whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp

@@ -59,6 +59,8 @@ void ggml_sycl_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
+void ggml_sycl_expm1(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
 void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_softplus(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
@@ -71,8 +73,6 @@ void ggml_sycl_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp

@@ -0,0 +1,56 @@
+//
+// MIT license
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "common.hpp"
+
+sycl::half * ggml_sycl_fattn_kv_buffers::kv_buffer::ensure_half(size_t n_elems) {
+    const size_t need_bytes = n_elems * sizeof(sycl::half);
+
+    if (capacity >= need_bytes) {
+        return ptr;
+    }
+
+    if (ptr) {
+        SYCL_CHECK(CHECK_TRY_ERROR(qptr->wait()));
+        SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
+        ptr = nullptr;
+        capacity = 0;
+    }
+
+    size_t cap = 0;
+    while (cap < need_bytes) {
+        cap += CHUNK_SIZE;
+    }
+
+    void * dev_ptr;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(dev_ptr = sycl::malloc_device(
+                        cap, *qptr)));
+
+    if (!dev_ptr) {
+        GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, cap);
+        GGML_ABORT("fattn buffer alloc failed");
+    }
+
+    ptr = static_cast<sycl::half *>(dev_ptr);
+    capacity = cap;
+    return ptr;
+}
+
+ggml_sycl_fattn_kv_buffers::kv_buffer::~kv_buffer() {
+#ifdef DEBUG_SYCL_POOL
+    GGML_LOG_INFO("ggml_sycl_fattn_kv_buffer[%d]: %.2f MiB\n", device, capacity / 1024.0 / 1024.0);
+#endif
+    if (ptr) {
+        SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
+    }
+}

data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp

@@ -0,0 +1,63 @@
+//
+// MIT license
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_FATTN_BUFFERS_HPP
+#define GGML_SYCL_FATTN_BUFFERS_HPP
+
+#include <sycl/sycl.hpp>
+
+typedef sycl::queue *queue_ptr;
+
+struct ggml_sycl_fattn_kv_buffers {
+    // buffers grow in chunks of this size
+    static constexpr size_t CHUNK_SIZE = 16ull << 20; // 16 MiB
+
+    struct kv_buffer {
+        kv_buffer(queue_ptr qptr_, int device_) : qptr(qptr_), device(device_) {}
+        ~kv_buffer();
+
+        kv_buffer(const kv_buffer &) = delete;
+        kv_buffer & operator=(const kv_buffer &) = delete;
+
+        sycl::half * ensure_half(size_t n_elems);
+
+    private:
+        sycl::half * ptr      = nullptr;
+        size_t       capacity = 0;
+        queue_ptr    qptr     = nullptr;
+        [[maybe_unused]] int device = 0;
+    };
+
+    kv_buffer K;
+    kv_buffer V;
+
+    ggml_sycl_fattn_kv_buffers(queue_ptr qptr, int device) : K(qptr, device), V(qptr, device) {}
+
+    ggml_sycl_fattn_kv_buffers(const ggml_sycl_fattn_kv_buffers &) = delete;
+    ggml_sycl_fattn_kv_buffers & operator=(const ggml_sycl_fattn_kv_buffers &) = delete;
+};
+
+/**
+ * Imitates `ggml_sycl_pool_alloc` to keep the code calling alloc unchanged.
+ */
+struct ggml_sycl_fattn_alloc {
+    ggml_sycl_fattn_kv_buffers::kv_buffer & buf;
+    sycl::half *                         ptr = nullptr;
+
+    explicit ggml_sycl_fattn_alloc(ggml_sycl_fattn_kv_buffers::kv_buffer & buf_) : buf(buf_) {}
+
+    sycl::half * alloc(size_t n_elems) {
+        ptr = buf.ensure_half(n_elems);
+        return ptr;
+    }
+};
+#endif

data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp

@@ -5,6 +5,7 @@
 #include "common.hpp"
 #include "convert.hpp"
 #include "vecdotq.hpp"
+#include "fattn-buffers.hpp"
 
 #include "ggml.h"
 
