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++
// Windows HIP auto-chunk cap (#1630) and the commit-limit verdict (#1607): no GPU, no Windows needed.
#include <cstdio>
#include "strata/platform/host_limits.hpp"
using namespace strata::platform::host_limits;
static int failed = 0;
#define CHECK(x) do { if (!(x)) { std::fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, #x); ++failed; } } while (0)
int main() {
// a bare auto on Windows HIP is capped below the 8192 that crashed; every other build keeps it
CHECK(auto_chunk_cap(true, true) == 6144);
CHECK(auto_chunk_cap(true, true) < 8192);
CHECK(auto_chunk_cap(false, true) == 0);
CHECK(auto_chunk_cap(true, false) == 0); // auto:N / STRATA_PREFILL_AUTO_MAX: the operator's pick
CHECK(auto_chunk_cap(false, false) == 0);
// a chunk the user names stays, with a warning from 8192 up (7500 was stable)
CHECK(warn_user_chunk(true, false, 8192));
CHECK(warn_user_chunk(true, false, 16384));
CHECK(!warn_user_chunk(true, false, 7500));
CHECK(!warn_user_chunk(true, true, 8192)); // the bare auto is capped instead
CHECK(!warn_user_chunk(false, false, 8192));
// commit: 188 GiB limit with 90 GiB left and a 10 GiB plan: quiet; 95 GiB limit with 6 GiB left: warns
const uint64_t G = 1ull << 30;
CHECK(!commit_verdict(90 * G, 188 * G, 10 * G, 4 * G).warn);
CHECK(!commit_verdict(14 * G, 96 * G, 10 * G, 4 * G).warn); // exactly enough
const CommitVerdict v = commit_verdict(6 * G, 96 * G, 10 * G, 4 * G);
CHECK(v.warn);
CHECK(v.text.find("1607") != std::string::npos);
CHECK(!commit_verdict(0, 0, 10 * G, 4 * G).warn); // unknown limit (not Windows)
if (failed == 0) std::printf("host_limits_test: ok\n");
return failed ? 1 : 0;
}