@novastera-oss/llamarn 0.4.0 → 0.4.3-beta4

package/cpp/llama.cpp/ggml/src/ggml-cuda/mmvq.cuh

@@ -3,7 +3,7 @@
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
 
 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 
 void ggml_cuda_op_mul_mat_vec_q(
     ggml_backend_cuda_context & ctx,

package/cpp/llama.cpp/ggml/src/ggml-cuda/norm.cu

@@ -105,29 +105,29 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
 }
 
 template <int block_size, bool do_multiply = false, bool do_add = false>
-static __global__ void rms_norm_f32(const float * x, float *       dst,
+static __global__ 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 float * mul                = nullptr,
-                                    const int64_t mul_stride_row     = 0,
-                                    const int64_t mul_stride_channel = 0,
-                                    const int64_t mul_stride_sample  = 0,
-                                    const int     mul_ncols          = 0,
-                                    const int     mul_nrows          = 0,
-                                    const int     mul_nchannels      = 0,
-                                    const int     mul_nsamples       = 0,
-                                    const float * add                = nullptr,
-                                    const int64_t add_stride_row     = 0,
-                                    const int64_t add_stride_channel = 0,
-                                    const int64_t add_stride_sample  = 0,
-                                    const int     add_ncols          = 0,
-                                    const int     add_nrows          = 0,
-                                    const int     add_nchannels      = 0,
-                                    const int     add_nsamples       = 0) {
-
+                                    const float * mul                  = nullptr,
+                                    const int64_t mul_stride_row       = 0,
+                                    const int64_t mul_stride_channel   = 0,
+                                    const int64_t mul_stride_sample    = 0,
+                                    const uint3   mul_ncols_packed     = make_uint3(0, 0, 0),
+                                    const uint3   mul_nrows_packed     = make_uint3(0, 0, 0),
+                                    const uint3   mul_nchannels_packed = make_uint3(0, 0, 0),
+                                    const uint3   mul_nsamples_packed  = make_uint3(0, 0, 0),
+                                    const float * add                  = nullptr,
+                                    const int64_t add_stride_row       = 0,
+                                    const int64_t add_stride_channel   = 0,
+                                    const int64_t add_stride_sample    = 0,
+                                    const uint3   add_ncols_packed     = make_uint3(0, 0, 0),
+                                    const uint3   add_nrows_packed     = make_uint3(0, 0, 0),
+                                    const uint3   add_nchannels_packed = make_uint3(0, 0, 0),
+                                    const uint3   add_nsamples_packed  = make_uint3(0, 0, 0)) {
     const int nrows     = gridDim.x;
     const int nchannels = gridDim.y;
 
@@ -142,16 +142,16 @@ static __global__ void rms_norm_f32(const float * x, float *       dst,
     dst += ((sample*nchannels + channel)*nrows + row)*ncols;
 
     if constexpr (do_multiply) {
-        const int mul_row = row % mul_nrows;
-        const int mul_channel = channel % mul_nchannels;
-        const int mul_sample = sample % mul_nsamples;
-        mul += mul_sample*mul_stride_sample + mul_channel*mul_stride_channel + mul_row*mul_stride_row;
+        const uint32_t mul_row     = fastmodulo(row, mul_nrows_packed);
+        const uint32_t mul_channel = fastmodulo(channel, mul_nchannels_packed);
+        const uint32_t mul_sample  = fastmodulo(sample, mul_nsamples_packed);
+        mul += mul_sample * mul_stride_sample + mul_channel * mul_stride_channel + mul_row * mul_stride_row;
     }
 
     if constexpr (do_add) {
-        const int add_row     = row % add_nrows;
-        const int add_channel = channel % add_nchannels;
-        const int add_sample  = sample % add_nsamples;
+        const int add_row     = fastmodulo(row, add_nrows_packed);
+        const int add_channel = fastmodulo(channel, add_nchannels_packed);
+        const int add_sample  = fastmodulo(sample, add_nsamples_packed);
         add += add_sample * add_stride_sample + add_channel * add_stride_channel + add_row * add_stride_row;
     }
 
