diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 2456f7dcc621..7423bd48fd8f 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -5259,6 +5259,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SUM:
             return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_TOP_K:
+#if defined(GGML_CUDA_USE_CUB) || defined(GGML_HIP_TOPK_HIPCUB)
+            return true;
+#else
+            return op->src[0]->ne[0] <= 1024;
+#endif
         case GGML_OP_ARGSORT:
 #ifndef GGML_CUDA_USE_CUB
             return op->src[0]->ne[0] <= 1024;
diff --git a/ggml/src/ggml-cuda/top-k.cu b/ggml/src/ggml-cuda/top-k.cu
index 9681cd293338..ef8385035b3b 100644
--- a/ggml/src/ggml-cuda/top-k.cu
+++ b/ggml/src/ggml-cuda/top-k.cu
@@ -1,6 +1,10 @@
 #include "argsort.cuh"
 #include "top-k.cuh"
 
+#ifdef GGML_HIP_TOPK_HIPCUB
+#    include <hipcub/hipcub.hpp>
+#endif  // GGML_HIP_TOPK_HIPCUB
+
 #ifdef GGML_CUDA_USE_CUB
 #    include <cub/cub.cuh>
 #    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2)
@@ -48,6 +52,89 @@ static int next_power_of_2(int x) {
 
 #endif                            // CUB_TOP_K_AVAILABLE
 
+#ifdef GGML_HIP_TOPK_HIPCUB
+
+// The HIP backend has no CUB, so the shared-memory bitonic argsort is the only
+// top-k path and it caps rows at 1024 elements.  Vocabulary-width top-k (e.g.
+// the DFlash2 candidate selector, k = 16 over 248,320 logits) therefore falls
+// back to the CPU, which costs a device-to-host copy of the whole logits
+// tensor on every draft step.  hipCUB's segmented radix sort keeps it on the
+// GPU.
+
+static __global__ void top_k_hipcub_init_indices(int * indices, const int ncols, const int nrows) {
+    const int col = blockIdx.x * blockDim.x + threadIdx.x;
+    const int row = blockIdx.y;
+
+    if (col < ncols && row < nrows) {
+        indices[row * ncols + col] = col;
+    }
+}
+
+static __global__ void top_k_hipcub_init_offsets(int * offsets, const int ncols, const int nrows_offset) {
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (i < nrows_offset) {
+        offsets[i] = i * ncols;
+    }
+}
+
+static void top_k_hipcub(ggml_cuda_pool & pool,
+                         const float *    src,
+                         int *            dst,
+                         const int        ncols,
+                         const int        nrows,
+                         const int        k,
+                         cudaStream_t     stream) {
+    const size_t n = (size_t) ncols * nrows;
+
+    ggml_cuda_pool_alloc<float> keys_in_alloc (pool, n);
+    ggml_cuda_pool_alloc<float> keys_out_alloc(pool, n);
+    ggml_cuda_pool_alloc<int>   vals_in_alloc (pool, n);
+    ggml_cuda_pool_alloc<int>   vals_out_alloc(pool, n);
+
+    float * keys_in  = keys_in_alloc .get();
+    float * keys_out = keys_out_alloc.get();
+    int *   vals_in  = vals_in_alloc .get();
+    int *   vals_out = vals_out_alloc.get();
+
+    static const int block_size = 256;
+
+    const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
+    top_k_hipcub_init_indices<<<grid_size, block_size, 0, stream>>>(vals_in, ncols, nrows);
+
+    const int                 nrows_offset = nrows + 1;
+    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows_offset);
+    int *                     offsets = offsets_alloc.get();
+    const dim3                offset_grid((nrows_offset + block_size - 1) / block_size);
+    top_k_hipcub_init_offsets<<<offset_grid, block_size, 0, stream>>>(offsets, ncols, nrows_offset);
+
