From 8608bf02f21774be0388d2aa3a9f886d009d0b4c Mon Sep 17 00:00:00 2001
From: Adrian Lizarraga <adlizarraga@microsoft.com>
Date: Mon, 10 Feb 2025 08:55:12 -0800
Subject: [PATCH] [ORT 1.20.2 Release] Cherry pick 1st round (#23574)

### Description
- Update ORT version to 1.20.2
- Cherry-pick commits:
  - https://github.com/microsoft/onnxruntime/pull/23243
  - https://github.com/microsoft/onnxruntime/pull/22738
  - https://github.com/microsoft/onnxruntime/pull/22868
  - https://github.com/microsoft/onnxruntime/pull/23281
  - https://github.com/microsoft/onnxruntime/pull/22543
  - https://github.com/microsoft/onnxruntime/pull/22566
  - https://github.com/microsoft/onnxruntime/pull/23308
  - https://github.com/microsoft/onnxruntime/pull/23017
  - [Main feature] https://github.com/microsoft/onnxruntime/pull/23368



### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

---------

Co-authored-by: Changming Sun <chasun@microsoft.com>
Co-authored-by: Tianlei Wu <tlwu@microsoft.com>
Co-authored-by: Jian Chen <cjian@microsoft.com>
Co-authored-by: Yi Zhang <zhanyi@microsoft.com>
Co-authored-by: Caroline Zhu <wolfivyaura@gmail.com>
---
 .github/codeql/codeql-config.yml              |   7 +
 .github/workflows/cffconvert.yml              |   2 +-
 .github/workflows/codeql.yml                  |   7 +-
 .../workflows/gradle-wrapper-validation.yml   |   2 +-
 .github/workflows/labeler.yml                 |   2 +-
 .github/workflows/lint.yml                    |   7 +-
 .github/workflows/linux_training.yml          |  55 ++++++
 .github/workflows/mac.yml                     |   2 +-
 .github/workflows/pr_checks.yml               |  13 +-
 .github/workflows/publish-c-apidocs.yml       |   2 +-
 .github/workflows/publish-csharp-apidocs.yml  |   2 +-
 .github/workflows/publish-gh-pages.yml        |   2 +-
 .github/workflows/publish-java-apidocs.yml    |   2 +-
 .github/workflows/publish-js-apidocs.yml      |   2 +-
 .github/workflows/publish-python-apidocs.yml  |   2 +-
 .github/workflows/skip-doc-change.yml.j2      |   2 +-
 .github/workflows/stale.yml                   |   2 +-
 .github/workflows/title-only-labeler.yml      |   2 +-
 VERSION_NUMBER                                |   2 +-
 ...Microsoft.ML.OnnxRuntime.Tests.MAUI.csproj |   4 +-
 docs/python/README.rst                        |   5 +
 .../core/session/onnxruntime_c_api.h          |   4 +-
 js/common/lib/version.ts                      |   2 +-
 js/common/package-lock.json                   |   4 +-
 js/common/package.json                        |   2 +-
 js/node/lib/version.ts                        |   2 +-
 js/node/package-lock.json                     |   6 +-
 js/node/package.json                          |   2 +-
 js/react_native/lib/version.ts                |   2 +-
 js/react_native/package.json                  |   2 +-
 js/react_native/yarn.lock                     |   2 +-
 js/web/lib/version.ts                         |   2 +-
 js/web/package-lock.json                      |   6 +-
 js/web/package.json                           |   2 +-
 js/web/test/data/ops/conv.jsonc               | 160 +++++++++---------
 js/web/test/data/ops/fused-conv.jsonc         | 152 ++++++++---------
 onnxruntime/__init__.py                       |   2 +-
 .../bert/flash_attention/flash_fwd_kernel.h   |  12 +-
 .../cuda/bert/flash_attention/mask.h          |  15 +-
 .../cuda/bert/flash_attention/softmax.h       |  13 +-
 .../bert/lean_attention/lean_fwd_kernel.h     |  10 +-
 .../cuda/bert/lean_attention/mask.h           |  16 +-
 .../cuda/bert/lean_attention/softmax.h        |  14 +-
 .../cuda/bert/ngram_repeat_block_impl.cu      |   4 +-
 .../cuda/transformers/beam_search_topk.cu     |   2 +-
 .../transformers/greedy_search_top_one.cu     |   6 +-
 .../core/optimizer/attention_fusion_helper.h  |   4 +-
 .../core/providers/cuda/math/topk_impl.cuh    |  10 +-
 .../providers/cuda/shared_inc/cuda_utils.h    |  44 ++---
 .../providers/qnn/qnn_execution_provider.cc   |  10 +-
 .../core/providers/xnnpack/detail/utils.cc    |   5 +-
 .../core/providers/xnnpack/math/gemm.cc       |   7 +-
 .../core/providers/xnnpack/math/matmul.cc     |   5 +-
 .../core/providers/xnnpack/nn/average_pool.cc |   4 +-
 .../core/providers/xnnpack/nn/conv_base.cc    |   4 +-
 .../core/providers/xnnpack/nn/max_pool.cc     |   6 +-
 onnxruntime/core/session/onnxruntime_c_api.cc |   2 +-
 .../apple/apple_package_test/Podfile.template |   2 +-
 .../test/providers/qnn/argmaxmin_op_test.cc   |   1 +
 .../test/providers/qnn/average_pool_test.cc   |   2 +
 .../test/providers/qnn/batch_norm_htp_test.cc |   2 +
 onnxruntime/test/providers/qnn/cast_test.cc   |   1 +
 .../test/providers/qnn/clip_op_test.cc        |   2 +
 onnxruntime/test/providers/qnn/conv_test.cc   |   5 +
 .../test/providers/qnn/flatten_op_test.cc     |   2 +
 .../providers/qnn/gather_elems_op_test.cc     |   3 +
 .../test/providers/qnn/gather_op_htp_test.cc  |   3 +-
 .../test/providers/qnn/gemm_op_test.cc        |   3 +
 .../providers/qnn/instance_norm_htp_test.cc   |   1 +
 .../test/providers/qnn/layer_norm_test.cc     |   2 +
 .../providers/qnn/leakyrelu_op_htp_test.cc    |   2 +
 .../providers/qnn/logical_comp_ops_test.cc    |   3 +
 onnxruntime/test/providers/qnn/lrn_op_test.cc |   2 +
 .../test/providers/qnn/matmul_test.cpp        |   2 +
 .../test/providers/qnn/max_min_op_test.cc     |   2 +
 .../test/providers/qnn/pad_op_test.cpp        |   2 +
 .../test/providers/qnn/pool_op_test.cpp       |   2 +
 .../test/providers/qnn/qnn_basic_test.cc      |  12 ++
 .../test/providers/qnn/qnn_ep_context_test.cc |  17 ++
 .../test/providers/qnn/qnn_test_utils.cc      |   4 +-
 .../test/providers/qnn/reduce_op_test.cc      |   2 +
 .../providers/qnn/reshape_expand_op_test.cc   |   3 +
 onnxruntime/test/providers/qnn/resize_test.cc |   3 +
 .../test/providers/qnn/simple_op_htp_test.cc  |   5 +
 .../test/providers/qnn/slice_htp_test.cc      |   1 +
 .../test/providers/qnn/split_op_test.cc       |   1 +
 .../qnn/squeeze_unsqueeze_op_test.cc          |   3 +
 .../test/providers/qnn/tile_op_test.cc        |   1 +
 .../test/providers/qnn/topk_op_test.cc        |   1 +
 .../test/providers/qnn/transpose_htp_test.cc  |   1 +
 .../test/providers/qnn/where_htp_test.cc      |   1 +
 ...t_full_apple_framework_build_settings.json |   2 +-
 ...training_ios_framework_build_settings.json |   2 +-
 .../npm-packaging-pipeline.yml                |   4 +-
 .../orttraining-linux-ci-pipeline.yml         |  95 -----------
 .../orttraining-linux-gpu-ci-pipeline.yml     |  55 ------
 .../azure-pipelines/post-merge-jobs.yml       |   2 +-
 .../templates/mac-cpu-packaging-pipeline.yml  |   3 -
 .../templates/react-native-ci.yml             |  24 +++
 .../azure-pipelines/templates/web-ci.yml      |   4 +-
 .../azure-pipelines/web-ci-pipeline.yml       |   4 +-
 .../github/linux/build_training_ci.sh         |   4 -
 .../docker/Dockerfile.ubuntu_gpu_training     |  60 -------
 103 files changed, 503 insertions(+), 526 deletions(-)
 create mode 100644 .github/codeql/codeql-config.yml
 create mode 100644 .github/workflows/linux_training.yml
 delete mode 100644 tools/ci_build/github/azure-pipelines/orttraining-linux-ci-pipeline.yml
 delete mode 100644 tools/ci_build/github/azure-pipelines/orttraining-linux-gpu-ci-pipeline.yml
 delete mode 100755 tools/ci_build/github/linux/build_training_ci.sh
 delete mode 100644 tools/ci_build/github/linux/docker/Dockerfile.ubuntu_gpu_training

