llama_cpp 0.10.0 → 0.10.1

data/ext/llama_cpp/src/ggml-cuda.cu

@@ -1,13 +1,15 @@
 #include <algorithm>
+#include <assert.h>
+#include <atomic>
+#include <cinttypes>
 #include <cstddef>
 #include <cstdint>
-#include <cinttypes>
 #include <float.h>
 #include <limits>
 #include <stdint.h>
 #include <stdio.h>
-#include <atomic>
-#include <assert.h>
+#include <vector>
+
 
 #if defined(GGML_USE_HIPBLAS)
 #include <hip/hip_runtime.h>
@@ -437,6 +439,7 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 
 #define CUDA_GELU_BLOCK_SIZE 256
 #define CUDA_SILU_BLOCK_SIZE 256
+#define CUDA_TANH_BLOCK_SIZE 256
 #define CUDA_RELU_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
 #define CUDA_CPY_BLOCK_SIZE 32
@@ -449,6 +452,11 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define CUDA_QUANTIZE_BLOCK_SIZE 256
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
+#define CUDA_UPSCALE_BLOCK_SIZE 256
+#define CUDA_CONCAT_BLOCK_SIZE 256
+#define CUDA_PAD_BLOCK_SIZE 256
+#define CUDA_ACC_BLOCK_SIZE 256
+#define CUDA_IM2COL_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_CUDA_DMMV_X
@@ -610,6 +618,24 @@ static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * s
     dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
 }
 
+static __global__ 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 int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= ne) {
+        return;
+    }
+    int src1_idx = i - offset;
+    int oz = src1_idx / nb2;
+    int oy = (src1_idx - (oz * nb2)) / nb1;
+    int ox = src1_idx % nb1;
+    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
+        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
+    } else {
+        dst[i] = x[i];
+    }
+}
+
 static __global__ void gelu_f32(const float * x, float * dst, const int k) {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
@@ -632,6 +658,23 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] / (1.0f + expf(-x[i]));
 }
 
+static __global__ void gelu_quick_f32(const float *x, float *dst, int k) {
+    const float GELU_QUICK_COEF = -1.702f;
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
+}
+
+static __global__ void tanh_f32(const float *x, float *dst, int k) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = tanhf(x[i]);
+}
+
 static __global__ void relu_f32(const float * x, float * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -641,6 +684,14 @@ static __global__ void relu_f32(const float * x, float * dst, const int k) {
     dst[i] = fmaxf(x[i], 0);
 }
 
+static __global__ void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
+}
+
 static __global__ void sqr_f32(const float * x, float * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -686,6 +737,132 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols, c
     }
 }
 
+static __global__ void concat_f32(const float  *x,const float  *y, float *dst, const int ne0, const int ne02) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (blockIdx.z < ne02) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            blockIdx.z * ne0 * gridDim.y;
+            dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            (blockIdx.z - ne02) * ne0 *  gridDim.y;
+            dst[offset_dst] = y[offset_src];
+    }
+}
+
+static __global__ void upscale_f32(const float  *x, float *dst, const int ne00, const int nb02, const int scale_factor) {
+    int ne0 = ne00 * scale_factor;
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    // operation
+    int i00 = nidx / scale_factor;
+    int i01 = blockIdx.y / scale_factor;
+    int offset_src =
+        i00 +
+        i01 * ne00 +
+        blockIdx.z * nb02;
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    dst[offset_dst] = x[offset_src];
+}
+
+static __global__ void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+            dst[offset_dst] = x[offset_src];
+    } else {
+        dst[offset_dst] = 0.0f;
+    }
+}
+
+template <int block_size>
+static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
+    int start = blockIdx.x * group_size;
+    int end = start + group_size;
+
+    start += threadIdx.x;
+
+    if (end >= ne_elements) {
+        end = ne_elements;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += block_size) {
+        tmp += x[j];
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += block_size) {
+        float xi = x[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float variance = tmp / group_size;
+    float scale = rsqrtf(variance + eps);
+    for (int j = start; j < end; j += block_size) {
+        dst[j] *= scale;
+    }
+}
+
 template <int block_size>
 static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
@@ -1684,31 +1861,65 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
 }
 
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void k_get_rows(const void * x, const int32_t * y, dst_t * dst, const int ncols) {
-    const int col = (blockIdx.x*blockDim.x + threadIdx.x)*2;
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-
-    if (col >= ncols) {
+static __global__ void k_get_rows(
+            const void * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
         return;
     }
 
