whispercpp 1.3.3 → 1.3.5

data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu

@@ -34,10 +34,7 @@ static  __global__ void conv_transpose_1d_kernel(
         }
     }
     dst[global_index] = accumulator;
-    GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3);
-    GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3);
-    GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1);
-    GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2);
+    GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2);
 }
 
 static void conv_transpose_1d_f32_f32_cuda(

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

@@ -0,0 +1,166 @@
+#include "conv2d.cuh"
+#include "convert.cuh"
+
+struct conv_params {
+    const int64_t IW, IH;
+    const int64_t OW, OH;
+    const int64_t KW, KH;
+    const int64_t ST_X, ST_Y;
+    const int64_t PD_X, PD_Y;
+    const int64_t DL_X, DL_Y;
+    const int64_t IC, OC;
+    const int64_t B;
+    const int64_t TOTAL;
+};
+
+struct kernel_bounds {
+    int64_t y_min, y_max;
+    int64_t x_min, x_max;
+};
+
+__device__ __forceinline__ int64_t max64(int64_t a, int64_t b) {
+    return (a > b) ? a : b;
+}
+
+__device__ __forceinline__ int64_t min64(int64_t a, int64_t b) {
+    return (a < b) ? a : b;
+}
+
+__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int64_t out_x, int64_t out_y, const conv_params & P) {
+    kernel_bounds bounds;
+    bounds.y_min = max64(0, (P.PD_Y - out_y * P.ST_Y + P.DL_Y - 1) / P.DL_Y);
+    bounds.y_max = min64(P.KH, (P.IH + P.PD_Y - out_y * P.ST_Y + P.DL_Y - 1) / P.DL_Y);
+    bounds.x_min = max64(0, (P.PD_X - out_x * P.ST_X + P.DL_X - 1) / P.DL_X);
+    bounds.x_max = min64(P.KW, (P.IW + P.PD_X - out_x * P.ST_X + P.DL_X - 1) / P.DL_X);
+    return bounds;
+}
+
+__device__ __forceinline__ int calculate_input_coord(int64_t out_coord,
+                                                     int64_t kern_coord,
+                                                     int64_t stride,
+                                                     int64_t dilation,
+                                                     int64_t padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+struct whcn_layout {
+    __device__ static int64_t input_index(int64_t n, int64_t c, int64_t y, int64_t x, const conv_params & P) {
+        return n * (P.IC * P.IW * P.IH) + c * P.IW * P.IH + y * P.IW + x;
+    }
+
+    __device__ static int64_t kernel_index(int64_t c_out, int64_t c_in, int64_t ky, int64_t kx, const conv_params & P) {
+        return c_out * (P.IC * P.KH * P.KW) + c_in * (P.KH * P.KW) + ky * P.KW + kx;
+    }
+
+    __device__ static int64_t output_index(int64_t n, int64_t c, int64_t y, int64_t x, const conv_params & P) {
+        return n * (P.OC * P.OW * P.OH) + c * P.OW * P.OH + y * P.OW + x;
+    }
+
+    __device__ static void unpack_indices(int64_t             global_idx,
+                                          const conv_params & P,
+                                          int64_t &           n,
+                                          int64_t &           c,
+                                          int64_t &           out_y,
+                                          int64_t &           out_x) {
+        out_x = global_idx % P.OW;
+        out_y = (global_idx / P.OW) % P.OH;
+        c     = (global_idx / (P.OW * P.OH)) % P.OC;
+        n     = global_idx / (P.OW * P.OH * P.OC);
+    }
+};
+
+template <typename T, typename Layout>
+static __global__ void conv2d_kernel(const float * __restrict__ input,
+                                     const T * __restrict__ kernel,
