llama_cpp 0.13.0 → 0.14.1

data/vendor/tmp/llama.cpp/ggml-cuda.cu

@@ -2,6 +2,15 @@
 #include "ggml.h"
 #include "ggml-backend-impl.h"
 
+#if defined(GGML_USE_HIPBLAS)
+#define GGML_COMMON_DECL_HIP
+#define GGML_COMMON_IMPL_HIP
+#else
+#define GGML_COMMON_DECL_CUDA
+#define GGML_COMMON_IMPL_CUDA
+#endif
+#include "ggml-common.h"
+
 #include <algorithm>
 #include <assert.h>
 #include <atomic>
@@ -63,6 +72,7 @@
 #define cudaEventCreateWithFlags hipEventCreateWithFlags
 #define cudaEventDisableTiming hipEventDisableTiming
 #define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
 #define cudaEvent_t hipEvent_t
 #define cudaEventDestroy hipEventDestroy
 #define cudaFree hipFree
@@ -72,6 +82,7 @@
 #define cudaGetDeviceProperties hipGetDeviceProperties
 #define cudaGetErrorString hipGetErrorString
 #define cudaGetLastError hipGetLastError
+#define cudaLaunchHostFunc hipLaunchHostFunc
 #ifdef GGML_HIP_UMA
 #define cudaMalloc hipMallocManaged
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
@@ -95,6 +106,7 @@
 #define cudaStreamCreateWithFlags hipStreamCreateWithFlags
 #define cudaStreamFireAndForget hipStreamFireAndForget
 #define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
 #define cudaStreamSynchronize hipStreamSynchronize
 #define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
 #define cudaStream_t hipStream_t
@@ -172,6 +184,7 @@
 #endif
 
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
 static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
     const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
     const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
@@ -196,6 +209,18 @@ static __device__ __forceinline__ int __vsub4(const int a, const int b) {
     return __vsubss4(a, b);
 }
 
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
 static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
@@ -343,66 +368,6 @@ typedef void (*ggml_cuda_op_flatten_t)(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream);
 
-// QK = number of values after dequantization
-// QR = QK / number of values before dequantization
-// QI = number of 32 bit integers before dequantization
-
-#define QK4_0 32
-#define QR4_0 2
-#define QI4_0 (QK4_0 / (4 * QR4_0))
-typedef struct {
-    half    d;              // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-#define QR4_1 2
-#define QI4_1 (QK4_1 / (4 * QR4_1))
-typedef struct {
-    half2   dm;             // dm.x = delta, dm.y = min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-#define QR5_0 2
-#define QI5_0 (QK5_0 / (4 * QR5_0))
-typedef struct {
-    half d;                 // delta
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2];  // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-#define QR5_1 2
-#define QI5_1 (QK5_1 / (4 * QR5_1))
-typedef struct {
-    half2 dm;               // dm.x = delta, dm.y = min
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2];  // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-#define QR8_0 1
-#define QI8_0 (QK8_0 / (4 * QR8_0))
-typedef struct {
-    half    d;              // delta
-    int8_t  qs[QK8_0];      // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-#define QR8_1 1
-#define QI8_1 (QK8_1 / (4 * QR8_1))
-typedef struct {
-    half2   ds;             // ds.x = delta, ds.y = sum
-    int8_t  qs[QK8_0];      // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
-
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
 typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
 typedef void (*load_tiles_cuda_t)(
@@ -412,130 +377,6 @@ typedef float (*vec_dot_q_mul_mat_cuda_t)(
     const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
     const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k);
 
-//================================= k-quants
-
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-#define QR2_K 4
-#define QI2_K (QK_K / (4*QR2_K))
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    half2 dm;                // super-block scale for quantized scales/mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-#define QR3_K 4
-#define QI3_K (QK_K / (4*QR3_K))
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-#ifdef GGML_QKK_64
-    uint8_t scales[2]; // scales, quantized with 8 bits
-#else
-    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
-#endif
-    half d;             // super-block scale
-} block_q3_K;
-//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
-
-#define QR4_K 2
-#define QI4_K (QK_K / (4*QR4_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    half    dm[2];             // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    half2 dm;                  // super-block scale for quantized scales/mins
-    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-#define QR5_K 2
-#define QI5_K (QK_K / (4*QR5_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    half d;                  // super-block scale
-    int8_t scales[QK_K/16];  // block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    half2 dm;                     // super-block scale for quantized scales/mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];           // quants, high bit
-    uint8_t qs[QK_K/2];           // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-#define QR6_K 2
-#define QI6_K (QK_K / (4*QR6_K))
-typedef struct {
-    uint8_t ql[QK_K/2];   // quants, lower 4 bits
-    uint8_t qh[QK_K/4];   // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales
-    half    d;         // delta
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
-
-#define QR2_XXS 8
-#define QI2_XXS (QK_K / (4*QR2_XXS))
-typedef struct {
-    half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-#define QR2_XS 8
-#define QI2_XS (QK_K / (4*QR2_XS))
-typedef struct {
-    half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-#define QR3_XXS 8
-#define QI3_XXS (QK_K / (4*QR3_XXS))
-typedef struct {
-    half d;
-    uint8_t qs[3*(QK_K/8)];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-#define QR1_S 8
-#define QI1_S (QK_K / (4*QR1_S))
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/8];
-    uint8_t scales[QK_K/16];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-#define QK4_NL 32
-#define QR4_NL 2
-#define QI4_NL (QK4_NL / (4*QR4_NL))
-typedef struct {
-    half d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -559,6 +400,8 @@ static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4
 #define CUDA_UPSCALE_BLOCK_SIZE 256
 #define CUDA_CONCAT_BLOCK_SIZE 256
 #define CUDA_PAD_BLOCK_SIZE 256
+#define CUDA_ARANGE_BLOCK_SIZE 256
+#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
 #define CUDA_ACC_BLOCK_SIZE 256
 #define CUDA_IM2COL_BLOCK_SIZE 256
 #define CUDA_POOL2D_BLOCK_SIZE 256
@@ -661,18 +504,20 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
     return a;
 }
 
-//static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-//#pragma unroll
-//    for (int mask = 16; mask > 0; mask >>= 1) {
-//        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
-//    }
-//    return a;
-//#else
-//    (void) a;
-//    NO_DEVICE_CODE;
-//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-//}
+#ifdef GGML_CUDA_F16
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#pragma unroll
+   for (int mask = 16; mask > 0; mask >>= 1) {
+       a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+   }
+   return a;
+#else
+   (void) a;
+   NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
+#endif // GGML_CUDA_F16
 
