BluScriptCP / logs /in2_train_csv /CP_Training_log11_config.txt
unparallelled's picture
Upload folder using huggingface_hub
d7922f9 verified
Raw History Blame Contribute Delete
571 Bytes
Configuration (bluscriptCP):
d_model: 1536
n_layers: 36
q_heads: 12
kv_heads: 4 (G=3, head_dim=128)
ffn_hidden: 4096
context_length: 16384
vocab_size: 50304 (tying=0)
B: 4
T: 16384
world_size: 4 (dp=1 x cp=4)
CP_ATTN_MODE: ring (load_balance=HeadTail)
global_batch: 524288
grad_accum_steps: 8
Parameters: 1060624896
Parameters per GPU: 1060624896
Max Learning Rate: 0.001
Min Learning Rate: 0.0001
max_steps: 400
warmup_steps: 1
checkpointing: 1
ckpt_freq: 100
run: 11
GPU Memory Used (rank 0): 129812.7 MB / 143155.9 MB