Cherry-picks for 1.16.1 release (#17741)

Cherry-pick the following PRs to the release branch：

Fix: Fail to skip disabledmodel in winml (#17728) 
Move dotnet build and test into docker in Linux CPU CI (#17417) 
Run Nuget_Test_Linux_GPU in container (#17452) 
Run Final_Jar_Testing_Linux_GPU in docker (#17533) 
TreeEnsemble speed up (#17449) 
Remove onnxruntime extensions from list of gitmodules (#17615) 
Include onnxruntime_float16.h in the package. (#17637) 
Fix static quantization for QDQ and Percentile distribution (#17649) 
[TensorRT EP] Back out the PerThreadContext (#17690) 
Update nodejs to 18.x (#17657) 
Update linux-wasm-ci.yml: remove the ln command (#17735)

.gitmodules

@@ -8,6 +8,3 @@
 	path = cmake/external/emsdk
 	url = https://github.com/emscripten-core/emsdk.git
 	branch = 3.1.44
-[submodule "cmake/external/onnxruntime-extensions"]
-	path = cmake/external/onnxruntime-extensions
-	url = https://github.com/microsoft/onnxruntime-extensions.git

VERSION_NUMBER

@@ -1 +1 @@
-1.16.0
+1.16.1

docs/python/README.rst

@@ -8,6 +8,11 @@ For more information on ONNX Runtime, please see `aka.ms/onnxruntime <https://ak
 Changes
 -------
 
+1.16.1
+^^^^^^
+
+Release Notes : https://github.com/Microsoft/onnxruntime/releases/tag/v1.16.1
+
 1.16.0
 ^^^^^^
 

js/common/lib/version.ts

@@ -4,4 +4,4 @@
 // This file is generated by /js/scripts/update-version.ts
 // Do not modify file content manually.
 
-export const version = '1.16.0';
+export const version = '1.16.1';

js/common/package.json

@@ -2,7 +2,7 @@
   "license": "MIT",
   "type": "module",
   "name": "onnxruntime-common",
-  "version": "1.16.0",
+  "version": "1.16.1",
   "repository": {
     "url": "https://github.com/Microsoft/onnxruntime.git",
     "type": "git"

js/node/lib/version.ts

@@ -4,4 +4,4 @@
 // This file is generated by /js/scripts/update-version.ts
 // Do not modify file content manually.
 
-export const version = '1.16.0';
+export const version = '1.16.1';

js/node/package.json

@@ -13,7 +13,7 @@
       3
     ]
   },
-  "version": "1.16.0",
+  "version": "1.16.1",
   "dependencies": {
     "onnxruntime-common": "file:../common"
   },

js/react_native/lib/backend.ts

@@ -66,12 +66,14 @@ class OnnxruntimeSessionHandler implements SessionHandler {
       let results: Binding.ModelLoadInfoType;
       // load a model
       if (typeof this.#pathOrBuffer === 'string') {
+        // load model from model path
         results = await this.#inferenceSession.loadModel(normalizePath(this.#pathOrBuffer), options);
       } else {
+        // load model from buffer
         if (!this.#inferenceSession.loadModelFromBlob) {
           throw new Error('Native module method "loadModelFromBlob" is not defined');
         }
-        const modelBlob = jsiHelper.storeArrayBuffer(this.#pathOrBuffer);
+        const modelBlob = jsiHelper.storeArrayBuffer(this.#pathOrBuffer.buffer);
         results = await this.#inferenceSession.loadModelFromBlob(modelBlob, options);
       }
       // resolve promise if onnxruntime session is successfully created

js/react_native/lib/version.ts

@@ -4,4 +4,4 @@
 // This file is generated by /js/scripts/update-version.ts
 // Do not modify file content manually.
 
-export const version = '1.16.0';
+export const version = '1.16.1';

js/react_native/package.json

@@ -36,7 +36,7 @@
     "registry": "https://registry.npmjs.org/"
   },
   "source": "lib/index",
-  "version": "1.16.0",
+  "version": "1.16.1",
   "main": "dist/commonjs/index",
   "homepage": "https://github.com/microsoft/onnxruntime/blob/main/js/react_native/README.md",
   "files": [

js/react_native/yarn.lock

@@ -5188,7 +5188,7 @@ onetime@^5.1.0, onetime@^5.1.2:
     mimic-fn "^2.1.0"
 
 "onnxruntime-common@file:../common":
-  version "1.16.0"
+  version "1.16.1"
 
 open@^6.2.0:
   version "6.4.0"

js/web/lib/version.ts

@@ -4,4 +4,4 @@
 // This file is generated by /js/scripts/update-version.ts
 // Do not modify file content manually.
 