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -931,17 +776,21 @@ static __global__ void concat_f32(const float * x,const float * y, float * dst,
             nidx +
             blockIdx.y * ne0 +
             blockIdx.z * ne0 * gridDim.y;
-            dst[offset_dst] = x[offset_src];
+        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];
+        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) {
+static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int ne00xne01, const int scale_factor) {
+    // blockIdx.z: idx of ne02*ne03
+    // blockIdx.y: idx of ne01*scale_factor， aka ne1
+    // blockIDx.x: idx of ne00*scale_factor / BLOCK_SIZE
+    // ne00xne01: ne00 * ne01
     int ne0 = ne00 * scale_factor;
     int nidx = threadIdx.x + blockIdx.x * blockDim.x;
     if (nidx >= ne0) {
@@ -953,7 +802,7 @@ static __global__ void upscale_f32(const float * x, float * dst, const int ne00,
     int offset_src =
         i00 +
         i01 * ne00 +
-        blockIdx.z * nb02;
+        blockIdx.z * ne00xne01;
     int offset_dst =
         nidx +
         blockIdx.y * ne0 +
@@ -961,7 +810,10 @@ static __global__ void upscale_f32(const float * x, float * dst, const int ne00,
     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) {
+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) {
         return;
@@ -972,19 +824,53 @@ static __global__ void pad_f32(const float * x, float * dst, const int ne0, cons
         nidx +
         blockIdx.y * ne0 +
         blockIdx.z * ne0 * gridDim.y;
-    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02) {
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
         int offset_src =
             nidx +
             blockIdx.y * ne00 +
             blockIdx.z * ne00 * ne01;
-            dst[offset_dst] = x[offset_src];
+        dst[offset_dst] = x[offset_src];
     } else {
         dst[offset_dst] = 0.0f;
     }
 }
 
+static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    dst[nidx] = start + step * nidx;
+}
+
+static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) {
+    // blockIDx.y: idx of timesteps->ne[0]
+    // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE
+    int i = blockIdx.y;
+    int j = threadIdx.x + blockIdx.x * blockDim.x;
+    float * embed_data = (float *)((char *)dst +  i*nb1);
+
+    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
+        embed_data[dim] = 0.f;
+    }
+
+    int half = dim / 2;
+    if (j >= half) {
+        return;
+    }
+
+    float timestep = timesteps[i];
+    float freq = (float)expf(-logf(max_period) * j / half);
+    float arg = timestep * freq;
+    embed_data[j] = cosf(arg);
+    embed_data[j + half] = sinf(arg);
+}
+
 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) {
+    // blockIdx.x: num_groups idx
+    // threadIdx.x: block_size idx
     int start = blockIdx.x * group_size;
     int end = start + group_size;
 
@@ -1467,420 +1353,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
 #endif
 }
 
-static const __device__ uint64_t iq2xxs_grid[256] = {
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
-    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
-    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
-    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
-    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
-    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
-    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
-    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
-    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
-    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
-    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
-    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
-    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
-    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
-    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
-    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
-    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
-    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
-    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
-    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
-    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
-    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
-    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
-    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
-    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
-    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
-    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
-    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
-    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
-    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
-    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
-    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
-    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
-    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
-    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
-    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
-    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
-    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
-    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
-    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
-    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
-    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
-    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
-    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
-    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
-    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
-    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
-    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
-    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
-    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
-    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
-    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
-    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
-    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
-    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
-    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
-    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
-    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
-    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
-};
-
-static const __device__ uint64_t iq2xs_grid[512] = {
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
-    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
-    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
-    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
-    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
-    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
-    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
-    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
-    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
-    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
-    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
-    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
-    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
-    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
-    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
-    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
-    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
-    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
-    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
-    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
-    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
-    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
-    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
-    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
-    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
-    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
-    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
-    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
-    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
-    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
-    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
-    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
-    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
-    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
-    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
-    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
-    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
-    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
-    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
-    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
-    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
-    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
-    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
-    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
-    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
-    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
-    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
-    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
-    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
-    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
-    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
-    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
-    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
-    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
-    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
-    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
-    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
-    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
-    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
-    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
-    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
-    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
-    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
-    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
-    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
-    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
-    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
-    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
-    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
-    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
-    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
-    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
-    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
-    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
-    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
-    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
-    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
-    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
-    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
-    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
-    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
-    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
-    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
-    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
-    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
-    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
-    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
-    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
-    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
-    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
-    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
-    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
-    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
-    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
-    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
-    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
-    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
-    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
-    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
-    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
-    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
-    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
-    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
-    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
-    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
-    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
-    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
-    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
-    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
-    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
-    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
-    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
-    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
-    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
-    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
-};
-
-static const __device__ uint32_t iq3xxs_grid[256] = {
-    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
-    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
-    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
-    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
-    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
-    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
-    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
-    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
-    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
-    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
-    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
-    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
-    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
-    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
-    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
-    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
-    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
-    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
-    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
-    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
-    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
-    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
-    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
-    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
-    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
-    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
-    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
-    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
-    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
-    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
-    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
-    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
-};
-
-static const __device__ uint64_t iq1s_grid[512] = {
-    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
-    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
-    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
-    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
-    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
-    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
-    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
-    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
-    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
-    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
-    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
-    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
-    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
-    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
-    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
-    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
-    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
-    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
-    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
-    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
-    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
-    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
-    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
-    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
-    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
-    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
-    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
-    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
-    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
-    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
-    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
-    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
-    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
-    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
-    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
-    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
-    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
-    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
-    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
-    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
-    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
-    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
-    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
-    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
-    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
-    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
-    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
-    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
-    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
-    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
-    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
-    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
-    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
-    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
-    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
-    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
-    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
-    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
-    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
-    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
-    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
-    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
-    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
-    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
-    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
-    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
-    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
-    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
-    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
-    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
-    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
-    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
-    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
-    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
-    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
-    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
-    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
-    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
-    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
-    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
-    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
-    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
-    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
-    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
-    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
-    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
-    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
-    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
-    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
-    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
-    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
-    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
-    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
-    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
-    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
-    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
-    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
-    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
-    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
-    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
-    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
-    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
-    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
-    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
-    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
-    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
-    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
-    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
-    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
-    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
-    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
-    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
-    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
-    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
-    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
-    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
-    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
-    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
-    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
-    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
-    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
-    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
-    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
-    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
-    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
-    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
-    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
-    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
-};
-
-static const __device__ uint8_t ksigns_iq2xs[128] = {
-      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
-    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
-    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
-     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
-    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
-     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
-     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
-    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
-};
-
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-static const __device__ uint64_t ksigns64[128] = {
-    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
-    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
-    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
-    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
-    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
-    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
-    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
-    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
-    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
-    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
-    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
-    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
-    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
-    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
-    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
-    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
-    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
-    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
-    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
-    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
-    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
-    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
-    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
-    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
-    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
-    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
-    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
-    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
-    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
-    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
-    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
-    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
-};
-//#endif
-
-static const __device__ uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
-
 inline bool ggml_cuda_supports_mmq(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -1945,6 +1417,27 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
 
