skipping CUDA graph capture

[quaternior]

Hello, thanks for your amazing work! While I was evaluating with this framework, I found some problems as below, with following configuations

• 1 A100 (80GB)
• LLaDA-8B
• Evaluating with GSM8K, benchmarking

  0%|          | 0/1319 [00:00<?, ?it/s]
 14%|█▍        | 185/1319 [00:00<00:00, 1840.44it/s]
 28%|██▊       | 370/1319 [00:00<00:00, 1829.80it/s]
 42%|████▏     | 555/1319 [00:00<00:00, 1835.87it/s]
 56%|█████▌    | 739/1319 [00:00<00:00, 726.18it/s] 
 70%|██████▉   | 923/1319 [00:00<00:00, 928.34it/s]
 84%|████████▍ | 1108/1319 [00:00<00:00, 1117.84it/s]
 98%|█████████▊| 1293/1319 [00:01<00:00, 1283.09it/s]
100%|██████████| 1319/1319 [00:01<00:00, 1191.79it/s]

[1088, 1056, 1152, 1120, 1184, 1216]

  0%|          | 0/6 [00:00<?, ?it/s]skipping cudagraphs due to skipping cudagraphs due to cpu device (arg3_1)
skipping cudagraphs due to skipping cudagraphs due to cpu device (arg7_1)

Still the torch.compile is applied, CUDA graph capture is skipped.