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++

launch header: exl3.cu includes it for every backend; exl3_cpu kGemmK only off x86

Niko12215ed8ae03be72
diff --git a/src/dsv41/vendor/exl3.cu b/src/dsv41/vendor/exl3.cu
index e94ffd7b6..2cdbc074b 100644
--- a/src/dsv41/vendor/exl3.cu
+++ b/src/dsv41/vendor/exl3.cu
@@ -19,8 +19,8 @@
// NVIDIA GPUs before sm_80 (V100, T4, RTX 20) decode with v1.1.8's kernels instead (exl3_legacy.cu, use_legacy
// below): this file's CUDA-core fallback was slower there (a V100 with EXL3-3.05bpw: decode 4.9 -> 4.4 tok/s). The
// prompt path below is this file's on every GPU.
-#if defined(STRATA_METAL)
#include "../ds4_launch.h" // strata_ds4_launch: <<<>>> on CUDA and HIP, a dispatch on Metal
+#if defined(STRATA_METAL)
#include <ds4_metal_host.h>
#endif
#include "strata/dsv41/exl3.hpp"
diff --git a/src/dsv41/vendor/exl3_cpu.cpp b/src/dsv41/vendor/exl3_cpu.cpp
index b7a313fbc..ae953586f 100644
--- a/src/dsv41/vendor/exl3_cpu.cpp
+++ b/src/dsv41/vendor/exl3_cpu.cpp
@@ -1,4 +1,4 @@
-#if !defined(__AVX512F__)
+#if !(defined(__x86_64__) || defined(_M_X64)) // (x86 defines it below, inside its guard)
// the band GEMM's chunk width, defined here because the uses are outside the AVX-512 guard
constexpr int kGemmK = 256;
#endif