model
Browse files- scripts/model.yaml +3 -4
scripts/model.yaml
CHANGED
|
@@ -56,8 +56,8 @@ train:
|
|
| 56 |
global_batch_size: 512
|
| 57 |
|
| 58 |
# Number of samples per data-parallel rank (type: int, default: 4)
|
| 59 |
-
|
| 60 |
-
micro_batch_size:
|
| 61 |
|
| 62 |
# Number of iterations with learning rate warmup active (type: int, default: 2000)
|
| 63 |
lr_warmup_steps: 2000
|
|
@@ -103,8 +103,7 @@ eval:
|
|
| 103 |
|
| 104 |
# Optimizer-related arguments
|
| 105 |
optimizer:
|
| 106 |
-
|
| 107 |
-
class_path: torch.optim.PagedAdamW
|
| 108 |
# class_path: bitsandbytes.optim.PagedAdamW
|
| 109 |
# class_path: bitsandbytes.optim.AdamW8bit
|
| 110 |
# class_path: bitsandbytes.optim.PagedAdamW8bit
|
|
|
|
| 56 |
global_batch_size: 512
|
| 57 |
|
| 58 |
# Number of samples per data-parallel rank (type: int, default: 4)
|
| 59 |
+
micro_batch_size: 16
|
| 60 |
+
# micro_batch_size: 15
|
| 61 |
|
| 62 |
# Number of iterations with learning rate warmup active (type: int, default: 2000)
|
| 63 |
lr_warmup_steps: 2000
|
|
|
|
| 103 |
|
| 104 |
# Optimizer-related arguments
|
| 105 |
optimizer:
|
| 106 |
+
class_path: torch.optim.AdamW
|
|
|
|
| 107 |
# class_path: bitsandbytes.optim.PagedAdamW
|
| 108 |
# class_path: bitsandbytes.optim.AdamW8bit
|
| 109 |
# class_path: bitsandbytes.optim.PagedAdamW8bit
|