grokadamw.GrokAdamW
Browse files- scripts/pretrain-core-model.yaml +20 -8
- scripts/requirements.in +5 -12
scripts/pretrain-core-model.yaml
CHANGED
@@ -63,8 +63,8 @@ train:
|
|
63 |
log_interval: 1
|
64 |
|
65 |
# Number of samples between optimizer steps across data-parallel ranks (type: int, default: 512)
|
66 |
-
|
67 |
-
global_batch_size: 256
|
68 |
|
69 |
# Number of samples per data-parallel rank (type: int, default: 4)
|
70 |
# micro_batch_size: 4
|
@@ -114,12 +114,24 @@ eval:
|
|
114 |
|
115 |
# Optimizer-related arguments
|
116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
optimizer:
|
118 |
-
|
119 |
-
# class_path: torchao.prototype.low_bit_optim.AdamW8bit
|
120 |
-
# class_path: torchao.prototype.low_bit_optim.AdamW4bit
|
121 |
-
# class_path: bitsandbytes.optim.AdamW8bit
|
122 |
-
class_path: bitsandbytes.optim.PagedAdamW8bit
|
123 |
init_args:
|
124 |
# (type: float, default: 0.001)
|
125 |
lr: 1e-4
|
@@ -128,7 +140,7 @@ optimizer:
|
|
128 |
# (type: tuple, default: (0.9,0.999))
|
129 |
betas:
|
130 |
- 0.9
|
131 |
-
- 0.
|
132 |
|
133 |
# How many devices/GPUs to use. Uses all GPUs by default. (type: Union[int, str], default: auto)
|
134 |
devices: auto
|
|
|
63 |
log_interval: 1
|
64 |
|
65 |
# Number of samples between optimizer steps across data-parallel ranks (type: int, default: 512)
|
66 |
+
global_batch_size: 512
|
67 |
+
# global_batch_size: 256
|
68 |
|
69 |
# Number of samples per data-parallel rank (type: int, default: 4)
|
70 |
# micro_batch_size: 4
|
|
|
114 |
|
115 |
# Optimizer-related arguments
|
116 |
|
117 |
+
# optimizer:
|
118 |
+
# # class_path: torch.optim.AdamW
|
119 |
+
# class_path: torchao.prototype.low_bit_optim.AdamW8bit
|
120 |
+
# # class_path: torchao.prototype.low_bit_optim.AdamW4bit
|
121 |
+
# # class_path: bitsandbytes.optim.AdamW8bit
|
122 |
+
# # class_path: bitsandbytes.optim.PagedAdamW8bit
|
123 |
+
# init_args:
|
124 |
+
# # (type: float, default: 0.001)
|
125 |
+
# lr: 1e-4
|
126 |
+
# # (type: float, default: 0.01)
|
127 |
+
# weight_decay: 0.01
|
128 |
+
# # (type: tuple, default: (0.9,0.999))
|
129 |
+
# betas:
|
130 |
+
# - 0.9
|
131 |
+
# - 0.99
|
132 |
+
|
133 |
optimizer:
|
134 |
+
class_path: grokadamw.GrokAdamW
|
|
|
|
|
|
|
|
|
135 |
init_args:
|
136 |
# (type: float, default: 0.001)
|
137 |
lr: 1e-4
|
|
|
140 |
# (type: tuple, default: (0.9,0.999))
|
141 |
betas:
|
142 |
- 0.9
|
143 |
+
- 0.999
|
144 |
|
145 |
# How many devices/GPUs to use. Uses all GPUs by default. (type: Union[int, str], default: auto)
|
146 |
devices: auto
|
scripts/requirements.in
CHANGED
@@ -1,29 +1,22 @@
|
|
1 |
# pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
|
2 |
torch>=2.5.0,<2.6.0
|
3 |
numpy<2.0
|
4 |
-
torchao
|
5 |
|
6 |
tqdm
|
|
|
7 |
datasets
|
8 |
jinja2
|
9 |
transformers
|
10 |
wandb
|
11 |
-
# litgpt[all]
|
12 |
litgpt[all] @ git+https://github.com/Lightning-AI/litgpt.git
|
13 |
mergekit @ git+https://github.com/arcee-ai/mergekit.git
|
14 |
-
# litgpt @ git+https://github.com/Lightning-AI/litgpt.git
|
15 |
-
# litdata
|
16 |
-
# litdata @ git+https://github.com/Lightning-AI/litdata.git
|
17 |
-
# lpmm @ git+https://github.com/thu-ml/low-bit-optimizers.git
|
18 |
# muon @ git+https://github.com/KellerJordan/Muon
|
19 |
# pytorch-optimizer
|
20 |
-
|
21 |
bitsandbytes
|
22 |
-
|
23 |
# sophia-opt
|
24 |
-
# bitsandbytes
|
25 |
# pyzstd
|
26 |
# zstd
|
27 |
-
unsloth
|
28 |
-
|
29 |
-
Pillow
|
|
|
1 |
# pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
|
2 |
torch>=2.5.0,<2.6.0
|
3 |
numpy<2.0
|
|
|
4 |
|
5 |
tqdm
|
6 |
+
Pillow
|
7 |
datasets
|
8 |
jinja2
|
9 |
transformers
|
10 |
wandb
|
|
|
11 |
litgpt[all] @ git+https://github.com/Lightning-AI/litgpt.git
|
12 |
mergekit @ git+https://github.com/arcee-ai/mergekit.git
|
|
|
|
|
|
|
|
|
13 |
# muon @ git+https://github.com/KellerJordan/Muon
|
14 |
# pytorch-optimizer
|
15 |
+
torchao
|
16 |
bitsandbytes
|
17 |
+
grokadamw
|
18 |
# sophia-opt
|
|
|
19 |
# pyzstd
|
20 |
# zstd
|
21 |
+
# unsloth
|
22 |
+
lm_eval[ifeval,math]
|
|