refactor(poly-norm): fix indentation

activation/activation_kernels.cu

@@ -130,60 +130,60 @@ __global__ void poly_norm_backward_kernel(
   acc_t sq_mean_4 = rsqrtf(mean_4) * mean_4;
   acc_t sq_mean_6 = rsqrtf(mean_6) * mean_6;
 
-	acc_t denom_2 = mean_2 * sq_mean_2;
-	acc_t denom_4 = mean_4 * sq_mean_4;
-	acc_t denom_6 = mean_6 * sq_mean_6;
-
-	acc_t sum_dw0 = 0;
-	acc_t sum_dw1 = 0;
-	acc_t sum_dw2 = 0;
-	acc_t sum_db = 0;
-
-	for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
-		acc_t dy = output_grad[token_idx * d + idx];
-		acc_t x_1 = input[token_idx * d + idx];
-		acc_t x_2 = x_1 * x_1;
-		acc_t x_3 = x_2 * x_1;
-
-		acc_t _dx_1 = w2 * dy;
-		acc_t _dx_2 = w1 * dy;
-		acc_t _dx_3 = w0 * dy;
-
-		acc_t dx_3 =
-			3 * x_2 * (_dx_3 / sq_mean_6 - x_3 * sum_dx_3 / (d * denom_6));
-		acc_t dx_2 =
-			2 * x_1 * (_dx_2 / sq_mean_4 - x_2 * sum_dx_2 / (d * denom_4));
-		acc_t dx_1 =
-			_dx_1 / sq_mean_2 - x_1 * sum_dx_1 / (d * denom_2);
-
-		if (input_grad) {
-			input_grad[token_idx * d + idx] = dx_1 + dx_2 + dx_3;
-		}
-
-		sum_dw0 += dy * (x_3 / sq_mean_6);
-		sum_dw1 += dy * (x_2 / sq_mean_4);
-		sum_dw2 += dy * (x_1 / sq_mean_2);
-		sum_db += dy;
-	}
-
-	if (weight_grad) {
-		sum_dw0 = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_dw0, d);
-		sum_dw1 = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_dw1, d);
-		sum_dw2 = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_dw2, d);
-
-		if (threadIdx.x == 0) {
-			atomic_add(&weight_grad[0], sum_dw0);
-			atomic_add(&weight_grad[1], sum_dw1);
-			atomic_add(&weight_grad[2], sum_dw2);
-		}
-	}
-
-	if (bias_grad) {
-		sum_db = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_db, d);
-		if (threadIdx.x == 0) {
-			atomic_add(&bias_grad[0], sum_db);
-		}
-	}
+  acc_t denom_2 = mean_2 * sq_mean_2;
+  acc_t denom_4 = mean_4 * sq_mean_4;
+  acc_t denom_6 = mean_6 * sq_mean_6;
+
+  acc_t sum_dw0 = 0;
+  acc_t sum_dw1 = 0;
+  acc_t sum_dw2 = 0;
+  acc_t sum_db = 0;
+
+  for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
+    acc_t dy = output_grad[token_idx * d + idx];
+    acc_t x_1 = input[token_idx * d + idx];
+    acc_t x_2 = x_1 * x_1;
+    acc_t x_3 = x_2 * x_1;
+
+    acc_t _dx_1 = w2 * dy;
+    acc_t _dx_2 = w1 * dy;
+    acc_t _dx_3 = w0 * dy;
+
+    acc_t dx_3 =
+      3 * x_2 * (_dx_3 / sq_mean_6 - x_3 * sum_dx_3 / (d * denom_6));
+    acc_t dx_2 =
+      2 * x_1 * (_dx_2 / sq_mean_4 - x_2 * sum_dx_2 / (d * denom_4));
+    acc_t dx_1 =
+      _dx_1 / sq_mean_2 - x_1 * sum_dx_1 / (d * denom_2);
+
+    if (input_grad) {
+      input_grad[token_idx * d + idx] = dx_1 + dx_2 + dx_3;
+    }
+
+    sum_dw0 += dy * (x_3 / sq_mean_6);
+    sum_dw1 += dy * (x_2 / sq_mean_4);
+    sum_dw2 += dy * (x_1 / sq_mean_2);
+    sum_db += dy;
+  }
+
+  if (weight_grad) {
+    sum_dw0 = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_dw0, d);
+    sum_dw1 = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_dw1, d);
+    sum_dw2 = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_dw2, d);
+
+    if (threadIdx.x == 0) {
+      atomic_add(&weight_grad[0], sum_dw0);
+      atomic_add(&weight_grad[1], sum_dw1);
+      atomic_add(&weight_grad[2], sum_dw2);
+    }
+  }
+
+  if (bias_grad) {
+    sum_db = _block_reduce_sum<acc_t, BLOCK_SIZE>(shared, sum_db, d);
+    if (threadIdx.x == 0) {
+      atomic_add(&bias_grad[0], sum_db);
+    }
+  }
 }
 
 }  // namespace motif
@@ -191,9 +191,9 @@ __global__ void poly_norm_backward_kernel(
 
 void poly_norm(torch::Tensor& out,          // [..., d]
                const torch::Tensor& input,  // [..., d]
-	       const torch::Tensor& weight, // [3]
-	       const torch::Tensor& bias,   // [1]
-	       double eps)
+               const torch::Tensor& weight, // [3]
+               const torch::Tensor& bias,   // [1]
+               double eps)
 {
   AssertTensorShapeEqual(input, out, "input", "out");
   AssertTensorNotNull(weight, "weight");
@@ -244,11 +244,12 @@ void poly_norm_backward(
 
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-	if (weight_grad.defined())
-		cudaMemsetAsync(weight_grad.data_ptr(), 0, weight_grad.numel() * weight_grad.element_size(), stream);
-	if (bias_grad.defined()) {
-		cudaMemsetAsync(bias_grad.data_ptr(), 0, bias_grad.numel() * bias_grad.element_size(), stream);
-	}
+  if (weight_grad.defined()) {
+    cudaMemsetAsync(weight_grad.data_ptr(), 0, weight_grad.numel() * weight_grad.element_size(), stream);
+  }
+  if (bias_grad.defined()) {
+    cudaMemsetAsync(bias_grad.data_ptr(), 0, bias_grad.numel() * bias_grad.element_size(), stream);
+  }
 
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   MOTIF_DISPATCH_FLOATING_TYPES(