@novastera-oss/llamarn 0.0.1-alpha.4

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

@@ -0,0 +1,345 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-mma-f16.cuh"
+#include "fattn-tile-f16.cuh"
+#include "fattn-tile-f32.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
+#include "fattn-wmma-f16.cuh"
+#include "fattn.cuh"
+
+template <int DKQ, int DV, int ncols2>
+static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    if constexpr (ncols2 <= 8) {
+        if (Q->ne[1] <= 8/ncols2) {
+            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
+            return;
+        }
+    }
+
+    if (Q->ne[1] <= 16/ncols2) {
+        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32/ncols2) {
+        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
+        return;
+    }
+
+    ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 64/ncols2, ncols2>(ctx, dst);
+}
+
+template <int DKQ, int DV>
+static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV  = dst;
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * mask = dst->src[3];
+
+    float max_bias = 0.0f;
+    memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+    const bool use_gqa_opt = mask && max_bias == 0.0f;
+
+    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+    const int gqa_ratio = Q->ne[2] / K->ne[2];
+
+    if (use_gqa_opt && gqa_ratio % 8 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
+        return;
+    }
+
+    if (use_gqa_opt && gqa_ratio % 4 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
+        return;
+    }
+
+    if (use_gqa_opt && gqa_ratio % 2 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
+        return;
+    }
+
+    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+}
+
+static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV  = dst;
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+
+    switch (Q->ne[0]) {
+        case 64:
+            GGML_ASSERT(V->ne[0] == 64);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 64,  64>(ctx, dst);
+            break;
+        case 80:
+            GGML_ASSERT(V->ne[0] == 80);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 80,  80>(ctx, dst);
+            break;
+        case 96:
+            GGML_ASSERT(V->ne[0] == 96);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 96,  96>(ctx, dst);
+            break;
+        case 112:
+            GGML_ASSERT(V->ne[0] == 112);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<112, 112>(ctx, dst);
+            break;
+        case 128:
+            GGML_ASSERT(V->ne[0] == 128);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<128, 128>(ctx, dst);
+            break;
+        case 256:
+            GGML_ASSERT(V->ne[0] == 256);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
+            break;
+        case 576: {
+            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+            GGML_ASSERT(V->ne[0] == 512);
+            float max_bias = 0.0f;
+            memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+            const bool use_gqa_opt = mask && max_bias == 0.0f;
+            GGML_ASSERT(use_gqa_opt);
+
+            GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+            const int gqa_ratio = Q->ne[2] / K->ne[2];
+            GGML_ASSERT(gqa_ratio % 16 == 0);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+        } break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+#define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
+        ggml_cuda_flash_attn_ext_vec_f16_case<D, type_K, type_V>(ctx, dst); \
+        return;                                                             \
+    }                                                                       \
+
+static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * Q = dst->src[0];
+    ggml_tensor * K = dst->src[1];
+    ggml_tensor * V = dst->src[2];
+
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16 )
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
+
+    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#else
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    on_no_fattn_vec_case(Q->ne[0]);
+}
+
+#define FATTN_VEC_F32_CASE(D, type_K, type_V)                               \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
+        ggml_cuda_flash_attn_ext_vec_f32_case<D, type_K, type_V>(ctx, dst); \
+        return;                                                             \
+    }                                                                       \
+
+static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * Q = dst->src[0];
+    ggml_tensor * K = dst->src[1];
+    ggml_tensor * V = dst->src[2];
+
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
+
+    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#else
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    on_no_fattn_vec_case(Q->ne[0]);
+}
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV  = dst;
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_cuda_set_device(ctx.device);
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+    const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
+
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
+#if defined(GGML_HIP_ROCWMMA_FATTN)
+        if (fp16_mma_available(cc)) {
+            ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+            return;
+        }
+#endif // defined(GGML_HIP_ROCWMMA_FATTN)
+
+        // On AMD the tile kernels perform poorly, use the vec kernel instead:
+        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fast_fp16_available(cc)) {
+        if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fp16_mma_available(cc)) {
+        if (prec == GGML_PREC_DEFAULT) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
+                ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+            } else {
+                ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+            }
+        } else {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
+                ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            } else {
+                ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+            }
+        }
+        return;
+    }
+
+    const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
+    const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
+    if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
+        if (prec == GGML_PREC_DEFAULT) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
+    // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
+    if (fp16_mma_available(cc) && !new_mma_available(cc)) {
+        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+        return;
+    }
+
+    ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
+}

