diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3d5c8359a847..f09a6086bc8d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -180,8 +180,8 @@ endif()
 
 # Customized release build type with assertions: TritonRelBuildWithAsserts
 if(NOT MSVC)
-  set(CMAKE_C_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g")
-  set(CMAKE_CXX_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g")
+  set(CMAKE_C_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g -Wunused")
+  set(CMAKE_CXX_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g -Wunused")
   set(CMAKE_C_FLAGS_TRITONBUILDWITHO1 "-O1")
   set(CMAKE_CXX_FLAGS_TRITONBUILDWITHO1 "-O1")
 else()
diff --git a/lib/Tools/GenericSwizzling.cpp b/lib/Tools/GenericSwizzling.cpp
index d90bae1a11c9..8457698c52c5 100644
--- a/lib/Tools/GenericSwizzling.cpp
+++ b/lib/Tools/GenericSwizzling.cpp
@@ -29,8 +29,8 @@ using namespace mlir::triton;
 
 namespace {
 
-void printBasis(const llvm::SmallVector<int32_t> &basis,
-                const std::string &name) {
+[[maybe_unused]] void printBasis(const llvm::SmallVector<int32_t> &basis,
+                                 const std::string &name) {
   llvm::errs() << name << ": ";
   for (int32_t b : basis)
     llvm::errs() << b << " ";
diff --git a/lib/Tools/LinearLayout.cpp b/lib/Tools/LinearLayout.cpp
index e9b3e668a1ca..91cc1529e64c 100644
--- a/lib/Tools/LinearLayout.cpp
+++ b/lib/Tools/LinearLayout.cpp
@@ -54,7 +54,7 @@ BasesT makeBasesMap(
 }
 
 // Dump the matrix to stderr in a human-readable format for debugging.
-void dumpMatrix(uint64_t *m, int numRows, int numCols) {
+[[maybe_unused]] void dumpMatrix(uint64_t *m, int numRows, int numCols) {
   assert(numCols <= 64);
   for (int r = 0; r < numRows; r++) {
     llvm::errs() << "0b";
diff --git a/python/src/gluon_ir.cc b/python/src/gluon_ir.cc
index ccd248bad768..bff5e811ea4f 100644
--- a/python/src/gluon_ir.cc
+++ b/python/src/gluon_ir.cc
@@ -65,7 +65,9 @@ template <typename AttrOrType>
 constexpr auto hasVerifier(AttrOrType t) -> decltype(t.verifyInvariants, true) {
   return true;
 }
-constexpr auto hasVerifier(...) { return false; }
+template <typename... ArgTs> constexpr auto hasVerifier(ArgTs...) {
+  return false;
+}
 
 // Print a diagnostic without its location. The frontend will attach the AST
 // location to the error message.
diff --git a/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM.cpp b/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM.cpp
index 71cc98c11a1a..ada235d8c1a6 100644
--- a/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM.cpp
+++ b/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM.cpp
@@ -30,8 +30,7 @@ struct ScaledDotOpConversion
     : public ConvertOpToLLVMPattern<triton::DotScaledOp> {
   using ConvertOpToLLVMPattern<triton::DotScaledOp>::ConvertOpToLLVMPattern;
 
-  ScaledDotOpConversion(LLVMTypeConverter &converter, int,
-                        PatternBenefit benefit)
+  ScaledDotOpConversion(LLVMTypeConverter &converter, PatternBenefit benefit)
       : ConvertOpToLLVMPattern<triton::DotScaledOp>(converter, benefit) {}
 
   LogicalResult
@@ -163,6 +162,5 @@ void mlir::triton::NVIDIA::populateDotOpToLLVMPatterns(
   patterns.add<DotOpConversion>(typeConverter, computeCapability, benefit);
   patterns.add<WarpGroupDotOpConversion>(typeConverter, benefit);
   patterns.add<WarpGroupDotWaitOpConversion>(typeConverter, benefit);
-  patterns.add<ScaledDotOpConversion>(typeConverter, computeCapability,
-                                      benefit);
+  patterns.add<ScaledDotOpConversion>(typeConverter, benefit);
 }
diff --git a/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM/MMAHelpers.h b/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM/MMAHelpers.h
index ce25cdf269b2..1b61dfa7e9e3 100644
--- a/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM/MMAHelpers.h
+++ b/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/DotOpToLLVM/MMAHelpers.h
@@ -329,10 +329,10 @@ class DotOpMmaSmemLoader : public DotOpMmaMemLoader {
   }
 };
 
-static Value getOffsetedBase(Value v, gpu::MemDescType memDescTy,
-                             const TypeConverter *typeConverter,
-                             ConversionPatternRewriter &rewriter,
-                             Location loc) {
+static inline Value getOffsetedBase(Value v, gpu::MemDescType memDescTy,
+                                    const TypeConverter *typeConverter,
+                                    ConversionPatternRewriter &rewriter,
+                                    Location loc) {
   TritonLLVMOpBuilder tb(loc, rewriter);
   auto llvmElemTy = typeConverter->convertType(memDescTy.getElementType());
   auto smemObj =