-    const int r = y[row];
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
 
-    // copy x[r*ncols + col] to dst[row*ncols + col]
-    const int xi = r*ncols + col;
-    const int di = row*ncols + col;
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
 
-    const int ib = xi/qk; // block index
-    const int iqs = (xi%qk)/qr; // quant index
-    const int iybs = di - di%qk; // y block start index
+    const int ib = i00/qk; // block index
+    const int iqs = (i00%qk)/qr; // quant index
+    const int iybs = i00 - i00%qk; // dst block start index
     const int y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
     dfloat2 v;
-    dequantize_kernel(x, ib, iqs, v);
+    dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst[iybs + iqs + 0]        = v.x;
-    dst[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = v.x;
+    dst_row[iybs + iqs + y_offset] = v.y;
+}
+
+template<typename src0_t, typename dst_t>
+static __global__ void k_get_rows_float(
+            const src0_t * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+    dst_row[i00] = src0_row[i00];
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -5035,29 +5246,98 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min,
 
 static  __global__ void im2col_f32_f16(
         const float * x, half * dst,
-        int ofs0, int ofs1, int IW, int IH, int CHW,
+        int offset_delta, int IW, int IH, int OW, int KW, int KH, int pelements, int CHW,
         int s0, int s1, int p0, int p1, int d0, int d1) {
-    const int iiw = blockIdx.z * s0 + threadIdx.z * d0 - p0;
-    const int iih = blockIdx.y * s1 + threadIdx.y * d1 - p1;
+    const int i = threadIdx.x + blockIdx.x * blockDim.x;
+    if (i >= pelements) {
+        return;
+    }
+
+    const int ksize = OW * (KH > 1 ? KW : 1);
+    const int kx = i / ksize;
+    const int kd = kx * ksize;
+    const int ky = (i - kd) / OW;
+    const int ix = i % OW;
+
+    const int iiw = ix * s0 + kx * d0 - p0;
+    const int iih = blockIdx.y * s1 + ky * d1 - p1;
 
     const int offset_dst =
-        (threadIdx.x * gridDim.y * gridDim.z + blockIdx.y * gridDim.z + blockIdx.z) * CHW +
-        (blockIdx.x * (blockDim.y * blockDim.z) + threadIdx.y * blockDim.z + threadIdx.z);
+        (blockIdx.y * OW + ix) * CHW +
+        (blockIdx.z * (KW * KH) + ky * KW + kx);
 
     if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
         dst[offset_dst] = __float2half(0.0f);
     } else {
-        const int offset_src =  threadIdx.x * ofs0 + blockIdx.x * ofs1;
+        const int offset_src = blockIdx.z * offset_delta;
         dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]);
     }
 }
 
 template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) {
+static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ncols + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
-    const dim3 block_nums(block_num_x, nrows, 1);
-    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols);
+    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    (void) dst;
+}
+
+template<typename src0_t>
+static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    (void) dst;
 }
 
 template<float (*bin_op)(const float, const float)>
@@ -5069,7 +5349,6 @@ struct bin_bcast_cuda {
 
         GGML_TENSOR_BINARY_OP_LOCALS
 
-
         int nr0 = ne10/ne0;
         int nr1 = ne11/ne1;
         int nr2 = ne12/ne2;
@@ -5117,26 +5396,28 @@ struct bin_bcast_cuda {
             int64_t ne12 = cne1[2];
             int64_t ne13 = cne1[3];
 
-            //size_t nb0 = cnb0[0];
+            size_t nb0 = cnb0[0];
             size_t nb1 = cnb0[1];
             size_t nb2 = cnb0[2];
             size_t nb3 = cnb0[3];
 
-            //size_t nb10 = cnb1[0];
+            size_t nb10 = cnb1[0];
             size_t nb11 = cnb1[1];
             size_t nb12 = cnb1[2];
             size_t nb13 = cnb1[3];
 
-            //size_t s0 = nb0 / sizeof(src1_t);
-            size_t s1 = nb1 / sizeof(src1_t);
-            size_t s2 = nb2 / sizeof(src1_t);
-            size_t s3 = nb3 / sizeof(src1_t);
+            size_t s0 = nb0 / sizeof(dst_t);
+            size_t s1 = nb1 / sizeof(dst_t);
+            size_t s2 = nb2 / sizeof(dst_t);
+            size_t s3 = nb3 / sizeof(dst_t);
 
