@fugood/llama.node 0.3.14 → 0.3.16

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

@@ -1,7 +1,7 @@
 #include "common.hpp"
 #include "element_wise.hpp"
 
-void acc_f32(const float * x, const float * y, float * dst, const int ne,
+static void acc_f32(const float * x, const float * y, float * dst, const int ne,
     const int ne10, const int ne11, const int ne12,
     const int nb1, const int nb2, int offset, const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
@@ -20,7 +20,7 @@ void acc_f32(const float * x, const float * y, float * dst, const int ne,
     }
 }
 
-void gelu_f32(const float * x, float * dst, const int k,
+static void gelu_f32(const float * x, float * dst, const int k,
                      const sycl::nd_item<3> &item_ct1) {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
@@ -37,7 +37,7 @@ void gelu_f32(const float * x, float * dst, const int k,
               sycl::tanh(SQRT_2_OVER_PI * xi * (1.0f + GELU_COEF_A * xi * xi)));
 }
 
-void silu_f32(const float * x, float * dst, const int k,
+static void silu_f32(const float * x, float * dst, const int k,
                      const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -48,7 +48,7 @@ void silu_f32(const float * x, float * dst, const int k,
     dst[i] = x[i] / (1.0f + sycl::native::exp(-x[i]));
 }
 
-void gelu_quick_f32(const float *x, float *dst, int k,
+static void gelu_quick_f32(const float *x, float *dst, int k,
                            const sycl::nd_item<3> &item_ct1) {
     const float GELU_QUICK_COEF = -1.702f;
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
@@ -59,7 +59,7 @@ void gelu_quick_f32(const float *x, float *dst, int k,
     dst[i] = x[i] * (1.0f / (1.0f + sycl::native::exp(GELU_QUICK_COEF * x[i])));
 }
 
-void tanh_f32(const float *x, float *dst, int k,
+static void tanh_f32(const float *x, float *dst, int k,
                      const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -69,7 +69,7 @@ void tanh_f32(const float *x, float *dst, int k,
     dst[i] = sycl::tanh((float)(x[i]));
 }
 
-void relu_f32(const float * x, float * dst, const int k,
+static void relu_f32(const float * x, float * dst, const int k,
                      const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -80,7 +80,7 @@ void relu_f32(const float * x, float * dst, const int k,
     dst[i] = sycl::fmax((float)(x[i]), (float)0);
 }
 
-void sigmoid_f32(const float * x, float * dst, const int k,
+static void sigmoid_f32(const float * x, float * dst, const int k,
                             const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -91,7 +91,7 @@ void sigmoid_f32(const float * x, float * dst, const int k,
     dst[i] = 1.0f / (1.0f + sycl::native::exp(-x[i]));
 }
 
-void sqrt_f32(const float * x, float * dst, const int k,
+static void sqrt_f32(const float * x, float * dst, const int k,
                             const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -102,7 +102,7 @@ void sqrt_f32(const float * x, float * dst, const int k,
     dst[i] = sycl::sqrt(x[i]);
 }
 
-void sin_f32(const float * x, float * dst, const int k,
+static void sin_f32(const float * x, float * dst, const int k,
                             const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -113,7 +113,7 @@ void sin_f32(const float * x, float * dst, const int k,
     dst[i] = sycl::sin(x[i]);
 }
 
-void cos_f32(const float * x, float * dst, const int k,
+static void cos_f32(const float * x, float * dst, const int k,
                             const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -124,7 +124,7 @@ void cos_f32(const float * x, float * dst, const int k,
     dst[i] = sycl::cos(x[i]);
 }
 
-void hardsigmoid_f32(const float * x, float * dst, const int k,
+static void hardsigmoid_f32(const float * x, float * dst, const int k,
                             const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -135,7 +135,7 @@ void hardsigmoid_f32(const float * x, float * dst, const int k,
     dst[i] = sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f));
 }
 
-void hardswish_f32(const float * x, float * dst, const int k,
+static void hardswish_f32(const float * x, float * dst, const int k,
                           const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -146,7 +146,7 @@ void hardswish_f32(const float * x, float * dst, const int k,
     dst[i] = x[i] * sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f));
 }
 
-void exp_f32(const float * x, float * dst, const int k,
+static void exp_f32(const float * x, float * dst, const int k,
                           const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -157,7 +157,7 @@ void exp_f32(const float * x, float * dst, const int k,
     dst[i] = sycl::exp(x[i]);
 }
 
-void log_f32(const float * x, float * dst, const int k,
+static void log_f32(const float * x, float * dst, const int k,
                           const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -173,7 +173,7 @@ void log_f32(const float * x, float * dst, const int k,
     }
 }
 
-void neg_f32(const float * x, float * dst, const int k,
+static void neg_f32(const float * x, float * dst, const int k,
                           const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -184,7 +184,7 @@ void neg_f32(const float * x, float * dst, const int k,
     dst[i] = -x[i];
 }
 
