Upload nemo_train_params.yaml with huggingface_hub
Browse files- nemo_train_params.yaml +82 -0
nemo_train_params.yaml
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
cfg:
|
2 |
+
micro_batch_size: 20
|
3 |
+
global_batch_size: 8000
|
4 |
+
tensor_model_parallel_size: 1
|
5 |
+
pipeline_model_parallel_size: 1
|
6 |
+
encoder_seq_length: 512
|
7 |
+
max_position_embeddings: 512
|
8 |
+
num_layers: 24
|
9 |
+
hidden_size: 1024
|
10 |
+
ffn_hidden_size: 4096
|
11 |
+
num_attention_heads: 16
|
12 |
+
init_method_std: 0.02
|
13 |
+
hidden_dropout: 0.1
|
14 |
+
kv_channels: null
|
15 |
+
apply_query_key_layer_scaling: true
|
16 |
+
layernorm_epsilon: 1.0e-05
|
17 |
+
make_vocab_size_divisible_by: 128
|
18 |
+
pre_process: true
|
19 |
+
post_process: true
|
20 |
+
bert_binary_head: true
|
21 |
+
tokenizer:
|
22 |
+
library: huggingface
|
23 |
+
type: KBLab/wordpiece-64k-pretok-small_data-tokenizer
|
24 |
+
model: null
|
25 |
+
vocab_file: null
|
26 |
+
merge_file: null
|
27 |
+
native_amp_init_scale: 4294967296
|
28 |
+
native_amp_growth_interval: 1000
|
29 |
+
fp32_residual_connection: false
|
30 |
+
fp16_lm_cross_entropy: false
|
31 |
+
megatron_amp_O2: false
|
32 |
+
grad_allreduce_chunk_size_mb: 125
|
33 |
+
grad_div_ar_fusion: false
|
34 |
+
seed: 666
|
35 |
+
use_cpu_initialization: false
|
36 |
+
onnx_safe: false
|
37 |
+
gradient_as_bucket_view: true
|
38 |
+
activations_checkpoint_granularity: null
|
39 |
+
activations_checkpoint_method: null
|
40 |
+
activations_checkpoint_num_layers: null
|
41 |
+
num_micro_batches_with_partial_activation_checkpoints: null
|
42 |
+
activations_checkpoint_layers_per_pipeline: null
|
43 |
+
sequence_parallel: false
|
44 |
+
data:
|
45 |
+
data_prefix:
|
46 |
+
- 1
|
47 |
+
- /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/wikipedia-wordpiece-64k-pretok-small_data_text_sentence
|
48 |
+
- 1
|
49 |
+
- /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/edepos_html-wordpiece-64k-pretok-small_data_text_sentence
|
50 |
+
- 1
|
51 |
+
- /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/oscar-wordpiece-64k-pretok-small_data_text_sentence
|
52 |
+
- 1
|
53 |
+
- /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/kw3-2017-wordpiece-64k-pretok-small_data_text_sentence
|
54 |
+
- 1
|
55 |
+
- /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/issues-wordpiece-64k-pretok-small_data_text_sentence
|
56 |
+
- 1
|
57 |
+
- /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/mc4-wordpiece-64k-pretok-small_data_text_sentence
|
58 |
+
index_mapping_dir: /project/scratch/$PID/data/wordpiece-64k-pretok-small_data/npy_files/
|
59 |
+
data_impl: mmap
|
60 |
+
splits_string: 980,10,10
|
61 |
+
seq_length: 512
|
62 |
+
skip_warmup: true
|
63 |
+
num_workers: 32
|
64 |
+
dataloader_type: single
|
65 |
+
reset_position_ids: false
|
66 |
+
reset_attention_mask: false
|
67 |
+
eod_mask_loss: false
|
68 |
+
masked_lm_prob: 0.15
|
69 |
+
short_seq_prob: 0.1
|
70 |
+
optim:
|
71 |
+
name: fused_adam
|
72 |
+
lr: 0.0006
|
73 |
+
weight_decay: 0.01
|
74 |
+
betas:
|
75 |
+
- 0.9
|
76 |
+
- 0.98
|
77 |
+
sched:
|
78 |
+
name: CosineAnnealing
|
79 |
+
warmup_steps: 500
|
80 |
+
constant_steps: 500
|
81 |
+
min_lr: 2.0e-05
|
82 |
+
precision: 16
|