whispercpp 1.3.6 → 1.3.7

data/ext/sources/ggml/src/ggml-cuda/concat.cu

@@ -1,102 +1,88 @@
 #include "concat.cuh"
 
-// contiguous kernels
-static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
+#include <stdint.h>
 
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (nidx < ne00) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            (nidx - ne00) +
-            blockIdx.y * (ne0 - ne00) +
-            blockIdx.z * (ne0 - ne00) * gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
+// contiguous kernels
+template <typename T, int dim>
+static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_cont(const T * x,
+                                                                             const T * y,
+                                                                             T *       dst,
+                                                                             int64_t   ne00,
+                                                                             int64_t   ne01,
+                                                                             int64_t   ne02,
+                                                                             int64_t   ne0,
+                                                                             int64_t   ne1,
+                                                                             int64_t   ne2) {
+    static_assert(dim >= 0 && dim <= 2, "dim must be in [0, 2]");
+
+    const int64_t n = ne0 * ne1 * ne2;
+
+    ggml_cuda_pdl_sync();
+    for (int64_t i = (int64_t) blockIdx.x * blockDim.x + threadIdx.x; i < n; i += (int64_t) blockDim.x * gridDim.x) {
+        if constexpr (dim == 0) {
+            const int64_t row = i / ne0;
+            const int64_t i0  = i - row * ne0;
+
+            if (i0 < ne00) {
+                dst[i] = x[row * ne00 + i0];
+            } else {
+                dst[i] = y[row * (ne0 - ne00) + (i0 - ne00)];
+            }
+        } else if constexpr (dim == 1) {
+            const int64_t dst_plane  = ne0 * ne1;
+            const int64_t src0_plane = ne0 * ne01;
+            const int64_t src1_plane = dst_plane - src0_plane;
+            const int64_t i2         = i / dst_plane;
+            const int64_t i01        = i - i2 * dst_plane;
+
+            if (i01 < src0_plane) {
+                dst[i] = x[i2 * src0_plane + i01];
+            } else {
+                dst[i] = y[i2 * src1_plane + (i01 - src0_plane)];
+            }
+        } else {
+            const int64_t src0_size = ne0 * ne1 * ne02;
 
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.y < (unsigned)ne01) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * ne01;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            (blockIdx.y - ne01) * ne0 +
-            blockIdx.z * ne0 * (gridDim.y - ne01);
-        dst[offset_dst] = y[offset_src];
+            if (i < src0_size) {
+                dst[i] = x[i];
+            } else {
+                dst[i] = y[i - src0_size];
+            }
+        }
     }
 }
 
-static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.z < (unsigned)ne02) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            (blockIdx.z - ne02) * ne0 *  gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
+template <typename T>
+static void concat_cont_cuda(const T * x,
+                             const T * y,
+                             T *       dst,
+                             int64_t   ne00,
+                             int64_t   ne01,
+                             int64_t   ne02,
+                             int64_t   ne0,
+                             int64_t   ne1,
+                             int64_t   ne2,
+                             int       dim,
+                             cudaStream_t stream) {
+    const int64_t n          = ne0 * ne1 * ne2;
+    const int     num_blocks = (n + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
 
-static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
     if (dim == 0) {
-        concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(concat_cont<T, 0>, launch_params, x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
     if (dim == 1) {
-        concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
+        concat_cont<T, 1><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
-    concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+    concat_cont<T, 2><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
 }
 
 // non-contiguous kernel (slow)
-template <int dim>
+template <typename T, int dim>
 static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
-    concat_f32_non_cont(
+    concat_non_cont(
         const char * src0,
         const char * src1,
               char * dst,
@@ -123,61 +109,49 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
           uint64_t   nb0,
           uint64_t   nb1,
           uint64_t   nb2,
-          uint64_t   nb3){
+          uint64_t   nb3) {
     static_assert(dim >= 0 && dim <= 3, "dim must be in [0, 3]");
 
     const int64_t i3 = blockIdx.z;
     const int64_t i2 = blockIdx.y;
     const int64_t i1 = blockIdx.x;
 
