[2026-05-07 14:43:44,144] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect) [2026-05-07 14:43:44,144] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect) [2026-05-07 14:43:44,145] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect) [2026-05-07 14:43:44,145] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect) [2026-05-07 14:43:52,850] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False [2026-05-07 14:43:52,850] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False [2026-05-07 14:43:52,850] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False [2026-05-07 14:43:52,851] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False Trainer._get_train_sampler replaced with custom implementation.Trainer._get_train_sampler replaced with custom implementation.Trainer._get_train_sampler replaced with custom implementation. Trainer._get_train_sampler replaced with custom implementation. [2026-05-07 14:43:53,466] [INFO] [comm.py:675:init_distributed] cdb=None [2026-05-07 14:43:53,467] [INFO] [comm.py:675:init_distributed] cdb=None [2026-05-07 14:43:53,467] [INFO] [comm.py:675:init_distributed] cdb=None [2026-05-07 14:43:53,468] [INFO] [comm.py:675:init_distributed] cdb=None [2026-05-07 14:43:53,468] [INFO] [comm.py:706:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl FlashAttention 3 is availableFlashAttention 3 is available FlashAttention 3 is available FlashAttention 3 is available You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`. You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`. You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`. You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`. Initializing PI3 Encoder... Initializing PI3 Encoder... Initializing PI3 Encoder... Initializing PI3 Encoder... Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version insteadWarning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version insteadWarning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead Loading checkpoint shards: 0%| | 0/4 [00:00 Traceback (most recent call last): File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 480, in matmul_ext_update_autotune_table Exception ignored in atexit callback: Traceback (most recent call last): File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 480, in matmul_ext_update_autotune_table Exception ignored in atexit callback: Traceback (most recent call last): File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 480, in matmul_ext_update_autotune_table Exception ignored in atexit callback: Traceback (most recent call last): File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 480, in matmul_ext_update_autotune_table fp16_matmul._update_autotune_table() fp16_matmul._update_autotune_table() fp16_matmul._update_autotune_table() fp16_matmul._update_autotune_table() File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 457, in _update_autotune_table File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 457, in _update_autotune_table File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 457, in _update_autotune_table File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 457, in _update_autotune_table TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel) TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel) File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 181, in _update_autotune_table TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel) TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel) File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 181, in _update_autotune_table File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 181, in _update_autotune_table File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 181, in _update_autotune_table autotune_table = cache_manager.load() ^^^^^^^^^^^^^^^^^^^^ File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 109, in load autotune_table = cache_manager.load() autotune_table = cache_manager.load() ^^^^^^^^^^^^^^^^^^^^ autotune_table = cache_manager.load() File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 109, in load ^^^^^^^^^^^^^^^^^^^^ File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 109, in load ^^^^^^^^^^^^^^^^^^^^ File "/scratch/Projects/CFP-03/CFP03-SF-107/zhangshihua/miniconda3/envs/DyMLLM/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 109, in load with open(self.file_path, 'rb') as handle: ^^^^^^^^^^^^^^^^^^^^^^^^^^ FileNotFoundError: [Errno 2] No such file or directory: '/home/svu/e1554749/.triton/autotune/Fp16Matmul_2d_kernel.pickle' with open(self.file_path, 'rb') as handle: with open(self.file_path, 'rb') as handle: with open(self.file_path, 'rb') as handle: ^^^^^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^^ FileNotFoundError: [Errno 2] No such file or directory: '/home/svu/e1554749/.triton/autotune/Fp16Matmul_2d_kernel.pickle' FileNotFoundError: [Errno 2] No such file or directory: '/home/svu/e1554749/.triton/autotune/Fp16Matmul_2d_kernel.pickle' ^^^^^^^^^^^^^^^^^^^^^^^^^^ FileNotFoundError: [Errno 2] No such file or directory: '/home/svu/e1554749/.triton/autotune/Fp16Matmul_2d_kernel.pickle' 100%|██████████| 4656/4656 [19:30:39<00:00, 19.33s/it] {'train_runtime': 70302.7025, 'train_samples_per_second': 4.239, 'train_steps_per_second': 0.066, 'train_loss': 7.148990503291494, 'epoch': 1.0} 100%|██████████| 4656/4656 [19:31:42<00:00, 19.33s/it] 100%|██████████| 4656/4656 [19:31:42<00:00, 15.10s/it] [rank0]:[W508 10:17:53.245718034 ProcessGroupNCCL.cpp:1496] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())