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
@@ -1973,6 +1466,32 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
 
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * qs = x[i].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
+    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
+    const uint8_t signs = x[i].signs[4*ib + il];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
 template<typename dst_t>
 static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
@@ -1984,11 +1503,15 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const int i8 = 4*ib+il;
-    uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
-    const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
-    const float d = (float)x[i].d * (2*(h & 7) + 1);
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
 #else
     assert(false);
 #endif
@@ -2016,6 +1539,25 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
 
 }
 
+#if QK_K != 64
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int i   = blockIdx.x;
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const int tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
+    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+#endif
+
 static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
@@ -2112,10 +1654,7 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
 #endif
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (threadIdx.x == 0) {
         dst[row] = tmp;
@@ -2216,10 +1755,7 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
 #endif
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (threadIdx.x == 0) {
         dst[row] = tmp;
@@ -2352,10 +1888,7 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 #endif
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (tid == 0) {
         dst[row] = tmp;
@@ -2468,10 +2001,7 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
 #endif
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (threadIdx.x == 0) {
         dst[row] = tmp;
@@ -2578,10 +2108,7 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 #endif
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (tid == 0) {
         dst[row] = tmp;
@@ -2616,11 +2143,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
     float amax = fabsf(xi);
     float sum = xi;
 
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        amax = fmaxf(amax, __shfl_xor_sync(0xffffffff, amax, mask, 32));
-        sum += __shfl_xor_sync(0xffffffff, sum, mask, 32);
-    }
+    amax = warp_reduce_max(amax);
+    sum = warp_reduce_sum(sum);
 
     const float d = amax / 127;
     const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
@@ -3827,7 +3351,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR2_K; ++ i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
     }
 
     return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
@@ -3949,7 +3473,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR3_K; ++i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
     }
 
     return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
@@ -4118,7 +3642,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
 
     for (int i = 0; i < QR4_K; ++i) {
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = __low2half(bq8i->ds);
+        d8[i] = __low2float(bq8i->ds);
 
         const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
         u[2*i+0] = q8[0];
@@ -4483,7 +4007,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR6_K; ++i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
+        d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
     }
 
     return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
@@ -4682,6 +4206,54 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
 #endif
 }
 
+// TODO
+static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
+
+    const int ib32 = iqs;
+    const int8_t  * q8 = bq8_1[ib32].qs;
+    const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32;
+    const uint8_t ls1 = bq2->scales[ib32] & 0xf;
+    const uint8_t ls2 = bq2->scales[ib32] >>  4;
+    int sumi1 = 0;
+    for (int l = 0; l < 2; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
+        const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
+        const uint32_t signs1 = __vcmpeq4(((signs[l] >>  4) * 0x01010101) & 0x08040201, 0x08040201);
+        const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
+        sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
+        sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
+        q8 += 8;
+    }
+    int sumi2 = 0;
+    for (int l = 2; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
+        const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
+        const uint32_t signs1 = __vcmpeq4(((signs[l] >>  4) * 0x01010101) & 0x08040201, 0x08040201);
+        const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
+        sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
+        sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
+    return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
+#else
+    (void) ksigns64;
+    assert(false);
+    return 0.f;
+#endif
+#else
+    (void) ksigns64;
+    assert(false);
+    return 0.f;
+#endif
+}
+
 static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
@@ -4717,43 +4289,70 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #endif
 }
 
+// TODO: don't use lookup table for signs
+static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
+
+    const int ib32 = iqs;
+    const uint8_t  * qs = bq2->qs + 8*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256));
+        const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256));
+        uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
+        uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >>  4) * 0x01010101) & 0x08040201, 0x08040201);
+        const int grid_l = __vsub4(grid1[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid2[0] ^ signs1, signs1);
+        sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
+        sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds);
+    return d * sumi;
+#else
+    assert(false);
+    return 0.f;
+#endif
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
 static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if QK_K == 256
     const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
 
     const int ib32 = iqs;
-    int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
-    const uint8_t h1 = bq1->scales[2*ib32+0];
-    const uint8_t h2 = bq1->scales[2*ib32+1];
+    int sumi = 0;
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
     const int * q8 = (const int *)bq8_1[ib32].qs;
-    const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
-    const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
-    const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
-    const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
-    for (int j = 0; j < 2; ++j) {
-        sumi1 = __dp4a(q8[j+0], grid1[j], sumi1);
-        sumi2 = __dp4a(q8[j+2], grid2[j], sumi2);
-        sumi3 = __dp4a(q8[j+4], grid3[j], sumi3);
-        sumi4 = __dp4a(q8[j+6], grid4[j], sumi4);
+    for (int l = 0; l < 4; ++l) {
+        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
+        sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi));
     }
 #else
-    const int8_t   * q8 = bq8_1[ib32].qs;
-    const int8_t * grid1 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
-    const int8_t * grid2 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
-    const int8_t * grid3 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
-    const int8_t * grid4 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
-    for (int j = 0; j < 8; ++j) {
-        sumi1 += q8[j+ 0] * grid1[j];
-        sumi2 += q8[j+ 8] * grid2[j];
-        sumi3 += q8[j+16] * grid3[j];
-        sumi4 += q8[j+24] * grid4[j];
+    const int8_t * q8 = bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        for (int j = 0; j < 4; ++j) {
+            sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
+        }
+        q8 += 8;
     }
 #endif
-    const float d = (float)bq1->d * __low2float(bq8_1[ib32].ds);
-    return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
-                sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
+    const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA;
+    const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1);
+    const float d = d1q * __low2float (bq8_1[ib32].ds);
+    const float m = d1q * __high2float(bq8_1[ib32].ds);
+    return d * sumi + m * delta;
 #else
     assert(false);
     return 0.f;
@@ -4810,6 +4409,75 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
     return d * (sumi1 + sumi2);
 }
 
