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++

sycl: the planar Q6_K head kernel gets the large-GRF launch at 2-8 columns on Arc A-series (#2340 + #2207)

Niko12219691ea2763fa
diff --git a/sycl/src/kernels/cuda/native_mmvq.dp.cpp b/sycl/src/kernels/cuda/native_mmvq.dp.cpp
index 1ba3c817e..3e2234434 100644
--- a/sycl/src/kernels/cuda/native_mmvq.dp.cpp
+++ b/sycl/src/kernels/cuda/native_mmvq.dp.cpp
@@ -2947,6 +2947,17 @@ void q6_planar_launch(const Q6Planar& w, const Q81Block* x, float* y, dpct::queu
const uint8_t *ql = w.ql, *qh = w.qh, *sc = w.sc;
const sycl::half* d = w.d;
const int n_in = w.n_in, n_out = w.n_out;
+ if constexpr (NCOLS > 1) { // same rule as the GGUF-layout wide kernel (#2207): the multi-column kernel spills at 128 GRF on Alchemist
+ if (mmvq_large_grf(s)) {
+ s->parallel_for<dpct_kernel_name<class native_mmvq_q6k_planar_grf256, dpct_kernel_scalar<NCOLS>>>(
+ sycl::nd_range<3>(sycl::range(1, 1, blocks) * sycl::range(1, WARPS, WARP), sycl::range(1, WARPS, WARP)),
+ sycl::ext::oneapi::experimental::properties{sycl::ext::intel::experimental::grf_size<256>},
+ [=](sycl::nd_item<3>) [[sycl::reqd_sub_group_size(32)]] {
+ native_mmvq_q6k_planar_kernel<NCOLS>(ql, qh, sc, d, x, y, n_in, n_out);
+ });
+ return;
+ }
+ }
s->parallel_for<dpct_kernel_name<class native_mmvq_q6k_planar, dpct_kernel_scalar<NCOLS>>>(
sycl::nd_range<3>(sycl::range(1, 1, blocks) * sycl::range(1, WARPS, WARP), sycl::range(1, WARPS, WARP)),
[=](sycl::nd_item<3>) [[sycl::reqd_sub_group_size(32)]] {