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++

wip (experiment): expert layouts compiled for every HIP arch

Niko12214c953b424dd4
diff --git a/src/kernels/cuda/iq_kernels.cu b/src/kernels/cuda/iq_kernels.cu
index f39e9a953..c1c02d97f 100644
--- a/src/kernels/cuda/iq_kernels.cu
+++ b/src/kernels/cuda/iq_kernels.cu
@@ -4360,7 +4360,7 @@ size_t native_expert_scratch_bytes(int64_t cap, int64_t n_ff) {
// The AMD layouts below (STRATA_EXP_MODE) also build for CUDA: Volta takes mode 8 by default (exp_mode), every
// other CUDA card keeps mode 0, the CUDA layout above.
-#if defined(STRATA_HIP_GFX906) || !defined(__HIPCC__)
+#if 1 // dflash2-kern: the layouts build for every HIP arch too (default mode stays 0 off gfx906; STRATA_EXP_MODE picks one)
#define STRATA_EXP_LAYOUTS 1
#else
#define STRATA_EXP_LAYOUTS 0