@novastera-oss/llamarn 0.2.1 → 0.2.2

package/cpp/llama.cpp/ggml/src/ggml-sycl/gla.cpp

@@ -76,6 +76,7 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B,
 }
 
 void ggml_sycl_op_gated_linear_attn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/5);
     const float * k_d  = static_cast<const float *>(dst->src[0]->data);
     const float * v_d  = static_cast<const float *>(dst->src[1]->data);
     const float * r_d  = static_cast<const float *>(dst->src[2]->data);

package/cpp/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp

@@ -24,6 +24,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
     const int     blocks_per_row              = ncols / block_traits::qk;
     constexpr int blocks_per_subgroup         = ceil_div(block_traits::vdr_mmvq * WARP_SIZE, block_traits::qi);
     constexpr int block_elements_per_subgroup = block_traits::qi / block_traits::vdr_mmvq;
+    const int     nblocks                     = nrows * (ncols / block_traits::qk);
 
     static_assert(blocks_per_subgroup > 0);
     static_assert(block_elements_per_subgroup > 0);
@@ -45,7 +46,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
             // x block quant index when casting the quants to int
             const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
 
-            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs);
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs, nblocks);
         }
     }
 
@@ -739,6 +740,27 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     }
 }
 
+static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+    const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+
+    const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y);
+    constexpr size_t num_subgroups = 16;
+    GGML_ASSERT(block_num_y % num_subgroups == 0);
+
+    const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
+    const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                            [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols,
+                                                                                            nrows, nd_item);
+                            });
+    });
+}
+
+
 static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -1035,7 +1057,14 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 break;
             case GGML_TYPE_Q4_K:
-                mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q4_k_q8_1_sycl\n");
+                    reorder_mul_mat_vec_q4_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                } else {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q4_K_q8_1_sycl\n");
+                    mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
                 break;
             case GGML_TYPE_Q5_K:
                 mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);

package/cpp/llama.cpp/ggml/src/ggml-sycl/norm.cpp

@@ -1,40 +1,50 @@
 #include "norm.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml-sycl/presets.hpp"
 
-static void norm_f32(const float* x, float* dst, const int ncols, const float eps,
-    const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
-    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
-        item_ct1.get_local_id(1);
-    const int tid = item_ct1.get_local_id(2);
+static void norm_f32(const float* x, float* dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
+        const int64_t stride_sample, const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
+
+    const int nrows = item_ct1.get_group_range(2);
+    const int nchannels = item_ct1.get_group_range(1);
 
     const int nthreads = item_ct1.get_local_range(2);
+    const int sample  = item_ct1.get_group(0);
+    const int channel = item_ct1.get_group(1);
+    const int row     = item_ct1.get_group(2);
+
+    const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
+
+    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
+    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+
+    x += strided_offset;
+    dst += packed_offset;
+
     sycl::float2 mean_var = sycl::float2(0.f, 0.f);
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row * ncols + col];
+        const float xi = x[col];
         mean_var.x() += xi;
         mean_var.y() += xi * xi;
     }
 
     // sum up partial sums
     mean_var = warp_reduce_sum(mean_var, item_ct1);
-    if (block_size > WARP_SIZE) {
-
-        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
-        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = mean_var;
+    if  (block_size > WARP_SIZE) {
+        const auto sub_group = item_ct1.get_sub_group();
+        const auto sg_id = sub_group.get_group_linear_id();
+        const auto wi_in_sg = sub_group.get_local_linear_id();
+        if (wi_in_sg == 0) {
+            s_sum[sg_id] = mean_var;
         }
-        /*
-        DPCT1118:0: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
         item_ct1.barrier(sycl::access::fence_space::local_space);
         mean_var = 0.f;
-        size_t nreduce = nwarps / WARP_SIZE;
+        const size_t nreduce = ceil_div(nwarps, WARP_SIZE);
         for (size_t i = 0; i < nreduce; i += 1)
         {
-            mean_var += s_sum[lane_id + i * WARP_SIZE];
+            mean_var += s_sum[wi_in_sg + i * WARP_SIZE];
         }
         mean_var = warp_reduce_sum(mean_var, item_ct1);
     }
@@ -44,7 +54,7 @@ static void norm_f32(const float* x, float* dst, const int ncols, const float ep
     const float inv_std = sycl::rsqrt(var + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row * ncols + col] = (x[row * ncols + col] - mean) * inv_std;
+        dst[col] = (x[col] - mean) * inv_std;
     }
 }
 
