diff --git a/paddle/phi/kernels/cpu/lu_kernel.cc b/paddle/phi/kernels/cpu/lu_kernel.cc
index 9c968a83b111c..b014d9c98845a 100644
--- a/paddle/phi/kernels/cpu/lu_kernel.cc
+++ b/paddle/phi/kernels/cpu/lu_kernel.cc
@@ -34,6 +34,12 @@ void LUKernel(const Context& dev_ctx,
                         "lu without pivoting is not implemented on the CPU, "
                         "but got pivots=False"));
 
+  if (x.numel() == 0) {
+    dev_ctx.template Alloc<int>(infos);
+    dev_ctx.template Alloc<int>(pivots);
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   *out = Transpose2DTo6D<Context, T>(dev_ctx, x);
 
   auto outdims = out->dims();
diff --git a/paddle/phi/kernels/cpu/weight_quantize_kernel.cc b/paddle/phi/kernels/cpu/weight_quantize_kernel.cc
index bb2f840953fe8..dd9a4a763ee87 100644
--- a/paddle/phi/kernels/cpu/weight_quantize_kernel.cc
+++ b/paddle/phi/kernels/cpu/weight_quantize_kernel.cc
@@ -156,6 +156,14 @@ void WeightQuantizeKernel(const Context& dev_ctx,
                           DenseTensor* out,
                           DenseTensor* scale) {
   dev_ctx.template Alloc<int8_t>(out);
+  if (out->numel() == 0) {
+    if (algo == "llm.int8") {
+      dev_ctx.template Alloc<float>(scale);
+    } else {
+      dev_ctx.template Alloc<T>(scale);
+    }
+    return;
+  }
   if (algo == "weight_only_int8") {
     dev_ctx.template Alloc<T>(scale);
     quant_compute<Context, T, int8_t, 8>(
diff --git a/paddle/phi/kernels/gpu/lu_kernel.cu b/paddle/phi/kernels/gpu/lu_kernel.cu
index da5b64b505a25..b04ef2be2e791 100644
--- a/paddle/phi/kernels/gpu/lu_kernel.cu
+++ b/paddle/phi/kernels/gpu/lu_kernel.cu
@@ -133,6 +133,12 @@ void LUKernel(const Context& dev_ctx,
       2,
       ::common::errors::PreconditionNotMet(
           "Invalid input x dimensionality: %d (expected ≥2)", x.dims().size()));
+  if (x.numel() == 0) {
+    dev_ctx.template Alloc<int>(infos);
+    dev_ctx.template Alloc<int>(pivots);
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   int64_t largest_matrix = (1LL << 31) - 1;
   int64_t last = x.dims()[x.dims().size() - 1],
           second_last = x.dims()[x.dims().size() - 2];
diff --git a/paddle/phi/kernels/gpu/weight_only_linear_grad_kernel.cu b/paddle/phi/kernels/gpu/weight_only_linear_grad_kernel.cu
index 300f5d2fc1095..9147ee3f092e9 100644
--- a/paddle/phi/kernels/gpu/weight_only_linear_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/weight_only_linear_grad_kernel.cu
@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/common/datatype_traits.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/matmul_kernel.h"
 
 #if defined(PADDLE_WITH_CUTLASS)
@@ -51,6 +52,11 @@ void WeightOnlyLinearGradKernel(const Context& dev_ctx,
   int n = weight_scale.dims()[0];
   int k = weight.dims()[1];
   dev_ctx.template Alloc<T>(x_grad);
+  if (x_grad->numel() == 0 || out_grad.numel() == 0) {
+    Full<T, Context>(
+        dev_ctx, phi::IntArray(common::vectorize(x_grad->dims())), 0, x_grad);
+    return;
+  }
   DenseTensor weight_dequantized;
   weight_dequantized.Resize({{n, k}});
   dev_ctx.template Alloc<T>(&weight_dequantized);
diff --git a/paddle/phi/kernels/gpu/weight_only_linear_kernel.cu b/paddle/phi/kernels/gpu/weight_only_linear_kernel.cu
index 270b02a5ba449..f46b45a831790 100644
--- a/paddle/phi/kernels/gpu/weight_only_linear_kernel.cu
+++ b/paddle/phi/kernels/gpu/weight_only_linear_kernel.cu
@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/common/datatype_traits.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/weight_only_gemv.h"
 #if defined(PADDLE_WITH_CUTLASS)
 #include "paddle/phi/kernels/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h"
@@ -46,6 +47,11 @@ void WeightOnlyLinearKernel(const Context& dev_ctx,
 #endif
 
