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: copy the prompt path's grouping tables instead of reading them from mapped memory
Niko1221e318f71e75a6
diff --git a/sycl/src/prefill/prefill.cpp b/sycl/src/prefill/prefill.cppindex 20b04cd5f..8206fb614 100644--- a/sycl/src/prefill/prefill.cpp+++ b/sycl/src/prefill/prefill.cpp@@ -1293,8 +1293,13 @@ bool Prefill::init(const core::WeightTable &wt, const core::ModelGeometry &g,m.ids_host.resize(T * K); m.slot_host.resize(T * K); m.src_host.resize(T * K); m.cnt.resize(m.g->n_expert); m.off.resize(m.g->n_expert + 1);{const size_t need = 3 * T * K + (size_t) (2 * (m.g->n_expert + m.g->n_expert / MMQ_GROUP + 2));+ // SYCL port: the *copy* arm is the default. On the mapped arm the GPU reads these tables straight out of+ // mapped host memory, per layer; copying them first is measurably faster on a PCIe-attached pair and+ // token-identical - 4,106 tokens 1,303 -> 1,335 tok/s, 16,380 tokens 1,622 -> 1,665, the 10,938-token chat+ // prompt 1,576 -> 1,617, decode unchanged (round 46). STRATA_GROUP_COPY=0 restores the mapped arm.const char* gc = std::getenv("STRATA_GROUP_COPY");- if (m.grp_n < need && !(gc && gc[0] == '1')) {+ const bool group_copy = gc == nullptr || gc[0] != '0';+ if (m.grp_n < need && !group_copy) {if (m.grp_host) sycl::free(m.grp_host, dpct::get_in_order_queue());m.grp_host = m.grp_dev = nullptr;m.grp_n = m.grp_tk = 0;