@@ -918,12 +919,13 @@ void launch_fattn(
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
 
     ggml_sycl_pool & pool = ctx.pool();
+    ggml_sycl_fattn_kv_buffers & fbuf = ctx.fattn_buffers();
     dpct::queue_ptr  main_stream = ctx.stream();
     const int id  = ggml_sycl_get_device();
     const int nsm = ggml_sycl_info().devices[id].nsm;
 
-    ggml_sycl_pool_alloc<sycl::half>   K_f16(pool);
-    ggml_sycl_pool_alloc<sycl::half>   V_f16(pool);
+    ggml_sycl_fattn_alloc        K_f16(fbuf.K);
+    ggml_sycl_fattn_alloc        V_f16(fbuf.V);
     ggml_sycl_pool_alloc<int>    KV_max(pool);
     ggml_sycl_pool_alloc<float>  dst_tmp(pool);
     ggml_sycl_pool_alloc<sycl::float2> dst_tmp_meta(pool);
@@ -1029,7 +1031,7 @@ void launch_fattn(
                 auto KV_max_ptr_ct1 = KV_max.ptr;
 
                 cgh.parallel_for(sycl::nd_range<3>(blocks_num_KV_max * block_dim_KV_max, block_dim_KV_max),
-                                 [=](sycl::nd_item<3> item_ct1) {
+                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(warp_size)]] {
                                      GGML_UNUSED(item_ct1);
                                      flash_attn_mask_to_KV_max<ncols1, warp_size>(
                                          mask_data_ct0, KV_max_ptr_ct1, iter_k, s31, s33,
@@ -1147,7 +1149,7 @@ void launch_fattn(
                 auto K_ne_ct6             = K->ne[2];
 
                 cgh.parallel_for(sycl::nd_range<3>(blocks_num_combine * block_dim_combine, block_dim_combine),
-                                 [=](sycl::nd_item<3> item_ct1) {
+                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(warp_size)]] {
                                      GGML_UNUSED(item_ct1);
                                      flash_attn_stream_k_fixup<DV, ncols1, ncols2>(KQV_data_ct0, dst_tmp_meta_ptr_ct1,
                                                                                    Q_ne_ct2, Q_ne_ct3, Q_ne_ct4,
@@ -1167,7 +1169,7 @@ void launch_fattn(
             auto KQV_data_ct2         = (float *) KQV->data;
 
             cgh.parallel_for(sycl::nd_range<3>(blocks_num_combine * block_dim_combine, block_dim_combine),
-                             [=](sycl::nd_item<3> item_ct1) {
+                             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(warp_size)]] {
                                  GGML_UNUSED(item_ct1);
                                  flash_attn_combine_results<DV>(
                                      dst_tmp_ptr_ct0, dst_tmp_meta_ptr_ct1, KQV_data_ct2, parallel_blocks,

data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp

@@ -44,6 +44,10 @@ void ggml_sycl_flash_attn_ext_tile(ggml_backend_sycl_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_sycl_flash_attn_ext_tile_case<256, 256>(ctx, dst);
         } break;
+        case 512: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_sycl_flash_attn_ext_tile_case<512, 512>(ctx, dst);
+        } break;
         case 576: {
             GGML_ASSERT(V->ne[0] == 512);
             ggml_sycl_flash_attn_ext_tile_case<576, 512>(ctx, dst);

data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp

@@ -67,9 +67,16 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp16(const int DKQ, co
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  2,  64, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2,  64,  64)
+
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2,  64,  64)
 
     return 0;
 }
@@ -123,6 +130,12 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp32(const int DKQ, co
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  2, 128, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
+    GGML_SYCL_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2,  64,  64)
+
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
@@ -130,134 +143,6 @@ static constexpr uint32_t ggml_sycl_fattn_tile_get_config_fp32(const int DKQ, co
     return 0;
 }
 