+                                     float * __restrict__ output,
+                                     const conv_params P) {
+    const int64_t global_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (global_idx >= P.TOTAL) {
+        return;
+    }
+
+    int64_t n, c_out, out_y, out_x;
+    Layout::unpack_indices(global_idx, P, n, c_out, out_y, out_x);
+
+    float acc = 0.0f;
+
+    for (int64_t c_in = 0; c_in < P.IC; ++c_in) {
+        kernel_bounds bounds = calculate_kernel_bounds(out_x, out_y, P);
+
+        for (int64_t ky = bounds.y_min; ky < bounds.y_max; ++ky) {
+            const int64_t in_y = calculate_input_coord(out_y, ky, P.ST_Y, P.DL_Y, P.PD_Y);
+
+            for (int64_t kx = bounds.x_min; kx < bounds.x_max; ++kx) {
+                const int64_t in_x = calculate_input_coord(out_x, kx, P.ST_X, P.DL_X, P.PD_X);
+
+                const float input_val = input[Layout::input_index(n, c_in, in_y, in_x, P)];
+                const T kernel_val = kernel[Layout::kernel_index(c_out, c_in, ky, kx, P)];
+                acc += (input_val * ggml_cuda_cast<float>(kernel_val));
+            }
+        }
+    }
+
+    // [N, OC, OH, OW]
+    output[Layout::output_index(n, c_out, out_y, out_x, P)] = acc;
+}
+
+template <typename T>
+static void conv2d_cuda(const float * X_D, const T * K_D, float * Y_D, const conv_params P, cudaStream_t st) {
+    const int blocks = (P.TOTAL + CUDA_CONV2D_BLOCK_SIZE - 1) / CUDA_CONV2D_BLOCK_SIZE;
+    conv2d_kernel<T, whcn_layout><<<blocks, CUDA_CONV2D_BLOCK_SIZE, 0, st>>>(X_D, K_D, Y_D, P);
+}
+
+static void conv2d_cuda_f16(const float * X_D, const half * K_D, float * Y_D, const conv_params P, cudaStream_t st) {
+    conv2d_cuda<half>(X_D, K_D, Y_D, P, st);
+}
+
+static void conv2d_cuda_f32(const float * X_D, const float * K_D, float * Y_D, const conv_params P, cudaStream_t st) {
+    conv2d_cuda<float>(X_D, K_D, Y_D, P, st);
+}
+
+void ggml_cuda_op_conv2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+    float *             K_D    = (float *) kernel->data;
+    const float *       X_D    = (const float *) input->data;
+    float *             Y_D    = (float *) dst->data;
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+
+    // same number of input channels
+    GGML_ASSERT(input->ne[2] == kernel->ne[2]);
+
+    cudaStream_t st = ctx.stream();
+
+    const int32_t * p    = (const int32_t *) dst->op_params;
+    const int       ST_X = p[0];  // stride_x
+    const int       ST_Y = p[1];  // stride_y
+    const int       PD_X = p[2];  // padding_x
+    const int       PD_Y = p[3];  // padding_y
+    const int       DL_X = p[4];  // dilation_x
+    const int       DL_Y = p[5];  // dilation_y
+
+    // No cwhn
+    GGML_ASSERT(p[6] == false);
+
+    const int IW = input->ne[0];   // input_w
+    const int IH = input->ne[1];   // input_h
+    const int OW = dst->ne[0];     // output_w
+    const int OH = dst->ne[1];     // output_h
+    const int KW = kernel->ne[0];  // kernel_w
+    const int KH = kernel->ne[1];  // kernel_h
+    const int IC = input->ne[2];   // input_channels
+    const int OC = kernel->ne[3];  // ouptut_chanles
+    const int B  = input->ne[3];   // n_batches
+
+    const int64_t total  = B * OC * OH * OW;
+    conv_params   params = { IW, IH, OW, OH, KW, KH, ST_X, ST_Y, PD_X, PD_Y, DL_X, DL_Y, IC, OC, B, total };
+
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_cuda_f16(X_D, (half *) K_D, Y_D, params, st);
+    } else {
+        conv2d_cuda_f32(X_D, K_D, Y_D, params, st);
+    }
+}