+static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+#if QK_K == 256
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+
+    const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
+    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
+
+    //// iqs is 0...7
+    //const int ib64 = iqs/2;
+    //const int il = iqs%2;
+    //const int32_t  * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il;
+    //const int32_t  * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il;
+    //const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il;
+    //const uint32_t * q4_2 = q4_1 + 4;
+    //const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4);
+    //const int8_t ls2 = (bq4->scales_l[ib64] >>  4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4);
+    //const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds);
+    //const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds);
+    //int v1, v2;
+    //int sumi1 = 0, sumi2 = 0;
+    //for (int j = 0; j < 2; ++j) {
+    //    get_int_from_table_16(q4_1[j], values, v1, v2);
+    //    sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1));
+    //    get_int_from_table_16(q4_2[j], values, v1, v2);
+    //    sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2));
+    //}
+    //return d1 * sumi1 + d2 * sumi2;
+
+    // iqs is 0...7
+    const int ib32 = iqs;
+    const int32_t  * q8 = (const int *)bq8_1[ib32].qs;
+    const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32;
+    const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
+    int v1, v2;
+    int sumi1 = 0, sumi2 = 0;
+    for (int j = 0; j < 4; ++j) {
+        get_int_from_table_16(q4[j], values, v1, v2);
+        sumi1 = __dp4a(v1, q8[j+0], sumi1);
+        sumi2 = __dp4a(v2, q8[j+4], sumi2);
+    }
+    return d * (sumi1 + sumi2);
+
+    //// iqs is 0...15
+    //const int ib32 = iqs/2;
+    //const int il = iqs%2;
+    //const int32_t  * q8 = (const int *)bq8_1[ib32].qs + 2*il;
+    //const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il;
+    //const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
+    //const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
+    //int v1, v2;
+    //int sumi1 = 0, sumi2 = 0;
+    //for (int j = 0; j < 2; ++j) {
+    //    get_int_from_table_16(q4[j], values, v1, v2);
+    //    sumi1 = __dp4a(v1, q8[j+0], sumi1);
+    //    sumi2 = __dp4a(v2, q8[j+4], sumi2);
+    //}
+    //return d * (sumi1 + sumi2);
+#else
+    assert(false);
+    return 0.f;
+#endif
+#else
+    return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
+#endif
+}
+
 template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
               allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
 static __device__ __forceinline__ void mul_mat_q(
@@ -4876,7 +4544,7 @@ static __device__ __forceinline__ void mul_mat_q(
                     *dsi_dst = *dsi_src;
                 } else {
                     float * dfi_dst = (float *) dsi_dst;
-                    *dfi_dst = __low2half(*dsi_src);
+                    *dfi_dst = __low2float(*dsi_src);
                 }
             }
 
@@ -5730,10 +5398,7 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
     }
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (tid == 0) {
 #ifdef GGML_CUDA_F16
@@ -5783,10 +5448,7 @@ static __global__ void mul_mat_p021_f16_f32(
     const int idst = channel*nrows_dst + row_dst;
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (threadIdx.x == 0) {
         dst[idst] = tmp;
@@ -5829,10 +5491,7 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     }
 
     // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
+    tmp = warp_reduce_sum(tmp);
 
     if (threadIdx.x == 0) {
         dst[idst] = tmp;
@@ -5872,7 +5531,7 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
                                    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
         return;
@@ -5880,17 +5539,17 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
 
     // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
     // then combine those indices with the corresponding byte offsets to get the total offsets
-    const int i03 = i/(ne00 * ne01 * ne02);
-    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
-    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
-    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
-    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
-
-    const int i13 = i/(ne10 * ne11 * ne12);
-    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
-    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
-    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
-    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
+    const int64_t i03 = i/(ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int64_t i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int64_t i13 = i/(ne10 * ne11 * ne12);
+    const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
 
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
@@ -6216,11 +5875,11 @@ static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int n
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
+                    if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
                         swap(dst_row[col], dst_row[ixj]);
                     }
                 } else {
-                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
+                    if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
                         swap(dst_row[col], dst_row[ixj]);
                     }
                 }
@@ -6328,6 +5987,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f
     // find the sum of exps in the block
     tmp = warp_reduce_sum(tmp);
     if (block_size > WARP_SIZE) {
+        __syncthreads();
         if (warp_id == 0) {
             buf_iw[lane_id] = 0.0f;
         }
@@ -6379,23 +6039,23 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min,
 
 template <typename T>
 static  __global__ void im2col_kernel(
-        const float * x, T * dst, int batch_offset,
-        int offset_delta, int IC, int IW, int IH, int OH, int OW, int KW, int KH, int pelements, int CHW,
+        const float * x, T * dst, int64_t batch_offset,
+        int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
         int s0, int s1, int p0, int p1, int d0, int d1) {
-    const int i = threadIdx.x + blockIdx.x * blockDim.x;
+    const int64_t 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 int64_t  ksize = OW * (KH > 1 ? KW : 1);
+    const int64_t  kx = i / ksize;
+    const int64_t  kd = kx * ksize;
+    const int64_t  ky = (i - kd) / OW;
+    const int64_t  ix = i % OW;
 
-    const int oh = blockIdx.y;
-    const int batch = blockIdx.z / IC;
-    const int ic = blockIdx.z % IC;
+    const int64_t  oh = blockIdx.y;
+    const int64_t  batch = blockIdx.z / IC;
+    const int64_t  ic = blockIdx.z % IC;
 
     const int64_t iiw = ix * s0 + kx * d0 - p0;
     const int64_t iih = oh * s1 + ky * d1 - p1;
@@ -6721,19 +6381,33 @@ static void concat_f32_cuda(const float * x, const float * y, float * dst, const
     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) {
+static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int ne03,
+                             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);
+    dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02*ne03);
     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) {
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    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);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02);
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
+    arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start,  step);
+}
+
+static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1,
+                                        const int dim, const int max_period, cudaStream_t stream) {
+    int half_ceil = (dim + 1) / 2;
+    int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne00, 1);
+    timestep_embedding_f32<<<gridDim, CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE, 0, stream>>>(x, dst, nb1, dim, max_period);
 }
 