@@ -165,15 +165,18 @@ static __global__ void rms_norm_f32(const float * x, float *       dst,
     // sum up partial sums
     tmp = warp_reduce_sum(tmp);
     if constexpr (block_size > WARP_SIZE) {
-        static_assert(block_size == 1024, "unexpected block_size");
+        static_assert((block_size <= 1024) && (block_size % 32 == 0), "unexpected block_size");
         __shared__ float s_sum[32];
-        const int warp_id = threadIdx.x / WARP_SIZE;
-        const int lane_id = threadIdx.x % WARP_SIZE;
+        const int        warp_id = tid / WARP_SIZE;
+        const int        lane_id = tid % WARP_SIZE;
         if (lane_id == 0) {
             s_sum[warp_id] = tmp;
         }
         __syncthreads();
-        tmp = s_sum[lane_id];
+        tmp = 0.0f;
+        if (lane_id < (block_size / WARP_SIZE)) {
+            tmp = s_sum[lane_id];
+        }
         tmp = warp_reduce_sum(tmp);
     }
 
@@ -182,12 +185,12 @@ static __global__ void rms_norm_f32(const float * x, float *       dst,
 
     for (int col = tid; col < ncols; col += block_size) {
         if constexpr (do_multiply && do_add) {
-            const int mul_col = col % mul_ncols;
-            const int add_col = col % add_ncols;
-            dst[col] = scale * x[col] * mul[mul_col] + add[add_col];
+            const int mul_col = fastmodulo(col, mul_ncols_packed);
+            const int add_col = fastmodulo(col, add_ncols_packed);
+            dst[col]          = scale * x[col] * mul[mul_col] + add[add_col];
         } else if constexpr (do_multiply) {
-            const int mul_col = col % mul_ncols;
-            dst[col] = scale * x[col] * mul[mul_col];
+            const int mul_col = fastmodulo(col, mul_ncols_packed);
+            dst[col]          = scale * x[col] * mul[mul_col];
         } else {
             dst[col] = scale * x[col];
         }
@@ -354,77 +357,86 @@ static void rms_norm_f32_cuda(
         const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) {
     const dim3 blocks_num(nrows, nchannels, nsamples);
     if (ncols < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        rms_norm_f32<WARP_SIZE, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        const dim3 block_dims(256, 1, 1);
+        rms_norm_f32<256, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
         rms_norm_f32<1024, false><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
     }
 }
 