diff --git a/.github/codeql/codeql-config.yml b/.github/codeql/codeql-config.yml
new file mode 100644
index 0000000000000..6a76f7bcdbcb0
--- /dev/null
+++ b/.github/codeql/codeql-config.yml
@@ -0,0 +1,7 @@
+name: "CodeQL config"
+queries: 
+  - uses: security-extended
+  - uses: security-and-quality
+paths-ignore:
+  - tests
+  - build
\ No newline at end of file
diff --git a/.github/workflows/cffconvert.yml b/.github/workflows/cffconvert.yml
index 7144363717749..0cbaf24059390 100644
--- a/.github/workflows/cffconvert.yml
+++ b/.github/workflows/cffconvert.yml
@@ -8,7 +8,7 @@ on:
 jobs:
   validate:
     name: "validate"
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - name: Check out a copy of the repository
         uses: actions/checkout@v4
diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml
index e4d1b91bab736..d3b51c0681a20 100644
--- a/.github/workflows/codeql.yml
+++ b/.github/workflows/codeql.yml
@@ -18,7 +18,7 @@ on:
 jobs:
   analyze:
     name: Analyze
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     permissions:
       actions: read
       contents: read
@@ -55,6 +55,11 @@ jobs:
         java-version: '11'
         distribution: 'microsoft'
 
+    - if: ${{ matrix.language == 'javascript' }}
+      uses: actions/setup-node@v4
+      with:
+        node-version: 20
+
     # Autobuild attempts to build any compiled languages  (C/C++, C#, or Java).
     # If this step fails, then you should remove it and run the build manually (see below)
     - if: ${{ matrix.language != 'cpp' }}
diff --git a/.github/workflows/gradle-wrapper-validation.yml b/.github/workflows/gradle-wrapper-validation.yml
index 32aed81092774..cf3bc598d02bb 100644
--- a/.github/workflows/gradle-wrapper-validation.yml
+++ b/.github/workflows/gradle-wrapper-validation.yml
@@ -8,7 +8,7 @@ on: [push, pull_request]
 jobs:
   validation:
     name: "Validation"
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
       - uses: gradle/actions/wrapper-validation@v4
diff --git a/.github/workflows/labeler.yml b/.github/workflows/labeler.yml
index a196226a4b836..00960c848b107 100644
--- a/.github/workflows/labeler.yml
+++ b/.github/workflows/labeler.yml
@@ -8,7 +8,7 @@ permissions:
 
 jobs:
   triage:
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
     - uses: github/issue-labeler@v3.4
       with:
diff --git a/.github/workflows/lint.yml b/.github/workflows/lint.yml
index 74fd5d7dfdb28..ec834b07b2c78 100644
--- a/.github/workflows/lint.yml
+++ b/.github/workflows/lint.yml
@@ -36,7 +36,7 @@ jobs:
   lint-python-format:
     # Required workflow
     name: Python format
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
       - name: Setup Python
@@ -114,9 +114,12 @@ jobs:
 
   lint-js:
     name: Lint JavaScript
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 20
       - uses: reviewdog/action-eslint@v1
         with:
           reporter: github-pr-check
diff --git a/.github/workflows/linux_training.yml b/.github/workflows/linux_training.yml
new file mode 100644
index 0000000000000..d382cdf476283
--- /dev/null
+++ b/.github/workflows/linux_training.yml
@@ -0,0 +1,55 @@
+name: orttraining-linux-ci-pipeline
+on:
+  push:
+    branches:
+      - main
+      - rel-*
+  pull_request:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+  
+jobs:
+  orttraining-linux-ci-pipeline:
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
+    permissions:
+      actions: read
+      contents: read
+      security-events: write
+    steps:
+      - uses: actions/checkout@v4
+      - run: |
+         python3 -m pip install --user -r tools/ci_build/github/linux/python/requirements.txt
+      - name: Initialize CodeQL
+        uses: github/codeql-action/init@v3
+        with:
+          config-file: ./.github/codeql/codeql-config.yml
+          languages: 'cpp'
+      - run: |
+         set -e -x
+         rm -rf build
+         python3 tools/ci_build/build.py --build_dir build --config Release --enable_training --skip_submodule_sync --parallel --update --build
+         
+      - name: Perform CodeQL Analysis
+        uses: github/codeql-action/analyze@v3
+        with:
+          category: "/language:cpp"
+          output: sarif-results
+          upload: failure-only
+        
+      - name: filter-sarif
+        uses: advanced-security/filter-sarif@v1
+        with:
+          patterns: |
+            +**/*.cc
+            +**/*.h
+            -tests/**/*.*
+            -build/**/*.*
+          input: sarif-results/cpp.sarif
+          output: sarif-results/cpp.sarif
+
+      - name: Upload SARIF
+        uses: github/codeql-action/upload-sarif@v3
+        with:
+          sarif_file: sarif-results/cpp.sarif
\ No newline at end of file
diff --git a/.github/workflows/mac.yml b/.github/workflows/mac.yml
index b36b0aa555940..f0669f94c2b1c 100644
--- a/.github/workflows/mac.yml
+++ b/.github/workflows/mac.yml
@@ -105,7 +105,7 @@ jobs:
           --use_xnnpack \
           --use_binskim_compliant_compile_flags \
           --ios \
-          --apple_deploy_target=13.0 \
+          --apple_deploy_target=15.1 \
           --apple_sysroot=iphonesimulator \
           --osx_arch=${{ matrix.target_arch }}
 
diff --git a/.github/workflows/pr_checks.yml b/.github/workflows/pr_checks.yml
index a711b753c492d..152a2bfc13941 100644
--- a/.github/workflows/pr_checks.yml
+++ b/.github/workflows/pr_checks.yml
@@ -19,7 +19,7 @@ concurrency:
 jobs:
   auto-apply-fixes:
     name: Suggest fixes
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     permissions:
       contents: read
       pull-requests: write
@@ -34,14 +34,13 @@ jobs:
         with:
           toolchain: stable
           components: rustfmt
-      - name: Install dependencies
+      - name: Install dependencies and run lintrunner on all files
         run: |
-          python -m pip install -r requirements-dev.txt
-          python -m pip install lintrunner lintrunner-adapters
+	  export PATH=$HOME/.local/bin:$PATH
+	  set -e
+          python -m pip install --user -r requirements-dev.txt
+          python -m pip install --user lintrunner lintrunner-adapters
           lintrunner init
-      - name: Run lintrunner on all files
-        run: |
-          set +e
           lintrunner f --all-files -v
           exit 0
       - uses: parkerbxyz/suggest-changes@v1
diff --git a/.github/workflows/publish-c-apidocs.yml b/.github/workflows/publish-c-apidocs.yml
index 6c4dc43847d0b..303a6f1df6a68 100644
--- a/.github/workflows/publish-c-apidocs.yml
+++ b/.github/workflows/publish-c-apidocs.yml
@@ -22,7 +22,7 @@ permissions:
 jobs:
   build:
     name: Generate C/C++ API docs
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
       - name: Install doxygen and dependencies
diff --git a/.github/workflows/publish-csharp-apidocs.yml b/.github/workflows/publish-csharp-apidocs.yml
index 862a7a70e33a2..ece0f6c534071 100644
--- a/.github/workflows/publish-csharp-apidocs.yml
+++ b/.github/workflows/publish-csharp-apidocs.yml
@@ -20,7 +20,7 @@ permissions:
 