 static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
@@ -6843,12 +6517,24 @@ static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template<typename dst_t>
 static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
     dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template<typename dst_t>
 static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
@@ -6861,6 +6547,16 @@ static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+#if QK_K == 64
+    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
+#else
+    dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
+#endif
+}
+
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
@@ -6898,12 +6594,18 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq2_xxs_cuda;
         case GGML_TYPE_IQ2_XS:
             return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
         case GGML_TYPE_IQ3_XXS:
             return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_cuda;
         case GGML_TYPE_IQ4_NL:
             return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
         default:
@@ -6937,12 +6639,18 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq2_xxs_cuda;
         case GGML_TYPE_IQ2_XS:
             return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
         case GGML_TYPE_IQ3_XXS:
             return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_cuda;
         case GGML_TYPE_IQ4_NL:
             return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half>;
         default:
@@ -7764,10 +7472,10 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
 
     const dim3 block_dims(ncols, 1, 1);
     const dim3 block_nums(1, nrows, 1);
-    if (order == GGML_SORT_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ASC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
-    } else if (order == GGML_SORT_DESC) {
-        k_argsort_f32_i32<GGML_SORT_DESC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    if (order == GGML_SORT_ORDER_ASC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    } else if (order == GGML_SORT_ORDER_DESC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
     } else {
         GGML_ASSERT(false);
     }
@@ -7832,8 +7540,8 @@ static void soft_max_f32_cuda(const float * x, const float * mask, const float *
 
 template <typename T>
 static void im2col_cuda(const float* x, T* dst,
-    int IW, int IH, int OW, int OH, int KW, int KH, int IC,
-    int batch, int batch_offset, int offset_delta,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t 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;
@@ -7916,8 +7624,8 @@ static void * ggml_cuda_pool_malloc_leg(int device, size_t size, size_t * actual
     *actual_size = look_ahead_size;
     g_cuda_pool_size[device] += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
-    fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
-            (uint32_t)(max_size/1024/1024), (uint32_t)(g_cuda_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
+    fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
+            (uint32_t)(max_size/1024/1024), (uint32_t)(g_cuda_pool_size[device]/1024/1024), (uint32_t)(size/1024/1024));
 #endif
     return ptr;
 }
@@ -8003,7 +7711,7 @@ static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual
     g_cuda_pool_used[device] += size;
 
 #ifdef DEBUG_CUDA_MALLOC
-    printf("cuda pool[%d]: allocated %llu bytes at %llx [%s]\n", id, (unsigned long long) size, ptr);
+    printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
 #endif
 
     return ptr;
@@ -8013,7 +7721,7 @@ static void ggml_cuda_pool_free_vmm(int device, void * ptr, size_t size) {
     scoped_spin_lock lock(g_cuda_pool_lock);
 
 #ifdef DEBUG_CUDA_MALLOC
-    printf("cuda pool[%d]: freed %llu bytes at %llx\n", id, (unsigned long long) size, ptr);
+    printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
 #endif
 
     g_cuda_pool_used[device] -= size;
@@ -8199,11 +7907,11 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
 
     cudaMemcpyKind kind;
     char * src_ptr;
-    if (src->backend == GGML_BACKEND_CPU) {
+    if (src->backend == GGML_BACKEND_TYPE_CPU) {
         kind = cudaMemcpyHostToDevice;
         src_ptr = (char *) src->data;
-    } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) {
-        GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
+    } else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
+        GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
         kind = cudaMemcpyDeviceToDevice;
         ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
         int id;
@@ -8512,7 +8220,7 @@ static void ggml_cuda_op_group_norm(
 
     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);
+    group_norm_f32_cuda(src0_dd, dst_dd, num_groups * src0->ne[3], group_size, ggml_nelements(src0), main_stream);
 
     (void) src1;
     (void) dst;
@@ -8545,7 +8253,7 @@ static void ggml_cuda_op_upscale(
 
     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);
+    upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], scale_factor, main_stream);
 
     (void) src1;
     (void) dst;
@@ -8561,8 +8269,49 @@ static void ggml_cuda_op_pad(
     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);
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+static void ggml_cuda_op_arange(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
+
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    float start;
+    float stop;
+    float step;
+    memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
+    memcpy(&stop,  (float *)dst->op_params + 1, sizeof(float));
+    memcpy(&step,  (float *)dst->op_params + 2, sizeof(float));
+
+    int64_t steps = (int64_t)ceil((stop - start) / step);
+    GGML_ASSERT(ggml_nelements(dst) == steps);
+
+    arange_f32_cuda(dst_dd, dst->ne[0], start, step, main_stream);
+
+    (void) src0;
+    (void) src1;
+    (void) src0_dd;
+    (void) src1_dd;
+}
+
+static void ggml_cuda_op_timestep_embedding(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int dim = dst->op_params[0];
+    const int max_period = dst->op_params[1];
+
+    timestep_embedding_f32_cuda(src0_dd, dst_dd, src0->ne[0], dst->nb[1], dim, max_period, main_stream);
 
     (void) src1;
     (void) dst;
@@ -8608,7 +8357,7 @@ static void ggml_cuda_op_mul_mat_q(
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
+    const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff;
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
@@ -8685,9 +8434,12 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
             return max_compute_capability >= CC_RDNA2 ? 128 : 64;
         default:
             GGML_ASSERT(false);
@@ -8710,9 +8462,12 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
             return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
@@ -8755,7 +8510,7 @@ static void ggml_cuda_op_mul_mat_vec_q(
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
+    const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff;
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
@@ -8806,6 +8561,10 @@ static void ggml_cuda_op_mul_mat_vec_q(
             mul_mat_vec_q_cuda<QK_K, QI2_XS, block_iq2_xs, 1, vec_dot_iq2_xs_q8_1>
                 (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_q_cuda<QK_K, QI2_S, block_iq2_s, 1, vec_dot_iq2_s_q8_1>
+                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_cuda<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
                 (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
@@ -8818,6 +8577,14 @@ static void ggml_cuda_op_mul_mat_vec_q(
             mul_mat_vec_q_cuda<QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
                 (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_q_cuda<QK_K, QI4_XS, block_iq4_xs, 1, vec_dot_iq4_xs_q8_1>
+                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_q_cuda<QK_K, QI3_XS, block_iq3_s, 1, vec_dot_iq3_s_q8_1>
+                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -8927,7 +8694,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
+    int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff;
 
     const int compute_capability = g_device_caps[id].cc;
 