-void step_f32(const float * x, float * dst, const int k,
+static void step_f32(const float * x, float * dst, const int k,
                           const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -195,7 +195,7 @@ void step_f32(const float * x, float * dst, const int k,
     dst[i] = x[i] > 0.0f;
 }
 
-void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope,
+static void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope,
                            const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -206,7 +206,7 @@ void leaky_relu_f32(const float *x, float *dst, const int k, const float negativ
              sycl::fmin((float)(x[i]), 0.0f) * negative_slope;
 }
 
-void sqr_f32(const float * x, float * dst, const int k,
+static void sqr_f32(const float * x, float * dst, const int k,
                     const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -217,7 +217,7 @@ void sqr_f32(const float * x, float * dst, const int k,
     dst[i] = x[i] * x[i];
 }
 
-void upscale_f32(const float  *x, float *dst, const int nb00, const int nb01,
+static void upscale_f32(const float  *x, float *dst, const int nb00, const int nb01,
                         const int nb02, const int nb03, const int ne10, const int ne11,
                         const int ne12, const int ne13, const float sf0, const float sf1,
                         const float sf2, const float sf3, const sycl::nd_item<1> &item_ct1) {
@@ -240,7 +240,7 @@ void upscale_f32(const float  *x, float *dst, const int nb00, const int nb01,
     dst[index] = *(const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
 
-void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02,
+static void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02,
                     const sycl::nd_item<3> &item_ct1) {
     int nidx = item_ct1.get_local_id(2) +
                item_ct1.get_group(2) * item_ct1.get_local_range(2);
@@ -262,7 +262,7 @@ void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const i
 
 
 
-void acc_f32_sycl(const float *x, const float *y, float *dst,
+static void acc_f32_sycl(const float *x, const float *y, float *dst,
                          const int n_elements, const int ne10, const int ne11,
                          const int ne12, const int nb1, const int nb2,
                          const int offset, queue_ptr stream) {
@@ -277,7 +277,7 @@ void acc_f32_sycl(const float *x, const float *y, float *dst,
         });
 }
 
-void gelu_f32_sycl(const float *x, float *dst, const int k,
+static void gelu_f32_sycl(const float *x, float *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
     stream->parallel_for(
@@ -289,7 +289,7 @@ void gelu_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void silu_f32_sycl(const float *x, float *dst, const int k,
+static void silu_f32_sycl(const float *x, float *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE;
     stream->parallel_for(
@@ -301,7 +301,7 @@ void silu_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void gelu_quick_f32_sycl(const float *x, float *dst, const int k,
+static void gelu_quick_f32_sycl(const float *x, float *dst, const int k,
                                 queue_ptr stream) {
     const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
     stream->parallel_for(
@@ -313,7 +313,7 @@ void gelu_quick_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void tanh_f32_sycl(const float *x, float *dst, const int k,
+static void tanh_f32_sycl(const float *x, float *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE;
     stream->parallel_for(
@@ -325,7 +325,7 @@ void tanh_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void relu_f32_sycl(const float *x, float *dst, const int k,
+static void relu_f32_sycl(const float *x, float *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
     stream->parallel_for(
@@ -337,7 +337,7 @@ void relu_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void hardsigmoid_f32_sycl(const float *x, float *dst, const int k,
+static void hardsigmoid_f32_sycl(const float *x, float *dst, const int k,
                                  queue_ptr stream) {
     const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE;
     stream->parallel_for(
@@ -349,7 +349,7 @@ void hardsigmoid_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void hardswish_f32_sycl(const float *x, float *dst, const int k,
+static void hardswish_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE;
     stream->parallel_for(
@@ -361,7 +361,7 @@ void hardswish_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void exp_f32_sycl(const float *x, float *dst, const int k,
+static void exp_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
     stream->parallel_for(
@@ -373,7 +373,7 @@ void exp_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void log_f32_sycl(const float *x, float *dst, const int k,
+static void log_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
     stream->parallel_for(
@@ -385,7 +385,7 @@ void log_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void neg_f32_sycl(const float *x, float *dst, const int k,
+static void neg_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
     stream->parallel_for(
@@ -397,7 +397,7 @@ void neg_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void step_f32_sycl(const float *x, float *dst, const int k,
+static void step_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
     stream->parallel_for(
@@ -409,7 +409,7 @@ void step_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void sigmoid_f32_sycl(const float *x, float *dst, const int k,
+static void sigmoid_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE;
     stream->parallel_for(
@@ -421,7 +421,7 @@ void sigmoid_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void sqrt_f32_sycl(const float *x, float *dst, const int k,
+static void sqrt_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE;
     stream->parallel_for(
@@ -433,7 +433,7 @@ void sqrt_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void sin_f32_sycl(const float *x, float *dst, const int k,
+static void sin_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
     stream->parallel_for(
@@ -445,7 +445,7 @@ void sin_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void cos_f32_sycl(const float *x, float *dst, const int k,
+static void cos_f32_sycl(const float *x, float *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
     stream->parallel_for(
@@ -457,7 +457,7 @@ void cos_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void leaky_relu_f32_sycl(const float *x, float *dst, const int k,
+static void leaky_relu_f32_sycl(const float *x, float *dst, const int k,
                                 const float negative_slope,
                                 queue_ptr stream) {
     const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
@@ -470,7 +470,7 @@ void leaky_relu_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void sqr_f32_sycl(const float *x, float *dst, const int k,
+static void sqr_f32_sycl(const float *x, float *dst, const int k,
                          queue_ptr stream) {
     const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE;
     stream->parallel_for(
@@ -482,7 +482,7 @@ void sqr_f32_sycl(const float *x, float *dst, const int k,
         });
 }
 