-            //size_t s10 = nb10 / sizeof(src1_t);
+            size_t s10 = nb10 / sizeof(src1_t);
             size_t s11 = nb11 / sizeof(src1_t);
             size_t s12 = nb12 / sizeof(src1_t);
             size_t s13 = nb13 / sizeof(src1_t);
 
+            GGML_ASSERT(s0 == 1);
+            GGML_ASSERT(s10 == 1);
 
             const int block_size = 128;
 
@@ -5174,6 +5455,13 @@ struct bin_bcast_cuda {
     }
 };
 
+static void acc_f32_cuda(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, cudaStream_t stream) {
+    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+}
+
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
     gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -5184,11 +5472,26 @@ static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
     silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
+    gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
+    tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
     relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
+}
+
 static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
     sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -5205,6 +5508,38 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
     }
 }
 
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
+    static const float eps = 1e-6f;
+    if (group_size < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    }
+}
+
+static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+}
+
+static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int scale_factor, cudaStream_t stream) {
+    int ne0 = (ne00 * scale_factor);
+    int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02);
+    upscale_f32<<<gridDim, CUDA_UPSCALE_BLOCK_SIZE, 0, stream>>>(x, dst, ne00, ne00 * ne01, scale_factor);
+}
+
+static void pad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02,
+    const int ne0, const int ne1, const int ne2, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02);
+}
+
 static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
@@ -6167,13 +6502,14 @@ static void soft_max_f32_cuda(const float * x, const float * y, float * dst, con
     soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
 }
 
-static void im2col_f32_f16_cuda(const float * x, half * dst,
-    int OH, int IW, int IH, int OW, int IC,
-    int KH, int KW, int N,  int ofs0, int ofs1,
-    int s0, int s1, int p0, int p1, int d0, int d1, cudaStream_t stream) {
-    dim3 block_nums(IC, OH, OW);
-    dim3 block_dims(N,  KH, KW);
-    im2col_f32_f16<<<block_nums, block_dims, 0, stream>>>(x, dst, ofs0, ofs1, IW, IH, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+static void im2col_f32_f16_cuda(const float* x, half* dst,
+    int IW, int IH, int OW, int OH, int KW, int KH, int IC,
+    int offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+    const int parallel_elements = OW * KW * KH;
+    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    dim3 block_nums(num_blocks, OH, IC);
+    im2col_f32_f16<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, offset_delta, IW, IH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
 }
 
 // buffer pool for cuda
@@ -6447,36 +6783,34 @@ static void ggml_cuda_op_get_rows(
 
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(src1));
-    GGML_ASSERT(ggml_is_contiguous(dst));
 
-    const int ncols = src0->ne[0];
-    const int nrows = ggml_nelements(src1);
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
 
     const int32_t * src1_i32 = (const int32_t *) src1_d;
 
     switch (src0->type) {
         case GGML_TYPE_F16:
-            get_rows_cuda<1, 1, convert_f16>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_F32:
-            get_rows_cuda<1, 1, convert_f32>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         default:
             // TODO: k-quants
@@ -6522,6 +6856,25 @@ inline void ggml_cuda_op_add(
     ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
 }
 
+inline void ggml_cuda_op_acc(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
+
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
+
+    acc_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
+
+    (void) dst;
+}
+
 inline void ggml_cuda_op_mul(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6564,6 +6917,34 @@ inline void ggml_cuda_op_silu(
     (void) src1_dd;
 }
 
+inline void ggml_cuda_op_gelu_quick(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_quick_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_cuda_op_tanh(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    tanh_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 inline void ggml_cuda_op_relu(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6578,6 +6959,23 @@ inline void ggml_cuda_op_relu(
     (void) src1_dd;
 }
 
+inline void ggml_cuda_op_leaky_relu(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    leaky_relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 inline void ggml_cuda_op_sqr(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6612,6 +7010,71 @@ inline void ggml_cuda_op_norm(
     (void) src1_dd;
 }
 