-    const float * x;
+    const T * x;
 
     for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
         if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+            x = (const T *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
         } else {
             if constexpr (dim == 0) {
-                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + i1 * nb11 + (i0 - ne00) * nb10);
+                x = (const T *)(src1 + i3*nb13 + i2*nb12 + i1*nb11 + (i0 - ne00)*nb10);
             } else if constexpr (dim == 1) {
-                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + (i1 - ne01) * nb11 + i0 * nb10);
+                x = (const T *)(src1 + i3*nb13 + i2*nb12 + (i1 - ne01)*nb11 + i0*nb10);
             } else if constexpr (dim == 2) {
-                x = (const float *) (src1 + i3 * nb13 + (i2 - ne02) * nb12 + i1 * nb11 + i0 * nb10);
+                x = (const T *)(src1 + i3*nb13 + (i2 - ne02)*nb12 + i1*nb11 + i0*nb10);
             } else if constexpr (dim == 3) {
-                x = (const float *) (src1 + (i3 - ne03) * nb13 + i2 * nb12 + i1 * nb11 + i0 * nb10);
+                x = (const T *)(src1 + (i3 - ne03)*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
             }
         }
 
-        float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+        T * y = (T *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
         *y = *x;
     }
 }
 
-
-void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    cudaStream_t stream = ctx.stream();
-
-    const int32_t dim = ((int32_t *) dst->op_params)[0];
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
-
+template <typename T>
+static void concat_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, int dim, cudaStream_t stream) {
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-        const float * src0_d = (const float *)src0->data;
-        const float * src1_d = (const float *)src1->data;
-
-        float * dst_d = (float *)dst->data;
+        const T * src0_d = (const T *) src0->data;
+        const T * src1_d = (const T *) src1->data;
+        T *       dst_d  = (T *) dst->data;
 
         if (dim != 3) {
-            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
-                concat_f32_cuda(
-                        src0_d + i3 * (src0->nb[3] / 4),
-                        src1_d + i3 * (src1->nb[3] / 4),
-                        dst_d + i3 * ( dst->nb[3] / 4),
+            for (int64_t i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_cont_cuda(
+                        src0_d + i3*(src0->nb[3] / sizeof(T)),
+                        src1_d + i3*(src1->nb[3] / sizeof(T)),
+                        dst_d  + i3*( dst->nb[3] / sizeof(T)),
                         src0->ne[0], src0->ne[1], src0->ne[2],
                         dst->ne[0],  dst->ne[1],  dst->ne[2], dim, stream);
             }
@@ -185,13 +159,13 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             const size_t size0 = ggml_nbytes(src0);
             const size_t size1 = ggml_nbytes(src1);
 
-            CUDA_CHECK(cudaMemcpyAsync(dst_d,           src0_d, size0, cudaMemcpyDeviceToDevice, stream));
-            CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
+            CUDA_CHECK(cudaMemcpyAsync((char *) dst->data,         src0->data, size0, cudaMemcpyDeviceToDevice, stream));
+            CUDA_CHECK(cudaMemcpyAsync((char *) dst->data + size0, src1->data, size1, cudaMemcpyDeviceToDevice, stream));
         }
     } else {
         dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
         auto launch_kernel = [&](auto dim) {
-            concat_f32_non_cont<dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+            concat_non_cont<T, dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
                 (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
                 src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
                 src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
@@ -219,3 +193,35 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         }
     }
 }
+
+void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    cudaStream_t stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(dst->type  == src0->type);
+    GGML_ASSERT(!ggml_is_quantized(src0->type));
+    GGML_ASSERT(ggml_blck_size(src0->type) == 1);
+
+    switch (ggml_type_size(src0->type)) {
+        case 1:
+            concat_cuda<uint8_t>(src0, src1, dst, dim, stream);
+            break;
+        case 2:
+            concat_cuda<uint16_t>(src0, src1, dst, dim, stream);
+            break;
+        case 4:
+            concat_cuda<uint32_t>(src0, src1, dst, dim, stream);
+            break;
+        case 8:
+            concat_cuda<uint64_t>(src0, src1, dst, dim, stream);
+            break;
+        default:
+            GGML_ABORT("Unsupported type size: %zu", ggml_type_size(src0->type));
+            break;
+    }
+}

data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu

@@ -1,12 +1,20 @@
-#include <algorithm>
-
 #include "conv2d-transpose.cuh"
-#include "ggml.h"
-
-__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel,
-                                        float * __restrict__ output, const int in_w, const int in_h, const int out_w,
-                                        const int out_h, const int kernel_w, const int kernel_h, const int stride,
-                                        const int c_in, const int c_out, const int batches) {
+#include "convert.cuh"
+
+template <typename kernel_t>
+static __global__ void conv2d_transpose_kernel(const float * __restrict__ input,
+                                               const kernel_t * __restrict__ kernel,
+                                               float * __restrict__ output,
+                                               const int in_w,
+                                               const int in_h,
+                                               const int out_w,
+                                               const int out_h,
+                                               const int kernel_w,
+                                               const int kernel_h,
+                                               const int stride,
+                                               const int c_in,
+                                               const int c_out,
+                                               const int batches) {
     const int global_idx = blockIdx.x * blockDim.x + threadIdx.x;
 