@@ -135,39 +145,51 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     }
 }
 
-static void rms_norm_f32(const float* x, float* dst, const int ncols, const float eps,
-    const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
-    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
-        item_ct1.get_local_id(1);
-    const int tid = item_ct1.get_local_id(2);
+static void rms_norm_f32(const float* x, float* dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
+        const int64_t stride_sample, const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
+
+    const int nrows = item_ct1.get_group_range(2);
+    const int nchannels = item_ct1.get_group_range(1);
+
+    const int sample  = item_ct1.get_group(0);
+    const int channel = item_ct1.get_group(1);
+    const int row     = item_ct1.get_group(2);
+
     const int nthreads = item_ct1.get_local_range(2);
+
+    const int tid = item_ct1.get_local_id(2);
     const int nwarps = nthreads / WARP_SIZE;
+
+    const auto strided_offset = calculate_offset<3>({stride_sample, stride_channel, stride_row}, {sample, channel, row});
+    const auto packed_offset = calculate_offset<3>({nchannels * nrows * ncols, nrows * ncols, ncols}, {sample, channel, row});
+
+    x   += strided_offset;
+    dst += packed_offset;
+
+
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row * ncols + col];
+        const float xi = x[col];
         tmp += xi * xi;
     }
 
     // sum up partial sums
     tmp = warp_reduce_sum(tmp, item_ct1);
     if (block_size > WARP_SIZE) {
-
-        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
-        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
+        const auto sub_group = item_ct1.get_sub_group();
+        const auto sg_id = sub_group.get_group_linear_id();
+        const auto wi_in_sg = sub_group.get_local_linear_id();
+        if (wi_in_sg == 0) {
+            s_sum[sg_id] = tmp;
         }
-        /*
-        DPCT1118:3: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
+
         item_ct1.barrier(sycl::access::fence_space::local_space);
-        size_t nreduce = nwarps / WARP_SIZE;
+        const size_t nreduce = ceil_div(nwarps, WARP_SIZE);
         tmp = 0.f;
         for (size_t i = 0; i < nreduce; i += 1)
         {
-            tmp += s_sum[lane_id + i * WARP_SIZE];
+            tmp += s_sum[wi_in_sg + i * WARP_SIZE];
         }
         tmp = warp_reduce_sum(tmp, item_ct1);
     }
@@ -176,7 +198,7 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa
     const float scale = sycl::rsqrt(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row * ncols + col] = scale * x[row * ncols + col];
+        dst[col] = scale * x[col];
     }
 }
 
@@ -224,20 +246,20 @@ static void l2_norm_f32(const float* x, float* dst, const int ncols, const float
     }
 }
 
-static void norm_f32_sycl(const float* x, float* dst, const int ncols,
-    const int nrows, const float eps,
-    queue_ptr stream, int device) {
+static void norm_f32_sycl(const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample,
+        const float eps, queue_ptr stream, int device) {
+
+    const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, eps, item_ct1,
-                        nullptr, WARP_SIZE);
+                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -252,15 +274,12 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
         */
         stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<sycl::float2, 1> s_sum_acc_ct1(
-                sycl::range<1>(work_group_size / WARP_SIZE), cgh);
-
+                            sycl::range<1>(work_group_size / WARP_SIZE), cgh);
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, eps, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
+                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -313,21 +332,20 @@ static void group_norm_f32_sycl(const float* x, float* dst,
     }
 }
 
-static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
-    const int nrows, const float eps,
-    queue_ptr stream, int device) {
+static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, queue_ptr stream, int device) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
+
+    const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, eps, item_ct1,
-                        nullptr, WARP_SIZE);
+                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
                 });
             });
     }
@@ -344,12 +362,10 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
             cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
                 [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, eps, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
+                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
                 });
             });
     }
@@ -398,12 +414,12 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
 }
 
 void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
+    const ggml_tensor * src0 = dst->src[0];
 