 jobs:
   build:
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     env:
       DOCFXVERSION: 2.62.2
     steps:
diff --git a/.github/workflows/publish-gh-pages.yml b/.github/workflows/publish-gh-pages.yml
index 1818261b4b766..11745ce24f9e5 100644
--- a/.github/workflows/publish-gh-pages.yml
+++ b/.github/workflows/publish-gh-pages.yml
@@ -8,7 +8,7 @@ on:
 
 jobs:
   placeholder:
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - name: Placeholder step to have workflow included in the GitHub web UI
         run: |
diff --git a/.github/workflows/publish-java-apidocs.yml b/.github/workflows/publish-java-apidocs.yml
index 9e42dca708a17..cb2c4cdedc8ae 100644
--- a/.github/workflows/publish-java-apidocs.yml
+++ b/.github/workflows/publish-java-apidocs.yml
@@ -21,7 +21,7 @@ permissions:
 jobs:
   build:
     name: Generate Java docs
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
       - name: Set up JDK 11
diff --git a/.github/workflows/publish-js-apidocs.yml b/.github/workflows/publish-js-apidocs.yml
index cec4a52d39c93..5aaac55be52c7 100644
--- a/.github/workflows/publish-js-apidocs.yml
+++ b/.github/workflows/publish-js-apidocs.yml
@@ -21,7 +21,7 @@ permissions:
 jobs:
   build:
     name: Generate JS API docs
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
       - name: Setup Node.js
diff --git a/.github/workflows/publish-python-apidocs.yml b/.github/workflows/publish-python-apidocs.yml
index 8b2f72d80bacf..7622416d314e1 100644
--- a/.github/workflows/publish-python-apidocs.yml
+++ b/.github/workflows/publish-python-apidocs.yml
@@ -22,7 +22,7 @@ permissions:
 jobs:
   build:
     name: Generate Python API docs
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - uses: actions/checkout@v4
       - name: Install tools
diff --git a/.github/workflows/skip-doc-change.yml.j2 b/.github/workflows/skip-doc-change.yml.j2
index 58f048122a87e..04f77e5d28713 100644
--- a/.github/workflows/skip-doc-change.yml.j2
+++ b/.github/workflows/skip-doc-change.yml.j2
@@ -14,7 +14,7 @@ jobs:
 {%- for name in job_names %}
   job{{ loop.index }}:
     name: {{ name }}
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
       - run: 'echo "No build required, only documentation changed"'
 {% endfor %}
diff --git a/.github/workflows/stale.yml b/.github/workflows/stale.yml
index 181f3fb17d332..14cf0825873a0 100644
--- a/.github/workflows/stale.yml
+++ b/.github/workflows/stale.yml
@@ -8,7 +8,7 @@ on:
 
 jobs:
   close-stale-issues:
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     permissions:
       issues: write
       pull-requests: write
diff --git a/.github/workflows/title-only-labeler.yml b/.github/workflows/title-only-labeler.yml
index e0af2dd06b1b7..7ee9f3917a901 100644
--- a/.github/workflows/title-only-labeler.yml
+++ b/.github/workflows/title-only-labeler.yml
@@ -8,7 +8,7 @@ permissions:
 
 jobs:
   triage:
-    runs-on: ubuntu-latest
+    runs-on: ["self-hosted", "1ES.Pool=onnxruntime-github-Ubuntu2204-AMD-CPU"]
     steps:
     - uses: github/issue-labeler@v3.4
       with:
diff --git a/VERSION_NUMBER b/VERSION_NUMBER
index 0044d6cb9691c..769e37e159d42 100644
--- a/VERSION_NUMBER
+++ b/VERSION_NUMBER
@@ -1 +1 @@
-1.20.1
+1.20.2
diff --git a/csharp/test/Microsoft.ML.OnnxRuntime.Tests.MAUI/Microsoft.ML.OnnxRuntime.Tests.MAUI.csproj b/csharp/test/Microsoft.ML.OnnxRuntime.Tests.MAUI/Microsoft.ML.OnnxRuntime.Tests.MAUI.csproj
index e07448daeea7f..b4f1ea5a4f7fe 100644
--- a/csharp/test/Microsoft.ML.OnnxRuntime.Tests.MAUI/Microsoft.ML.OnnxRuntime.Tests.MAUI.csproj
+++ b/csharp/test/Microsoft.ML.OnnxRuntime.Tests.MAUI/Microsoft.ML.OnnxRuntime.Tests.MAUI.csproj
@@ -7,7 +7,9 @@
 
     <!-- General app properties -->
     <PropertyGroup>
-        <TargetFrameworks>net8.0-android;net8.0-ios;net8.0-maccatalyst</TargetFrameworks>
+        <!-- TODO: Add net8.0-android target back in once IndexOutOfRangeException for MAUI net8.0-android issue is
+        resolved. -->
+        <TargetFrameworks>net8.0-ios;net8.0-maccatalyst</TargetFrameworks>
         <TargetFrameworks Condition="$([MSBuild]::IsOSPlatform('windows'))">$(TargetFrameworks);net8.0-windows10.0.19041.0</TargetFrameworks>
 
         <!-- Note for MacCatalyst:
diff --git a/docs/python/README.rst b/docs/python/README.rst
index 82c2fbde1d1d8..8632e4b2cb127 100644
--- a/docs/python/README.rst
+++ b/docs/python/README.rst
@@ -8,6 +8,11 @@ For more information on ONNX Runtime, please see `aka.ms/onnxruntime <https://ak
 Changes
 -------
 
+1.20.2
+^^^^^^
+
+Release Notes : https://github.com/Microsoft/onnxruntime/releases/tag/v1.20.2
+
 1.20.1
 ^^^^^^
 