     const int total_elements = out_w * out_h * c_out * batches;
@@ -26,24 +34,32 @@ __global__ void conv2d_transpose_kernel(const float * __restrict__ input, const
     for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) {
         for (int kh = 0; kh < kernel_h; ++kh) {
             int in_y = out_y_idx - kh;
-            if (in_y < 0 || in_y % stride) continue;
+            if (in_y < 0 || in_y % stride) {
+                continue;
+            }
             in_y /= stride;
-            if (in_y >= in_h) continue;
+            if (in_y >= in_h) {
+                continue;
+            }
 
             for (int kw = 0; kw < kernel_w; ++kw) {
                 int in_x = out_x_idx - kw;
-                if (in_x < 0 || in_x % stride) continue;
+                if (in_x < 0 || in_x % stride) {
+                    continue;
+                }
                 in_x /= stride;
-                if (in_x >= in_w) continue;
+                if (in_x >= in_w) {
+                    continue;
+                }
 
                 const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x;
                 const int kernel_idx =
                     (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw;
 
-                float input_val = input[input_idx];
-                half  kern_val  = kernel[kernel_idx];
+                float    input_val = input[input_idx];
+                kernel_t kern_val  = kernel[kernel_idx];
 
-                accumulator += input_val * (float) kern_val;
+                accumulator += input_val * ggml_cuda_cast<float>(kern_val);
             }
         }
     }
@@ -56,11 +72,12 @@ void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor
     const ggml_tensor * kernel = dst->src[0];
     const ggml_tensor * input  = dst->src[1];
 
-    GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
 
     const float * input_data  = (const float *) input->data;
     float *       output_data = (float *) dst->data;
-    const half * kernel_data = (const half *) kernel->data;
+    const void *  kernel_data = kernel->data;
 
     const int input_w      = input->ne[0];
     const int input_h      = input->ne[1];
@@ -82,10 +99,17 @@ void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor
     GGML_ASSERT(ggml_is_contiguous(kernel));
     GGML_ASSERT(ggml_is_contiguous(dst));
 
-    const int total  = (output_w * output_h * channels_out * batches);
+    const int total  = output_w * output_h * channels_out * batches;
     const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE;
 
-    conv2d_transpose_kernel<<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
-        input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride,
-        channels_in, channels_out, batches);
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_transpose_kernel<half><<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+            input_data, (const half *) kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w,
+            kernel_h, stride, channels_in, channels_out, batches);
+
+    } else {
+        conv2d_transpose_kernel<float><<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+            input_data, (const float *) kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w,
+            kernel_h, stride, channels_in, channels_out, batches);
+    }
 }

data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh

@@ -1,4 +1,5 @@
 #include "common.cuh"
 
 #define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+
 void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

data/ext/sources/ggml/src/ggml-cuda/convert.cu

@@ -617,6 +617,45 @@ static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t
     dequantize_block_mxfp4<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template <typename dst_t>
+static __global__ void dequantize_block_nvfp4(
+        const void * __restrict__ vx,
+        dst_t * __restrict__ yy,
+        const int64_t ne) {
+    const int64_t i = blockIdx.x;
+    const int     tid = threadIdx.x;
+
+    const int64_t base = i * QK_NVFP4;
+    if (base >= ne) {
+        return;
+    }
+
+    const block_nvfp4 * x = (const block_nvfp4 *) vx;
+    const block_nvfp4 & xb = x[i];
+
+    const int sub = tid / (QK_NVFP4_SUB / 2);
+    const int j = tid % (QK_NVFP4_SUB / 2);
+
+    const float d = ggml_cuda_ue4m3_to_fp32(xb.d[sub]);
+    const uint8_t q = xb.qs[sub * (QK_NVFP4_SUB / 2) + j];
+
+    const int64_t y0 = base + sub * QK_NVFP4_SUB + j;
+    const int64_t y1 = y0 + QK_NVFP4_SUB / 2;
+
+    yy[y0] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q & 0x0F]);
+    yy[y1] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q >> 4]);
+}
+
+template <typename dst_t>
+static void dequantize_row_nvfp4_cuda(
+        const void * vx,
+        dst_t * y,
+        const int64_t k,
+        cudaStream_t stream) {
+    GGML_ASSERT(k % QK_NVFP4 == 0);
+    const int nb = k / QK_NVFP4;
+    dequantize_block_nvfp4<<<nb, 32, 0, stream>>>(vx, y, k);
+}
 template <typename src_t, typename dst_t>
 static __global__ void convert_unary(
         const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
@@ -672,6 +711,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
 
 to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -715,6 +756,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_MXFP4:
             return dequantize_row_mxfp4_cuda;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
@@ -726,6 +769,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
 