+    CUDA_CHECK(cudaMemcpyAsync(keys_in, src, n * sizeof(float), cudaMemcpyDeviceToDevice, stream));
+
+    size_t temp_storage_bytes = 0;
+    CUDA_CHECK(hipcub::DeviceSegmentedRadixSort::SortPairsDescending(
+        nullptr, temp_storage_bytes, keys_in, keys_out, vals_in, vals_out, n, nrows, offsets, offsets + 1, 0,
+        sizeof(float) * 8, stream));
+
+    ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
+    CUDA_CHECK(hipcub::DeviceSegmentedRadixSort::SortPairsDescending(
+        temp_storage_alloc.get(), temp_storage_bytes, keys_in, keys_out, vals_in, vals_out, n, nrows, offsets,
+        offsets + 1, 0, sizeof(float) * 8, stream));
+
+    // the op only requires the k largest indices, in any order
+    CUDA_CHECK(cudaMemcpy2DAsync(dst, k * sizeof(int), vals_out, ncols * sizeof(int), k * sizeof(int), nrows,
+                                 cudaMemcpyDeviceToDevice, stream));
+}
+
+// process at most this many bytes of keys+values per segmented sort call
+static int top_k_hipcub_chunk_nrows(const int ncols, const int nrows) {
+    const size_t chunk_bytes = 1 << 26;  // 64 MiB
+    const size_t row_bytes   = (size_t) ncols * (sizeof(float) + sizeof(int)) * 2;
+
+    return (int) std::min((int64_t) std::max<size_t>(chunk_bytes / std::max<size_t>(row_bytes, 1), 1), (int64_t) nrows);
+}
+
+#endif  // GGML_HIP_TOPK_HIPCUB
+
 void ggml_cuda_op_top_k(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0   = dst->src[0];
     const float *       src0_d = (const float *) src0->data;
@@ -96,6 +183,22 @@ void ggml_cuda_op_top_k(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         dst_d  += k     * iter_nrows;
     }
 #else                             // GGML_CUDA_USE_CUB
+#    ifdef GGML_HIP_TOPK_HIPCUB
+    // the bitonic argsort below needs the whole row in shared memory
+    if (ncols > 1024) {
+        const int chunk_nrows = top_k_hipcub_chunk_nrows(ncols, nrows);
+
+        for (int64_t i = 0; i < nrows; i += chunk_nrows) {
+            const int iter_nrows = std::min((int64_t) chunk_nrows, nrows - i);
+
+            top_k_hipcub(pool, src0_d, dst_d, ncols, iter_nrows, k, stream);
+
+            src0_d += (size_t) ncols * iter_nrows;
+            dst_d  += (size_t) k     * iter_nrows;
+        }
+        return;
+    }
+#    endif  // GGML_HIP_TOPK_HIPCUB
     ggml_cuda_pool_alloc<int> temp_dst_alloc(pool, ncols * nrows);
     int *                     tmp_dst = temp_dst_alloc.get();
     argsort_f32_i32_cuda_bitonic(src0_d, tmp_dst, ncols, nrows, GGML_SORT_ORDER_DESC, stream);
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index 47f16f56c470..c2b1ce276123 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -47,6 +47,14 @@ find_package(hip     REQUIRED)
 find_package(hipblas REQUIRED)
 find_package(rocblas REQUIRED)
 
+# hipCUB gives the top-k op a device-side path for rows wider than the shared
+# memory bitonic argsort can handle (1024).  Without it, vocabulary-width top-k
+# is not supported by this backend and falls back to the CPU.
+option(GGML_HIP_TOPK_HIPCUB "ggml: use hipCUB for wide-row top-k" ON)
+if (GGML_HIP_TOPK_HIPCUB)
+    find_package(hipcub)
+endif()
+
 if (GGML_HIP_RCCL)
     find_package(rccl REQUIRED)
 endif()
@@ -151,3 +159,9 @@ if (GGML_HIP_RCCL)
 endif()
 
 target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)
+
+if (GGML_HIP_TOPK_HIPCUB AND hipcub_FOUND)
+    message(STATUS "HIP: using hipCUB for wide-row top-k")
+    target_compile_definitions(ggml-hip PRIVATE GGML_HIP_TOPK_HIPCUB)
+    target_link_libraries(ggml-hip PRIVATE hip::hipcub)
+endif()