": 1,
|
| 26 |
+
"wall_clock_breakdown": False,
|
| 27 |
+
"steps_per_print": 50
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
DS_CONFIG_ft = {
|
| 31 |
+
"bf16": {
|
| 32 |
+
"enabled": True,
|
| 33 |
+
},
|
| 34 |
+
"optimizer": {
|
| 35 |
+
"type": "AdamW",
|
| 36 |
+
"params": {
|
| 37 |
+
"lr": 3e-5,
|
| 38 |
+
"betas": [0.98, 0.999],
|
| 39 |
+
"eps": 1e-9
|
| 40 |
+
}
|
| 41 |
+
},
|
| 42 |
+
"scheduler": {
|
| 43 |
+
"type": "WarmupLR",
|
| 44 |
+
"params": {
|
| 45 |
+
"warmup_min_lr": 2e-5,
|
| 46 |
+
"warmup_max_lr": 5e-5,
|
| 47 |
+
"warmup_num_steps": 300
|
| 48 |
+
}
|
| 49 |
+
},
|
| 50 |
+
"zero_optimization": {
|
| 51 |
+
"stage": 1,
|
| 52 |
+
"allgather_partitions": True,
|
| 53 |
+
"allgather_bucket_size": 2e8,
|
| 54 |
+
"overlap_comm": True,
|
| 55 |
+
"reduce_scatter": True,
|
| 56 |
+
"reduce_bucket_size": 2e8,
|
| 57 |
+
"contiguous_gradients": True,
|
| 58 |
+
"stage3_gather_16bit_weights_on_model_save": True,
|
| 59 |
+
"offload_optimizer": {
|
| 60 |
+
"device": "cpu"
|
| 61 |
+
}
|
| 62 |
+
},
|
| 63 |
+
"gradient_accumulation_steps": 16,
|
| 64 |
+
"train_micro_batch_size_per_gpu": 1,
|
| 65 |
+
"wall_clock_breakdown": False,
|
| 66 |
+
"steps_per_print": 50
|
| 67 |
+
}
|
| 68 |
+
|
| 69 |
+
TARGET_MODULES = [
|
| 70 |
+
"q_proj",
|
| 71 |
+
"v_proj",
|
| 72 |
+
]
|