-export const version = '1.16.0';
+export const version = '1.16.1';

js/web/package.json

@@ -8,7 +8,7 @@
     "type": "git"
   },
   "author": "fs-eire",
-  "version": "1.16.0",
+  "version": "1.16.1",
   "jsdelivr": "dist/ort.min.js",
   "dependencies": {
     "flatbuffers": "^1.12.0",

onnxruntime/__init__.py

@@ -7,7 +7,7 @@
 For more information on ONNX Runtime, please see `aka.ms/onnxruntime <https://aka.ms/onnxruntime/>`_
 or the `Github project <https://github.com/microsoft/onnxruntime/>`_.
 """
-__version__ = "1.16.0"
+__version__ = "1.16.1"
 __author__ = "Microsoft"
 
 # we need to do device version validation (for example to check Cuda version for an onnxruntime-training package).

onnxruntime/contrib_ops/cuda/bert/attention.cc

@@ -140,27 +140,29 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
 #endif
 
   if (!use_flash_attention) {
-    if (is_unidirectional_ && enable_fused_causal_attention_) {  // GPT
-      // GPT fused kernels requires left side padding. mask can be:
-      //     none (no padding), 1D sequence lengths or 2d mask.
-      // Fused kernels don't support different sequence lengths of q and kv, so only apply to the first token
-      // where past state is empty.
-      bool is_mask_2d_key_padding = parameters.mask_type == AttentionMaskType::MASK_2D_KEY_PADDING;
-      bool use_causal_fused_runner = (nullptr == mask_index || is_mask_1d_seq_len || is_mask_2d_key_padding) &&
-                                     nullptr == relative_position_bias &&
-                                     parameters.past_sequence_length == 0 &&
-                                     parameters.hidden_size == parameters.v_hidden_size &&
-                                     FusedMHARunnerFP16v2::is_supported(sm, parameters.head_size, sequence_length,
-                                                                        enable_trt_flash_attention_, true);
-      if (use_causal_fused_runner) {
-        // Here we assume that num_heads, head_size and is_unidirectional does not change for an Attention node.
-        if (nullptr == fused_fp16_runner_.get()) {
-          fused_fp16_runner_ = FusedMHARunnerFP16v2::Create(num_heads_, parameters.head_size, sm, is_unidirectional_,
-                                                            enable_trt_flash_attention_, parameters.scale);
+    if (is_unidirectional_) {  // GPT
+      if (enable_fused_causal_attention_) {
+        // GPT fused kernels requires left side padding. mask can be:
+        //     none (no padding), 1D sequence lengths or 2d mask.
+        // Fused kernels don't support different sequence lengths of q and kv, so only apply to the first token
+        // where past state is empty.
+        bool is_mask_2d_key_padding = parameters.mask_type == AttentionMaskType::MASK_2D_KEY_PADDING;
+        bool use_causal_fused_runner = (nullptr == mask_index || is_mask_1d_seq_len || is_mask_2d_key_padding) &&
+                                       nullptr == relative_position_bias &&
+                                       parameters.past_sequence_length == 0 &&
+                                       parameters.hidden_size == parameters.v_hidden_size &&
+                                       FusedMHARunnerFP16v2::is_supported(sm, parameters.head_size, sequence_length,
+                                                                          enable_trt_flash_attention_, true);
+        if (use_causal_fused_runner) {
+          // Here we assume that num_heads, head_size and is_unidirectional does not change for an Attention node.
+          if (nullptr == fused_fp16_runner_.get()) {
+            fused_fp16_runner_ = FusedMHARunnerFP16v2::Create(num_heads_, parameters.head_size, sm, is_unidirectional_,
+                                                              enable_trt_flash_attention_, parameters.scale);
+          }
+
+          // Here we assume all causal kernels can be loaded into shared memory. TODO: add a function to check.
+          fused_runner = fused_fp16_runner_.get();
         }
-
-        // Here we assume all causal kernels can be loaded into shared memory. TODO: add a function to check.
-        fused_runner = fused_fp16_runner_.get();
       }
     } else {  // BERT
       bool use_fused_runner = !disable_fused_self_attention_ &&

onnxruntime/core/providers/cpu/ml/tree_ensemble_aggregator.h

@@ -64,31 +64,38 @@ enum MissingTrack : uint8_t {
   kFalse = 0
 };
 
+template <typename T>
+struct TreeNodeElement;
+
+template <typename T>
+union PtrOrWeight {
+  TreeNodeElement<T>* ptr;
+  struct WeightData {
+    int32_t weight;
+    int32_t n_weights;
+  } weight_data;
+};
+
 template <typename T>
 struct TreeNodeElement {
   int feature_id;
 
   // Stores the node threshold or the weights if the tree has one target.
   T value_or_unique_weight;
 
-  // onnx specification says hitrates is used to store information about the node,
+  // The onnx specification says hitrates is used to store information about the node,
   // but this information is not used for inference.
   // T hitrates;
 
