Potential workaround for "cpy.cu: GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX) failed" (CUDA/HIP)

[stduhpf]

Sometimes, especially when generating or decoding long high resolution videos on CUDA or HIP backends, it crashes with this assert in ggml_cuda_cpy() kernel.

Out of frustration, I just commented out the asserts and somehow, it seems to run just fine.
Looking at the code, I don't immediately see what's the purpose of these asserts. So I'd say it might be safe to remove them if they're in your way.

Am I missing something? Should I just propose the change upstream?

[CarlGao4]

llama.cpp has encountered this issue when using Qwen3-Next-80B, so it has been fixed in ggml-org/llama.cpp#18433
Waiting ggml repo to follow up this change and then sd.cpp could update ggml so this would be fixed