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t nrows = ggml_nrows(dst->src[0]);
+    GGML_TENSOR_UNARY_OP_LOCALS
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
@@ -411,8 +427,14 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
-
-    norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+    GGML_ASSERT(eps >= 0.0f);
+    const size_t ts0 = ggml_type_size(src0->type);
+    GGML_ASSERT(nb00 == ts0);
+    const int64_t s01 = nb01 / ts0;
+    const int64_t s02 = nb02 / ts0;
+    const int64_t s03 = nb03 / ts0;
+
+    norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
@@ -436,11 +458,10 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
+    const ggml_tensor * src0 = dst->src[0];
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t nrows = ggml_nrows(dst->src[0]);
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
 
@@ -450,7 +471,13 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+    GGML_TENSOR_UNARY_OP_LOCALS
+    const size_t ts0 = ggml_type_size(src0->type);
+    GGML_ASSERT(nb00 == ts0);
+    const int64_t s01 = nb01 / ts0;
+    const int64_t s02 = nb02 / ts0;
+    const int64_t s03 = nb03 / ts0;
+    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
 }
 
 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {

package/cpp/llama.cpp/ggml/src/ggml-sycl/outprod.cpp

@@ -1,6 +1,7 @@
 #include "outprod.hpp"
 
 void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     const ggml_tensor *src0 = dst->src[0];
     const ggml_tensor *src1 = dst->src[1];
 

package/cpp/llama.cpp/ggml/src/ggml-sycl/quants.hpp

@@ -56,6 +56,28 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q4_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI4_K;
+        static constexpr uint32_t qr       = QR4_K;
+        static constexpr uint32_t vdr_mmvq = 2;
+    };
+
+    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+
+    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks = (nrows * (ncols / traits::qk));
+        return (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2));
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+
+    constexpr size_t get_total_qs_bytes(int nblocks) { return nblocks * QK_K / 2; }
+
+    constexpr size_t get_dm_offset(int nblocks) { return get_total_qs_bytes(nblocks) + nblocks * K_SCALE_SIZE; }
+};
+
 }  // namespace ggml_sycl_reordered
 
 #endif  // GGML_SYCL_QUANTS_HPP

package/cpp/llama.cpp/ggml/src/ggml-sycl/rope.cpp

@@ -355,8 +355,7 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
 }
 
 void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    GGML_SYCL_DEBUG("call %s\n", __func__);
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/3);
     ggml_sycl_op_rope(ctx, dst);
-    GGML_SYCL_DEBUG("call %s done\n", __func__);
 }
 

package/cpp/llama.cpp/ggml/src/ggml-sycl/softmax.cpp

@@ -225,7 +225,7 @@ static void soft_max_f32_sycl(const float * x, const T * mask,
 }
 
 void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -249,16 +249,13 @@ void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
     if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F16) {
         const sycl::half * src1_dd = static_cast<sycl::half *>(dst->src[1]->data);
-        GGML_SYCL_DEBUG("%s: F16 mask\n", __func__);
         soft_max_f32_sycl<sycl::half>(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias,
                           main_stream, ctx.device);
     } else if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F32) {
         const float * src1_dd = static_cast<const float *>(dst->src[1]->data);
-        GGML_SYCL_DEBUG("%s: F32 mask\n", __func__);
         soft_max_f32_sycl<float>(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
     } else {
         /* mask unavailable */
-        GGML_SYCL_DEBUG("%s: No mask\n", __func__);
         soft_max_f32_sycl<float>(src0_dd, nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-sycl/tsembd.cpp

@@ -56,8 +56,8 @@ static void timestep_embedding_f32_sycl(
 }
 
 void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor *src0 = dst->src[0];
-    const ggml_tensor *src1 = dst->src[1];
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    const ggml_tensor *  src0   = dst->src[0];
     const float * src0_d = (const float *)src0->data;
     float * dst_d = (float *)dst->data;
     dpct::queue_ptr stream = ctx.stream();
@@ -69,5 +69,4 @@ void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tenso
     const int max_period = dst->op_params[1];
 
     timestep_embedding_f32_sycl(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
-    GGML_UNUSED(src1);
 }

package/cpp/llama.cpp/ggml/src/ggml-sycl/vecdotq.hpp

@@ -285,7 +285,7 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     }
 
     __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int /* nblocks */) {
         const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
         const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
         int             v[q4_0_traits::vdr_mmvq];
@@ -303,6 +303,67 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     };
 };
 
