SGLang issue 25670: init_process_group crashes with cudaErrorNotSupported on single-GPU virtualized/MIG slices
repo
Fields
id
sources/mgpu-sglang-25670-mig-eager-initPublisher
sgl-project/sglang
Kind
repo
Published
2026-05-18
Accessed
2026-10-03
notes
Exact NCCL/CUDA failure strings on a vGPU/MIG slice at world_size=1; root cause is PyTorch's eager device-connection path probing UVM/peer maps the hypervisor sandboxes.
Notes
Exact NCCL/CUDA failure strings on a vGPU/MIG slice at world_size=1; root cause is PyTorch's eager device-connection path probing UVM/peer maps the hypervisor sandboxes.
Cross-references
This record references no other records.