   dev_ctx.template Alloc<T>(out);
+  Full<T, Context>(
+      dev_ctx, phi::IntArray(common::vectorize(out->dims())), 0, out);
+  if (out->numel() == 0 || x.numel() == 0 || weight.numel() == 0) {
+    return;
+  }
   const T* x_data = x.data<T>();
   const int8_t* weight_data = weight.data<int8_t>();
   const T* bias_data = bias ? bias.get().data<T>() : nullptr;
diff --git a/paddle/phi/kernels/gpu/weight_quantize_kernel.cu b/paddle/phi/kernels/gpu/weight_quantize_kernel.cu
index 81f706a9c9441..c3ef20171d6a7 100644
--- a/paddle/phi/kernels/gpu/weight_quantize_kernel.cu
+++ b/paddle/phi/kernels/gpu/weight_quantize_kernel.cu
@@ -46,6 +46,14 @@ void WeightQuantizeKernel(const Context& dev_ctx,
 
   DenseTensor quanted_x;
   dev_ctx.template Alloc<int8_t>(out);
+  if (out->numel() == 0) {
+    if (algo == "llm.int8") {
+      dev_ctx.template Alloc<float>(scale);
+    } else {
+      dev_ctx.template Alloc<T>(scale);
+    }
+    return;
+  }
   quanted_x.Resize({m, n});
   dev_ctx.template Alloc<int8_t>(&quanted_x);
   std::vector<int64_t> weight_shape{m, n};
diff --git a/paddle/phi/kernels/impl/lu_grad_kernel_impl.h b/paddle/phi/kernels/impl/lu_grad_kernel_impl.h
index 71747addfcdbd..3fc553d365a12 100644
--- a/paddle/phi/kernels/impl/lu_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/lu_grad_kernel_impl.h
@@ -31,7 +31,7 @@ void LUGradKernel(const Context& dev_ctx,
                   bool pivot UNUSED,
                   DenseTensor* x_grad) {
   dev_ctx.template Alloc<T>(x_grad);
-
+  if (x_grad->numel() == 0) return;
   auto blas = phi::funcs::GetBlas<Context, T>(dev_ctx);
 
   auto xdims = x.dims();
diff --git a/paddle/phi/kernels/impl/strided_slice_grad_kernel_impl.h b/paddle/phi/kernels/impl/strided_slice_grad_kernel_impl.h
index 3824e301e2ec2..27a4d465eb74b 100644
--- a/paddle/phi/kernels/impl/strided_slice_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/strided_slice_grad_kernel_impl.h
@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #pragma once
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/strided_slice.h"
 #include "paddle/phi/kernels/strided_slice_grad_kernel.h"
 