diff --git a/include/onnxruntime/core/session/onnxruntime_c_api.h b/include/onnxruntime/core/session/onnxruntime_c_api.h
index bde27df94ed1c..ab344d367a426 100644
--- a/include/onnxruntime/core/session/onnxruntime_c_api.h
+++ b/include/onnxruntime/core/session/onnxruntime_c_api.h
@@ -3660,8 +3660,8 @@ struct OrtApi {
    *     - "1": Enabled.
    *   "offload_graph_io_quantization": Offload graph input quantization and graph output dequantization to another
    *   execution provider (typically CPU EP).
-   *     - "0": Default. Disabled. QNN EP will handle quantization and dequantization of graph I/O.
-   *     - "1": Enabled.
+   *     - "0": Disabled. QNN EP will handle quantization and dequantization of graph I/O.
+   *     - "1": Enabled. This is the default value.
    *
    * SNPE supported keys:
    *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",
diff --git a/js/common/lib/version.ts b/js/common/lib/version.ts
index 784b80f603acf..d48aaf1427f96 100644
--- a/js/common/lib/version.ts
+++ b/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 = '1.20.1';
+export const version = '1.20.2';
diff --git a/js/common/package-lock.json b/js/common/package-lock.json
index 03b8b4f0cc9a7..634f172c8d5c6 100644
--- a/js/common/package-lock.json
+++ b/js/common/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "onnxruntime-common",
-  "version": "1.20.1",
+  "version": "1.20.2",
   "lockfileVersion": 2,
   "requires": true,
   "packages": {
     "": {
       "name": "onnxruntime-common",
-      "version": "1.20.1",
+      "version": "1.20.2",
       "license": "MIT",
       "devDependencies": {
         "typedoc": "^0.25.7"
diff --git a/js/common/package.json b/js/common/package.json
index c483b41dfdce9..84daa98bf1c57 100644
--- a/js/common/package.json
+++ b/js/common/package.json
@@ -2,7 +2,7 @@
   "license": "MIT",
   "type": "module",
   "name": "onnxruntime-common",
-  "version": "1.20.1",
+  "version": "1.20.2",
   "repository": {
     "url": "https://github.com/Microsoft/onnxruntime.git",
     "type": "git"
diff --git a/js/node/lib/version.ts b/js/node/lib/version.ts
index 784b80f603acf..d48aaf1427f96 100644
--- a/js/node/lib/version.ts
+++ b/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 = '1.20.1';
+export const version = '1.20.2';
diff --git a/js/node/package-lock.json b/js/node/package-lock.json
index 633c7cd62f9f6..0dc500e89f613 100644
--- a/js/node/package-lock.json
+++ b/js/node/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "onnxruntime-node",
-  "version": "1.20.1",
+  "version": "1.20.2",
   "lockfileVersion": 2,
   "requires": true,
   "packages": {
     "": {
       "name": "onnxruntime-node",
-      "version": "1.20.1",
+      "version": "1.20.2",
       "hasInstallScript": true,
       "license": "MIT",
       "os": [
@@ -29,7 +29,7 @@
     },
     "../common": {
       "name": "onnxruntime-common",
-      "version": "1.20.1",
+      "version": "1.20.2",
       "license": "MIT",
       "devDependencies": {
         "typedoc": "^0.25.7"
diff --git a/js/node/package.json b/js/node/package.json
index 3842df7edf522..3fb8b0071ee99 100644
--- a/js/node/package.json
+++ b/js/node/package.json
@@ -13,7 +13,7 @@
       3
     ]
   },
-  "version": "1.20.1",
+  "version": "1.20.2",
   "dependencies": {
     "onnxruntime-common": "file:../common",
     "tar": "^7.0.1"
diff --git a/js/react_native/lib/version.ts b/js/react_native/lib/version.ts
index 784b80f603acf..d48aaf1427f96 100644
--- a/js/react_native/lib/version.ts
+++ b/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 = '1.20.1';
+export const version = '1.20.2';
diff --git a/js/react_native/package.json b/js/react_native/package.json
index 1acfd69ec84f2..e14be3bc0ae58 100644
--- a/js/react_native/package.json
+++ b/js/react_native/package.json
@@ -36,7 +36,7 @@
     "registry": "https://registry.npmjs.org/"
   },
   "source": "lib/index",
-  "version": "1.20.1",
+  "version": "1.20.2",
   "main": "dist/commonjs/index",
   "homepage": "https://github.com/microsoft/onnxruntime/blob/main/js/react_native/README.md",
   "files": [
diff --git a/js/react_native/yarn.lock b/js/react_native/yarn.lock
index c9eba883944d7..12c6fd5c12afd 100644
--- a/js/react_native/yarn.lock
+++ b/js/react_native/yarn.lock
@@ -5254,7 +5254,7 @@ onetime@^5.1.0, onetime@^5.1.2:
     mimic-fn "^2.1.0"
 
 "onnxruntime-common@file:../common":
-  version "1.20.1"
+  version "1.20.2"
 