@@ -9275,7 +9042,7 @@ static void ggml_cuda_op_soft_max(
     const bool use_src2 = src2 != nullptr;
 
     if (use_src2) {
-        const bool src2_on_device = src2->backend == GGML_BACKEND_GPU;
+        const bool src2_on_device = src2->backend == GGML_BACKEND_TYPE_GPU;
 
         if (src2_on_device) {
             ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
@@ -9333,16 +9100,16 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s
     const bool use_src1 = src1 != nullptr;
     const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
 
-    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(              dst->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
 
     ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
     ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
 
-    const bool src0_on_device =             src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
-    const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
-    const bool  dst_on_device =              dst->backend == GGML_BACKEND_GPU;
+    const bool src0_on_device =             src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU;
+    const bool  dst_on_device =              dst->backend == GGML_BACKEND_TYPE_GPU;
 
     // dd = data device
     float * src0_ddf = nullptr;
@@ -9386,7 +9153,7 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s
         CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream));
     }
 
-    if (dst->backend == GGML_BACKEND_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
         CUDA_CHECK(cudaDeviceSynchronize());
     }
 }
@@ -9467,8 +9234,8 @@ static void ggml_cuda_op_mul_mat(
     const int nb2 = dst->nb[2];
     const int nb3 = dst->nb[3];
 
-    GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
 
     GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
@@ -9484,20 +9251,20 @@ static void ggml_cuda_op_mul_mat(
     ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
     ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
 
-    const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
+    const bool src0_on_device = src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
     const bool src1_is_contiguous = ggml_is_contiguous(src1);
 
     const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
 
-    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
+    const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
     GGML_ASSERT(!(split && ne02 > 1));
     GGML_ASSERT(!(split && ne03 > 1));
     GGML_ASSERT(!(split && ne02 < ne12));
 
     std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
     if (split) {
-        // TODO: check that src0->buffer->buft is a split buffer type, replace GGML_BACKEND_GPU_SPLIT check
+        // TODO: check that src0->buffer->buft is a split buffer type, replace GGML_BACKEND_TYPE_GPU_SPLIT check
         // GGML_ASSERT(src0->buffer != nullptr && src0->buffer->buft == ...);
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
         tensor_split = buft_ctx->tensor_split;
@@ -9555,8 +9322,8 @@ static void ggml_cuda_op_mul_mat(
 
         used_devices++;
 
-        const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
-        const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
+        const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
+        const bool  dst_on_device =  dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
 
         ggml_cuda_set_device(id);
         cudaStream_t stream = g_cudaStreams[id][0];
@@ -9607,8 +9374,8 @@ static void ggml_cuda_op_mul_mat(
                 continue;
             }
 
-            const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
-            const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
+            const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
+            const bool  dst_on_device =  dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
             const int64_t row_diff = dev[id].row_high - dev[id].row_low;
 
             ggml_cuda_set_device(id);
@@ -9633,12 +9400,12 @@ static void ggml_cuda_op_mul_mat(
 
                 // the main device memory buffer can be on VRAM scratch, with space for all partial results
                 // in that case an offset on dst_ddf_i is needed
-                if (dst->backend == GGML_BACKEND_GPU && id == g_main_device) {
+                if (dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device) {
                     dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split
                 }
 
                 // copy src0, src1 to device if necessary
-                if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
+                if (src1->backend == GGML_BACKEND_TYPE_GPU && src1_is_contiguous) {
                     if (id != g_main_device) {
                         if (convert_src1_to_q8_1) {
                             char * src1_ddq_i_source = dev[g_main_device].src1_ddq + src1_ddq_i_offset;
@@ -9651,14 +9418,14 @@ static void ggml_cuda_op_mul_mat(
                                                             src1_ncols*ne10*sizeof(float), stream));
                         }
                     }
-                } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) {
+                } else if (src1->backend == GGML_BACKEND_TYPE_CPU || (src1_on_device && !src1_is_contiguous)) {
                     CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
                                 src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                 } else {
                     GGML_ASSERT(false);
                 }
 
-                if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_CPU || !src1_is_contiguous)) {
+                if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_TYPE_CPU || !src1_is_contiguous)) {
                     quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
@@ -9676,10 +9443,10 @@ static void ggml_cuda_op_mul_mat(
                 if (!dst_on_device) {
                     void * dst_off_device;
                     cudaMemcpyKind kind;
-                    if (dst->backend == GGML_BACKEND_CPU) {
+                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
                         dst_off_device = dst->data;
                         kind = cudaMemcpyDeviceToHost;
-                    } else if (dst->backend == GGML_BACKEND_GPU) {
+                    } else if (dst->backend == GGML_BACKEND_TYPE_GPU) {
                         dst_off_device = dst_extra->data_device[g_main_device];
                         kind = cudaMemcpyDeviceToDevice;
                     } else {
@@ -9744,7 +9511,7 @@ static void ggml_cuda_op_mul_mat(
         }
     }
 
-    if (dst->backend == GGML_BACKEND_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
         ggml_cuda_set_device(g_main_device);
         CUDA_CHECK(cudaDeviceSynchronize());
     }
@@ -9829,6 +9596,45 @@ static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, gg
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad);
 }
 
+static void ggml_cuda_arange(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
+
+    const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU;
+
+    // dd = data device
+    float * src0_ddf = nullptr;
+    float * src1_ddf = nullptr;
+    float *  dst_ddf = nullptr;
+
+    cuda_pool_alloc<float>  dst_f;
+
+    ggml_cuda_set_device(g_main_device);
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
+
+    if (dst_on_device) {
+        dst_ddf = (float *) dst_extra->data_device[g_main_device];
+    } else {
+        dst_ddf = dst_f.alloc(ggml_nelements(dst));
+    }
+
+    // do the computation
+    ggml_cuda_op_arange(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    // copy dst to host if necessary
+    if (!dst_on_device) {
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream));
+    }
+
+    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+        CUDA_CHECK(cudaDeviceSynchronize());
+    }
+}
+
+static void ggml_cuda_timestep_embedding(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_timestep_embedding);
+}
+
 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);
 }
@@ -9850,7 +9656,7 @@ GGML_CALL bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const stru
 
 static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
     GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -9881,7 +9687,7 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -9940,7 +9746,7 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
     GGML_TENSOR_BINARY_OP_LOCALS
@@ -10086,11 +9892,11 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm
 
 static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool all_on_device =
-        (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
-        (src1->backend == GGML_BACKEND_GPU) &&
-        ( dst->backend == GGML_BACKEND_GPU);
+        (src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT) &&
+        (src1->backend == GGML_BACKEND_TYPE_GPU) &&
+        ( dst->backend == GGML_BACKEND_TYPE_GPU);
 
