PyTorch torch.cuda.nccl.reduce denial of service with invalid operation codes
PyTorch's torch.cuda.nccl.reduce function crashed with a core dump when called with invalid operation codes instead of validating inputs or raising an error. PyTorch developers added input validation checks and tests to address the issue.
What happened
When torch.cuda.nccl.reduce was called with invalid operation codes, the program aborted with a core dump instead of raising a RuntimeError or validating the input.
Reported impact
The function crash prevented normal program operation when invalid operation codes were passed to torch.cuda.nccl.reduce.
Index analysis
The issue was caused by missing input validation for the operation parameter in torch.cuda.nccl functions.
Governance controls
PyTorch developers added validation checks for the operation parameter in nccl functions and added related tests to prevent invalid operation codes from causing crashes.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.