sycl: fix use-after-return of the SDPA scale in the oneDNN flash-attention path

docs/backend/SYCL.md

@@ -794,6 +794,8 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_USE_LEVEL_ZERO_API | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON at build time. SYCL backend always runs on Level Zero running time even if it's set as OFF (The SYCL api will be usage for memory allocation).|
 | GGML_SYCL_ENABLE_DNN | 0 or 1 (default)| Enable running computations through oneDNN and always use oneMKL. |
+| GGML_SYCL_FA_ONEDNN | 1 (default) or 0 | Enable the oneDNN fused SDPA (flash-attention) path on supported GPUs. Set to 0 to always use the native SYCL flash-attention kernel. |
+| GGML_SYCL_FA_ONEDNN_MAX_KV | 0 (default, disabled) or positive integer | By default (0), all sequences are handled by the oneDNN fused SDPA path, regardless of KV length; a positive value caps that length, past which sequences fall back to the native kernel. If GPU driver watchdog resets (DEVICE_LOST) occur during long-context inference, set this near the context depth where they start, e.g. 24576. |
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute (currently top-k MoE gating). |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |

ggml/src/ggml-sycl/common.hpp

@@ -65,6 +65,7 @@ extern int g_ggml_sycl_prioritize_dmmv;
 extern int g_ggml_sycl_enable_flash_attention;
 extern int g_ggml_sycl_dev2dev_memcpy;
 extern int g_ggml_sycl_fa_onednn;
+extern int g_ggml_sycl_fa_onednn_max_kv;
 
 
 #if defined(__clang__) && __has_builtin(__builtin_expect)

ggml/src/ggml-sycl/fattn-onednn.cpp

@@ -1,5 +1,6 @@
 #include <cstdint>
 #include <cstdio>
+#include <cstdlib>
 #include <cstring>
 #include <string>
 #include <unordered_map>
@@ -38,6 +39,12 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
     if (K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16) {
         return false;
     }
+    // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
+    // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
+    // some stacks; past the cap we fall back to the native FA kernel instead.
+    if (g_ggml_sycl_fa_onednn_max_kv > 0 && K->ne[1] > g_ggml_sycl_fa_onednn_max_kv) {
+        return false;
+    }
     // gate for the following cases
     // 1. if the oneDNN graph Add node has no input --> skip
     // 2. types other than f16 need different logical_tensor declaration
@@ -208,9 +215,27 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
     cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf.get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
 
     // divide-by-(1/scale) reproduces ggml's score *= kq_scale on the proven probe graph.
+    //
+    // The scale upload must not be an async memcpy from a stack local: on the in-order queue
+    // the copy waits behind the K/V staging kernels, and once those take long enough
+    // (n_kv >= ~26k on B70) the frame is recycled before the copy runs, feeding the SDPA a
+    // garbage scale (output collapses to a repeated token). Cache one device scalar per
+    // (device, value) -- the scale is constant per model -- and upload it synchronously once.
     const sycl::half scale_h = (sycl::half) (1.0f / kq_scale);
-    ggml_sycl_pool_alloc<sycl::half> scbuf(ctx.pool(), 1);
-    stream->memcpy(scbuf.get(), &scale_h, sizeof(sycl::half));
+    static std::unordered_map<std::string, sycl::half *> scale_cache;
+    char sckey[64];
+    snprintf(sckey, sizeof(sckey), "%d:%a", ggml_sycl_get_device(), (double) (float) scale_h);
+    sycl::half * scale_dev;
+    {
+        auto sit = scale_cache.find(sckey);
+        if (sit == scale_cache.end()) {
+            scale_dev = sycl::malloc_device<sycl::half>(1, *stream);
+            stream->memcpy(scale_dev, &scale_h, sizeof(sycl::half));
+            scale_cache.emplace(sckey, scale_dev);
+        } else {
+            scale_dev = sit->second;
+        }
+    }
 
     ggml_sycl_pool_alloc<sycl::half> outf(ctx.pool(), (size_t) H * q * d);   // f16 contiguous SDPA out [mb,H,q,d]
 
@@ -232,7 +257,7 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
         if (r == E.id_q)     return Qf.get();
         if (r == E.id_k)     return Kf.get();
         if (r == E.id_v)     return Vf.get();
-        if (r == E.id_scale) return scbuf.get();
+        if (r == E.id_scale) return scale_dev;
         if (r == E.id_mask)  return (void *) mask->data;
         return nullptr;
     };
@@ -245,14 +270,12 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
     E.cp.execute(strm, ti, {to});
 
