diag: force block_m=128 to test UMMA_N=192 validity for mxf4nvf4

csrc/jit_kernels/heuristics/mega_moe.hpp

@@ -82,6 +82,13 @@ static std::tuple<int, int, int, int> get_block_config_for_mega_moe(
         }
     }();
 
+    // DIAGNOSTIC: Force block_m=128 for NVFP4 mxf4nvf4 testing
+    // 192 may not be a valid UMMA_N for mxf4nvf4 hardware
+    const int forced_block_m = 128;
+    const int forced_store_block_m = 32;
+    const int forced_epilogue_threads = 256;
+    return {2, forced_block_m, forced_store_block_m, forced_epilogue_threads};
+
     // Check whether our `block_m` lies in `kCandidateBlockM`
     DG_HOST_ASSERT(std::any_of(
         layout::kCandidateBlockM, layout::kCandidateBlockM + layout::kNumCandidateBlockMs,