Move reduce shared memory to the back, query max. shmem size dynamically

Fix build error

optimum/quanto/library/extensions/cuda/marlin/marlin_cuda_kernel.cu

@@ -374,9 +374,9 @@ __global__ void Marlin(
   int4* sh_a = sh;
   int4* sh_b = sh_a + (stages * a_sh_stage);
   int4* sh_s = sh_b + (stages * b_sh_stage);
-  int4* sh_red = sh_s + (stages * s_sh_stage);
   // ADDED: shared memory storage for scaled zero points
-  int4* sh_sz = sh_red + (stages * s_sh_stage);
+  int4* sh_sz = sh_s + (stages * s_sh_stage);
+  int4* sh_red = sh_sz + (stages * s_sh_stage);
 
   // Register storage for double buffer of shared memory reads.
   FragA frag_a[2][thread_m_blocks];
@@ -728,7 +728,6 @@ __global__ void Marlin(
 // latency hiding. At the same time, we want relatively few warps to have many registers per warp and small tiles.
 const int THREADS = 256;
 const int STAGES = 4; // 4 pipeline stages fit into shared memory
-const int SHARED_MEM = 164 * 1000; // max shared memory on compute capability 8.0
 
 // ADDED: add scaled zero pointer
 #define CALL_IF(THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, GROUP_BLOCKS) \
@@ -739,11 +738,11 @@ const int SHARED_MEM = 164 * 1000; // max shared memory on compute capability 8.
     cudaFuncSetAttribute( \
       Marlin<THREADS, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, STAGES, GROUP_BLOCKS>, \
       cudaFuncAttributeMaxDynamicSharedMemorySize, \
-      SHARED_MEM \
+      max_shared_mem \
     ); \
     Marlin< \
       THREADS, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, STAGES, GROUP_BLOCKS \
-    ><<<blocks, THREADS, SHARED_MEM, stream>>>( \
+    ><<<blocks, THREADS, max_shared_mem, stream>>>( \
       A_ptr, B_ptr, C_ptr, s_ptr, sz_ptr,\
       prob_m, prob_n, prob_k, \
       locks \
@@ -789,6 +788,9 @@ int marlin_cuda(
     }
   }
 
+  int max_shared_mem = 0;
+  cudaDeviceGetAttribute(&max_shared_mem, cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
+
   int thread_k_blocks = thread_k / 16;
   int thread_n_blocks = thread_n / 16;
   int group_blocks = (groupsize == -1) ? -1 : groupsize / 16;