LoRA training really slow

around 1 token per second, each evaluation take 30 minutes
There are long lines (2000 words) in custom train.jsonl
Please help, give a clue. Thanks!

log:

Iter 1: Val loss 2.612, Val took 1822.826s
Iter 10: Train loss 3.608, It/sec 0.015, Tokens/sec 1.425
Iter 20: Train loss 3.508, It/sec 0.013, Tokens/sec 1.522
Iter 30: Train loss 3.524, It/sec 0.014, Tokens/sec 1.024
Iter 40: Train loss 3.352, It/sec 0.014, Tokens/sec 1.383
Iter 50: Train loss 3.298, It/sec 0.014, Tokens/sec 1.062
Iter 60: Train loss 3.141, It/sec 0.013, Tokens/sec 1.233
Iter 70: Train loss 2.889, It/sec 0.013, Tokens/sec 1.389
Iter 80: Train loss 2.708, It/sec 0.013, Tokens/sec 0.986
Iter 90: Train loss 2.332, It/sec 0.013, Tokens/sec 1.236
Iter 100: Train loss 2.168, It/sec 0.013, Tokens/sec 1.405
Iter 110: Train loss 2.099, It/sec 0.013, Tokens/sec 1.363
Iter 120: Train loss 1.986, It/sec 0.013, Tokens/sec 3.075
Iter 130: Train loss 1.606, It/sec 0.013, Tokens/sec 1.458
Iter 140: Train loss 1.749, It/sec 0.013, Tokens/sec 1.634
Iter 150: Train loss 1.804, It/sec 0.012, Tokens/sec 2.393
Iter 160: Train loss 1.583, It/sec 0.013, Tokens/sec 1.546
Iter 170: Train loss 1.434, It/sec 0.013, Tokens/sec 1.145
Iter 180: Train loss 1.559, It/sec 0.013, Tokens/sec 1.303
Iter 190: Train loss 1.774, It/sec 0.013, Tokens/sec 1.381
Iter 200: Train loss 1.555, It/sec 0.013, Tokens/sec 1.169

Start by locating the LoRA training entry point and inspecting how custom train.jsonl is loaded and evaluated, using the reported validation duration and tokens-per-second logs as the baseline. Reproduce the slowdown with the long 2000-word lines, then identify and verify a documented or code-level improvement that brings evaluation and training performance to an expected range.