NDR_patch_size: 16
accum_freq: 1
aug_cfg: {}
batch_size: 1024
beta1: 0.9
beta2: 0.98
checkpoint_path: ./logs-lr1e-3-datacomp/clip_vit_b16_s512m_bs16k_mix0_8/checkpoints
coca_caption_loss_weight: 2.0
coca_contrastive_loss_weight: 1.0
copy_codebase: False
csv_caption_key: title
csv_img_key: filepath
csv_separator: 	
dataset_resampled: False
dataset_type: webdataset
ddp_static_graph: True
debug: False
delete_prev_step_ckpt: True
delete_previous_checkpoint: False
device: cuda:0
dist_backend: nccl
dist_url: env://
distill: False
distill_model: None
distill_pretrained: None
distributed: True
epochs: 4
epochs_cooldown: None
eps: 1e-06
force_custom_text: False
force_image_size: 224
force_patch_dropout: None
force_quick_gelu: False
gather_with_grad: True
global_batch_size: 16384
grad_checkpointing: True
grad_clip_norm: None
horovod: False
image_interpolation: None
image_mean: None
image_resize_mode: None
image_std: None
imagenet_v2: None
imagenet_val: /mnt/bn/zilongdata-hl/dataset/imagenet/val
is_cls_token: True
local_loss: True
local_rank: 0
lock_image: False
lock_image_freeze_bn_stats: False
lock_image_unlocked_groups: 0
lock_text: False
lock_text_freeze_layer_norm: False
lock_text_unlocked_layers: 0
log_every_n_steps: 128
log_level: 20
log_local: False
log_path: ./logs-lr1e-3-datacomp/clip_vit_b16_s512m_bs16k_mix0_8/out.log
logs: ./logs-lr1e-3-datacomp
lr: 0.001
lr_cooldown_end: 0.0
lr_cooldown_power: 1.0
lr_scheduler: cosine
max_seq_len: 15000
model: ViT-B-16
name: clip_vit_b16_s512m_bs16k_mix0_8
native_dynamic_resolution: False
no_set_device_rank: False
only_packing: False
precision: amp
pretrained: 
pretrained_image: 
pretrained_text: 
rank: 0
remote_sync: None
remote_sync_frequency: 300
remote_sync_protocol: s3
report_to: wandb
resume: None
rope_attn_num_heads: 12
rope_model_width: 768
save_every_n_steps: 6104
save_frequency: 1
save_most_recent: False
seed: 0
siglip: False
skip_scheduler: False
tensorboard: False
tensorboard_path: 
torchcompile: False
torchscript: False
trace: False
train_data: /mnt/bn/zilongdata-hl/dataset/Recap-DataComp-1B-Dataset/{000000..140146}.tar
train_data_upsampling_factors: None
train_num_samples: 128000000
use_bn_sync: False
use_bnb_linear: None
val_data: None
val_frequency: 1
val_num_samples: None
val_steps: 0
wandb: True
wandb_notes: 
wandb_project_name: cls-clip-NDR
warmup: 500
wd: 0.2
workers: 1
world_size: 16
zeroshot_frequency: 4
zeroshot_steps: 0