 open@^6.2.0:
   version "6.4.0"
diff --git a/js/web/lib/version.ts b/js/web/lib/version.ts
index 784b80f603acf..d48aaf1427f96 100644
--- a/js/web/lib/version.ts
+++ b/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 = '1.20.1';
+export const version = '1.20.2';
diff --git a/js/web/package-lock.json b/js/web/package-lock.json
index 7f289cc914d42..291ed72e23f13 100644
--- a/js/web/package-lock.json
+++ b/js/web/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "onnxruntime-web",
-  "version": "1.20.1",
+  "version": "1.20.2",
   "lockfileVersion": 2,
   "requires": true,
   "packages": {
     "": {
       "name": "onnxruntime-web",
-      "version": "1.20.1",
+      "version": "1.20.2",
       "license": "MIT",
       "dependencies": {
         "flatbuffers": "^1.12.0",
@@ -51,7 +51,7 @@
     },
     "../common": {
       "name": "onnxruntime-common",
-      "version": "1.20.1",
+      "version": "1.20.2",
       "license": "MIT",
       "devDependencies": {
         "typedoc": "^0.25.7"
diff --git a/js/web/package.json b/js/web/package.json
index d5dba18c14a59..858fc96f57afc 100644
--- a/js/web/package.json
+++ b/js/web/package.json
@@ -7,7 +7,7 @@
     "type": "git"
   },
   "author": "fs-eire",
-  "version": "1.20.1",
+  "version": "1.20.2",
   "jsdelivr": "dist/ort.min.js",
   "dependencies": {
     "flatbuffers": "^1.12.0",
diff --git a/js/web/test/data/ops/conv.jsonc b/js/web/test/data/ops/conv.jsonc
index 262503214a50a..f514ae5fa75e6 100644
--- a/js/web/test/data/ops/conv.jsonc
+++ b/js/web/test/data/ops/conv.jsonc
@@ -391,48 +391,48 @@
       }
     ]
   },
-  {
-    "name": "conv - vectorize group - B",
-    "operator": "Conv",
-    "inputShapeDefinitions": "rankOnly",
-    "opset": { "domain": "", "version": 17 },
-    "attributes": [
-      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
-      { "name": "group", "data": 3, "type": "int" }
-    ],
-    "cases": [
-      {
-        "name": "T[0]",
-        "inputs": [
-          {
-            "data": [
-              0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0,
-              19.0, 20.0, 21.0, 22.0, 23.0, 0, 0, 0
-            ],
-            "dims": [1, 3, 3, 3],
-            "type": "float32"
-          },
-          {
-            "data": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0],
-            "dims": [3, 1, 2, 2],
-            "type": "float32"
-          },
-          {
-            "data": [0.1, 0.2, 0.3],
-            "dims": [3],
-            "type": "float32"
-          }
-        ],
-        "outputs": [
-          {
-            "data": [27.1, 37.1, 57.1, 67.1, 293.2, 319.2, 371.2, 397.2, 847.3, 889.3, 409.3, 428.3],
-            "dims": [1, 3, 2, 2],
-            "type": "float32"
-          }
-        ]
-      }
-    ]
-  },
+  // {
+  //   "name": "conv - vectorize group - B",
+  //   "operator": "Conv",
+  //   "inputShapeDefinitions": "rankOnly",
+  //   "opset": { "domain": "", "version": 17 },
+  //   "attributes": [
+  //     { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+  //     { "name": "group", "data": 3, "type": "int" }
+  //   ],
+  //   "cases": [
+  //     {
+  //       "name": "T[0]",
+  //       "inputs": [
+  //         {
+  //           "data": [
+  //             0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0,
+  //             19.0, 20.0, 21.0, 22.0, 23.0, 0, 0, 0
+  //           ],
+  //           "dims": [1, 3, 3, 3],
+  //           "type": "float32"
+  //         },
+  //         {
+  //           "data": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0],
+  //           "dims": [3, 1, 2, 2],
+  //           "type": "float32"
+  //         },
+  //         {
+  //           "data": [0.1, 0.2, 0.3],
+  //           "dims": [3],
+  //           "type": "float32"
+  //         }
+  //       ],
+  //       "outputs": [
+  //         {
+  //           "data": [27.1, 37.1, 57.1, 67.1, 293.2, 319.2, 371.2, 397.2, 847.3, 889.3, 409.3, 428.3],
+  //           "dims": [1, 3, 2, 2],
+  //           "type": "float32"
+  //         }
+  //       ]
+  //     }
+  //   ]
+  // },
   {
     "name": "conv - vectorize group - C",
     "operator": "Conv",
@@ -470,44 +470,44 @@
       }
     ]
   },
-  {
-    "name": "conv - vectorize group - D",
-    "operator": "Conv",
-    "inputShapeDefinitions": "rankOnly",
-    "opset": { "domain": "", "version": 17 },
-    "attributes": [
-      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
-      { "name": "group", "data": 3, "type": "int" },
-      { "name": "strides", "data": [2, 2], "type": "ints" }
-    ],
-    "cases": [
-      {
-        "name": "T[0] strides = [2, 2]",
-        "inputs": [
-          {
-            "data": [
-              0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0,
-              19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0, 29.0, 30.0, 31.0, 32.0, 33.0, 34.0, 35.0
-            ],
-            "dims": [1, 3, 3, 4],
-            "type": "float32"
-          },
-          {
-            "data": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0],
-            "dims": [3, 1, 2, 2],
-            "type": "float32"
-          }
-        ],
-        "outputs": [
-          {
-            "data": [34, 54, 386, 438, 1122, 1206],
-            "dims": [1, 3, 1, 2],
-            "type": "float32"
-          }
-        ]
-      }
-    ]
-  },
+  // {
+  //   "name": "conv - vectorize group - D",
+  //   "operator": "Conv",
+  //   "inputShapeDefinitions": "rankOnly",
+  //   "opset": { "domain": "", "version": 17 },
+  //   "attributes": [
+  //     { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+  //     { "name": "group", "data": 3, "type": "int" },
+  //     { "name": "strides", "data": [2, 2], "type": "ints" }
+  //   ],
+  //   "cases": [
+  //     {
+  //       "name": "T[0] strides = [2, 2]",
+  //       "inputs": [
+  //         {
+  //           "data": [
+  //             0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0,
+  //             19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0, 29.0, 30.0, 31.0, 32.0, 33.0, 34.0, 35.0
+  //           ],
+  //           "dims": [1, 3, 3, 4],
+  //           "type": "float32"
+  //         },
+  //         {
+  //           "data": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0],
+  //           "dims": [3, 1, 2, 2],
+  //           "type": "float32"
+  //         }
+  //       ],
+  //       "outputs": [
+  //         {
+  //           "data": [34, 54, 386, 438, 1122, 1206],
+  //           "dims": [1, 3, 1, 2],
+  //           "type": "float32"
+  //         }
+  //       ]
+  //     }
+  //   ]
+  // },
   {
     "name": "conv - pointwise",
     "operator": "Conv",
diff --git a/js/web/test/data/ops/fused-conv.jsonc b/js/web/test/data/ops/fused-conv.jsonc
index d88c91ebc9de7..ebb0b5d3e1f58 100644
--- a/js/web/test/data/ops/fused-conv.jsonc
+++ b/js/web/test/data/ops/fused-conv.jsonc
@@ -249,44 +249,44 @@
       }
     ]
   },
-  {
-    "name": "NHWC group-conv with HardSigmoid",
-    "operator": "Conv",
-    "attributes": [
-      { "name": "activation", "data": "HardSigmoid", "type": "string" },
-      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
-      { "name": "group", "data": 3, "type": "int" },
-      { "name": "activation_params", "data": [2.0, 5.0], "type": "floats" }
-    ],
-    "opset": { "domain": "com.ms.internal.nhwc", "version": 1 },
-    "cases": [
-      {
-        "name": "T[0]",
-        "inputs": [
-          {
-            "data": [
-              0.0, 1.0, 2.0, -3.0, 4.0, -5.0, 6.0, 7.0, 8.0, -9.0, -10.0, 11.0, -12.0, 13.0, -14.0, 15.0, 16.0, 17.0,
-              18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0
-            ],
-            "dims": [1, 3, 3, 3],
-            "type": "float32"
-          },
-          {
-            "data": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
-            "dims": [3, 1, 2, 2],
-            "type": "float32"
-          }
-        ],
-        "outputs": [
-          {
-            "data": [0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-            "dims": [1, 2, 2, 3],
-            "type": "float32"
-          }
-        ]
-      }
-    ]
-  },
+  // {
+  //   "name": "NHWC group-conv with HardSigmoid",
+  //   "operator": "Conv",
+  //   "attributes": [
+  //     { "name": "activation", "data": "HardSigmoid", "type": "string" },
+  //     { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+  //     { "name": "group", "data": 3, "type": "int" },
+  //     { "name": "activation_params", "data": [2.0, 5.0], "type": "floats" }
+  //   ],
+  //   "opset": { "domain": "com.ms.internal.nhwc", "version": 1 },
+  //   "cases": [
+  //     {
+  //       "name": "T[0]",
+  //       "inputs": [
+  //         {
+  //           "data": [
+  //             0.0, 1.0, 2.0, -3.0, 4.0, -5.0, 6.0, 7.0, 8.0, -9.0, -10.0, 11.0, -12.0, 13.0, -14.0, 15.0, 16.0, 17.0,
+  //             18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0
+  //           ],
+  //           "dims": [1, 3, 3, 3],
+  //           "type": "float32"
+  //         },
+  //         {
+  //           "data": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
+  //           "dims": [3, 1, 2, 2],
+  //           "type": "float32"
+  //         }
+  //       ],
+  //       "outputs": [
+  //         {
+  //           "data": [0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+  //           "dims": [1, 2, 2, 3],
+  //           "type": "float32"
+  //         }
+  //       ]
+  //     }
+  //   ]
+  // },
   {
     "name": "fused group-conv with LeakyRelu",
     "operator": "FusedConv",
@@ -325,44 +325,44 @@
       }
     ]
   },
-  {
-    "name": "NHWC group-conv with LeakyRelu",
-    "operator": "Conv",
-    "attributes": [
-      { "name": "activation", "data": "LeakyRelu", "type": "string" },
-      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
-      { "name": "group", "data": 3, "type": "int" },
-      { "name": "activation_params", "data": [2.0], "type": "floats" }
-    ],
-    "opset": { "domain": "com.ms.internal.nhwc", "version": 1 },
-    "cases": [
-      {
-        "name": "T[0]",
-        "inputs": [
-          {
-            "data": [
-              0.0, 1.0, 2.0, -3.0, 4.0, -5.0, 6.0, 7.0, 8.0, -9.0, -10.0, 11.0, -12.0, 13.0, -14.0, 15.0, 16.0, 17.0,
-              18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0
-            ],
-            "dims": [1, 3, 3, 3],
-            "type": "float32"
-          },
-          {
-            "data": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
-            "dims": [3, 1, 2, 2],
-            "type": "float32"
-          }
-        ],
-        "outputs": [
-          {
-            "data": [-162, 63, -158, 33, 281, 85, 105, 337, 455, 177, 515, 609],
-            "dims": [1, 2, 2, 3],
-            "type": "float32"
-          }
-        ]
-      }
-    ]
-  },
+  // {
+  //   "name": "NHWC group-conv with LeakyRelu",
+  //   "operator": "Conv",
+  //   "attributes": [
+  //     { "name": "activation", "data": "LeakyRelu", "type": "string" },
+  //     { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+  //     { "name": "group", "data": 3, "type": "int" },
+  //     { "name": "activation_params", "data": [2.0], "type": "floats" }
+  //   ],
+  //   "opset": { "domain": "com.ms.internal.nhwc", "version": 1 },
+  //   "cases": [
+  //     {
+  //       "name": "T[0]",
+  //       "inputs": [
+  //         {
+  //           "data": [
+  //             0.0, 1.0, 2.0, -3.0, 4.0, -5.0, 6.0, 7.0, 8.0, -9.0, -10.0, 11.0, -12.0, 13.0, -14.0, 15.0, 16.0, 17.0,
+  //             18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0
+  //           ],
+  //           "dims": [1, 3, 3, 3],
+  //           "type": "float32"
+  //         },
+  //         {
+  //           "data": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
+  //           "dims": [3, 1, 2, 2],
+  //           "type": "float32"
+  //         }
+  //       ],
+  //       "outputs": [
+  //         {
+  //           "data": [-162, 63, -158, 33, 281, 85, 105, 337, 455, 177, 515, 609],
+  //           "dims": [1, 2, 2, 3],
+  //           "type": "float32"
+  //         }
+  //       ]
+  //     }
+  //   ]
+  // },
   {
     "name": "fused conv with LeakyRelu",
     "operator": "FusedConv",
diff --git a/onnxruntime/__init__.py b/onnxruntime/__init__.py
index cded663706ff6..0a0384d571711 100644
--- a/onnxruntime/__init__.py
+++ b/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__ = "1.20.1"
+__version__ = "1.20.2"
 __author__ = "Microsoft"
 
