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++
[Engine] SYCL: the lane-cell prompt attention default is for Battlemage (Xe2) cards only
Niko1221f2bf7a3ce3cc
diff --git a/sycl/src/kernels/cuda/qsa_decode_attn.dp.cpp b/sycl/src/kernels/cuda/qsa_decode_attn.dp.cppindex 9caef220d..18abd96a6 100644--- a/sycl/src/kernels/cuda/qsa_decode_attn.dp.cpp+++ b/sycl/src/kernels/cuda/qsa_decode_attn.dp.cpp@@ -668,12 +668,18 @@ void qsa_decode_attn_batch(const float* q, const QsaAttnPools& pools, const int3// S25: the score phase one cell per thread (bit-identical scores, no shuffle reductions). Prompt batches take// it by default: on 2x Arc Pro B65 it reads a 4,106-token prompt 5.2% faster, 8,191 tokens 5.2% and 16,380// tokens 4.8%, with the same tokens; the decode window's few queries per call are 1.4% slower with it (56.6-57.0- // against 55.9-56.3 tok/s), so the choice follows the query count. STRATA_ATTN_LANECELL=1/0 forces either way.+ // against 55.9-56.3 tok/s), so the choice follows the query count. It is a Battlemage (Xe2) default only: on an Arc+ // A750 (Alchemist) the same switch reads a 4,106-token prompt 0.7% slower (three interleaved runs, 4 of 5 pairs+ // faster with it off), so the generation comes from the device's architecture id (strata::intel_gpu_gen: the+ // runtime's architecture enum first, the name only as a fallback), never from the name alone.+ // STRATA_ATTN_LANECELL=1/0 forces either way on any device.static const int lane_cell_env = [] {const char* v = std::getenv("STRATA_ATTN_LANECELL");return v != nullptr && v[0] != '\0' ? std::atoi(v) : -1;}();- const bool lane_cell = lane_cell_env >= 0 ? lane_cell_env != 0 : n_q >= 16;+ const bool lane_cell = lane_cell_env >= 0+ ? lane_cell_env != 0+ : (n_q >= 16 && strata::intel_gpu_gen(st->get_device()) == strata::IntelGpuGen::Battlemage);#define STRATA_ATTN_LC(M) \{ \auto exp_props = sycl::ext::oneapi::experimental::properties{ \