CUDA runtime errors on H20 GPUs — possible missing Hopper architecture in compilation

[lvjx04]

Hi,

I tried running TilERT v0.1.1 on 8 × NVIDIA H20 GPUs. I got these errors:

/workspace/include/core/executor/piped_prefetch/entry_ex.cuh(196): CUDA error: no kernel image is available for execution on the device.
/workspace/src/lib/ops/tilert_init_op.cu(65): CUDA error: operation not permitted when stream is capturing.

The project was mainly tested on B200 GPUs. Could it be that the current build doesn’t include support for Hopper (H20) architecture?

Is there a recommended way to rebuild or set flags to support H20 GPUs?

Thanks!

[xysmlx]

Hi,

Thanks for reaching out.

The current implementation of TileRT is strictly optimized for the Blackwell architecture and leverages some Blackwell-specific hardware instructions.

Therefore, it cannot be rebuilt or run on the Hopper architecture at this moment, as the implementation relies on some specific hardware features.

Thanks for your understanding!