diff --git a/paddle/phi/kernels/impl/strided_slice_kernel_impl.h b/paddle/phi/kernels/impl/strided_slice_kernel_impl.h
index f8dc298f47e60..87e6c3797e72e 100644
--- a/paddle/phi/kernels/impl/strided_slice_kernel_impl.h
+++ b/paddle/phi/kernels/impl/strided_slice_kernel_impl.h
@@ -28,6 +28,10 @@ void StridedSliceRawKernel(const Context& dev_ctx,
                            const std::vector<int>& infer_flags,
                            const std::vector<int>& decrease_axis,
                            DenseTensor* out) {
+  if (out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   int rank = x.dims().size();
 #define SLICE_CASE(Rank)                                        \
   case Rank:                                                    \
diff --git a/paddle/phi/kernels/stride/strided_slice_grad_kernel.cc b/paddle/phi/kernels/stride/strided_slice_grad_kernel.cc
index ee1c2cd2ea236..0b4738509037f 100644
--- a/paddle/phi/kernels/stride/strided_slice_grad_kernel.cc
+++ b/paddle/phi/kernels/stride/strided_slice_grad_kernel.cc
@@ -44,6 +44,7 @@ void StridedSliceRawGradStridedKernel(const Context& dev_ctx,
   PD_VISIT_ALL_TYPES(x_grad->dtype(), "StridedSliceRawGradStridedKernel", ([&] {
                        phi::StridedTensorFill<data_t>(*x_grad, 0, x_grad);
                      }));
+  if (x_grad->numel() == 0) return;
   DenseTensor tmp;
   tmp.set_layout(out_grad.layout());
   tmp.set_lod(out_grad.lod());
diff --git a/paddle/phi/kernels/xpu/weight_quantize_kernel.cc b/paddle/phi/kernels/xpu/weight_quantize_kernel.cc
index 2e9e779d8aaf4..e8bde6b932a8a 100644
--- a/paddle/phi/kernels/xpu/weight_quantize_kernel.cc
+++ b/paddle/phi/kernels/xpu/weight_quantize_kernel.cc
@@ -39,6 +39,10 @@ void WeightQuantizeKernel(const Context& dev_ctx,
   scale->Resize({static_cast<int64_t>(n)});
 
   dev_ctx.template Alloc<float>(scale);
+  if (out->numel() == 0) {
+    dev_ctx.template Alloc<int8_t>(out);
+    return;
+  }
 
   if (algo == "weight_only_int8") {
     out->Resize({static_cast<int64_t>(k), static_cast<int64_t>(n)});
diff --git a/test/legacy_test/test_lu_op.py b/test/legacy_test/test_lu_op.py
index 7f039fad59e90..af338a084d8fa 100644
--- a/test/legacy_test/test_lu_op.py
+++ b/test/legacy_test/test_lu_op.py
@@ -301,18 +301,37 @@ def run_lu_static(shape, dtype):
             run_lu_static(tensor_shape, dtype)
 
 
-class TestLUAPIError(unittest.TestCase):
-    def test_errors(self):
-        with paddle.base.dygraph.guard():
-            # The size of input in lu should not be 0.
-            def test_0_size():
-                array = np.array([], dtype=np.float32)
-                x = paddle.to_tensor(
-                    np.reshape(array, [0, 0, 0]), dtype='float32'
-                )
-                paddle.linalg.lu(x, get_infos=True)
-
-            self.assertRaises(ValueError, test_0_size)
+# class TestLUAPIError(unittest.TestCase):
+#     def test_errors(self):
+#         with paddle.base.dygraph.guard():
+#             # The size of input in lu should not be 0.
+#             def test_0_size():
+#                 array = np.array([], dtype=np.float32)
+#                 x = paddle.to_tensor(
+#                     np.reshape(array, [0, 0, 0]), dtype='float32'
+#                 )
+#                 paddle.linalg.lu(x, get_infos=True)
+
+#             self.assertRaises(ValueError, test_0_size)
+
+
+class TestLUAPIZeroSize(unittest.TestCase):
+    def init_test_case(self):
+        self.x_shape = [1, 0, 10]
+        self.dtype = "float32"
+
+    def _test_dygraph(self):
+        paddle.disable_static()
+        array = np.ones(self.x_shape).astype(self.dtype)
+        x = paddle.to_tensor(array, stop_gradient=False)
+        lu, p, info = paddle.linalg.lu(x, get_infos=True)
+        loss = lu.sum()
+        loss.backward()
+        self.assertEqual(x.grad.shape, x.shape)
+
+    def test_zero_size(self):
+        self.init_test_case()
+        self._test_dygraph()
 
 
 if __name__ == "__main__":
diff --git a/test/legacy_test/test_strided_slice_op.py b/test/legacy_test/test_strided_slice_op.py
index 98e7fee9fc455..1682f7d661414 100644
--- a/test/legacy_test/test_strided_slice_op.py
+++ b/test/legacy_test/test_strided_slice_op.py
@@ -327,6 +327,17 @@ def test_check_grad(self):
         pass
 