     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).
-    // Any future multi-GPU refactor MUST re-measure this single-device path and keep the best
-    // single-device PP speed. Otherwise (multiple devices/streams can race the reuse):
+    // Single device needs no sync: the dnnl stream wraps this same in-order queue, so the SDPA
+    // serializes with the staging kernels before it and the permute/pool reuse after it. The
+    // garbage output formerly blamed on the missing sync here was the scale use-after-return
+    // fixed above. Keep the conservative wait for multi-GPU, where other devices' streams can
+    // race the pool:
     if (ggml_sycl_info().device_count > 1) {
-        // cont_to_f16 -> oneDNN execute -> permute is async on this stream, but the
-        // pool_alloc*s above free their device buffers at host return. Without this wait the next
-        // scheduler op re-acquires those bytes while the GPU is still computing the SDPA, turning
-        // it into garbage and collapsing multi-turn output to a single repeated token ("GGGGG...").
         stream->wait_and_throw();
     }
 }

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -85,6 +85,7 @@ int g_ggml_sycl_enable_optimize = 1;
 int g_ggml_sycl_enable_graph = 0;
 int g_ggml_sycl_enable_dnn = 1;
 int g_ggml_sycl_fa_onednn = 1;
+int g_ggml_sycl_fa_onednn_max_kv = 0;
 int g_ggml_sycl_enable_vmm = 1;
 int g_ggml_sycl_enable_fusion = 1;
 int g_ggml_sycl_prioritize_dmmv = 0;
@@ -287,6 +288,7 @@ static void ggml_check_sycl() try {
         g_ggml_sycl_enable_graph = ggml_sycl_get_env("GGML_SYCL_ENABLE_GRAPH", 0);
         g_ggml_sycl_enable_dnn = ggml_sycl_get_env("GGML_SYCL_ENABLE_DNN", 1);
         g_ggml_sycl_fa_onednn = ggml_sycl_get_env("GGML_SYCL_FA_ONEDNN", 1);
+        g_ggml_sycl_fa_onednn_max_kv = ggml_sycl_get_env("GGML_SYCL_FA_ONEDNN_MAX_KV", 0);
         g_ggml_sycl_enable_vmm = ggml_sycl_get_env("GGML_SYCL_ENABLE_VMM", 1);
         g_ggml_sycl_enable_fusion = ggml_sycl_get_env("GGML_SYCL_ENABLE_FUSION", 1);
         g_ggml_sycl_prioritize_dmmv = ggml_sycl_get_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
@@ -359,6 +361,7 @@ static void ggml_check_sycl() try {
         GGML_LOG_INFO("  GGML_SYCL_ENABLE_DNN: DNN disabled by compile flag\n");
         GGML_LOG_INFO("  GGML_SYCL_FA_ONEDNN: %d\n", g_ggml_sycl_fa_onednn);
 #endif
+        GGML_LOG_INFO("  GGML_SYCL_FA_ONEDNN_MAX_KV: %d\n", g_ggml_sycl_fa_onednn_max_kv);
 #ifdef SYCL_FLASH_ATTN
         GGML_LOG_INFO("  GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
 #else

tests/test-backend-ops.cpp

@@ -9542,6 +9542,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_flash_attn_ext(64, 128, 4, {1, 1}, 128, 2, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q4_0, GGML_TYPE_Q1_0));
     test_cases.emplace_back(new test_flash_attn_ext(128, 64, 4, {1, 1}, 64, 2, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q1_0, GGML_TYPE_F16));
 
+    // large-KV F16 cases (Qwen3.6-27B geometry and a llama-class control): the upstream matrix
+    // stops at kv=1024, blind to long-context FA bugs (e.g. the oneDNN SDPA ordering race on BMG).
+    for (int64_t kv : { 4096, 16384 }) {
+        test_cases.emplace_back(new test_flash_attn_ext(256, 256, 4, {6, 1}, kv, 512, true, false, 0, 0,
+                                                        GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+        test_cases.emplace_back(new test_flash_attn_ext(128, 128, 8, {4, 1}, kv, 512, true, false, 0, 0,
+                                                        GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    }
+
     test_cases.emplace_back(new test_cross_entropy_loss     (GGML_TYPE_F32, {   10, 5, 4, 3}));
     test_cases.emplace_back(new test_cross_entropy_loss     (GGML_TYPE_F32, {30000, 1, 1, 1}));
     test_cases.emplace_back(new test_cross_entropy_loss_back(GGML_TYPE_F32, {   10, 5, 4, 3}));