-static void rms_norm_mul_f32_cuda(const float * x,
-                                  const float * mul,
-                                  const float * add,
-                                  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 int64_t mul_stride_row,
-                                  const int64_t mul_stride_channel,
-                                  const int64_t mul_stride_sample,
-                                  const int     mul_ncols,
-                                  const int     mul_nrows,
-                                  const int     mul_nchannels,
-                                  const int     mul_nsamples,
-                                  const int64_t add_stride_row,
-                                  const int64_t add_stride_channel,
-                                  const int64_t add_stride_sample,
-                                  const int     add_ncols,
-                                  const int     add_nrows,
-                                  const int     add_nchannels,
-                                  const int     add_nsamples,
-                                  const float   eps,
-                                  cudaStream_t  stream) {
+static void rms_norm_mul_f32_cuda(const float *  x,
+                                  const float *  mul,
+                                  const float *  add,
+                                  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 int64_t  mul_stride_row,
+                                  const int64_t  mul_stride_channel,
+                                  const int64_t  mul_stride_sample,
+                                  const uint32_t mul_ncols,
+                                  const uint32_t mul_nrows,
+                                  const uint32_t mul_nchannels,
+                                  const uint32_t mul_nsamples,
+                                  const int64_t  add_stride_row,
+                                  const int64_t  add_stride_channel,
+                                  const int64_t  add_stride_sample,
+                                  const uint32_t add_ncols,
+                                  const uint32_t add_nrows,
+                                  const uint32_t add_nchannels,
+                                  const uint32_t add_nsamples,
+                                  const float    eps,
+                                  cudaStream_t   stream) {
     const dim3 blocks_num(nrows, nchannels, nsamples);
     if (mul == nullptr) {
         rms_norm_f32_cuda(x, dst, ncols, nrows, nchannels, nsamples, stride_row, stride_channel, stride_sample, eps, stream);
         return;
     }
     if (add == nullptr) {
+        const uint3 mul_ncols_packed     = init_fastdiv_values(mul_ncols);
+        const uint3 mul_nrows_packed     = init_fastdiv_values(mul_nrows);
+        const uint3 mul_nchannels_packed = init_fastdiv_values(mul_nchannels);
+        const uint3 mul_nsamples_packed  = init_fastdiv_values(mul_nsamples);
         if (ncols < 1024) {
-            const dim3 block_dims(WARP_SIZE, 1, 1);
-            rms_norm_f32<WARP_SIZE, true><<<blocks_num, block_dims, 0, stream>>>(x, dst,
-                ncols, stride_row, stride_channel, stride_sample, eps,
-                mul, mul_stride_row, mul_stride_channel, mul_stride_sample,
-                mul_ncols, mul_nrows, mul_nchannels, mul_nsamples);
+            const dim3 block_dims(256, 1, 1);
+            rms_norm_f32<256, true><<<blocks_num, block_dims, 0, stream>>>(
+                x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
+                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed);
         } else {
             const dim3 block_dims(1024, 1, 1);
-            rms_norm_f32<1024, true><<<blocks_num, block_dims, 0, stream>>>(x, dst,
-                ncols, stride_row, stride_channel, stride_sample, eps,
-                mul, mul_stride_row, mul_stride_channel, mul_stride_sample,
-                mul_ncols, mul_nrows, mul_nchannels, mul_nsamples);
+            rms_norm_f32<1024, true><<<blocks_num, block_dims, 0, stream>>>(
+                x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
+                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed);
         }
     } else {
+        const uint3 mul_ncols_packed     = init_fastdiv_values(mul_ncols);
+        const uint3 mul_nrows_packed     = init_fastdiv_values(mul_nrows);
+        const uint3 mul_nchannels_packed = init_fastdiv_values(mul_nchannels);
+        const uint3 mul_nsamples_packed  = init_fastdiv_values(mul_nsamples);
+
+        const uint3 add_ncols_packed     = init_fastdiv_values(add_ncols);
+        const uint3 add_nrows_packed     = init_fastdiv_values(add_nrows);
+        const uint3 add_nchannels_packed = init_fastdiv_values(add_nchannels);
+        const uint3 add_nsamples_packed  = init_fastdiv_values(add_nsamples);
         if (ncols < 1024) {
-            const dim3 block_dims(WARP_SIZE, 1, 1);
-            rms_norm_f32<WARP_SIZE, true, true><<<blocks_num, block_dims, 0, stream>>>(x, dst,
-                ncols, stride_row, stride_channel, stride_sample, eps,
-                mul, mul_stride_row, mul_stride_channel, mul_stride_sample,
-                mul_ncols, mul_nrows, mul_nchannels, mul_nsamples,
-                add, add_stride_row, add_stride_channel, add_stride_sample,
-                add_ncols, add_nrows, add_nchannels, add_nsamples);
+            const dim3 block_dims(256, 1, 1);
+            rms_norm_f32<256, true, true><<<blocks_num, block_dims, 0, stream>>>(
+                x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
+                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed, add,
+                add_stride_row, add_stride_channel, add_stride_sample, add_ncols_packed, add_nrows_packed,
+                add_nchannels_packed, add_nsamples_packed);
         } else {
             const dim3 block_dims(1024, 1, 1);
-            rms_norm_f32<1024, true, true><<<blocks_num, block_dims, 0, stream>>>(x, dst,
-                ncols, stride_row, stride_channel, stride_sample, eps,
-                mul, mul_stride_row, mul_stride_channel, mul_stride_sample,
-                mul_ncols, mul_nrows, mul_nchannels, mul_nsamples,
-                add, add_stride_row, add_stride_channel, add_stride_sample,
-                add_ncols, add_nrows, add_nchannels, add_nsamples);
+            rms_norm_f32<1024, true, true><<<blocks_num, block_dims, 0, stream>>>(
+                x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
+                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed, add,
+                add_stride_row, add_stride_channel, add_stride_sample, add_ncols_packed, add_nrows_packed,
+                add_nchannels_packed, add_nsamples_packed);
         }
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-cuda/pad.cu

