use pointer

csrc/multi_tensor_lamb.cu

@@ -52,7 +52,7 @@ struct LAMBStage1Functor
     const float epsilon,
     adamMode_t mode,
     const float decay,
-    const float global_grad_norm,
+    const float* global_grad_norm,
     const float max_global_grad_norm)
   {
     // I'd like this kernel to propagate infs/nans.
@@ -63,7 +63,7 @@ struct LAMBStage1Functor
     int chunk_idx = tl.block_to_chunk[blockIdx.x];
     int n = tl.sizes[tensor_loc];
 
-    float clipped_global_grad_norm = global_grad_norm > max_global_grad_norm ? global_grad_norm / max_global_grad_norm : 1.0f;
+    float clipped_global_grad_norm = (*global_grad_norm) > max_global_grad_norm ? (*global_grad_norm) / max_global_grad_norm : 1.0f;
 
     T* g = (T*)tl.addresses[0][tensor_loc];
     g += chunk_idx*chunk_size;

csrc/multi_tensor_lamb_stage_1.cu

@@ -124,7 +124,7 @@ void multi_tensor_lamb_stage1_cuda(
   using namespace at;
 
   auto g_grad_norm = global_grad_norm.data<float>();
-  float clipped_global_grad_norm = g_grad_norm > max_global_grad_norm ? g_grad_norm / max_global_grad_norm : 1.0f;
+  float clipped_global_grad_norm = *(g_grad_norm) > max_global_grad_norm ? *(g_grad_norm) / max_global_grad_norm : 1.0f;
   float next_step = float(step+1);
   float beta1_correction = 1.0f - std::pow(beta1, next_step);
   float beta2_correction = 1.0f - std::pow(beta2, next_step);