-static constexpr uint32_t ggml_sycl_fattn_tile_get_config_amd(const int DKQ, const int DV, const int ncols) {
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  2,  64, 3,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  4, 128, 3,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  8, 128, 2,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  2,  64, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  4, 128, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  8, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 16, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 32, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 64, 256, 2,  32,  48)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  2,  64, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  4, 128, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  8, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2,  32,  56)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  2, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  4, 128, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  8, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 2,  64,  32)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  2, 256, 2, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  4, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  8, 256, 2,  64, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
-
-    return 0;
-}
-
-static constexpr uint32_t ggml_sycl_fattn_tile_get_config_amd_rdna(const int DKQ, const int DV, const int ncols) {
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 40,  40, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  4,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64,  8, 128, 5, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 128, 5, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 128, 4,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 128, 5,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 16, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 32, 256, 2,  32,  40)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 80,  80, 64, 256, 2,  32,  40)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  2,  64, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  4, 128, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96,  8, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 16, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 32, 256, 2,  32,  48)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE( 96,  96, 64, 256, 2,  32,  48)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  2,  64, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  4, 128, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112,  8, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2,  32,  56)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2,  32,  56)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  4, 128, 8,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128,  8, 128, 8,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 3, 128, 128)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 3, 128,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 3,  64,  64)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  2,  64, 8,  32,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  4, 128, 6,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256,  8, 128, 6,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
-
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
-    GGML_SYCL_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
-
-    return 0;
-}
-
 static constexpr uint32_t ggml_sycl_fattn_tile_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
     if(fast_fp16_available(cc))
         return ggml_sycl_fattn_tile_get_config_fp16(DKQ, DV, ncols);
@@ -310,11 +195,11 @@ static __dpct_inline__ void flash_attn_tile_load_tile(const sycl::half2 * const
                                                       sycl::half2 * const __restrict__ tile_KV,
                                                       const int stride_KV,
                                                       const int i_sup) {
+    auto      item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
     constexpr int cpy_nb = ggml_sycl_get_max_cpy_bytes();
     constexpr int cpy_ne = cpy_nb / 4;
 
     auto load = [&] (const int n) {
-        auto      item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
         const int stride_j = warp_size >> n;
 
         if (stride_j == 0) {
@@ -455,7 +340,7 @@ static __dpct_inline__ void flash_attn_tile_iter_KQ(T_vec_dot * const Q_tmp,
 
     flash_attn_tile_load_tile<warp_size, nwarps, nbatch_fa, nbatch_K, cpy_ne, oob_check>
         (K_h2 + int64_t(k_VKQ_0)*stride_K2 + k_KQ_0/2, KV_tmp, stride_K2, k_VKQ_sup);
-    item_ct1.barrier();
+    item_ct1.barrier(sycl::access::fence_space::local_space);
 
 #ifdef SYCL_FAST_FP16
     static_assert((nbatch_K/2) % cpy_ne == 0, "bad nbatch_K");
@@ -505,7 +390,7 @@ static __dpct_inline__ void flash_attn_tile_iter_KQ(T_vec_dot * const Q_tmp,
     }
 
     if (k_KQ_0 + nbatch_K < DKQ) {
-        item_ct1.barrier();  // Sync not needed on last iteration.
+        item_ct1.barrier(sycl::access::fence_space::local_space);  // Sync not needed on last iteration.
     }
 }
 
@@ -545,7 +430,7 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
                                                  const int         k_VKQ_max,
                                                  const int         col_Q_0,
                                                  float *           KQ_max_new_shared) {
-    auto          item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
     constexpr int cpy_nb   = ggml_sycl_get_max_cpy_bytes();
     constexpr int cpy_ne = cpy_nb / 4;
 
@@ -620,14 +505,14 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
     }
 
     if constexpr (np == 1) {
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
     } else {
         static_assert(cpw == 1, "bad cpw");
 
         if (item_ct1.get_local_id(2) == 0) {
             KQ_max_new_shared[item_ct1.get_local_id(1)] = KQ_max_new[0];
         }
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
         KQ_max_new[0] = KQ_max_new_shared[(item_ct1.get_local_id(1) & ~(np - 1)) + item_ct1.get_local_id(2) % np];
         KQ_max_new[0] = warp_reduce_max<np>(KQ_max_new[0]);
     }
@@ -697,7 +582,7 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
     for (int k0 = 0; k0 < nbatch_fa; k0 += nbatch_V) {
         flash_attn_tile_load_tile<warp_size, nwarps, nbatch_V, DV, 0, oob_check>
             (V_h2 + int64_t(k_VKQ_0 + k0)*stride_V2, KV_tmp, stride_V2, k_VKQ_sup - k0);
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
 
