diff --git a/TopK.mlu b/TopK.mlu
new file mode 100644
index 0000000..0b57122
--- /dev/null
+++ b/TopK.mlu
@@ -0,0 +1,77 @@
+#include <bang.h>
+#include <torch/extension.h>
+#include <cnrt.h>
+#include <vector>
+
+#define BLOCK_SIZE 1024
+#define K_MAX 128
+#define TASKS 16
+
+__mlu_entry__ void topk_dim1_kernel(const half *x, half *values, int64_t *indices,
+                                    int batch, int cols, int k) {
+    int row = taskId;
+    if (row >= batch) {
+        return;
+    }
+
+    __nram__ half row_buf[BLOCK_SIZE];
+    __nram__ half top_vals[K_MAX];
+    __nram__ int top_idx[K_MAX];
+
+    const half *row_ptr = x + row * cols;
+    __memcpy(row_buf, row_ptr, cols * sizeof(half), GDRAM2NRAM);
+
+    for (int i = 0; i < k; ++i) {
+        half best = row_buf[0];
+        int best_idx = 0;
+
+        for (int j = 1; j < cols; ++j) {
+            half cur = row_buf[j];
+            if (cur > best || (cur == best && j < best_idx)) {
+                best = cur;
+                best_idx = j;
+            }
+        }
+
+        top_vals[i] = best;
+        top_idx[i] = best_idx;
+        row_buf[best_idx] = (half)-65504.0f;
+    }
+
+    __memcpy(values + row * k, top_vals, k * sizeof(half), NRAM2GDRAM);
+
+    int64_t *idx_out = indices + row * k;
+    for (int i = 0; i < k; ++i) {
+        idx_out[i] = (int64_t)top_idx[i];
+    }
+}
+
+std::vector<torch::Tensor> bang_func(torch::Tensor x, int k, int dim) {
+    TORCH_CHECK(x.is_contiguous(), "x must be contiguous");
+    TORCH_CHECK(x.scalar_type() == torch::kFloat16, "x must be float16");
+    TORCH_CHECK(x.dim() == 2, "x must be 2D");
+    TORCH_CHECK(dim == 1 || dim == -1, "only dim=1 is supported");
+    TORCH_CHECK(k > 0 && k <= K_MAX, "k must be in (0, K_MAX]");
+
+    int batch = x.size(0);
+    int cols = x.size(1);
+    TORCH_CHECK(cols <= BLOCK_SIZE, "cols must be <= BLOCK_SIZE");
+    TORCH_CHECK(k <= cols, "k must be <= cols");
+
+    auto values = torch::empty({batch, k}, x.options());
+    auto indices = torch::empty({batch, k}, x.options().dtype(torch::kInt64));
+
+    cnrtQueue_t queue = torch_mlu::getCurMLUStream();
+    cnrtDim3_t task_dim = {batch, 1, 1};
+    cnrtFunctionType_t ktype = cnrtFuncTypeUnion1;
+
+    topk_dim1_kernel<<<task_dim, ktype, queue>>>(
+        reinterpret_cast<const half *>(x.data_ptr<at::Half>()),
+        reinterpret_cast<half *>(values.data_ptr<at::Half>()),
+        reinterpret_cast<int64_t *>(indices.data_ptr<int64_t>()),
+        batch,
+        cols,
+        k);
+
+    return {values, indices};
+}
diff --git a/config b/config
index 49a5bb2..920a6ea 100644
--- a/config
+++ b/config
@@ -1 +1 @@
-070
+075