Skip to content

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++

tq_pascal_bench: other expert geometries (a longer hidden row, shorter experts) against the per-column kernels

Niko12219b5b758358c0
diff --git a/src/kernels/tq_pascal_bench.cpp b/src/kernels/tq_pascal_bench.cpp
index e6d2c7cfc..bbed71aca 100644
--- a/src/kernels/tq_pascal_bench.cpp
+++ b/src/kernels/tq_pascal_bench.cpp
@@ -292,6 +292,13 @@ int main(int argc, char** argv) {
}
for (int gu : {144, 145})
for (int dt : {144, 146}) check_grouped(gu, dt, 2560, 640, s, rng);
+ // other geometries: a longer hidden row (the staged gate/up falls back from its 80-block activation buffer), a shorter
+ // expert row (the 8-lane down rows need n_ff 640)
+ for (int gu : {144, 145}) {
+ check_grouped(gu, 146, 5120, 640, s, rng);
+ check_grouped(gu, 144, 2560, 512, s, rng);
+ check_grouped(gu, 146, 2048, 256, s, rng);
+ }
if (do_bench) bench(s, rng);
std::printf("tq_pascal_bench: %d failures\n", g_fail);
cudaStreamDestroy(s);