@@ -1,36 +1,50 @@
 #include "pad.cuh"
 
-static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
-    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
-    // blockIdx.y: idx of ne1
-    // blockIDx.x: idx of ne0 / BLOCK_SIZE
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
+static __global__ void pad_f32(const float * src, float * dst,
+                               const int lp0, const int rp0, const int lp1, const int rp1,
+                               const int lp2, const int rp2, const int lp3, const int rp3,
+                               const int ne0, const int ne1, const int ne2, const int ne3) {
+    // blockIdx.z: i3*ne2+i2
+    // blockIdx.y: i1
+    // blockIDx.x: i0 / CUDA_PAD_BLOCK_SIZE
+    // gridDim.y:  ne1
+    int i0 = threadIdx.x + blockIdx.x * blockDim.x;
+    int i1 = blockIdx.y;
+    int i2 = blockIdx.z % ne2;
+    int i3 = blockIdx.z / ne2;
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
         return;
     }
 
     // operation
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    if (nidx < ne00 && blockIdx.y < (unsigned)ne01 && blockIdx.z < (unsigned)(ne02*ne03)) {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * ne01;
-        dst[offset_dst] = x[offset_src];
+    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+    if ((i0 >= lp0 && i0 < ne0 - rp0) &&
+        (i1 >= lp1 && i1 < ne1 - rp1) &&
+        (i2 >= lp2 && i2 < ne2 - rp2) &&
+        (i3 >= lp3 && i3 < ne3 - rp3)) {
+        const int64_t i00 = i0 - lp0;
+        const int64_t i01 = i1 - lp1;
+        const int64_t i02 = i2 - lp2;
+        const int64_t i03 = i3 - lp3;
+        const int64_t ne02 = ne2 - lp2 - rp2;
+        const int64_t ne01 = ne1 - lp1 - rp1;
+        const int64_t ne00 = ne0 - lp0 - rp0;
+
+        const int64_t src_idx = i03*(ne00*ne01*ne02) + i02*(ne00*ne01) + i01*ne00 + i00;
+
+        dst[dst_idx] = src[src_idx];
     } else {
-        dst[offset_dst] = 0.0f;
+        dst[dst_idx] = 0.0f;
     }
 }
 
-static void pad_f32_cuda(const float * x, float * dst,
-    const int ne00, const int ne01, const int ne02, const int ne03,
+static void pad_f32_cuda(const float * src, float * dst,
+    const int lp0, const int rp0, const int lp1, const int rp1,
+    const int lp2, const int rp2, const int lp3, const int rp3,
     const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
     int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
     dim3 gridDim(num_blocks, ne1, ne2*ne3);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, ne0, ne1, ne2, ne3);
 }
 
 void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -41,9 +55,18 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     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
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int32_t lp0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t rp0 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t lp1 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t rp1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t lp2 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t rp2 = ((const int32_t*)(dst->op_params))[5];
+    const int32_t lp3 = ((const int32_t*)(dst->op_params))[6];
+    const int32_t rp3 = ((const int32_t*)(dst->op_params))[7];
 
     pad_f32_cuda(src0_d, dst_d,
-        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
-        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+                 lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
+                 dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
 }

package/cpp/llama.cpp/ggml/src/ggml-cuda/pad_reflect_1d.cu

@@ -1,82 +1,91 @@
 #include "pad_reflect_1d.cuh"
 
