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++
[Engine] SYCL: native_expert_bench refuses to time a launch that wrote nothing
Niko12218475011ac1e3
diff --git a/sycl/src/kernels/native_expert_bench.cpp b/sycl/src/kernels/native_expert_bench.cppindex 1c1b6566e..c75726c91 100644--- a/sycl/src/kernels/native_expert_bench.cpp+++ b/sycl/src/kernels/native_expert_bench.cpp@@ -135,6 +135,23 @@ int main(int argc, char** argv) {if (!sync("reference window")) return 1;s->memcpy(ref.data(), dout, ref.size() * 4).wait();+ // SYCL port: a launch that returns without enqueueing leaves the 0xff fill intact, and then the mode and the+ // reference agree bitwise on an untouched buffer - a "0 differing" that measures nothing. The fill is NaN as+ // a float, so a reference window with no finite value anywhere did no work: say so instead of timing it.+ {+ size_t written = 0;+ for (size_t i = 0; i < ref.size(); ++i) {+ float v;+ std::memcpy(&v, &ref[i], 4);+ if (std::isfinite(v)) ++written;+ }+ if (written == 0) {+ std::printf("layer %d: the expert launch wrote nothing (%zu outputs, none finite) - refusing to time it\n",+ (int) l, ref.size());+ ++failures;+ continue;+ }+ }s->memset(dout, 0xff, got.size() * 4).wait();run(mode, 0);if (!sync("mode window")) return 1;