 # we need to do device version validation (for example to check Cuda version for an onnxruntime-training package).
diff --git a/onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_fwd_kernel.h b/onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_fwd_kernel.h
index e961bab399326..d46d9597a758f 100644
--- a/onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_fwd_kernel.h
+++ b/onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_fwd_kernel.h
@@ -98,7 +98,7 @@ inline __device__ void compute_attn_1rowblock(const Params& params, const int bi
       for (int m = 0; m < size<1>(tOgO); ++m) {
         const int row = get<0>(tOcO(0, m, 0));
         if (row < binfo.actual_seqlen_q - m_block * kBlockM && get<1>(tOcO(0, m, 0)) == 0) {
-          gLSE(row) = INFINITY;
+          gLSE(row) = std::numeric_limits<ElementAccum>::infinity();
         }
       }
       return;
@@ -499,7 +499,7 @@ inline __device__ void compute_attn_1rowblock_splitkv(const Params& params, cons
     for (int m = 0; m < size<1>(tOgOaccum); ++m) {
       const int row = get<0>(tOcO(0, m, 0));
       if (row < binfo.actual_seqlen_q - m_block * kBlockM && get<1>(tOcO(0, m, 0)) == 0) {
-        gLSEaccum(row) = Split ? -INFINITY : INFINITY;
+        gLSEaccum(row) = Split ? -std::numeric_limits<ElementAccum>::infinity() : std::numeric_limits<ElementAccum>::infinity();
       }
     }
     return;
@@ -1061,7 +1061,7 @@ inline __device__ void combine_attn_seqk_parallel(const Params& params) {
   for (int l = 0; l < kNLsePerThread; ++l) {
     const int row = l * kRowsPerLoadLSE + tidx / kBlockM;
     const int col = tidx % kBlockM;
-    ElementAccum lse = (row < params.num_splits && col < params.b * params.h * params.seqlen_q - bidx * kBlockM) ? gLSEaccum(row, col) : -INFINITY;
+    ElementAccum lse = (row < params.num_splits && col < params.b * params.h * params.seqlen_q - bidx * kBlockM) ? gLSEaccum(row, col) : -std::numeric_limits<ElementAccum>::infinity();
     if (row < kMaxSplits) {
       sLSE[row][col] = lse;
     }
@@ -1082,7 +1082,7 @@ inline __device__ void combine_attn_seqk_parallel(const Params& params) {
   for (int l = 0; l < kNLsePerThread; ++l) {
     const int row = l * kRowsPerLoadTranspose + tidx % kRowsPerLoadTranspose;
     const int col = tidx / kRowsPerLoadTranspose;
-    lse_accum(l) = (row < kMaxSplits && col < kBlockM) ? sLSE[row][col] : -INFINITY;
+    lse_accum(l) = (row < kMaxSplits && col < kBlockM) ? sLSE[row][col] : -std::numeric_limits<ElementAccum>::infinity();
     // if (bidx == 0 && tidx < 32) { printf("tidx = %d, row = %d, col = %d, lse = %f\n", tidx, row, col, lse_accum(l)); }
   }
 
@@ -1094,7 +1094,7 @@ inline __device__ void combine_attn_seqk_parallel(const Params& params) {
   }
   MaxOp<float> max_op;
   lse_max = Allreduce<kRowsPerLoadTranspose>::run(lse_max, max_op);
-  lse_max = lse_max == -INFINITY ? 0.0f : lse_max;  // In case all local LSEs are -inf
+  lse_max = lse_max == -std::numeric_limits<ElementAccum>::infinity() ? 0.0f : lse_max;  // In case all local LSEs are -inf
   float lse_sum = expf(lse_accum(0) - lse_max);
 #pragma unroll
   for (int l = 1; l < kNLsePerThread; ++l) {
@@ -1104,7 +1104,7 @@ inline __device__ void combine_attn_seqk_parallel(const Params& params) {
   lse_sum = Allreduce<kRowsPerLoadTranspose>::run(lse_sum, sum_op);
   // For the case where all local lse == -INFINITY, we want to set lse_logsum to INFINITY. Otherwise
   // lse_logsum is log(0.0) = -INFINITY and we get NaN when we do lse_accum(l) - lse_logsum.
-  ElementAccum lse_logsum = (lse_sum == 0.f || lse_sum != lse_sum) ? INFINITY : logf(lse_sum) + lse_max;
+  ElementAccum lse_logsum = (lse_sum == 0.f || lse_sum != lse_sum) ? std::numeric_limits<ElementAccum>::infinity() : logf(lse_sum) + lse_max;
   // if (bidx == 0 && tidx < 32) { printf("tidx = %d, lse = %f, lse_max = %f, lse_logsum = %f\n", tidx, lse_accum(0), lse_max, lse_logsum); }
   if (tidx % kRowsPerLoadTranspose == 0 && tidx / kRowsPerLoadTranspose < kBlockM) {
     gLSE(tidx / kRowsPerLoadTranspose) = lse_logsum;
diff --git a/onnxruntime/contrib_ops/cuda/bert/flash_attention/mask.h b/onnxruntime/contrib_ops/cuda/bert/flash_attention/mask.h
index 0998155eba635..71434002f8df1 100644
--- a/onnxruntime/contrib_ops/cuda/bert/flash_attention/mask.h
+++ b/onnxruntime/contrib_ops/cuda/bert/flash_attention/mask.h
@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <limits>
 #include <cute/tensor.hpp>
 
 namespace onnxruntime {
@@ -28,7 +29,7 @@ __forceinline__ __device__ void apply_mask(Tensor<Engine, Layout>& tensor, const
 // Without the "make_coord" we get wrong results
 #pragma unroll
         for (int mi = 0; mi < size<0>(tensor); ++mi) {
-          tensor(mi, make_coord(j, nj)) = -INFINITY;
+          tensor(mi, make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
         }
       }
     }
@@ -59,7 +60,7 @@ __forceinline__ __device__ void apply_mask_local(Tensor<Engine, Layout>& tensor,
         for (int j = 0; j < size<1, 0>(tensor); ++j) {
           const int col_idx = col_idx_base + j;
           if (col_idx >= col_idx_limit_right || (HasWSLeft && col_idx < col_idx_limit_left)) {
-            tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+            tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
           }
         }
       }
@@ -96,7 +97,7 @@ __forceinline__ __device__ void apply_mask_causal_w_idx(
 #pragma unroll
     for (int ni = 0; ni < size<1, 1>(tensor); ++ni) {
       if (col_idx_offset_ + get<1>(idx_rowcol(0, ni)) >= col_idx_limit) {
-        tensor(mi, ni) = -INFINITY;
+        tensor(mi, ni) = -std::numeric_limits<float>::infinity();
       }
     }
     // if (cute::thread0()) {
@@ -151,7 +152,7 @@ struct Mask {
               }
               if constexpr (!Is_even_MN) {
                 if (col_idx >= max_seqlen_k) {
-                  tensor(mi, make_coord(j, nj)) = -INFINITY;
+                  tensor(mi, make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                 }
               }
             }
@@ -181,18 +182,18 @@ struct Mask {
                 }
                 if constexpr (Causal_mask) {
                   if (col_idx >= col_idx_limit_right) {
-                    tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+                    tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                   }
                 }
                 if constexpr (Is_local) {
                   if (col_idx >= col_idx_limit_right || col_idx < col_idx_limit_left) {
-                    tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+                    tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                   }
                 }
                 if constexpr (!Causal_mask && !Is_local && !Is_even_MN) {
                   // Causal and Local already handles MN masking
                   if (col_idx >= max_seqlen_k) {
-                    tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+                    tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                   }
                 }
               }
diff --git a/onnxruntime/contrib_ops/cuda/bert/flash_attention/softmax.h b/onnxruntime/contrib_ops/cuda/bert/flash_attention/softmax.h
index 7e0095cb39bd9..7fe506e01a9b9 100644
--- a/onnxruntime/contrib_ops/cuda/bert/flash_attention/softmax.h
+++ b/onnxruntime/contrib_ops/cuda/bert/flash_attention/softmax.h
@@ -4,6 +4,7 @@
 #pragma once
 