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

@@ -0,0 +1,5 @@
+#pragma once
+#include "common.cuh"
+
+#define CUDA_CONV2D_BLOCK_SIZE 256
+void ggml_cuda_op_conv2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -6,24 +6,33 @@
 #define CUDA_Q8_0_NE_ALIGN 2048
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x);
+static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    const int64_t i00 = 2 * (int64_t(blockDim.x)*blockIdx.x + threadIdx.x);
 
-    if (i >= k) {
+    if (i00 >= ne00) {
         return;
     }
 
-    const int64_t ib = i/qk; // block index
-    const int64_t iqs = (i%qk)/qr; // quant index
-    const int64_t iybs = i - i%qk; // y block start index
+    const int64_t i01 = blockIdx.y;
+    const int64_t i02 = blockIdx.z % ne02;
+    const int64_t i03 = blockIdx.z / ne02;
+
+    const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+
+    const int64_t ib = ibx0 + i00/qk; // block index
+    const int64_t iqs = (i00%qk)/qr; // quant index
+    const int64_t iybs = i00 - i00%qk; // y block start index
     const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    dfloat2 v;
+    float2 v;
     dequantize_kernel(vx, ib, iqs, v);
 
-    y[iybs + iqs + 0]        = v.x;
-    y[iybs + iqs + y_offset] = v.y;
+    const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
+    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template <bool need_check>
@@ -62,9 +71,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
         y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
     }
 #else
-    GGML_UNUSED(vx);
-    GGML_UNUSED(y);
-    GGML_UNUSED(k);
+    GGML_UNUSED_VARS(vx, y, k);
     NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 }
@@ -456,10 +463,36 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
     }
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_mxfp4(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_mxfp4 * x = (const block_mxfp4 *) vx + i*(QK_K/QK_MXFP4);
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = ggml_cuda_e8m0_to_fp32(x[ib].e);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_mxfp4[q4[j] & 0xf]*0.5f;
+        y[j+16] = d * kvalues_mxfp4[q4[j] >>  4]*0.5f;
+    }
+}
+
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
-    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+static void dequantize_block_cuda(const void * vx, dst_t * y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
+    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, ne02*ne03);
+    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
+        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_cont_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    dequantize_block_cuda<qk, qr, dequantize_kernel, dst_t>(vx, y, k, 1, 1, 1, k/qk, k/qk, k/qk, stream);
 }
 
 static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) {
@@ -571,6 +604,12 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_mxfp4<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 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, const int64_t ne02,
@@ -589,7 +628,7 @@ static __global__ void convert_unary(
 
     const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
     const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = float(x[ix]);
+    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
 }
 
 template <typename src_t, typename dst_t>
@@ -624,14 +663,14 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_Q4_1:
             return dequantize_row_q4_1_cuda;
         case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+            return dequantize_block_cont_cuda<QK5_0, QR5_0, dequantize_q5_0>;
         case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+            return dequantize_block_cont_cuda<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
             if (fp16_available(ggml_cuda_info().devices[ggml_cuda_get_device()].cc)) {
                 return dequantize_block_q8_0_f16_cuda;
             }
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+            return dequantize_block_cont_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_cuda;
         case GGML_TYPE_Q3_K:
@@ -660,6 +699,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq4_xs_cuda;
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
@@ -676,11 +717,11 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
         case GGML_TYPE_Q4_1:
             return dequantize_row_q4_1_cuda;
         case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+            return dequantize_block_cont_cuda<QK5_0, QR5_0, dequantize_q5_0>;
         case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+            return dequantize_block_cont_cuda<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+            return dequantize_block_cont_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_cuda;
         case GGML_TYPE_Q3_K:
@@ -709,6 +750,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq4_xs_cuda;
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cont_cuda<half>;
         case GGML_TYPE_BF16:
@@ -722,6 +765,16 @@ 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_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16>;
         default:
@@ -733,6 +786,16 @@ 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_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half, nv_bfloat16>;
         default:
@@ -744,6 +807,16 @@ 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_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16, float>;
         default:

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

@@ -1,3 +1,4 @@
+#pragma once
 #include "common.cuh"
 