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

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -0,0 +1,275 @@
+#include "getrows.cuh"
+#include "dequantize.cuh"
+
+template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void k_get_rows(
+        const void * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
+        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
+        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
+        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
+
+    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+    const int i00 = (blockIdx.y * blockDim.x + threadIdx.x)*2;
+    const int i10 =  blockIdx.x;
+    const int i11 =  blockIdx.z / ne12;
+    const int i12 =  blockIdx.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 void * src0_row = (const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03;
+
+    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(src0_row, ib, iqs, v);
+
+    dst_row[iybs + iqs + 0]        = float(v.x);
+    dst_row[iybs + iqs + y_offset] = float(v.y);
+}
+
+template<typename src0_t, typename dst_t>
+static __global__ void k_get_rows_float(
+        const src0_t * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
+        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
+        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
+        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
+
+    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+    const int i00 = blockIdx.y * blockDim.x + threadIdx.x;
+    const int i10 = blockIdx.x;
+    const int i11 = blockIdx.z / ne12;
+    const int i12 = blockIdx.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] = float(src0_row[i00]);
+}
+
+template<typename grad_t, typename dst_t>
+static __global__ void k_get_rows_back_float(
+        const grad_t * __restrict__ grad, const int32_t * __restrict__ rows, dst_t * __restrict__ dst, const int64_t ncols, const int64_t nrows_grad) {
+    const int col = blockIdx.x*blockDim.x + threadIdx.x;
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int dst_row = blockIdx.y*blockDim.y + threadIdx.y;
+
+    float sum = 0.0f;
+
+    for (int64_t i = 0; i < nrows_grad; ++i) {
+        if (rows[i] != dst_row) {
+            continue;
+        }
+        sum += grad[i*ncols + col];
+    }
+
+    dst[dst_row*ncols + col] = sum;
+}
+
+template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
+static void get_rows_cuda_q(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_y = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(ne10, block_num_y, ne11*ne12);
+
+    // strides in elements
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
+
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
+        src0_d, src1_d, dst_d,
+        ne00, /*ne01, ne02, ne03,*/
+        /*ne10, ne11,*/ ne12, /*ne13,*/
+        /* s0,*/ s1, s2, s3,
+        /* nb00,*/ nb01, nb02, nb03,
+        s10, s11, s12/*, s13*/);
+}
+
+template<typename src0_t, typename dst_t>
+static void get_rows_cuda_float(
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(ne10, block_num_y, ne11*ne12);
+
+    // strides in elements
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
+
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
+
+    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+        src0_d, src1_d, dst_d,
+        ne00, /*ne01, ne02, ne03,*/
+        /*ne10, ne11,*/ ne12, /*ne13,*/
+        /* s0,*/ s1, s2, s3,
+        /* nb00,*/ nb01, nb02, nb03,
+        s10, s11, s12/*, s13*/);
+}
+
+template <typename dst_t>
+static void ggml_cuda_get_rows_switch_src0_type(
+        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    switch (src0_type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
+            break;
+    }
+}
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream) {
+    switch (dst_type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
+            break;
+    }
+}
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(ne13 == 1);
+
+    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));
+
+    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
+        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+}
+
+void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // gradients of forward pass output
+    const ggml_tensor * src1 = dst->src[1]; // src1 in forward pass
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const float   * src0_d = (const float   *) src0->data;
+    const int32_t * src1_d = (const int32_t *) src1->data;
+    float         * dst_d  = (float         *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    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));
+
+    GGML_ASSERT(ne02*ne03 == 1);
+    GGML_ASSERT(ne12*ne13 == 1);
+    GGML_ASSERT(ne2*ne3 == 1);
+
+    const dim3 block_dims(CUDA_GET_ROWS_BACK_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BACK_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BACK_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne1, 1);
+
+    k_get_rows_back_float<<<block_nums, block_dims, 0, stream>>>(src0_d, src1_d, dst_d, ne00, ne10);
+}

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

@@ -0,0 +1,15 @@
+#include "common.cuh"
+
+#define CUDA_GET_ROWS_BLOCK_SIZE 256
+#define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream);
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);