-static __global__ void pad_reflect_1d_kernel_f32(
-    const void * __restrict__ src0,
-    void * __restrict__ dst,
-    const int64_t ne0,
-    const int64_t ne00,
-    const int64_t ne01,
-    const int64_t ne02,
-    const int64_t ne03,
-    const int64_t nb00,
-    const int64_t nb01,
-    const int64_t nb02,
-    const int64_t nb03,
-    const int64_t nb0,
-    const int64_t nb1,
-    const int64_t nb2,
-    const int64_t nb3,
-    const int p0,
-    const int p1) {
-
+static __global__ __launch_bounds__(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1) void
+    pad_reflect_1d_kernel_f32(
+        const void * __restrict__ src0,
+        void * __restrict__       dst,
+        const int64_t             ne0,
+        const int64_t             ne00,
+        const uint3               ne01,
+        const int64_t             ne02,
+        const int64_t             ne03,
+        const int64_t             nb00,
+        const int64_t             nb01,
+        const int64_t             nb02,
+        const int64_t             nb03,
+        const int64_t             nb0,
+        const int64_t             nb1,
+        const int64_t             nb2,
+        const int64_t             nb3,
+        const int                 p0,
+        const int                 p1) {
     const int64_t i3 = blockIdx.z;
     const int64_t i2 = blockIdx.y;
-    const int64_t i1 = blockIdx.x;
 
-    if (i1 >= ne01 || i2 >= ne02 || i3 >= ne03) {
+    const uint2   div_mod_packed = fast_div_modulo(blockIdx.x, ne01);
+    const int64_t tile1          = div_mod_packed.y;  // i1
+    const int64_t tile0          = div_mod_packed.x;  // nth i0 tile
+    const int64_t i1             = tile1;
+    const int64_t i0             = threadIdx.x + tile0 * blockDim.x;
+
+    // ne01.z is original value of unpacked ne01 (see init_fastdiv_values in common.cuh)
+    if (i0 >= ne0 || i1 >= ne01.z || i2 >= ne02 || i3 >= ne03) {
         return;
     }
 
-    const char * src0_ptr = (const char *)src0 + i3*nb03 + i2*nb02 + i1*nb01;
-    char * dst_ptr = (char *)dst + i3*nb3 + i2*nb2 + i1*nb1;
+    const char * src0_ptr = (const char *) src0 + i3 * nb03 + i2 * nb02 + i1 * nb01;
+    char *       dst_ptr  = (char *) dst + i3 * nb3 + i2 * nb2 + i1 * nb1;
 
-    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
-        float value;
+    const int64_t rel_i0 = i0 - p0;  // relative i0 in src0
+    int64_t src_idx;
 
-        if (i0 < p0) {
-            // Left padding - reflect
-            value = *(const float *)(src0_ptr + (p0 - i0) * nb00);
-        } else if (i0 < ne0 - p1) {
-            // Middle - copy
-            value = *(const float *)(src0_ptr + (i0 - p0) * nb00);
-        } else {
-            // Right padding - reflect
-            int64_t src_idx = (ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1;
-            value = *(const float *)(src0_ptr + src_idx * nb00);
-        }
-
-        *(float *)(dst_ptr + i0 * nb0) = value;
+    if (rel_i0 < 0) {
+        // Left padding - reflect
+        src_idx = -rel_i0;
+    } else if (rel_i0 < ne00) {
+        // Middle - copy
+        src_idx = rel_i0;
+    } else {
+        // Right padding - reflect
+        src_idx = 2 * ne00 - 2 - rel_i0;
     }
+    const float value               = *(const float *) (src0_ptr + src_idx * nb00);
+    *(float *) (dst_ptr + i0 * nb0) = value;
+
+    GGML_UNUSED(p1);
 }
 
 void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    cudaStream_t stream = ctx.stream();
+    const ggml_tensor * src0   = dst->src[0];
+    cudaStream_t        stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     const int32_t * opts = (const int32_t *) dst->op_params;
-    const int p0 = opts[0];
-    const int p1 = opts[1];
+    const int       p0   = opts[0];
+    const int       p1   = opts[1];
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
+    const int64_t ne00        = src0->ne[0];
+    const int64_t ne01        = src0->ne[1];
+    const uint3   ne01_packed = init_fastdiv_values(ne01);
+    const int64_t ne02        = src0->ne[2];
+    const int64_t ne03        = src0->ne[3];
 
     const int64_t ne0 = dst->ne[0];
 