+
+inline void ggml_cuda_op_group_norm(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int num_groups = dst->op_params[0];
+    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    group_norm_f32_cuda(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_cuda_op_concat(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+        concat_f32_cuda(src0_dd + i3 * (src0->nb[3] / 4), src1_dd + i3 * (src1->nb[3] / 4), dst_dd + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], main_stream);
+    }
+
+    (void) src1;
+    (void) dst;
+}
+
+inline void ggml_cuda_op_upscale(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    const int scale_factor = dst->op_params[0];
+
+    upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
+
+    (void) src1;
+    (void) dst;
+}
+
+inline void ggml_cuda_op_pad(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    pad_f32_cuda(src0_dd, dst_dd,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
+
+    (void) src1;
+    (void) dst;
+}
+
 inline void ggml_cuda_op_rms_norm(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -7126,7 +7589,6 @@ inline void ggml_cuda_op_im2col(
 
     const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
 
-    const int64_t N  = src1->ne[is_2D ? 3 : 2];
     const int64_t IC = src1->ne[is_2D ? 2 : 1];
     const int64_t IH = is_2D ? src1->ne[1] : 1;
     const int64_t IW =         src1->ne[0];
@@ -7137,17 +7599,15 @@ inline void ggml_cuda_op_im2col(
     const int64_t OH = is_2D ? dst->ne[2] : 1;
     const int64_t OW =         dst->ne[1];
 
-    const size_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
-    const size_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
 
-    im2col_f32_f16_cuda(src1_dd, (half*) dst_dd,
-        OH, IW, IH, OW, IC, KH, KW, N,
-        ofs0, ofs1, s0, s1, p0, p1, d0, d1, main_stream);
+    im2col_f32_f16_cuda(src1_dd, (half*) dst_dd, IW, IH, OW, OH, KW, KH, IC, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
 
     (void) src0;
     (void) src0_dd;
 }
 
+
 inline void ggml_cuda_op_sum_rows(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -7696,6 +8156,10 @@ static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, gg
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add);
 }
 
+static void ggml_cuda_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_acc);
+}
+
 static void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul);
 }
@@ -7712,10 +8176,22 @@ static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, g
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu);
 }
 
+static void ggml_cuda_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu_quick);
+}
+
+static void ggml_cuda_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_tanh);
+}
+
 static void ggml_cuda_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_relu);
 }
 
+static void ggml_cuda_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_leaky_relu);
+}
+
 static void ggml_cuda_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sqr);
 }
@@ -7724,6 +8200,22 @@ static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, g
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm);
 }
 
+static void ggml_cuda_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_group_norm);
+}
+
+static void ggml_cuda_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_concat);
+}
+
+static void ggml_cuda_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_upscale);
+}
+
+static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad);
+}
+
 static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm);
 }
@@ -8234,36 +8726,69 @@ static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) {
 }
 #endif
 
-static void ggml_cuda_mul_mat_id(const ggml_tensor * _src0, const ggml_tensor * _src1, ggml_tensor * dst) {
+static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #if 0
-//#ifdef CUDA_USE_TENSOR_CORES
-//    const bool use_tensor_cores = true;
-//#else
-//    const bool use_tensor_cores = false;
-//#endif
-
     ggml_cuda_mul_mat_id_cublas(dst);
-
     // TODO: mmq/mmv support
-#else
-    const struct ggml_tensor * ids = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-    const int id = dst->op_params[0];
+#endif
 
-    int32_t * ids_dev = (int32_t *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device];
+    GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
 
-    int32_t a_id;
-    CUDA_CHECK(cudaMemcpyAsync(&a_id, ids_dev + id, sizeof(int32_t), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
-    CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
+    const struct ggml_tensor * ids = src0;
+    const int32_t id = ((int32_t *) dst->op_params)[0];
+    const int32_t n_as = ((int32_t *) dst->op_params)[1];
 
-    GGML_ASSERT(a_id >= 0 && a_id < ids->ne[0]);
-    const struct ggml_tensor * src0 = dst->src[a_id + 2];
+    std::vector<char> ids_host(ggml_nbytes(ids));
 
