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++
// --remote-expert-opt (Remote Expert Decode Optimization) needs a second CUDA device as an expert tier; the SYCL port
// has no such tier (the peer-expert and remote-expert helpers are single-device stubs). The class exists so the shared
// headers and generate.cpp build; init() refuses with a message, so the option reports itself unsupported.
#include "strata/core/remote_expert_opt.hpp"
namespace strata::core {
RemoteExpertOpt::~RemoteExpertOpt() = default;
void RemoteExpertOpt::attach(RemoteExperts&) {}
bool RemoteExpertOpt::init(std::string& err) {
err = "--remote-expert-opt is not supported by the SYCL engine (no second-device expert tier)";
return false;
}
bool RemoteExpertOpt::owns(int64_t, int32_t) const { return false; }
bool RemoteExpertOpt::adapt(const std::vector<float>&, const std::vector<int32_t>&,
const std::vector<std::pair<int32_t, int32_t>>&, int, ExpertSource&) {
return false;
}
void RemoteExpertOpt::begin(const float*, int, int) {}
void RemoteExpertOpt::copy_rows(float*, const float*, int, int, const int32_t*, const int32_t*, void*) const {}
void RemoteExpertOpt::combine(float*, float*, int, int, const uint32_t*, uint32_t, void*) const {}
size_t RemoteExpertOpt::metadata_bytes() { return 0; }
void RemoteExpertOpt::prepare(const RemoteExperts&, void*) const {}
bool RemoteExpertOpt::reduce(RemoteExperts&, const void*, std::string&) { return true; }
void RemoteExpertOpt::accumulate(const RemoteExperts&) {}
} // namespace strata::core