-void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01,
+static void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01,
                              const int nb02, const int nb03, const int ne10, const int ne11,
                              const int ne12, const int ne13, const float sf0, const float sf1,
                              const float sf2, const float sf3, queue_ptr stream) {
@@ -496,7 +496,7 @@ void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01
         });
 }
 
-void pad_f32_sycl(const float *x, float *dst, const int ne00,
+static void pad_f32_sycl(const float *x, float *dst, const int ne00,
                          const int ne01, const int ne02, const int ne0,
                          const int ne1, const int ne2, queue_ptr stream) {
     int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;

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

@@ -13,9 +13,6 @@
 #ifndef GGML_SYCL_GEMM_HPP
 #define GGML_SYCL_GEMM_HPP
 
-#include <fstream>
-#include <iostream>
-
 #include "ggml-sycl.h"
 
 #if GGML_SYCL_DNNL
@@ -35,62 +32,34 @@ public:
         else static_assert(0);
     }
 
-    static inline void row_gemm(sycl::queue& q, bool a_trans,
-        bool b_trans, int m, int n, int k,
-        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
-    {
-        // Get the device associated with the queue
-        sycl::device dev = q.get_device();
-        // Get the context associated with the queue
-        sycl::context ctx = q.get_context();
-        const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx);
-        const dnnl::stream stream = dnnl::sycl_interop::make_stream(eng, q);
+    static inline void row_gemm(ggml_backend_sycl_context & ctx, bool a_trans, bool b_trans, int m, int n, int k,
+                                const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+        auto stream = ctx.stream_dnnl(q);
+        auto eng = ctx.engine_dnnl(q);
         dnnl::memory::dims a_dims = { m, k };
         dnnl::memory::dims b_dims = { k, n };
         dnnl::memory::dims c_dims = { m, n };
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
         const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
-        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
-        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
-        auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::ab);
 
-        // Create the primitive.
-        auto matmul_prim = dnnl::matmul(matmul_pd);
-        // Primitive arguments.
-        std::unordered_map<int, dnnl::memory> matmul_args;
-        matmul_args.insert({ DNNL_ARG_SRC, a_mem });
-        matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
-        matmul_args.insert({ DNNL_ARG_DST, c_mem });
+        dnnl::primitive_attr primitive_attr;
+        primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
 
-        matmul_prim.execute(stream, matmul_args);
-    }
-
-
-    static inline void row_gemm(const dnnl::stream& stream, bool a_trans,
-        bool b_trans, int m, int n, int k,
-        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
-    {
-        auto const eng = stream.get_engine();
-        dnnl::memory::dims a_dims = { m, k };
-        dnnl::memory::dims b_dims = { k, n };
-        dnnl::memory::dims c_dims = { m, n };
-        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
         auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
         auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
-        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
+        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md, primitive_attr);
         auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
 
-        // Create the primitive.
+        auto scratchpad_md = matmul_pd.scratchpad_desc();
+        auto scratchpad_mem = ctx.get_scratchpad_mem(scratchpad_md, eng, q);
         auto matmul_prim = dnnl::matmul(matmul_pd);
-        // Primitive arguments.
+
         std::unordered_map<int, dnnl::memory> matmul_args;
         matmul_args.insert({ DNNL_ARG_SRC, a_mem });
         matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
         matmul_args.insert({ DNNL_ARG_DST, c_mem });
+        matmul_args.insert({ DNNL_ARG_SCRATCHPAD, scratchpad_mem });
 
         matmul_prim.execute(stream, matmul_args);
     }

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

@@ -207,7 +207,7 @@ static void get_rows_sycl_reorder(ggml_backend_sycl_context & ctx, const ggml_te
     const size_t nrows = ne01;
     const sycl::half* src0_dq = (const sycl::half*)(src0_q + nrows * ncols / 2);
     stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                         [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]]{
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              k_get_rows_reorder<qk, qr, dq_reorder>(
                                  src0_dd, src0_dq, src1_dd, dst_dd, ne00, ne12, s1, s2,
                                  s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
@@ -302,7 +302,6 @@ void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *s
             // TODO: k-quants
             GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
             GGML_ABORT("fatal error");
-            break;
     }
 }