Upload ablation-ncp-log16 ncp config
Browse files
log16/baseline/ncp/training-config.yaml
ADDED
|
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
task: next-concept
|
| 2 |
+
experiment: ablation-ncp-log16
|
| 3 |
+
experiment_dir: /workspace/ncm/runs/ablation-ncp-log16/ncp
|
| 4 |
+
model:
|
| 5 |
+
n_layer: 16
|
| 6 |
+
n_head: 16
|
| 7 |
+
bias: true
|
| 8 |
+
dropout: 0.1
|
| 9 |
+
n_embd: 1024
|
| 10 |
+
context_length: 1024
|
| 11 |
+
vocab_size:
|
| 12 |
+
- 16384
|
| 13 |
+
- 16384
|
| 14 |
+
- 16384
|
| 15 |
+
- 16384
|
| 16 |
+
- 16384
|
| 17 |
+
- 16384
|
| 18 |
+
- 16384
|
| 19 |
+
- 16384
|
| 20 |
+
- 16384
|
| 21 |
+
- 16384
|
| 22 |
+
- 16384
|
| 23 |
+
- 16384
|
| 24 |
+
- 16384
|
| 25 |
+
- 16384
|
| 26 |
+
- 16384
|
| 27 |
+
- 16384
|
| 28 |
+
attn_pattern: block_diagonal
|
| 29 |
+
use_positional_encoding: true
|
| 30 |
+
use_level_encoding: true
|
| 31 |
+
prefix_len: 0
|
| 32 |
+
use_rope: true
|
| 33 |
+
rope_base: 100000
|
| 34 |
+
use_qk_norm: true
|
| 35 |
+
post_upsample_conv:
|
| 36 |
+
enabled: false
|
| 37 |
+
kernel_size: 3
|
| 38 |
+
use_relu2: true
|
| 39 |
+
use_flex_attention: true
|
| 40 |
+
shared_output_head: true
|
| 41 |
+
shared_head_mlp: false
|
| 42 |
+
shared_head_mlp_mult: 2.0
|
| 43 |
+
shared_head_mlp_residual: true
|
| 44 |
+
shared_head_mlp_blocks: 2
|
| 45 |
+
n_kv_heads: 4
|
| 46 |
+
mlp_ratio: 3.5
|
| 47 |
+
in_vocab_size: 50304
|
| 48 |
+
predict_level0: true
|
| 49 |
+
bos_token_id: 50256
|
| 50 |
+
tokenizer:
|
| 51 |
+
pre_tokenizer:
|
| 52 |
+
name: hf
|
| 53 |
+
tokenizer: hf
|
| 54 |
+
model_id: gpt2
|
| 55 |
+
special_tokens:
|
| 56 |
+
eos_token: <|endoftext|>
|
| 57 |
+
vqvae:
|
| 58 |
+
n_layers: 6
|
| 59 |
+
vqvae_vocab_size:
|
| 60 |
+
- 16384
|
| 61 |
+
- 16384
|
| 62 |
+
- 16384
|
| 63 |
+
- 16384
|
| 64 |
+
- 16384
|
| 65 |
+
- 16384
|
| 66 |
+
- 16384
|
| 67 |
+
- 16384
|
| 68 |
+
- 16384
|
| 69 |
+
- 16384
|
| 70 |
+
- 16384
|
| 71 |
+
- 16384
|
| 72 |
+
- 16384
|
| 73 |
+
- 16384
|
| 74 |
+
- 16384
|
| 75 |
+
- 16384
|
| 76 |
+
embed_dim: 384
|
| 77 |
+
compression_factor: 1
|
| 78 |
+
in_vocab_size: 50304
|
| 79 |
+
context_length: 1024
|
| 80 |
+
# Exact pinned codec. Never a directory: the sweep must be able to state
|
| 81 |
+
# which VQ-VAE every NCP result was measured through.
|
| 82 |
+
checkpoint_path: ncm/runs/ablation-vqvae-log16/vqvae/checkpoint-iter-10000.pt
|
| 83 |
+
vector_quantizer_config:
|
| 84 |
+
clss: multiscale_residual_vector_quantizer
|
| 85 |
+
decay: 0.99
|
| 86 |
+
epsilon: 1.0e-05
|
| 87 |
+
commitment_cost: 0.25
|
| 88 |
+
learned_l1_sampling: true
|
| 89 |
+
learned_l1_cascade: true
|
| 90 |
+
learned_l1_cascade_base: 4
|
| 91 |
+
learned_all_sampling: false
|
| 92 |
+
quant_resi:
|
| 93 |
+
enabled: true
|
| 94 |
+
ratio: 0.5
|
| 95 |
+
share_mode: 0
|
| 96 |
+
learnable_ratio: false
|
| 97 |
+
levels:
|
| 98 |
+
use_manual_levels: true
|
| 99 |
+
manual_levels:
|
| 100 |
+
- 1
|
| 101 |
+
- 16
|
| 102 |
+
- 64
|
| 103 |
+
- 160
|
| 104 |
+
- 288
|
| 105 |
+
- 410
|
| 106 |
+
- 512
|
| 107 |
+
- 600
|
| 108 |
+
- 676
|
| 109 |
+
- 742
|
| 110 |
+
- 800
|
| 111 |
+
- 852
|
| 112 |
+
- 898
|
| 113 |
+
- 940
|
| 114 |
+
- 984
|
| 115 |
+
- 1024
|
| 116 |
+
aux:
|
| 117 |
+
fine_drop:
|
| 118 |
+
prob: 0.5
|
| 119 |
+
min_keep: 1
|
| 120 |
+
dropout: 0.2
|
| 121 |
+
pre_quant_groupnorm: 4
|
| 122 |
+
pre_quant_dropout: null
|
| 123 |
+
dataset:
|
| 124 |
+
train_source: data/full_owt/train_gpt2.bin
|
| 125 |
+
validate_source: data/full_owt/valid_gpt2.bin
|
| 126 |
+
wandb:
|
| 127 |
+
project: ncp-colm
|
| 128 |
+
group: ablation-ncp-log16
|
| 129 |
+
enable: false
|
| 130 |
+
resume: null
|
| 131 |
+
gradients_and_params:
|
| 132 |
+
enable: false
|
| 133 |
+
log: all
|
| 134 |
+
log_freq: 1000
|
| 135 |
+
optimization:
|
| 136 |
+
level_loss_alpha: 0.25
|
| 137 |
+
lr: 0.0005
|
| 138 |
+
reset_lr: false
|
| 139 |
+
warmup_iters: 300
|
| 140 |
+
lr_decay_iters: null
|
| 141 |
+
min_lr: 1.0e-05
|
| 142 |
+
beta_1: 0.9
|
| 143 |
+
beta_2: 0.99
|
| 144 |
+
weight_decay: 0.01
|
| 145 |
+
max_grad_norm: 1.0
|
| 146 |
+
grad_accumulation_steps: 8
|
| 147 |
+
corruption:
|
| 148 |
+
mode: none
|
| 149 |
+
training:
|
| 150 |
+
seed: 55
|
| 151 |
+
log_interval: 10
|
| 152 |
+
eval_interval: 1000
|
| 153 |
+
checkpoint_interval: 500
|
| 154 |
+
val_iters: 100
|
| 155 |
+
total_iters: null
|
| 156 |
+
tqdm_interval: 10
|
| 157 |
+
n_epochs: 3
|
| 158 |
+
batch_size: 8
|
| 159 |
+
eval_batch_size: 16
|
| 160 |
+
demo_batch_size: 4
|
| 161 |
+
stop_after_iters: 2000
|
| 162 |
+
keep_last: 100
|
| 163 |
+
torch_compile:
|
| 164 |
+
enable: true
|
| 165 |
+
mode: default
|
| 166 |
+
dynamic: false
|
| 167 |
+
fullgraph: false
|
| 168 |
+
backend: inductor
|
| 169 |
+
mixed_precision:
|
| 170 |
+
enable: true
|