minor code refactoring

Signed-off-by: Alex Kogan <[email protected]>

Author: sakogan

csrc/ops.h

@@ -298,10 +298,11 @@ void per_token_group_quant_int8(const torch::Tensor& input,
                                 torch::Tensor& output_s, int64_t group_size,
                                 double eps, double int8_min, double int8_max);
 
-torch::Tensor rtn_marlin_gemm(const torch::Tensor& a, const torch::Tensor& b_q_weight,
+torch::Tensor rtn_marlin_gemm(const torch::Tensor& a,
+                              const torch::Tensor& b_q_weight,
                               const torch::Tensor& b_scales,
-                              torch::Tensor& workspace, int64_t size_m, int64_t size_n,
-                              int64_t size_k);
+                              torch::Tensor& workspace, int64_t size_m,
+                              int64_t size_n, int64_t size_k);
 
 #endif
 

csrc/quantization/rtn_marlin/rtn_marlin_cuda_kernel.cu

@@ -968,7 +968,7 @@ torch::Tensor rtn_marlin_gemm(const torch::Tensor& a,
   // Switch to a Dequantize-and-GEMM path for long inputs
   // (see https://arxiv.org/pdf/2505.15909 for details)
   const int SLOW_PATH_MATMUL_HEURISTIC_CONDITION = 1024;
-  if (size_m >= SLOW_PATH_MATMUL_HEURISTIC_CONDITION) {    
+  if (size_m >= SLOW_PATH_MATMUL_HEURISTIC_CONDITION) {
     auto weight = rtn_marlin_dequantize(b_q_weight, b_scales, qbits);
     return at::matmul(a, weight.t());
   }