 #ifdef SYCL_FAST_FP16
 #pragma unroll
@@ -765,7 +650,7 @@ static __dpct_inline__ void flash_attn_tile_iter(T_vec_dot * const Q_tmp,
             }
         }
 #endif // SYCL_FAST_FP16
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
     }
 }
 
@@ -972,7 +857,7 @@ static void flash_attn_tile(const char *  Q,
         }
     }
 
-    item_ct1.barrier();
+    item_ct1.barrier(sycl::access::fence_space::local_space);
 
     // Main loop over KV cache:
     const int k_VKQ_max = KV_max ? KV_max[sequence * item_ct1.get_group_range(2) + item_ct1.get_group(2)] : ne11;
@@ -1051,7 +936,7 @@ static void flash_attn_tile(const char *  Q,
             return;
         }
 
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
 
 #pragma unroll
         for (int ip = 1; ip < np; ++ip) {
@@ -1193,37 +1078,39 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_sycl_context & ctx, ggm
 
     constexpr size_t nbytes_shared = 0;
 
-    if constexpr (DV <= 256) {
-        if (Q->ne[1] > 16/ncols2) {
-            constexpr int cols_per_block = 32;
-            const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
-            const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
-            launch_fattn<DV, cols_per_block/ncols2, ncols2,
-                flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
-                (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
-            return;
+    if (DV < 512 && Q->ne[1] < 32) {
+        if constexpr (ncols2 <= 32) {
+            if (Q->ne[1] > 16/ncols2) {
+                constexpr int cols_per_block = 32;
+                const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+                const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+                launch_fattn<DV, cols_per_block/ncols2, ncols2,
+                    flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+                    (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+                return;
+            }
         }
-    }
-
-    if (Q->ne[1] > 8/ncols2) {
-        constexpr int cols_per_block = 16;
-        const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
-        const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
-        launch_fattn<DV, cols_per_block/ncols2, ncols2,
-            flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
-            (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
-        return;
-    }
-
-    if constexpr (ncols2 <= 8) {
-        if (Q->ne[1] > 4/ncols2) {
-            constexpr int cols_per_block = 8;
-            const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
-            const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
-            launch_fattn<DV, cols_per_block/ncols2, ncols2,
-                flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
-                (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
-            return;
+        if constexpr (ncols2 <= 16) {
+            if (Q->ne[1] > 8/ncols2) {
+                constexpr int cols_per_block = 16;
+                const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+                const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+                launch_fattn<DV, cols_per_block/ncols2, ncols2,
+                    flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+                    (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+                return;
+            }
+        }
+        if constexpr (ncols2 <= 8) {
+            if (Q->ne[1] > 4/ncols2) {
+                constexpr int cols_per_block = 8;
+                const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+                const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+                launch_fattn<DV, cols_per_block/ncols2, ncols2,
+                    flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+                    (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+                return;
+            }
         }
     }
 
@@ -1249,6 +1136,16 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_sycl_context & ctx, ggm
         return;
     }
 
+    {
+        constexpr int cols_per_block = ncols2*2;
+        const int nwarps    = ggml_sycl_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+        const int nbatch_fa = ggml_sycl_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+        launch_fattn<DV, cols_per_block/ncols2, ncols2,
+            flash_attn_tile<DKQ, DV, cols_per_block / ncols2, ncols2, use_logit_softcap, warp_size>, warp_size>
+            (ctx, dst, nwarps, nbytes_shared, nbatch_fa, true, true, false);
+        return;
+    }
+
     GGML_ABORT("fatal error");
 }
 