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -766,6 +811,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_MXFP4:
             return dequantize_row_mxfp4_cuda;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cont_cuda<half>;
         case GGML_TYPE_BF16:
@@ -779,6 +826,8 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -800,6 +849,8 @@ to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -821,6 +872,8 @@ to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F16:
             return convert_unary_cuda<half, float>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:

data/ext/sources/ggml/src/ggml-cuda/convert.cuh

@@ -41,6 +41,16 @@ template<typename dst_t, typename src_t>
         return __bfloat162float(x);
     } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, half2>) {
         return __float22half2_rn(x);
+    } else if constexpr(std::is_same_v<src_t, nv_bfloat162> && std::is_same_v<dst_t, float2>) {
+#ifdef GGML_USE_HIP
+        return make_float2(__bfloat162float(__low2bfloat16(x)), __bfloat162float(__high2bfloat16(x)));
+#else
+#if __CUDA_ARCH__ >= 800
+        return __bfloat1622float2(x);
+#else
+        return make_float2(__bfloat162float(x.x), __bfloat162float(x.y));
+#endif // __CUDA_ARCH__ >= 800
+#endif // GGML_USE_HIP
     } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, nv_bfloat162>) {
         // bypass compile error on cuda 12.0.1
 #ifdef GGML_USE_HIP

data/ext/sources/ggml/src/ggml-cuda/cpy.cu

@@ -16,6 +16,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int64_t ne
                                   const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t nb00, const int64_t nb01, const int64_t nb02,
                                   const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11,
                                   const int64_t nb12, const int64_t nb13) {
+    ggml_cuda_pdl_lc();
     const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -36,6 +37,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int64_t ne
     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;
 
+    ggml_cuda_pdl_sync();
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
@@ -59,6 +61,7 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
     __shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
     int cur_tile_buf = 0;
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
 
@@ -142,6 +145,7 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int64_t ne,
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
@@ -168,6 +172,7 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int64_t ne,
     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;
 
+    ggml_cuda_pdl_sync();
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
@@ -182,6 +187,7 @@ static __global__ void cpy_scalar_contiguous(const char * cx, char * cdst, const
     const src_t * x = (const src_t *) cx;
     dst_t *     dst = (dst_t *) cdst;
 
+    ggml_cuda_pdl_sync();
     dst[i] = ggml_cuda_cast<dst_t>(x[i]);
 }
 
@@ -192,8 +198,8 @@ cudaStream_t stream) {
 
     const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     GGML_ASSERT(num_blocks < UINT_MAX);
-    cpy_scalar_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(cpy_scalar_contiguous<src_t, dst_t>, launch_params, cx, cdst, ne);
 }
 
 template<typename src_t, typename dst_t, bool transposed = false>
@@ -223,13 +229,15 @@ static void ggml_cpy_scalar_cuda(
         GGML_ASSERT(grid_z < USHRT_MAX);
         dim3 dimGrid(grid_x, grid_y, grid_z);
         dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
-        cpy_scalar_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
-            (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
+            cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     } else {
         const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
         GGML_ASSERT(num_blocks < UINT_MAX);
-        cpy_scalar<cpy_1_scalar<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-            (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
+            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     }
 }
 

data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh

@@ -1,5 +1,27 @@
 #include "common.cuh"
 
+static __device__ __forceinline__ void dequantize_q1_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
+    const block_q1_0 * x = (const block_q1_0 *) vx;
+
+    const float d = x[ib].d;
+
+    const int bit_index_0 = iqs;
+    const int bit_index_1 = iqs + 1;
+
+    const int byte_index_0 = bit_index_0 / 8;
+    const int bit_offset_0 = bit_index_0 % 8;
+
+    const int byte_index_1 = bit_index_1 / 8;
+    const int bit_offset_1 = bit_index_1 % 8;
+
+    // Extract bits: 1 = +d, 0 = -d (branchless)
+    const int bit_0 = (x[ib].qs[byte_index_0] >> bit_offset_0) & 1;
+    const int bit_1 = (x[ib].qs[byte_index_1] >> bit_offset_1) & 1;
+
+    v.x = (2*bit_0 - 1) * d;
+    v.y = (2*bit_1 - 1) * d;
+}
+
 static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;