Compare commits

...

1 Commits

Author SHA1 Message Date
Kawrakow
b18f658a7d CUDA: set current device in compute_forward 2025-12-05 15:40:48 +00:00

View File

@ -2950,6 +2950,13 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
int64_t tim1 = ggml_time_us();
#endif
if (ggml_is_noop(dst)) {
return true;
}
// In case we forget to do that in some kernel.
ggml_cuda_set_device(ctx.device);
auto next = i < cgraph->n_nodes - 1 ? cgraph->nodes[i+1] : nullptr;
auto fusion = ctx.fusion;