 #include <cmath>
+#include <limits>
 
 #include <cute/tensor.hpp>
 
@@ -71,7 +72,9 @@ __forceinline__ __device__ void scale_apply_exp2(Tensor<Engine0, Layout0>& tenso
     // If max is -inf, then all elements must have been -inf (possibly due to masking).
     // We don't want (-inf - (-inf)) since that would give NaN.
     // If we don't have float around M_LOG2E the multiplication is done in fp64.
-    const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * (Scale_max ? scale : float(M_LOG2E));
+    const float max_scaled = max(mi) == -std::numeric_limits<float>::infinity()
+                                 ? 0.f
+                                 : max(mi) * (Scale_max ? scale : float(M_LOG2E));
 #pragma unroll
     for (int ni = 0; ni < size<1>(tensor); ++ni) {
       // Instead of computing exp(x - max), we compute exp2(x * log_2(e) -
@@ -99,7 +102,7 @@ __forceinline__ __device__ void max_scale_exp2_sum(Tensor<Engine0, Layout0>& ten
     max(mi) = Allreduce<4>::run(max(mi), max_op);
     // If max is -inf, then all elements must have been -inf (possibly due to masking).
     // We don't want (-inf - (-inf)) since that would give NaN.
-    const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * scale;
+    const float max_scaled = max(mi) == -std::numeric_limits<float>::infinity() ? 0.f : max(mi) * scale;
     sum(mi) = 0;
 #pragma unroll
     for (int ni = 0; ni < size<1>(tensor); ++ni) {
@@ -143,7 +146,7 @@ struct Softmax {
       for (int mi = 0; mi < size(row_max); ++mi) {
         float scores_max_cur = !Check_inf
                                    ? row_max(mi)
-                                   : (row_max(mi) == -INFINITY ? 0.0f : row_max(mi));
+                                   : (row_max(mi) == -std::numeric_limits<float>::infinity() ? 0.0f : row_max(mi));
         float scores_scale = exp2f((scores_max_prev(mi) - scores_max_cur) * softmax_scale_log2);
         row_sum(mi) *= scores_scale;
 #pragma unroll
@@ -169,7 +172,9 @@ struct Softmax {
     for (int mi = 0; mi < size<0>(acc_o_rowcol); ++mi) {
       float sum = smooth_softmax ? row_sum(mi) + expf(-row_max(mi) * softmax_scale) : row_sum(mi);
       float inv_sum = (sum == 0.f || sum != sum) ? 1.f : 1.f / sum;
-      lse(mi) = (sum == 0.f || sum != sum) ? (Split ? -INFINITY : INFINITY) : row_max(mi) * softmax_scale + __logf(sum);
+      lse(mi) = (sum == 0.f || sum != sum)
+                    ? (Split ? -std::numeric_limits<float>::infinity() : std::numeric_limits<float>::infinity())
+                    : row_max(mi) * softmax_scale + __logf(sum);
       float scale = inv_sum;
 #pragma unroll
       for (int ni = 0; ni < size<1>(acc_o_rowcol); ++ni) {
diff --git a/onnxruntime/contrib_ops/cuda/bert/lean_attention/lean_fwd_kernel.h b/onnxruntime/contrib_ops/cuda/bert/lean_attention/lean_fwd_kernel.h
index 5be69ea0af55c..bd54b404420e5 100644
--- a/onnxruntime/contrib_ops/cuda/bert/lean_attention/lean_fwd_kernel.h
+++ b/onnxruntime/contrib_ops/cuda/bert/lean_attention/lean_fwd_kernel.h
@@ -825,7 +825,7 @@ inline __device__ void lean_compute_attn_impl_ver3(const Params& params, const i
         const int row = l * kRowsPerLoadLSE + tidx / kBlockM;
         const int col = tidx % kBlockM;
         // We skip the first row = 0, as we already populated it in shared memory.
-        ElementAccum lse = (row > 0 && row < total_splits && col < params.b * params.h * (index_t)params.seqlen_q - row_offset_lseaccum) ? gLSEaccumRead(row, col) : -INFINITY;
+        ElementAccum lse = (row > 0 && row < total_splits && col < params.b * params.h * (index_t)params.seqlen_q - row_offset_lseaccum) ? gLSEaccumRead(row, col) : -std::numeric_limits<ElementAccum>::infinity();
         if (row > 0 && row < kMaxSplits) {
           sLSE(row, col) = lse;
 
@@ -857,7 +857,7 @@ inline __device__ void lean_compute_attn_impl_ver3(const Params& params, const i
       for (int l = 0; l < kNLsePerThread; ++l) {
         const int row = l * kRowsPerLoadTranspose + tidx % kRowsPerLoadTranspose;
         const int col = tidx / kRowsPerLoadTranspose;
-        lse_accum(l) = (row < kMaxSplits && col < kBlockM) ? sLSE(row, col) : -INFINITY;
+        lse_accum(l) = (row < kMaxSplits && col < kBlockM) ? sLSE(row, col) : -std::numeric_limits<ElementAccum>::infinity();
 
 #if defined(DEBUG_LEAN_ATTENTION)
         if (threadIdx.x == 0 && blockIdx.z == tracing_block) {
@@ -874,7 +874,7 @@ inline __device__ void lean_compute_attn_impl_ver3(const Params& params, const i
       }
       MaxOp<float> max_op;
       lse_max = Allreduce<kRowsPerLoadTranspose>::run(lse_max, max_op);
-      lse_max = lse_max == -INFINITY ? 0.0f : lse_max;  // In case all local LSEs are -inf
+      lse_max = lse_max == -std::numeric_limits<ElementAccum>::infinity() ? 0.0f : lse_max;  // In case all local LSEs are -inf
       float lse_sum = expf(lse_accum(0) - lse_max);
 #pragma unroll
       for (int l = 1; l < kNLsePerThread; ++l) {
@@ -884,7 +884,9 @@ inline __device__ void lean_compute_attn_impl_ver3(const Params& params, const i
       lse_sum = Allreduce<kRowsPerLoadTranspose>::run(lse_sum, sum_op);
       // For the case where all local lse == -INFINITY, we want to set lse_logsum to INFINITY. Otherwise
       // lse_logsum is log(0.0) = -INFINITY and we get NaN when we do lse_accum(l) - lse_logsum.
-      ElementAccum lse_logsum = (lse_sum == 0.f || lse_sum != lse_sum) ? INFINITY : logf(lse_sum) + lse_max;
+      ElementAccum lse_logsum = (lse_sum == 0.f || lse_sum != lse_sum)
+                                    ? std::numeric_limits<ElementAccum>::infinity()
+                                    : logf(lse_sum) + lse_max;
 // if (tidx % kRowsPerLoadTranspose == 0 && tidx / kRowsPerLoadTranspose < kBlockM) { gLSE(tidx / kRowsPerLoadTranspose) = lse_logsum; }
 // Store the scales exp(lse - lse_logsum) in shared memory.
 #pragma unroll
diff --git a/onnxruntime/contrib_ops/cuda/bert/lean_attention/mask.h b/onnxruntime/contrib_ops/cuda/bert/lean_attention/mask.h
index d63c80b012de6..2d33418d69667 100644
--- a/onnxruntime/contrib_ops/cuda/bert/lean_attention/mask.h
+++ b/onnxruntime/contrib_ops/cuda/bert/lean_attention/mask.h
@@ -3,7 +3,7 @@
  ******************************************************************************/
 
