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++
verify: the DFlash2 tap kernel only where the DFlash2 sources are built (Metal and SYCL keep the per-row dump path)
Niko1221ba6966d83f81
diff --git a/src/core/verify.cpp b/src/core/verify.cppindex 5b898f9da..13d1aac4d 100644--- a/src/core/verify.cpp+++ b/src/core/verify.cpp@@ -56,7 +56,9 @@#include "strata/core/progress.hpp"#include "strata/kernels/shared_expert.hpp"#include "strata/kernels/verify_kernels.hpp"+#if defined(STRATA_HAVE_DFLASH2)#include "strata/kernels/dflash2.hpp"+#endif#include <algorithm>#include <atomic>@@ -2190,8 +2192,16 @@ bool Verifier::record_window(int T, cudaStream_t cs, std::string& err) {// the layer's completed output: R with this FFN half's write applied (the next layer's read would fold it), the mean of// the hc streams, all rows of the group in one launchconst int64_t nt = (int64_t) tap_layers_.size();+#if defined(STRATA_HAVE_DFLASH2)strata::kernels::df2_tap_gr(Rt(tb), HC * N, bo_ + tb * N, N, inj2_ + tb * HC, HC, tap_dev_ + ((size_t) tb * nt + (size_t) tap_slot_[(size_t) l]) * N,nt * N, n, N, (int) HC, cs);+#else // builds without the DFlash2 kernels (Metal, SYCL): the research dump's per-row path+ for (int t = tb; t < te; ++t) {+ cudaMemcpyAsync(tap_scratch_, Rt(t), (size_t) HC * N * 4, cudaMemcpyDeviceToDevice, cs);+ gr_write(tap_scratch_, bo_ + t * N, inj2_ + t * HC, gs, tap_scratch_, cs);+ hc_mean_streams(tap_scratch_, tap_dev_ + ((size_t) t * nt + (size_t) tap_slot_[(size_t) l]) * N, N, (int) HC, cs);+ }+#endif}if (l == g.n_layers - 1) {if (!fuse_head_gr) {