Fix Attention addLayer, make cmake to work with TRT 10.14

Author: yizhuoz004

mlir-tensorrt/build_tools/cmake/MTRTDependencies.cmake

@@ -57,8 +57,8 @@ macro(configure_tensorrt_python_plugin_header)
     find_file(
       trt_python_plugin_header
       NAMES NvInferPythonPlugin.h plugin.h
-      HINTS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/python/include/impl
-      PATHS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/python/include/impl
+      HINTS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/include/impl
+      PATHS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/include/impl
       REQUIRED
       NO_CMAKE_PATH NO_DEFAULT_PATH
       NO_CACHE

mlir-tensorrt/build_tools/cmake/TensorRTDownloadURL.cmake

@@ -80,6 +80,10 @@ function(mtrt_get_tensorrt_download_url ARG_VERSION OS_NAME TARGET_ARCH ARG_OUT_
     set(ARG_VERSION "10.12.0.36")
   endif()
 
+  if(ARG_VERSION VERSION_EQUAL "10.14")
+    set(ARG_VERSION "10.14.1.48")
+  endif()
+
   set(downloadable_versions
     "8.6.1.6"
     "9.0.1.4" "9.1.0.4" "9.2.0.5"
@@ -97,6 +101,7 @@ function(mtrt_get_tensorrt_download_url ARG_VERSION OS_NAME TARGET_ARCH ARG_OUT_
     "10.8.0.43"
     "10.9.0.34"
     "10.12.0.36"
+    "10.14.1.48"
   )
 
   if(NOT ARG_VERSION IN_LIST downloadable_versions)
@@ -164,6 +169,8 @@ function(mtrt_get_tensorrt_download_url ARG_VERSION OS_NAME TARGET_ARCH ARG_OUT_
   elseif(ARG_VERSION VERSION_GREATER 10.10
       AND ARG_VERSION VERSION_LESS 10.13)
     set(TRT_CUDA_VERSION 12.9)
+  elseif(ARG_VERSION VERSION_GREATER 10.13)
+    set(TRT_CUDA_VERSION 13.0)
   endif()
 
   # Handle TRT 8 versions.

mlir-tensorrt/tensorrt/include/mlir-tensorrt-dialect/TensorRT/IR/TensorRTOps.td

@@ -4539,7 +4539,7 @@ def TensorRT_AttentionOp : TensorRT_Op<"attention",
     TensorRT_RankedTensorOf<[F16, BF16, F32]>:$value,
     Optional<TensorRT_Tensor>:$mask,
     Optional<TensorRT_RankedTensorOf<[F16, BF16, F32]>>:$normalization_quantize_scale,
-    OptionalAttr<TensorRT_AttentionNormalizationOpAttr>:$normalization_operation,
+    DefaultValuedAttr<TensorRT_AttentionNormalizationOpAttr, "tensorrt::AttentionNormalizationOp::kSOFTMAX">:$normalization_operation,
     DefaultValuedAttr<BoolAttr, "false">:$causal,
     DefaultValuedAttr<BoolAttr, "false">:$decomposable,
     OptionalAttr<TensorRT_DataTypeAttr>:$normalization_quantize_to_type
@@ -4565,12 +4565,7 @@ def TensorRT_AttentionOp : TensorRT_Op<"attention",
   }] # baseClassDeclaration;
 
   let trtLayerAdd = [{
-    // Get normalization operation, default to kSOFTMAX
-    nvinfer1::AttentionNormalizationOp normOp = $normalization_operation 
-      ? *$normalization_operation 
-      : nvinfer1::AttentionNormalizationOp::kSOFTMAX;
-    
-    nvinfer1::IAttention *layer = $net->addAttention(*$query, *$key, *$value, normOp, $causal);
+    nvinfer1::IAttention *layer = $net->addAttention(*$query, *$key, *$value, *$normalization_operation, $causal);
     if (!layer)
       return failure();
 
@@ -4584,19 +4579,22 @@ def TensorRT_AttentionOp : TensorRT_Op<"attention",
     }
 
     if ($normalization_quantize_to_type) {
-      layer->setNormalizationQuantizeToType(*$normalization_quantize_to_type);
+      auto convertedDataType = ::mlir::tensorrt::convertDataTypeToNvInferEnum(*$normalization_quantize_to_type);
+      if (!convertedDataType)
+        return emitError($op->getLoc()) << "failed to convert DataType to nvinfer enum";
+      layer->setNormalizationQuantizeToType(*convertedDataType);
     }
 
     if (!$e.isStronglyTyped()){
       FailureOr<nvinfer1::DataType> outputTrtType = getNvInferDataType($op.getLoc(),
                                                           $op.getType().getElementType());
       if (failed(outputTrtType))
         return failure();
-      layer->setOutputType(0, *outputTrtType);
     }
 
     $results.push_back(layer->getOutput(0));
-    $e.setMetadata(layer, $op);
+    // TODO: nvinfer1::IAttention does not have setMetadata API in 10.14
+    // layer->setMetadata($op);
   }];
 }