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++

dsv41 iq: HIP keeps the release grids and kernels (no persistent work-item grid there)

Niko122186bdfbd0dafe
diff --git a/src/dsv41/ds4_iq.cu b/src/dsv41/ds4_iq.cu
index fd97be241..40a6ff9bf 100644
--- a/src/dsv41/ds4_iq.cu
+++ b/src/dsv41/ds4_iq.cu
@@ -515,8 +515,13 @@ template <int T, bool F = true> constexpr int kTabWords = F && Fast<T>::WORDS >
// Persistent grids (blocks looping over work items) only for the types with a codebook worth keeping in the block (8 KiB and
// more); the others run a block (tile, group) as before - their blocks leave at once when the group is empty.
+#if defined(STRATA_USE_HIP)
+template <int T>
+constexpr bool kPersist = false; // (HIP keeps the release grids)
+#else
template <int T>
constexpr bool kPersist = kTabWords<T> * 4 >= 8192;
+#endif
template <int T, bool F = true>
__device__ __forceinline__ void load_tab_fast(uint32_t* tab) {