You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@benhg the github UI won't let me formally assign this one to you. It'd be great if you could take a look at the added xfail. Are they OK, or is there an underlying issue we should fix?
Currently I don't have a Windows or Jetson system to reproduce those xfails (the systems I had were from a machine pool and got re-imaged in the meantime).
I think it's good to leave the TODOs introduced with #1302 in the code, but what should we do about this issue? Mark as P2?
I agree with you. I also don't have any understanding of VMM on Jetson systems, so I think it's fair to leave the xfails with TODO and follow up when we can text better. I'll defer to Leo's expertise as well.
I took a look at this and found one possible underlying issue behind the Jetson xfail.
test_vmm_allocator_policy_configuration requests gpu_direct_rdma=True, which sets CUmemAllocationProp.allocFlags.gpuDirectRDMACapable for cuMemCreate. The test and VirtualMemoryResource.__init__ currently gate that path on Device.properties.gpu_direct_rdma_supported, but cuda.core also exposes the more specific Device.properties.gpu_direct_rdma_with_cuda_vmm_supported property, documented as support for specifying the GPUDirect RDMA flag with cuMemCreate.
Would a small PR that switches the VMM RDMA guard/test skip to gpu_direct_rdma_with_cuda_vmm_supported be useful here?
Activity
[-]Review test_memory.py `xfail`[/-][+]Review test_memory.py::test_vmm_allocator_policy_configuration `xfail`[/+]rwgk commentedon Dec 2, 2025
@benhg the github UI won't let me formally assign this one to you. It'd be great if you could take a look at the added
xfail. Are they OK, or is there an underlying issue we should fix?benhg commentedon Dec 5, 2025
Hi @rwgk the xfail is okay with me. I don't think there's an underlying issue but I am also not an expert on VMM for windows.
rwgk commentedon Dec 5, 2025
There is also the Jetson failure.
Currently I don't have a Windows or Jetson system to reproduce those
xfails(the systems I had were from a machine pool and got re-imaged in the meantime).I think it's good to leave the TODOs introduced with #1302 in the code, but what should we do about this issue? Mark as P2?
@leofang for advice
benhg commentedon Dec 5, 2025
I agree with you. I also don't have any understanding of VMM on Jetson systems, so I think it's fair to leave the xfails with TODO and follow up when we can text better. I'll defer to Leo's expertise as well.
aryanputta commentedon Jun 23, 2026
I took a look at this and found one possible underlying issue behind the Jetson xfail.
test_vmm_allocator_policy_configurationrequestsgpu_direct_rdma=True, which setsCUmemAllocationProp.allocFlags.gpuDirectRDMACapableforcuMemCreate. The test andVirtualMemoryResource.__init__currently gate that path onDevice.properties.gpu_direct_rdma_supported, butcuda.corealso exposes the more specificDevice.properties.gpu_direct_rdma_with_cuda_vmm_supportedproperty, documented as support for specifying the GPUDirect RDMA flag withcuMemCreate.Would a small PR that switches the VMM RDMA guard/test skip to
gpu_direct_rdma_with_cuda_vmm_supportedbe useful here?