-    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
+    const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
 
     int64_t min_compute_capability = INT_MAX;
 
@@ -10240,7 +10046,7 @@ static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) {
     GGML_ASSERT(!ggml_is_transposed(src00));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src00->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
     const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
@@ -10384,7 +10190,7 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
 
     cudaStream_t stream = g_cudaStreams[g_main_device][0];
 
-    if (ids->backend == GGML_BACKEND_GPU) {
+    if (ids->backend == GGML_BACKEND_TYPE_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, stream));
         CUDA_CHECK(cudaStreamSynchronize(stream));
@@ -10401,20 +10207,20 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
     ggml_tensor src1_row = *src1;
     ggml_tensor dst_row = *dst;
 
-    src1_row.backend = GGML_BACKEND_GPU;
-    dst_row.backend  = GGML_BACKEND_GPU;
+    src1_row.backend = GGML_BACKEND_TYPE_GPU;
+    dst_row.backend  = GGML_BACKEND_TYPE_GPU;
 
     src1_row.extra = &src1_row_extra;
     dst_row.extra = &dst_row_extra;
 
-    char * src1_original = src1->backend == GGML_BACKEND_CPU ?
+    char * src1_original = src1->backend == GGML_BACKEND_TYPE_CPU ?
         (char *) src1->data : (char *) src1_extra->data_device[g_main_device];
-    char * dst_original  =  dst->backend == GGML_BACKEND_CPU ?
+    char * dst_original  =  dst->backend == GGML_BACKEND_TYPE_CPU ?
         (char *)  dst->data : (char *)  dst_extra->data_device[g_main_device];
 
     if (src1->ne[1] == 1) {
-        GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
-        GGML_ASSERT(dst->backend  == GGML_BACKEND_GPU);
+        GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+        GGML_ASSERT(dst->backend  == GGML_BACKEND_TYPE_GPU);
 
         for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
             //int32_t row_id;
@@ -10442,9 +10248,9 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
         src1_row_extra.data_device[g_main_device] = src1_contiguous.get();
         dst_row_extra.data_device[g_main_device]  =  dst_contiguous.get();
 
-        const cudaMemcpyKind src1_kind = src1->backend == GGML_BACKEND_CPU ?
+        const cudaMemcpyKind src1_kind = src1->backend == GGML_BACKEND_TYPE_CPU ?
             cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice;
-        const cudaMemcpyKind dst_kind  =  dst->backend == GGML_BACKEND_CPU ?
+        const cudaMemcpyKind dst_kind  =  dst->backend == GGML_BACKEND_TYPE_CPU ?
             cudaMemcpyDeviceToHost : cudaMemcpyDeviceToDevice;
 
         for (int32_t row_id = 0; row_id < n_as; ++row_id) {
@@ -10499,7 +10305,7 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
         }
     }
 
-    if (dst->backend == GGML_BACKEND_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
         CUDA_CHECK(cudaStreamSynchronize(stream));
     }
 }
@@ -10516,8 +10322,8 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
-    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
-    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
 
     GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
     GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
@@ -10648,9 +10454,9 @@ GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, st
     if (!g_cublas_loaded) return false;
 
     ggml_cuda_func_t func;
-    const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
-        || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
-        || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
+    const bool any_on_device = tensor->backend == GGML_BACKEND_TYPE_GPU
+        || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
+        || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU);
 
     if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) {
         return false;
@@ -10729,6 +10535,12 @@ GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, st
         case GGML_OP_PAD:
             func = ggml_cuda_pad;
             break;
+        case GGML_OP_ARANGE:
+            func = ggml_cuda_arange;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            func = ggml_cuda_timestep_embedding;
+            break;
         case GGML_OP_LEAKY_RELU:
             func = ggml_cuda_leaky_relu;
             break;
@@ -10797,14 +10609,14 @@ GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, st
             return false;
     }
 
-    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT) {
+    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
         ggml_cuda_set_peer_access(tensor->src[1]->ne[1]);
     }
 
     if (params->ith != 0) {
         return true;
     }
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return true;
     }
     func(tensor->src[0], tensor->src[1], tensor);
@@ -10832,8 +10644,20 @@ GGML_CALL void ggml_cuda_get_device_description(int device, char * description,
 #define UNUSED GGML_UNUSED
 
 struct ggml_backend_cuda_context {
+    explicit ggml_backend_cuda_context(int device) :
+        device(device),
+        name(GGML_CUDA_NAME + std::to_string(device)) {
+    }
+
+    ~ggml_backend_cuda_context() {
+        if (copy_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(copy_event));
+        }
+    }
+
     int device;
     std::string name;
+    cudaEvent_t copy_event = nullptr;
 };
 
 // cuda buffer
@@ -10903,7 +10727,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
 
     extra->data_device[ctx->device] = tensor->data;
 
-    tensor->backend = GGML_BACKEND_GPU;
+    tensor->backend = GGML_BACKEND_TYPE_GPU;
     tensor->extra = extra;
 
     if (ggml_is_quantized(tensor->type)) {
@@ -10918,42 +10742,40 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
 }
 
 GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     if (ggml_backend_buffer_is_cuda(src->buffer)) {
         ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
-        ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
-
-        ggml_cuda_set_device(src_ctx->device);
-        CUDA_CHECK(cudaDeviceSynchronize());
-        ggml_cuda_set_device(dst_ctx->device);
-        CUDA_CHECK(cudaDeviceSynchronize());
-        CUDA_CHECK(cudaMemcpy((char *)dst->data, (const char *)src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice));
-        CUDA_CHECK(cudaDeviceSynchronize());
-
+        ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context;
+        if (src_ctx->device == dst_ctx->device) {
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
+        } else {
+            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread));
+        }
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
         return true;
     }
     return false;
+
+    UNUSED(buffer);
 }
 
 GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
@@ -11164,7 +10986,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu
             CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
         }
     }
-    tensor->backend = GGML_BACKEND_GPU_SPLIT;
+    tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT;
     tensor->extra = extra;
 }
 
