rm CastOutputToFloat when finetuning

[maybeluo]

1. CastOutputToFloat seems unnecessary when finetuning. When computing loss, the code lm_logits = lm_logits.to(torch.float32) will cast half to float32. I also compare the result w/wo the CastOutputToFloat op and run finetuning two times, the loss curves are similar(I only test with first 500 steps with a small proportion of data from alpaca.json).
2. rm an invisible space right after the backslash in README.md.

With CastOutputToFloat:

{'loss': 2.2166, 'learning_rate': 9.140000000000001e-05, 'epoch': 5.0}
{'loss': 0.9893, 'learning_rate': 8.14e-05, 'epoch': 10.0}
{'loss': 0.394, 'learning_rate': 7.14e-05, 'epoch': 15.0}
{'loss': 0.1696, 'learning_rate': 6.14e-05, 'epoch': 20.0}
{'loss': 0.063, 'learning_rate': 5.14e-05, 'epoch': 25.0}
{'loss': 0.0224, 'learning_rate': 4.14e-05, 'epoch': 30.0}
{'loss': 0.0126, 'learning_rate': 3.1400000000000004e-05, 'epoch': 35.0}
{'loss': 0.0093, 'learning_rate': 2.1400000000000002e-05, 'epoch': 40.0}
{'loss': 0.0077, 'learning_rate': 1.1400000000000001e-05, 'epoch': 45.0}
{'loss': 0.0068, 'learning_rate': 1.4000000000000001e-06, 'epoch': 50.0}
{'train_runtime': 250.4525, 'train_samples_per_second': 3.993, 'train_steps_per_second': 1.996, 'train_loss': 0.389122870862484, 'epoch': 50.0}

2nd run:

{'loss': 2.2138, 'learning_rate': 9.120000000000001e-05, 'epoch': 5.0}
{'loss': 0.9816, 'learning_rate': 8.120000000000001e-05, 'epoch': 10.0}
{'loss': 0.4238, 'learning_rate': 7.12e-05, 'epoch': 15.0}
{'loss': 0.2162, 'learning_rate': 6.12e-05, 'epoch': 20.0}
{'loss': 0.0774, 'learning_rate': 5.1200000000000004e-05, 'epoch': 25.0}
{'loss': 0.0294, 'learning_rate': 4.12e-05, 'epoch': 30.0}
{'loss': 0.0156, 'learning_rate': 3.12e-05, 'epoch': 35.0}
{'loss': 0.0106, 'learning_rate': 2.12e-05, 'epoch': 40.0}
{'loss': 0.0086, 'learning_rate': 1.1200000000000001e-05, 'epoch': 45.0}
{'loss': 0.0075, 'learning_rate': 1.2000000000000002e-06, 'epoch': 50.0}
{'train_runtime': 250.4801, 'train_samples_per_second': 3.992, 'train_steps_per_second': 1.996, 'train_loss': 0.3984642353057861, 'epoch': 50.0}

Without CastOutputToFloat: 16.4G
1st run:

{'loss': 2.2132, 'learning_rate': 9.120000000000001e-05, 'epoch': 5.0}
{'loss': 0.9828, 'learning_rate': 8.120000000000001e-05, 'epoch': 10.0}
{'loss': 0.3797, 'learning_rate': 7.12e-05, 'epoch': 15.0}
{'loss': 0.1552, 'learning_rate': 6.12e-05, 'epoch': 20.0}
{'loss': 0.0552, 'learning_rate': 5.1200000000000004e-05, 'epoch': 25.0}
{'loss': 0.0218, 'learning_rate': 4.12e-05, 'epoch': 30.0}
{'loss': 0.0123, 'learning_rate': 3.12e-05, 'epoch': 35.0}
{'loss': 0.0091, 'learning_rate': 2.12e-05, 'epoch': 40.0}
{'loss': 0.0076, 'learning_rate': 1.1200000000000001e-05, 'epoch': 45.0}
{'loss': 0.0067, 'learning_rate': 1.2000000000000002e-06, 'epoch': 50.0}
{'train_runtime': 251.2769, 'train_samples_per_second': 3.98, 'train_steps_per_second': 1.99, 'train_loss': 0.38434655278921126, 'epoch': 50.0}

2nd run:

{'loss': 2.2095, 'learning_rate': 9.120000000000001e-05, 'epoch': 5.0}
{'loss': 0.9806, 'learning_rate': 8.120000000000001e-05, 'epoch': 10.0}
{'loss': 0.3772, 'learning_rate': 7.12e-05, 'epoch': 15.0}
{'loss': 0.1662, 'learning_rate': 6.12e-05, 'epoch': 20.0}
{'loss': 0.0543, 'learning_rate': 5.1200000000000004e-05, 'epoch': 25.0}
{'loss': 0.0235, 'learning_rate': 4.12e-05, 'epoch': 30.0}
{'loss': 0.0128, 'learning_rate': 3.12e-05, 'epoch': 35.0}
{'loss': 0.0094, 'learning_rate': 2.12e-05, 'epoch': 40.0}
{'loss': 0.0078, 'learning_rate': 1.1200000000000001e-05, 'epoch': 45.0}
{'loss': 0.007, 'learning_rate': 1.2000000000000002e-06, 'epoch': 50.0}

[datalee]

mark