Repro: Memory Savings at What Cost? (ICML 2026, QS0ZROPx8u)
Collection
Reproduction of BP vs checkpointed-BP vs FmAD vs ZO for LLM fine-tuning. Paper QS0ZROPx8u / arXiv 2506.21833. • 4 items • Updated
args dict | results dict |
|---|---|
{
"model_name": "FacebookAI/roberta-base",
"num_labels": 4,
"max_seq_len": 64,
"batch_size": 40,
"lora_r": 1,
"lora_alpha": 1,
"bp_lr": 0.001,
"zo_lr": 0.0001,
"eps": 0.001,
"n_pert": 1,
"max_steps": 2000,
"eval_every": 50,
"target_acc": 0.85,
"n_train": 8000,
"n_test": 2000,
"seed": 0,
... | {
"bp_vanilla": {
"method": "bp_vanilla",
"lr": 0.001,
"n_trainable": 630532,
"flops_per_step": 879526633088,
"peak_mem_bytes": 1787896832,
"final_acc": 0.92,
"best_acc": 0.9285,
"total_time_s": 238.68994450569153,
"steps": 2000,
"steps_to_target": 50,
"time_to_target_s": 7... |
No dataset card yet