fix: remove unnecessary code and comments

Author: Chamberlain0w0

infini_train/src/kernels/cuda/linear.cu

@@ -69,10 +69,7 @@ std::shared_ptr<Tensor> MatmulForward(const std::shared_ptr<Tensor> &input, cons
     int64_t stride_a = n * k;
     int64_t stride_b = k * m;
     int64_t stride_c = m * n;
-    // TODO(zbl): check GEMM algo
-    // CUBLAS_GEMM_DEFAULT might requires TensorCore
-    // Use CUBLAS_GEMM_ALGO0 to disable TensorCore algos
-
+    // NOTE(zbl): the last cublasGemmAlgo_t param has no effect on GPU arch >= sm_80(Ampere)
     CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_N, n, m, k, &alpha, other->DataPtr(),
                                             CUDA_R_32F, lda, stride_a, input->DataPtr(), CUDA_R_32F, ldb, stride_b,
                                             &beta, output->DataPtr(), CUDA_R_32F, ldc, stride_c, bs, CUDA_R_32F,

infini_train/src/kernels/cuda/softmax.cu

@@ -10,15 +10,6 @@
 #include "infini_train/include/tensor.h"
 
 namespace infini_train::kernels::cuda {
-
-#define CUDA_CHECK(call)                                                                                               \
-    do {                                                                                                               \
-        cudaError_t status = call;                                                                                     \
-        if (status != cudaSuccess) {                                                                                   \
-            LOG(FATAL) << "CUDA Error: " << cudaGetErrorString(status) << " at " << __FILE__ << ":" << __LINE__;       \
-        }                                                                                                              \
-    } while (0)
-
 template <size_t BLOCK_SIZE, typename T>
 __global__ void SoftmaxForwardKernel(T *output, const T *input, int64_t outer_size, int64_t axis_size,
                                      int64_t inner_size) {
@@ -116,7 +107,6 @@ std::shared_ptr<Tensor> SoftmaxForward(const std::shared_ptr<Tensor> &input, int
     default:
         LOG(FATAL) << "CUDA softmax forward: 'Unsupported data type' at " << __FILE__ << ":" << __LINE__;
     }
-    CUDA_CHECK(cudaDeviceSynchronize());
     return output;
 }