+    // sanity: padded length matches
     GGML_ASSERT(ne0 == ne00 + p0 + p1);
 
-    const dim3 block_dims(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1, 1);
-    const dim3 grid_dims(ne01, ne02, ne03);
+    constexpr int64_t bx     = CUDA_PAD_REFLECT_1D_BLOCK_SIZE;  // threads per block (x)
+    const int64_t     tiles0 = (ne0 + bx - 1) / bx;             // number of tiles along i0
+    // grid.x covers i1 and all tiles of i0: [ne01 * tiles0]
+    // grid.y covers i2: [ne02]
+    // grid.z covers i3: [ne03]
+    const dim3        grid_dims((unsigned) (ne01 * tiles0), (unsigned) ne02, (unsigned) ne03);
+    const dim3        block_dims((unsigned) bx, 1, 1);
 
     pad_reflect_1d_kernel_f32<<<grid_dims, block_dims, 0, stream>>>(
-        src0->data, dst->data,
-        ne0, ne00, ne01, ne02, ne03,
-        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
-        dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
-        p0, p1
-    );
+        src0->data, dst->data, ne0, ne00, ne01_packed, ne02, ne03, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+        dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], p0, p1);
 }

package/cpp/llama.cpp/ggml/src/ggml-cuda/quantize.cu

@@ -1,26 +1,27 @@
 #include "quantize.cuh"
 #include <cstdint>
 
+__launch_bounds__(CUDA_QUANTIZE_BLOCK_SIZE, 1)
 static __global__ void quantize_q8_1(
         const float * __restrict__ x, void * __restrict__ vy,
         const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-        const int64_t ne0, const int ne1, const int ne2) {
+        const int64_t ne0, const uint32_t ne1, const uint3 ne2) {
     const int64_t i0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i0 >= ne0) {
         return;
     }
 
+    const int64_t i3 = fastdiv(blockIdx.z, ne2);
+    const int64_t i2 = blockIdx.z - i3*ne2.z;
     const int64_t i1 = blockIdx.y;
-    const int64_t i2 = blockIdx.z % ne2;
-    const int64_t i3 = blockIdx.z / ne2;
 
     const int64_t & i00 = i0;
     const int64_t & i01 = i1;
     const int64_t & i02 = i2;
     const int64_t & i03 = i3;
 
-    const int64_t i_cont = ((i3*ne2 + i2) * ne1 + i1) * ne0 + i0;
+    const int64_t i_cont = ((i3*ne2.z + i2) * ne1 + i1) * ne0 + i0;
 
     block_q8_1 * y = (block_q8_1 *) vy;
 
@@ -31,10 +32,10 @@ static __global__ void quantize_q8_1(
     float amax = fabsf(xi);
     float sum = xi;
 
-    amax = warp_reduce_max(amax);
-    sum  = warp_reduce_sum(sum);
+    amax = warp_reduce_max<QK8_1>(amax);
+    sum  = warp_reduce_sum<QK8_1>(sum);
 
-    const float  d = amax / 127;
+    const float  d = amax / 127.0f;
     const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
 
     y[ib].qs[iqs] = q;
@@ -43,8 +44,7 @@ static __global__ void quantize_q8_1(
         return;
     }
 
-    reinterpret_cast<half&>(y[ib].ds.x) = d;
-    reinterpret_cast<half&>(y[ib].ds.y) = sum;
+    y[ib].ds = make_half2(d, sum);
 }
 
 template <mmq_q8_1_ds_layout ds_layout>
@@ -152,10 +152,12 @@ void quantize_row_q8_1_cuda(
     GGML_ASSERT(!ids);
     GGML_ASSERT(ne0 % QK8_1 == 0);
 
+    const uint3 ne2_fastdiv = init_fastdiv_values(ne2);
+
     const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
     const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2_fastdiv);
     GGML_UNUSED(type_src0);
 }