From 07cfbddadbe72591235298037f05cc257dd64b60 Mon Sep 17 00:00:00 2001 From: Katsumi Takeuchi Date: Fri, 17 Jul 2026 21:59:42 +0900 Subject: [PATCH] sycl: fix SDPA executing of fattn-onednn and ggml queue not synced. --- ggml/src/ggml-sycl/fattn-onednn.cpp | 2 ++ 1 file changed, 2 insertions(+) diff --git a/ggml/src/ggml-sycl/fattn-onednn.cpp b/ggml/src/ggml-sycl/fattn-onednn.cpp index f2e12ef1aeff..c1a4f2be17b8 100644 --- a/ggml/src/ggml-sycl/fattn-onednn.cpp +++ b/ggml/src/ggml-sycl/fattn-onednn.cpp @@ -242,7 +242,9 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso ti.emplace_back(lt, eng, id2ptr(lt.get_id())); } tensor to(E.out, eng, outf.get()); + stream->wait_and_throw(); E.cp.execute(strm, ti, {to}); + strm.wait(); permute_sdpa_out_sycl(outf.get(), (float *) dst->data, mb, H, q, d, stream); // Single device: no sync is required, and actually PP perf is ~6% > wait_and_throw() (tested on llama-3.1-8b & qwen3.6-27b, both Q8_0, with Arc B70).