RuntimeError: CUDA error: no kernel image is available for execution on the device

[uncle-sann]

[rank0]: Traceback (most recent call last): [rank0]: File "dlrm_main.py", line 729, in <module> [rank0]: invoke_main() # pragma: no cover [rank0]: File "dlrm_main.py", line 725, in invoke_main [rank0]: main(sys.argv[1:]) [rank0]: File "dlrm_main.py", line 710, in main [rank0]: train_val_test( [rank0]: File "dlrm_main.py", line 482, in train_val_test [rank0]: _train( [rank0]: File "dlrm_main.py", line 429, in _train [rank0]: pipeline.progress(batched_iterator) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/train_pipeline/train_pipelines.py", line 306, in progress [rank0]: self.fill_pipeline(dataloader_iter) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/train_pipeline/train_pipelines.py", line 290, in fill_pipeline [rank0]: self._init_pipelined_modules( [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/train_pipeline/train_pipelines.py", line 388, in _init_pipelined_modules [rank0]: self._pipeline_model(batch, context, pipelined_forward) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/train_pipeline/train_pipelines.py", line 367, in _pipeline_model [rank0]: self.start_sparse_data_dist(batch, context) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/train_pipeline/train_pipelines.py", line 441, in start_sparse_data_dist [rank0]: _start_data_dist(self._pipelined_modules, batch, context) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/train_pipeline/utils.py", line 435, in _start_data_dist [rank0]: context.input_dist_splits_requests[forward.name] = module.input_dist( [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/embeddingbag.py", line 1021, in input_dist [rank0]: awaitables.append(input_dist(features_by_shard)) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl [rank0]: return self._call_impl(*args, **kwargs) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl [rank0]: return forward_call(*args, **kwargs) [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/sharding/rw_sharding.py", line 316, in forward [rank0]: ) = bucketize_kjt_before_all2all( [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torchrec/distributed/embedding_sharding.py", line 241, in bucketize_kjt_before_all2all [rank0]: ) = torch.ops.fbgemm.block_bucketize_sparse_features( [rank0]: File "/home/john/miniconda3/envs/tfrecsys/lib/python3.8/site-packages/torch/_ops.py", line 1061, in __call__ [rank0]: return self_._op(*args, **(kwargs or {})) [rank0]: RuntimeError: CUDA error: no kernel image is available for execution on the device [rank0]: CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. [rank0]: For debugging consider passing CUDA_LAUNCH_BLOCKING=1 [rank0]: Compile withTORCH_USE_CUDA_DSAto enable device-side assertions.
However pytorch can see gpus:

[sarckk]

Hi @uncle-sann, I think the CUDA error: no kernel image is available for execution on the device error you are running into indicates that your versions of pytorch and CUDA are incompatible. Please refer to pytorch/pytorch#31285. Thanks!