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++
dsv41: DS4_PROF expert-kernel timers on both set paths
Niko12218fda83538362
diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cuindex 41304d8af..1d724de87 100644--- a/src/dsv41/ds4_model.cu+++ b/src/dsv41/ds4_model.cu@@ -3441,34 +3441,35 @@ struct Model::Impl {sa.n_embd = D;sa.n_ff = FF;sa.limit = g.lim;+ EkMark ek{};+ const bool ekp = prof_chunk;+ if (ekp) {+ std::lock_guard<std::mutex> lk(prof_mu);+ ek.e0 = ek_ev();+ ek.e1 = ek_ev();+ ek.e2 = ek_ev();+ ek.tg = gg::expert_type(sa.lay, 0);+ ek.td = gg::expert_type(sa.lay, 2);+ ek.nrows = nrows;+ ek.nexp = j - i0;+ ek.maxrows = max_rows;+ ek.ff = FF;+ cudaGetDevice(&ek.dev);+ cudaEventRecord(ek.e0, d.st);+ }if (sa.lay.fmt == gg::kFmtMXFP4) {gg::set_gate_up(sa, c.x, D, c.row_tok + a, c.h16, FF, d.st);+ if (ekp) cudaEventRecord(ek.e1, d.st);gg::set_down(sa, c.h16, FF, c.y, D, d.st);} else {- EkMark ek{};- const bool ekp = prof_chunk;- if (ekp) {- std::lock_guard<std::mutex> lk(prof_mu);- ek.e0 = ek_ev();- ek.e1 = ek_ev();- ek.e2 = ek_ev();- ek.tg = gg::expert_type(sa.lay, 0);- ek.td = gg::expert_type(sa.lay, 2);- ek.nrows = nrows;- ek.nexp = j - i0;- ek.maxrows = max_rows;- ek.ff = FF;- cudaGetDevice(&ek.dev);- cudaEventRecord(ek.e0, d.st);- }x3::set_gate_up(sa, c.x, D, c.row_tok + a, c.h16, FF, d.st);if (ekp) cudaEventRecord(ek.e1, d.st);x3::set_down(sa, c.h16, FF, c.y, D, d.st);- if (ekp) {- cudaEventRecord(ek.e2, d.st);- std::lock_guard<std::mutex> lk(prof_mu);- ek_marks.push_back(ek);- }+ }+ if (ekp) {+ cudaEventRecord(ek.e2, d.st);+ std::lock_guard<std::mutex> lk(prof_mu);+ ek_marks.push_back(ek);}if (dump_y != nullptr && (dump_y_layer < 0 || l.il == dump_y_layer)) {std::vector<int> bb(b + i0, b + j + 1);