 
+# zero size tensor
+class TestStrideSliceOp22(TestStrideSliceOp):
+    def initTestCase(self):
+        self.input = np.random.rand(10, 0, 100)
+        self.axes = [1]
+        self.starts = [-101]
+        self.ends = [-101]
+        self.strides = [1]
+        self.infer_flags = [1]
+
+
 class TestStrideSliceOpBool(TestStrideSliceOp):
     def test_check_grad(self):
         pass
diff --git a/test/quantization/test_weight_only_linear.py b/test/quantization/test_weight_only_linear.py
index 164ad49ae7e0f..1c54d0d8bf34c 100644
--- a/test/quantization/test_weight_only_linear.py
+++ b/test/quantization/test_weight_only_linear.py
@@ -878,6 +878,52 @@ def test_weightonly_linear_backward(
                 quant_x.grad, x.grad, rtol=1e-3, atol=1e-3
             )
 
+    def test_weightonly_linear_backward_int4_zerosize(self):
+        def test_weightonly_linear_backward(
+            self, algo='weight_only_int4', weight_dtype='int4'
+        ):
+            x = (
+                paddle.rand(shape=(0, 4096), dtype='float16')
+                * 1
+                / math.sqrt(4096)
+            )
+            x.stop_gradient = False
+            quant_x = copy.deepcopy(x)
+            quant_x.stop_gradient = False
+            weight = (
+                paddle.rand(shape=(0, 12288), dtype='float16')
+                * 1
+                / math.sqrt(4096)
+            )
+
+            quant_weight, quant_scale = Q.weight_quantize(
+                x=weight.cuda(), algo=algo
+            )
+            quant_weight = quant_weight.view(
+                [quant_weight.shape[0] * 2, quant_weight.shape[1] // 2]
+            )
+            dequant_weight = Q.weight_dequantize(
+                quant_weight.cuda(), quant_scale, algo=algo
+            )
+            np.testing.assert_allclose(
+                weight, dequant_weight, rtol=1e-2, atol=1e-2
+            )
+
+            quant_out = Q.weight_only_linear(
+                x=quant_x,
+                weight=quant_weight,
+                weight_scale=quant_scale,
+                weight_dtype=weight_dtype,
+            )
+            out = paddle.matmul(x=x, y=weight)
+            np.testing.assert_allclose(quant_out, out, rtol=1e-3, atol=1e-3)
+
+            quant_out.backward()
+            out.backward()
+            np.testing.assert_allclose(
+                quant_x.grad, x.grad, rtol=1e-3, atol=1e-3
+            )
+
 
 if __name__ == '__main__':
     unittest.main()
diff --git a/test/quantization/test_weight_quantize.py b/test/quantization/test_weight_quantize.py
index edf14ce60809a..29bc5195abe6f 100644
--- a/test/quantization/test_weight_quantize.py
+++ b/test/quantization/test_weight_quantize.py
@@ -113,5 +113,22 @@ def run_test(self):
         baseline = np.array(ref_out).reshape([-1])
         np.allclose(baseline, out.astype("int32").numpy(), atol=1e-2)
 
+class WeightQuantizeW4a8ZeroSizeTensorTestCase(unittest.TestCase):
+    def setUp(self):
+        self.x_shape = [1, 0, 10]
+        self.dtype = "float32"
+    def _test_dygraph(self):
+        paddle.disable_static()
+        x = paddle.rand(shape= self.x_shape, dtype=paddle.float16)
+        x.stop_gradient=False
+        out, scale = weight_quantize(x, algo='weight_only_int8')
+        loss = out.sum()
+        loss.backward()
+        np.assert_allclose(x.grad.shape,x.shape)
+    def run_test(self):
+        self.setUp()
+        self._test_dygraph()
+
+
 if __name__ == '__main__':
     unittest.main()