-  // True node, false node are obtained by computing `this + truenode_inc_or_first_weight`,
-  // `this + falsenode_inc_or_n_weights` if the node is not a leaf.
-  // In case of a leaf, these attributes are used to indicate the position of the weight
-  // in array `TreeEnsembleCommon::weights_`. If the number of targets or classes is one,
-  // the weight is also stored in `value_or_unique_weight`.
-  // This implementation assumes a tree has less than 2^31 nodes,
-  // and the total number of leave in the set of trees is below 2^31.
-  // A node cannot point to itself.
-  int32_t truenode_inc_or_first_weight;
-  // In case of a leaf, the following attribute indicates the number of weights
-  // in array `TreeEnsembleCommon::weights_`. If not a leaf, it indicates
-  // `this + falsenode_inc_or_n_weights` is the false node.
-  // A node cannot point to itself.
-  int32_t falsenode_inc_or_n_weights;
+  // PtrOrWeight acts as a tagged union, with the "tag" being whether the node is a leaf or not (see `is_not_leaf`).
+
+  // If it is not a leaf, it is a pointer to the true child node when traversing the decision tree. The false branch is
+  // always 1 position away from the TreeNodeElement in practice in `TreeEnsembleCommon::nodes_` so it is not stored.
+
+  // If it is a leaf, it contains `weight` and `n_weights` attributes which are used to indicate the position of the
+  // weight in array `TreeEnsembleCommon::weights_`. If the number of targets or classes is one, the weight is also
+  // stored in `value_or_unique_weight`.
+  PtrOrWeight<T> truenode_or_weight;
   uint8_t flags;
 
   inline NODE_MODE mode() const { return NODE_MODE(flags & 0xF); }
@@ -189,8 +196,8 @@ class TreeAggregatorSum : public TreeAggregator<InputType, ThresholdType, Output
   void ProcessTreeNodePrediction(InlinedVector<ScoreValue<ThresholdType>>& predictions,
                                  const TreeNodeElement<ThresholdType>& root,
                                  gsl::span<const SparseValue<ThresholdType>> weights) const {
-    auto it = weights.begin() + root.truenode_inc_or_first_weight;
-    for (int32_t i = 0; i < root.falsenode_inc_or_n_weights; ++i, ++it) {
+    auto it = weights.begin() + root.truenode_or_weight.weight_data.weight;
+    for (int32_t i = 0; i < root.truenode_or_weight.weight_data.n_weights; ++i, ++it) {
       ORT_ENFORCE(it->i < (int64_t)predictions.size());
       predictions[onnxruntime::narrow<size_t>(it->i)].score += it->value;
       predictions[onnxruntime::narrow<size_t>(it->i)].has_score = 1;
@@ -292,8 +299,8 @@ class TreeAggregatorMin : public TreeAggregator<InputType, ThresholdType, Output
   void ProcessTreeNodePrediction(InlinedVector<ScoreValue<ThresholdType>>& predictions,
                                  const TreeNodeElement<ThresholdType>& root,
                                  gsl::span<const SparseValue<ThresholdType>> weights) const {
-    auto it = weights.begin() + root.truenode_inc_or_first_weight;
-    for (int32_t i = 0; i < root.falsenode_inc_or_n_weights; ++i, ++it) {
+    auto it = weights.begin() + root.truenode_or_weight.weight_data.weight;
+    for (int32_t i = 0; i < root.truenode_or_weight.weight_data.n_weights; ++i, ++it) {
       predictions[onnxruntime::narrow<size_t>(it->i)].score =
           (!predictions[onnxruntime::narrow<size_t>(it->i)].has_score || it->value < predictions[onnxruntime::narrow<size_t>(it->i)].score)
               ? it->value
@@ -349,8 +356,8 @@ class TreeAggregatorMax : public TreeAggregator<InputType, ThresholdType, Output
   void ProcessTreeNodePrediction(InlinedVector<ScoreValue<ThresholdType>>& predictions,
                                  const TreeNodeElement<ThresholdType>& root,
                                  gsl::span<const SparseValue<ThresholdType>> weights) const {
-    auto it = weights.begin() + root.truenode_inc_or_first_weight;
-    for (int32_t i = 0; i < root.falsenode_inc_or_n_weights; ++i, ++it) {
+    auto it = weights.begin() + root.truenode_or_weight.weight_data.weight;
+    for (int32_t i = 0; i < root.truenode_or_weight.weight_data.n_weights; ++i, ++it) {
       predictions[onnxruntime::narrow<size_t>(it->i)].score =
           (!predictions[onnxruntime::narrow<size_t>(it->i)].has_score || it->value > predictions[onnxruntime::narrow<size_t>(it->i)].score)
               ? it->value