-    ggml_cuda_mul_mat(src0, src1, dst);
-#endif
+    if (ids->backend == GGML_BACKEND_GPU) {
+        const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device];
+        CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
+        CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
+    } else {
+        memcpy(ids_host.data(), ids->data, ggml_nbytes(ids));
+    }
+
+    const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra;
+    const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra;
+
+    ggml_tensor_extra_gpu src1_row_extra;
+    ggml_tensor_extra_gpu dst_row_extra;
+
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    src1_row.ne[1] = 1;
+    dst_row.ne[1] = 1;
+
+    src1_row.nb[2] = src1_row.nb[1];
+    dst_row.nb[2] = dst_row.nb[1];
+
+    src1_row.nb[3] = src1_row.nb[1];
+    dst_row.nb[3] = dst_row.nb[1];
+
+    src1_row.extra = &src1_row_extra;
+    dst_row.extra = &dst_row_extra;
+
+
+    for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+        //int32_t row_id;
+        //CUDA_CHECK(cudaMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
+        //CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
+
+        const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
+
+        GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+        const struct ggml_tensor * src0_row = dst->src[row_id + 2];
+
+        src1_row_extra.data_device[g_main_device] = (char *) src1_extra->data_device[g_main_device] + i01*src1->nb[1];
+        src1_row.data = (char *) src1->data + i01*src1->nb[1];
 
-    (void) _src0;
-    (void) _src1;
+        dst_row_extra.data_device[g_main_device] = (char *) dst_extra->data_device[g_main_device] + i01*dst->nb[1];
+        dst_row.data = (char *) dst->data + i01*dst->nb[1];
+
+        ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
+    }
 }
 
 static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -8373,6 +8898,12 @@ static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, gg
     (void) dst;
 }
 
+static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
+}
+
 void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
     const int64_t nrows = ggml_nrows(tensor);
 
@@ -8422,8 +8953,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
 
         // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
         if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
-                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
 
         char * buf;
@@ -8683,6 +9213,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_ADD:
             func = ggml_cuda_add;
             break;
+        case GGML_OP_ACC:
+            func = ggml_cuda_acc;
+            break;
         case GGML_OP_MUL:
             func = ggml_cuda_mul;
             break;
@@ -8697,6 +9230,12 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
                 case GGML_UNARY_OP_SILU:
                     func = ggml_cuda_silu;
                     break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    func = ggml_cuda_gelu_quick;
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    func = ggml_cuda_tanh;
+                    break;
                 case GGML_UNARY_OP_RELU:
                     func = ggml_cuda_relu;
                     break;
@@ -8707,6 +9246,21 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_NORM:
             func = ggml_cuda_norm;
             break;
+        case GGML_OP_GROUP_NORM:
+            func = ggml_cuda_group_norm;
+            break;
+        case GGML_OP_CONCAT:
+            func = ggml_cuda_concat;
+            break;
+        case GGML_OP_UPSCALE:
+            func = ggml_cuda_upscale;
+            break;
+        case GGML_OP_PAD:
+            func = ggml_cuda_pad;
+            break;
+        case GGML_OP_LEAKY_RELU:
+            func = ggml_cuda_leaky_relu;
+            break;
         case GGML_OP_RMS_NORM:
             func = ggml_cuda_rms_norm;
             break;
@@ -8729,9 +9283,6 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             func = ggml_cuda_sqr;
             break;
         case GGML_OP_CLAMP:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_clamp;
             break;
         case GGML_OP_CPY:
@@ -8740,6 +9291,7 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_CONT:
             func = ggml_cuda_dup;
             break;
+        case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
@@ -8938,8 +9490,7 @@ static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_t
 
     if (ggml_is_quantized(tensor->type)) {
         if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
-                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
     }
 
@@ -9159,6 +9710,8 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
                     return true;
                 default:
                     return false;
@@ -9181,6 +9734,45 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
                 }
                 return true;
             } break;
+        case GGML_OP_GET_ROWS:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                        return true;
+                    default:
+                        return false;
+                }
+            } break;
+        case GGML_OP_CPY:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                return false;
+            } break;
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -9188,7 +9780,6 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
         case GGML_OP_REPEAT:
-        case GGML_OP_GET_ROWS:
         case GGML_OP_DUP:
         case GGML_OP_ADD:
         case GGML_OP_MUL:
@@ -9197,7 +9788,6 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
-        case GGML_OP_CPY:
         case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
@@ -9206,6 +9796,12 @@ static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_ten
         case GGML_OP_IM2COL:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_CONCAT:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_LEAKY_RELU:
             return true;
         default:
             return false;
@@ -9264,7 +9860,9 @@ static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * use
     UNUSED(params);
 }
 
-extern "C" int ggml_backend_cuda_reg_devices() {
+extern "C" int ggml_backend_cuda_reg_devices();
+
+int ggml_backend_cuda_reg_devices() {
     int device_count = ggml_cuda_get_device_count();
     //int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
     for (int i = 0; i < device_count; i++) {