Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
#pragma once
// Multi-GPU: make `device` current for a scope and restore the caller's device after it. An object that owns CUDA
// streams, graphs and buffers on one device (a verify stage, the drafter) wraps its public calls in this, so the
// caller's thread may be on any device. A negative device, or the device already current, does nothing.
#include <cuda_runtime.h>
namespace strata::core {
struct OnDevice {
int previous = -1;
explicit OnDevice(int device) {
int cur = 0;
if (device >= 0 && cudaGetDevice(&cur) == cudaSuccess && cur != device && cudaSetDevice(device) == cudaSuccess)
previous = cur;
}
~OnDevice() { if (previous >= 0) cudaSetDevice(previous); }
OnDevice(const OnDevice&) = delete;
OnDevice& operator=(const OnDevice&) = delete;
};
} // namespace strata::core