@@ -11198,7 +11020,11 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buf
         }
 
         const char * buf_host = (const char *)data + offset_split;
-        CUDA_CHECK(cudaMemcpy(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice));
+        CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    }
+
+    for (int id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
     }
 }
 
@@ -11232,7 +11058,11 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buf
         }
 
         char * buf_host = (char *)data + offset_split;
-        CUDA_CHECK(cudaMemcpy(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost));
+        CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    }
+
+    for (int id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
     }
 }
 
@@ -11411,6 +11241,10 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
     return &ggml_backend_cuda_buffer_type_host;
 }
 
+//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) {
+//    return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
+//}
+
 // backend
 
 GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
@@ -11434,31 +11268,71 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer
 
 GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
 }
 
 GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
 }
 
-GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
 
-    if (dst->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && ggml_backend_buffer_is_cuda(src->buffer)) {
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx->device][0]));
-        return true;
+    ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
+    ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
+        return false;
     }
 
-    return false;
+    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
+        return false;
+    }
+
+    // device -> device
+    ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
+    ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
+
+    if (backend_src != backend_dst) {
+        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+
+        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
+        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
+
+        if (!cuda_ctx_src->copy_event) {
+            ggml_cuda_set_device(cuda_ctx_src->device);
+            CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
+        }
+
+        // copy on src stream
+        if (cuda_ctx_src->device == cuda_ctx_dst->device) {
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx_dst->device][0]));
+        } else {
+            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), g_cudaStreams[cuda_ctx_src->device][0]));
+        }
+
+        // record event on src stream
+        CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, g_cudaStreams[cuda_ctx_src->device][0]));
+
+        // wait on dst stream for the copy to complete
+        CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[cuda_ctx_dst->device][0], cuda_ctx_src->copy_event, 0));
+    } else {
+        // src and dst are on the same backend
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx_dst->device][0]));
+    }
+    return true;
 }
 
 GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
@@ -11469,13 +11343,13 @@ GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     ggml_cuda_set_main_device(cuda_ctx->device);
 
     ggml_compute_params params = {};
-    params.type = GGML_TASK_COMPUTE;
+    params.type = GGML_TASK_TYPE_COMPUTE;
     params.ith = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -11485,13 +11359,13 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg
         }
 
 #ifndef NDEBUG
-        assert(node->backend == GGML_BACKEND_GPU || node->backend == GGML_BACKEND_GPU_SPLIT);
+        assert(node->backend == GGML_BACKEND_TYPE_GPU || node->backend == GGML_BACKEND_TYPE_GPU_SPLIT);
         assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
         assert(node->extra != nullptr);
 
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             if (node->src[j] != nullptr) {
-                assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT);
+                assert(node->src[j]->backend == GGML_BACKEND_TYPE_GPU || node->src[j]->backend == GGML_BACKEND_TYPE_GPU_SPLIT);
                 assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
                 assert(node->src[j]->extra != nullptr);
             }
@@ -11505,7 +11379,7 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg
         GGML_ASSERT(ok);
     }
 
-    return true;
+    return GGML_STATUS_SUCCESS;
 }
 
 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
@@ -11541,7 +11415,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 }
                 ggml_type a_type = a->type;
                 if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
-                    a_type == GGML_TYPE_IQ1_S   || a_type == GGML_TYPE_IQ4_NL) {
+                    a_type == GGML_TYPE_IQ1_S   || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S   ||
+                    a_type == GGML_TYPE_IQ2_S   || a_type == GGML_TYPE_IQ4_XS) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }
@@ -11623,6 +11498,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_GROUP_NORM:
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
             return true;
         default:
@@ -11632,6 +11509,52 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
     UNUSED(backend);
 }
 
+static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    ggml_cuda_set_device(cuda_ctx->device);
+
+    cudaEvent_t event;
+    CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
+
+    return new ggml_backend_event {
+        /* .backend = */ backend,
+        /* .context = */ event,
+    };
+}
+
+static void ggml_backend_cuda_event_free(ggml_backend_event_t event) {
+    CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
+
+    delete event;
+}
+
+static void ggml_backend_cuda_event_record(ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)event->backend->context;
+
+    CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, g_cudaStreams[cuda_ctx->device][0]));
+}
+
+static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    if (ggml_backend_is_cuda(event->backend)) {
+        CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[cuda_ctx->device][0], (cudaEvent_t)event->context, 0));
+    } else {
+        // untested
+        auto wait_fn = [](void * user_data) {
+            ggml_backend_event_t event = (ggml_backend_event_t)user_data;
+            ggml_backend_event_synchronize(event);
+        };
+
+        CUDA_CHECK(cudaLaunchHostFunc(g_cudaStreams[cuda_ctx->device][0], wait_fn, event));
+    }
+}
+
+static void ggml_backend_cuda_event_synchronize(ggml_backend_event_t event) {
+    CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
+}
+
 static ggml_backend_i ggml_backend_cuda_interface = {
     /* .get_name                = */ ggml_backend_cuda_name,
     /* .free                    = */ ggml_backend_cuda_free,
@@ -11645,8 +11568,18 @@ static ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .supports_op             = */ ggml_backend_cuda_supports_op,
+    /* .event_new               = */ ggml_backend_cuda_event_new,
+    /* .event_free              = */ ggml_backend_cuda_event_free,
+    /* .event_record            = */ ggml_backend_cuda_event_record,
+    /* .event_wait              = */ ggml_backend_cuda_event_wait,
+    /* .event_synchronize       = */ ggml_backend_cuda_event_synchronize,
 };
 
+static ggml_guid_t ggml_backend_cuda_guid() {
+    static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
+    return &guid;
+}
+
 GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
     ggml_init_cublas(); // TODO: remove from ggml.c
 
@@ -11658,12 +11591,14 @@ GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
     // not strictly necessary, but it may reduce the overhead of the first graph_compute
     ggml_cuda_set_main_device(device);
 
-    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context {
-        /* .device = */ device,
-        /* .name   = */ GGML_CUDA_NAME + std::to_string(device),
-    };
+    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
+    if (ctx == nullptr) {
+        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        return nullptr;
+    }
 
     ggml_backend_t cuda_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cuda_guid(),
         /* .interface = */ ggml_backend_cuda_interface,
         /* .context   = */ ctx
     };
@@ -11672,7 +11607,7 @@ GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
 }
 
 GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) {
-    return backend && backend->iface.get_name == ggml_backend_cuda_name;
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
 }
 
 GGML_CALL int ggml_backend_cuda_get_device_count() {