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++
layer-major prefill: stays off on AMD builds (#3976)
Niko1221d02478b7a10c
diff --git a/src/prefill/prefill.cpp b/src/prefill/prefill.cppindex fd75b6ab3..91663284e 100644--- a/src/prefill/prefill.cpp+++ b/src/prefill/prefill.cpp@@ -5639,6 +5639,11 @@ void Prefill::lm_release() {int64_t Prefill::lm_prepare(int64_t n, std::string& why, int64_t cap) {Impl& m = *impl_;if (m.g == nullptr) { why = "not initialized"; return 0; }+#if defined(STRATA_USE_HIP) || defined(STRATA_HIP_GFX906)+ // only NVIDIA CUDA has run this path; on an AMD build the prompt stays on the chunk-major path+ why = "layer-major prefill has only been run on NVIDIA CUDA";+ return 0;+#endifconst core::ModelGeometry& g = *m.g;const core::OnDevice on_device(m.device); // a layer split's later stage is sized on its own GPU// layer split (2026-10-24): each stage runs its own layers layer-major over the same super-chunk and hands the