Fix fsdp2 example (#3657)

Author: shimizust

examples/fsdp2/README.md

@@ -32,5 +32,5 @@ In our example, we use a 8B Llama3.1 model, which has a hidden dimension of 4096
 The figures above were generated on 8x H100 SXM GPUs, with 8192 sequence length and 1000 steps. To run the example, you can use the following command, where you can specify the precision to train in:
 
 ```bash
-accelerate launch --fsdp2_fp8.py --sequence_length 8192 --num_steps 1000 --log_with wandb --precision [fp8 | bf16]
+accelerate launch fsdp2_fp8.py --sequence-length 8192 --num-steps 1000 --log_with wandb --precision [fp8 | bf16]
 ```