Upload config.yaml with huggingface_hub
Browse files- config.yaml +128 -0
config.yaml
ADDED
|
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
checkpoints:
|
| 2 |
+
checkpoint_interval: 10000
|
| 3 |
+
checkpoints_path: checkpoints/360M_UT_original_withcos_early0.01
|
| 4 |
+
checkpoints_path_is_shared_file_system: false
|
| 5 |
+
load_lr_scheduler: false
|
| 6 |
+
load_optimizer: false
|
| 7 |
+
resume_checkpoint_path: /mnt/bn/ridger1/yxy/SmolLM2-nanotron-ckpt/360M/pre-decay
|
| 8 |
+
save_final_state: false
|
| 9 |
+
save_initial_state: false
|
| 10 |
+
data_stages:
|
| 11 |
+
- data:
|
| 12 |
+
dataset:
|
| 13 |
+
dataset_folder:
|
| 14 |
+
- /mnt/bn/ridger1/yxy/datasets/finewebedu-dedup
|
| 15 |
+
- /mnt/bn/ridger1/datasets/cosmopedia-v2
|
| 16 |
+
- /mnt/bn/ridger1/yxy/datasets/megamath/megamath-text-code-block
|
| 17 |
+
- /mnt/bn/ridger1/yxy/datasets/megamath/megamath-qa
|
| 18 |
+
- /mnt/bn/ridger1/datasets/megamath/megamath-translated-code
|
| 19 |
+
- /mnt/bn/ridger1/yxy/datasets/megamath/megamath-web-pro
|
| 20 |
+
dataset_weights:
|
| 21 |
+
- 0.545
|
| 22 |
+
- 0.08
|
| 23 |
+
- 0.25
|
| 24 |
+
- 0.035
|
| 25 |
+
- 0.035
|
| 26 |
+
- 0.075
|
| 27 |
+
token_size_in_bytes: 2
|
| 28 |
+
tokenizer_name: HuggingFaceTB/cosmo2-tokenizer
|
| 29 |
+
vocab_size: 49152
|
| 30 |
+
num_loading_workers: 1
|
| 31 |
+
seed: 42
|
| 32 |
+
name: stable phase
|
| 33 |
+
start_training_step: 1
|
| 34 |
+
general:
|
| 35 |
+
benchmark_csv_path: null
|
| 36 |
+
consumed_train_samples: 51200000
|
| 37 |
+
ignore_sanity_checks: true
|
| 38 |
+
project: diffUT
|
| 39 |
+
run: 360M_32_with_cos_constant_and_decay_early0.1
|
| 40 |
+
seed: 8
|
| 41 |
+
step: 100000
|
| 42 |
+
lighteval: null
|
| 43 |
+
logging:
|
| 44 |
+
iteration_step_info_interval: 1
|
| 45 |
+
log_level: info
|
| 46 |
+
log_level_replica: info
|
| 47 |
+
model:
|
| 48 |
+
ddp_bucket_cap_mb: 25
|
| 49 |
+
dtype: bfloat16
|
| 50 |
+
init_method:
|
| 51 |
+
std: 0.041666666666666664
|
| 52 |
+
make_vocab_size_divisible_by: 1
|
| 53 |
+
model_config:
|
| 54 |
+
_attn_implementation: flash_attention_2
|
| 55 |
+
attention_bias: false
|
| 56 |
+
bos_token_id: 0
|
| 57 |
+
early_ratio: 0.01
|
| 58 |
+
eos_token_id: 0
|
| 59 |
+
hidden_act: silu
|
| 60 |
+
hidden_size: 960
|
| 61 |
+
initializer_range: 0.02
|
| 62 |
+
intermediate_size: 2560
|
| 63 |
+
is_llama_config: true
|
| 64 |
+
max_position_embeddings: 2048
|
| 65 |
+
num_attention_heads: 15
|
| 66 |
+
num_hidden_layers: 32
|
| 67 |
+
num_key_value_heads: 5
|
| 68 |
+
pad_token_id: null
|
| 69 |
+
pretraining_tp: 1
|
| 70 |
+
rms_norm_eps: 1.0e-05
|
| 71 |
+
rope_interleaved: false
|
| 72 |
+
rope_scaling: null
|
| 73 |
+
rope_theta: 10000.0
|
| 74 |
+
tie_word_embeddings: true
|
| 75 |
+
unroll: true
|
| 76 |
+
unroll_end: 25
|
| 77 |
+
unroll_start: 5
|
| 78 |
+
unroll_type: 1
|
| 79 |
+
use_cache: true
|
| 80 |
+
vocab_size: 49152
|
| 81 |
+
z_loss_coefficient: 0.0001
|
| 82 |
+
z_loss_enabled: false
|
| 83 |
+
optimizer:
|
| 84 |
+
accumulate_grad_in_fp32: true
|
| 85 |
+
clip_grad: 1.0
|
| 86 |
+
learning_rate_scheduler:
|
| 87 |
+
learning_rate: 0.003
|
| 88 |
+
lr_decay_starting_step: 10000
|
| 89 |
+
lr_decay_steps: 100000
|
| 90 |
+
lr_decay_style: linear
|
| 91 |
+
lr_warmup_steps: 1024
|
| 92 |
+
lr_warmup_style: linear
|
| 93 |
+
min_decay_lr: 0.0003
|
| 94 |
+
optimizer_factory:
|
| 95 |
+
adam_beta1: 0.9
|
| 96 |
+
adam_beta2: 0.95
|
| 97 |
+
adam_eps: 1.0e-08
|
| 98 |
+
name: adamW
|
| 99 |
+
torch_adam_is_fused: true
|
| 100 |
+
weight_decay: 0.01
|
| 101 |
+
weight_decay_exclude_named_params: []
|
| 102 |
+
zero_stage: 1
|
| 103 |
+
parallelism:
|
| 104 |
+
context_parallel_size: 1
|
| 105 |
+
dp: 64
|
| 106 |
+
expert_parallel_size: 1
|
| 107 |
+
moe_layer_recompute: false
|
| 108 |
+
pp: 1
|
| 109 |
+
pp_engine: 1f1b
|
| 110 |
+
recompute_layer: false
|
| 111 |
+
tp: 1
|
| 112 |
+
tp_linear_async_communication: true
|
| 113 |
+
tp_mode: REDUCE_SCATTER
|
| 114 |
+
tp_recompute_allgather: true
|
| 115 |
+
profiler: null
|
| 116 |
+
s3_upload: null
|
| 117 |
+
tokenizer:
|
| 118 |
+
tokenizer_max_length: null
|
| 119 |
+
tokenizer_name_or_path: HuggingFaceTB/cosmo2-tokenizer
|
| 120 |
+
tokenizer_revision: null
|
| 121 |
+
tokens:
|
| 122 |
+
batch_accumulation_per_replica: 1
|
| 123 |
+
limit_test_batches: 0
|
| 124 |
+
limit_val_batches: 0
|
| 125 |
+
micro_batch_size: 8
|
| 126 |
+
sequence_length: 2048
|
| 127 |
+
train_steps: 102400
|
| 128 |
+
val_check_interval: 10000
|