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
@@ -29,3 +30,27 @@ typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
 to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
 to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);
+
+template<typename dst_t, typename src_t>
+ __host__ __device__ inline dst_t ggml_cuda_cast(src_t x) {
+    if constexpr (std::is_same_v<dst_t, src_t>) {
+        return x;
+    } else if constexpr(std::is_same_v<dst_t, nv_bfloat16>) {
+        return __float2bfloat16(float(x));
+    } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
+        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, float2> && std::is_same_v<dst_t, nv_bfloat162>) {
+        // bypass compile error on cuda 12.0.1
+#ifdef GGML_USE_HIP
+        return __float22bfloat162_rn(x);
+#else
+        return {x.x, x.y};
+#endif // GGML_USE_HIP
+    } else if constexpr(std::is_same_v<dst_t, int32_t>) {
+        return int32_t(x);
+    } else {
+        return float(x);
+    }
+}

data/ext/sources/ggml/src/ggml-cuda/cpy-utils.cuh

@@ -0,0 +1,217 @@
+#pragma once
+
+#include "ggml-common.h"
+#include "convert.cuh"
+
+static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static __device__ void quantize_f32_q4_0_block(const float * __restrict__ x, block_q4_0 * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = x[0       + j]*id;
+        const float x1 = x[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
+
+        y->qs[j]  = xi0;
+        y->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void quantize_f32_q4_1_block(const float * __restrict__ x, block_q4_1 * __restrict__ y) {
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = x[j];
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->dm.x = d;
+    y->dm.y = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (x[0       + j] - vmin)*id;
+        const float x1 = (x[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
+
+        y->qs[j]  = xi0;
+        y->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void quantize_f32_q5_0_block(const float * __restrict__ x, block_q5_0 * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK5_0; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = x[0       + j]*id;
+        const float x1 = x[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
+
+        y->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+    memcpy(y->qh, &qh, sizeof(qh));
+}
+
+static __device__ void quantize_f32_q5_1_block(const float * __restrict__ x, block_q5_1 * __restrict__ y) {
+    float min = x[0];
+    float max = x[0];
+
+    for (int j = 1; j < QK5_1; ++j) {
+        const float v = x[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->dm.x = d;
+    y->dm.y = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (x[0       + j] - min)*id;
+        const float x1 = (x[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        y->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+    memcpy(y->qh, &qh, sizeof(qh));
+}
+
+static __device__ void quantize_f32_q8_0_block(const float * __restrict__ x, block_q8_0 * __restrict__ y) {
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = x[j];
+        amax = fmaxf(amax, fabsf(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = x[j]*id;
+        y->qs[j] = roundf(x0);
+    }
+}
+
+static __device__ void quantize_f32_iq4_nl_block(const float * __restrict__ x, block_iq4_nl * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_NL; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    float d = vmax / kvalues_iq4nl[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = x[0        + j]*id;
+        const float x1 = x[QK4_NL/2 + j]*id;
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
+        y->qs[j] = xi0 | (xi1 << 4);
+        const float v0 = kvalues_iq4nl[xi0];
+        const float v1 = kvalues_iq4nl[xi1];
+        const float w0 = x[0        + j]*x[0        + j];
+        const float w1 = x[QK4_NL/2 + j]*x[QK4_NL/2 + j];
+        sumqx += w0*v0*x[j] + w1*v1*x[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+    }
+
+    y->d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+// Wrapper functions for cpy.cu compatibility
+static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    quantize_f32_q4_0_block((const float *)cxi, (block_q4_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    quantize_f32_q4_1_block((const float *)cxi, (block_q4_1 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
+    quantize_f32_q5_0_block((const float *)cxi, (block_q5_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
+    quantize_f32_q5_1_block((const float *)cxi, (block_q5_1 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    quantize_f32_q8_0_block((const float *)cxi, (block_q8_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
+    quantize_f32_iq4_nl_block((const float *)cxi, (block_iq4_nl *)cdsti);
+}
+
+template<typename src_t, typename dst_t>
+static __device__ void cpy_1_scalar(const char * cxi, char * cdsti) {
+    *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
+}