+static inline float vec_dot_q4_K_q8_1_common(const int * __restrict__ q4, const uint16_t * __restrict__ scales,
+                                             const ggml_half2 & dm, const block_q8_1 * __restrict__ bq8_1,
+                                             const int &        iqs) {
+    int   v[2];
+    int   u[2 * QR4_K];
+    float d8[QR4_K];
+
+    v[0] = q4[0];
+    v[1] = q4[4];
+
+    uint16_t  aux[2];
+    const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+    if (j < 2) {
+        aux[0] = scales[j + 0] & 0x3f3f;
+        aux[1] = scales[j + 2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
+    }
+
+    const uint8_t * sc = (const uint8_t *) aux;
+    const uint8_t * m  = sc + 2;
+
+    const int bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+
+    for (int i = 0; i < QR4_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i]                   = bq8i->ds[0];
+
+        const int * q8 = (const int *) bq8i->qs + ((iqs / 2) % 4);
+        u[2 * i + 0]   = q8[0];
+        u[2 * i + 1]   = q8[4];
+    }
+
+    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, dm, d8);
+}
+
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q4_K;
+
+    using q4_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_K>;
+    using q4_k_traits = typename q4_k_block::traits;
+
+    float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
+                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int nblocks) {
+        const int ib = ibx_offset / (QK_K / 2);
+
+        const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
+        const uint8_t *    qs             = base + ibx_offset;
+        const int          total_qs_bytes = nblocks * (QK_K / 2);
+        const uint8_t *    scs            = base + total_qs_bytes + ib * K_SCALE_SIZE;
+        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset);
+
+        const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+        const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
+        const uint16_t * scales     = (const uint16_t *) scs;
+
+        return vec_dot_q4_K_q8_1_common(q4, scales, *dms, bq8_1, iqs);
+    }
+};
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -649,52 +710,17 @@ vec_dot_q3_K_q8_1(const void *__restrict__ vbq,
     return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
 }
 
-static __dpct_inline__ float
-vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
-                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
-
+static __dpct_inline__ float vec_dot_q4_K_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1,
+                                               const int & iqs) {
 #ifndef GGML_QKK_64
-    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
-
-    int    v[2];
-    int    u[2*QR4_K];
-    float d8[QR4_K];
 
-    // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
-    const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
-
-    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
-    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
-    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
-    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
-
-    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
-    v[0] = q4[0];
-    v[1] = q4[4];
-
-    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
-    uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
-    const uint8_t * sc = (const uint8_t *)aux;
-    const uint8_t * m  = sc + 2;
-
-    for (int i = 0; i < QR4_K; ++i) {
-        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds[0];
+    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
-        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
-        u[2*i+0] = q8[0];
-        u[2*i+1] = q8[4];
-    }
+    const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+    const int *      q4         = (const int *) (bq4_K->qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
+    const uint16_t * scales     = (const uint16_t *) bq4_K->scales;
 
-    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
+    return vec_dot_q4_K_q8_1_common(q4, scales, bq4_K->dm, bq8_1, iqs);
 
 #else
 

package/cpp/llama.cpp/ggml/src/ggml-sycl/wkv.cpp

@@ -180,10 +180,7 @@ static void rwkv_wkv7_f32_kernel(
 }
 
 void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
-
-    const ggml_tensor *src0 = dst->src[0];
-    const ggml_tensor *src1 = dst->src[1];
-
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/6);
     const float* k_d = (const float*)dst->src[0]->data;
     const float* v_d = (const float*)dst->src[1]->data;
     const float* r_d = (const float*)dst->src[2]->data;
@@ -236,16 +233,10 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     }
-
-    GGML_UNUSED(src0);
-    GGML_UNUSED(src1);
 }
 
 void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
-
-    const ggml_tensor *src0 = dst->src[0];
-    const ggml_tensor *src1 = dst->src[1];
-
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/7);
     const float* r_d = (const float*)dst->src[0]->data;
     const float* w_d = (const float*)dst->src[1]->data;
     const float* k_d = (const float*)dst->src[2]->data;
@@ -299,7 +290,4 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     }
-
-    GGML_UNUSED(src0);
-    GGML_UNUSED(src1);
 }