@@ -1280,6 +1177,16 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_sycl_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
             return;
         }
+        // ncols2=2 and ncols2=1 fallbacks only for cases where ncols=2 config exists (DKQ == DV).
+        // For DKQ == 576, DV == 512 only GQA-optimized variants are implemented.
+        if constexpr (DKQ == DV) {
+            if (use_gqa_opt && gqa_ratio % 2 == 0) {
+                launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
+                return;
+            }
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
+            return;
+        }
     }
 
     if constexpr (DV <= 256) {
@@ -1334,5 +1241,6 @@ extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);
 extern DECL_FATTN_TILE_CASE(128, 128);
 extern DECL_FATTN_TILE_CASE(256, 256);
+extern DECL_FATTN_TILE_CASE(512, 512);
 extern DECL_FATTN_TILE_CASE(576, 512);
 

data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp

@@ -664,4 +664,11 @@ EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q8_0)
 
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_F16)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q4_0)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q4_1)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q5_0)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q5_1)
+EXTERN_DECL_FATTN_VEC_CASES(512, GGML_TYPE_Q8_0)
+
 #endif // GGML_SYCL_FATTN_VEC_HPP

data/ext/sources/ggml/src/ggml-sycl/fattn.cpp

@@ -34,6 +34,7 @@
     FATTN_VEC_CASE( 64, type_K, type_V)       \
     FATTN_VEC_CASE(128, type_K, type_V)       \
     FATTN_VEC_CASE(256, type_K, type_V)       \
+    FATTN_VEC_CASE(512, type_K, type_V)       \
 
 static void ggml_sycl_flash_attn_ext_vec(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_tensor * Q = dst->src[0];
@@ -141,6 +142,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
         case 128:
         case 112:
         case 256:
+        case 512:
             if (V->ne[0] != K->ne[0]) {
                 return BEST_FATTN_KERNEL_NONE;
             }
@@ -185,7 +187,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
     }
 
     // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
-    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     // Todo: Use the XMX kernel if possible:
 

data/ext/sources/ggml/src/ggml-sycl/fill.cpp

@@ -0,0 +1,55 @@
+#include "fill.hpp"
+#include "common.hpp"
+
+#define SYCL_FILL_BLOCK_SIZE 256
+
+template <typename T>
+static void fill_kernel(T * dst, const int64_t k, const T value,
+                        const sycl::nd_item<1> & item) {
+    const int64_t i = (int64_t)item.get_global_id(0);
+    if (i >= k) {
+        return;
+    }
+    dst[i] = value;
+}
+
+inline void ggml_sycl_op_fill(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    dpct::queue_ptr stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    float value;
+    memcpy(&value, dst->op_params, sizeof(float));
+
+    const int64_t k = ggml_nelements(dst);
+    const int64_t num_blocks = (k + SYCL_FILL_BLOCK_SIZE - 1) / SYCL_FILL_BLOCK_SIZE;
+    void * dst_d = dst->data;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            stream->parallel_for(
+                sycl::nd_range<1>(num_blocks * SYCL_FILL_BLOCK_SIZE, SYCL_FILL_BLOCK_SIZE),
+                [=](sycl::nd_item<1> item) {
+                    fill_kernel(static_cast<float *>(dst_d), k, value, item);
+                });
+            break;
+        case GGML_TYPE_F16:
+            {
+                sycl::half h_value = sycl::half(value);
+                stream->parallel_for(
+                    sycl::nd_range<1>(num_blocks * SYCL_FILL_BLOCK_SIZE, SYCL_FILL_BLOCK_SIZE),
+                    [=](sycl::nd_item<1> item) {
+                        fill_kernel(static_cast<sycl::half *>(dst_d), k, h_value, item);
+                    });
+            }
+            break;
+        default:
+            GGML_ABORT("unsupported type");
+    }
+}
+
+void ggml_sycl_fill(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/0);
+    ggml_sycl_op_fill(ctx, dst);
+}

data/ext/sources/ggml/src/ggml-sycl/fill.hpp

@@ -0,0 +1,5 @@
+#pragma once
+
+#include "common.hpp"
+
+void ggml_sycl_fill(ggml_backend_sycl_context & ctx, ggml_tensor * dst);