 #pragma once
-
+#include <limits>
 #include <cute/tensor.hpp>
 
 namespace onnxruntime {
@@ -28,7 +28,7 @@ __forceinline__ __device__ void apply_mask(Tensor<Engine, Layout>& tensor, const
 // Without the "make_coord" we get wrong results
 #pragma unroll
         for (int mi = 0; mi < size<0>(tensor); ++mi) {
-          tensor(mi, make_coord(j, nj)) = -INFINITY;
+          tensor(mi, make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
         }
       }
     }
@@ -59,7 +59,7 @@ __forceinline__ __device__ void apply_mask_local(Tensor<Engine, Layout>& tensor,
         for (int j = 0; j < size<1, 0>(tensor); ++j) {
           const int col_idx = col_idx_base + j;
           if (col_idx >= col_idx_limit_right || (HasWSLeft && col_idx < col_idx_limit_left)) {
-            tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+            tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
           }
         }
       }
@@ -96,7 +96,7 @@ __forceinline__ __device__ void apply_mask_causal_w_idx(
 #pragma unroll
     for (int ni = 0; ni < size<1, 1>(tensor); ++ni) {
       if (col_idx_offset_ + get<1>(idx_rowcol(0, ni)) >= col_idx_limit) {
-        tensor(mi, ni) = -INFINITY;
+        tensor(mi, ni) = -std::numeric_limits<float>::infinity();
       }
     }
     // if (cute::thread0()) {
@@ -152,7 +152,7 @@ struct Mask {
               }
               if constexpr (!Is_even_MN) {
                 if (col_idx >= max_seqlen_k) {
-                  tensor(mi, make_coord(j, nj)) = -INFINITY;
+                  tensor(mi, make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                 }
               }
             }
@@ -182,18 +182,18 @@ struct Mask {
                 }
                 if constexpr (Causal_mask) {
                   if (col_idx >= col_idx_limit_right) {
-                    tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+                    tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                   }
                 }
                 if constexpr (Is_local) {
                   if (col_idx >= col_idx_limit_right || col_idx < col_idx_limit_left) {
-                    tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+                    tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                   }
                 }
                 if constexpr (!Causal_mask && !Is_local && !Is_even_MN) {
                   // Causal and Local already handles MN masking
                   if (col_idx >= max_seqlen_k) {
-                    tensor(make_coord(i, mi), make_coord(j, nj)) = -INFINITY;
+                    tensor(make_coord(i, mi), make_coord(j, nj)) = -std::numeric_limits<float>::infinity();
                   }
                 }
               }
diff --git a/onnxruntime/contrib_ops/cuda/bert/lean_attention/softmax.h b/onnxruntime/contrib_ops/cuda/bert/lean_attention/softmax.h
index ad66389848e6e..0b6ffb3f1985a 100644
--- a/onnxruntime/contrib_ops/cuda/bert/lean_attention/softmax.h
+++ b/onnxruntime/contrib_ops/cuda/bert/lean_attention/softmax.h
@@ -3,7 +3,7 @@
  ******************************************************************************/
 
 #pragma once
-
+#include <limits>
 #include <cmath>
 
 #include <cute/tensor.hpp>
@@ -72,7 +72,9 @@ __forceinline__ __device__ void scale_apply_exp2(Tensor<Engine0, Layout0>& tenso
     // If max is -inf, then all elements must have been -inf (possibly due to masking).
     // We don't want (-inf - (-inf)) since that would give NaN.
     // If we don't have float around M_LOG2E the multiplication is done in fp64.
-    const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * (Scale_max ? scale : float(M_LOG2E));
+    const float max_scaled = max(mi) == -std::numeric_limits<float>::infinity()
+                                 ? 0.f
+                                 : max(mi) * (Scale_max ? scale : float(M_LOG2E));
 #pragma unroll
     for (int ni = 0; ni < size<1>(tensor); ++ni) {
 // Instead of computing exp(x - max), we compute exp2(x * log_2(e) -
@@ -107,7 +109,7 @@ __forceinline__ __device__ void max_scale_exp2_sum(Tensor<Engine0, Layout0>& ten
     max(mi) = Allreduce<4>::run(max(mi), max_op);
     // If max is -inf, then all elements must have been -inf (possibly due to masking).
     // We don't want (-inf - (-inf)) since that would give NaN.
-    const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * scale;
+    const float max_scaled = max(mi) == -std::numeric_limits<float>::infinity() ? 0.f : max(mi) * scale;
     sum(mi) = 0;
 #pragma unroll
     for (int ni = 0; ni < size<1>(tensor); ++ni) {
@@ -151,7 +153,7 @@ struct Softmax {
       for (int mi = 0; mi < size(row_max); ++mi) {
         float scores_max_cur = !Check_inf
                                    ? row_max(mi)
-                                   : (row_max(mi) == -INFINITY ? 0.0f : row_max(mi));
+                                   : (row_max(mi) == -std::numeric_limits<float>::infinity() ? 0.0f : row_max(mi));
         float scores_scale = exp2f((scores_max_prev(mi) - scores_max_cur) * softmax_scale_log2);
         row_sum(mi) *= scores_scale;
 #pragma unroll
@@ -181,7 +183,9 @@ struct Softmax {
       //     printf("sum: %f, inv_sum: %f\n", sum, inv_sum);
       //     printf("mi %d row_max %f softmax_scale %f\n", mi, row_max(mi), softmax_scale);
       // }
-      lse(mi) = (sum == 0.f || sum != sum) ? (Split ? -INFINITY : INFINITY) : row_max(mi) * softmax_scale + __logf(sum);
+      lse(mi) = (sum == 0.f || sum != sum)
+                    ? (Split ? -std::numeric_limits<float>::infinity() : std::numeric_limits<float>::infinity())
+                    : row_max(mi) * softmax_scale + __logf(sum);
       float scale = !Is_dropout ? inv_sum : inv_sum * rp_dropout;
 #pragma unroll
       for (int ni = 0; ni < size<1>(acc_o_rowcol); ++ni) {
diff --git a/onnxruntime/contrib_ops/cuda/bert/ngram_repeat_block_impl.cu b/onnxruntime/contrib_ops/cuda/bert/ngram_repeat_block_impl.cu
index 8a04ede231a27..ab809d12a89ad 100644
--- a/onnxruntime/contrib_ops/cuda/bert/ngram_repeat_block_impl.cu
+++ b/onnxruntime/contrib_ops/cuda/bert/ngram_repeat_block_impl.cu
@@ -6,7 +6,7 @@ Licensed under the MIT License.
 /*
 Kernel implementation for blocking repeated n-grams.
 */
-
+#include <limits>
 #include "core/providers/cuda/cu_inc/common.cuh"
 #include "contrib_ops/cuda/bert/ngram_repeat_block_impl.h"
 
@@ -48,7 +48,7 @@ __global__ void banRepeatedTokens(const int64_t* __restrict__ tokens,
   }
   if (is_banned == true) {
     auto token_to_be_banned = tokens_shm[col + no_repeat_ngram_size - 1];
-    lprobs[lprob_start + token_to_be_banned] = -INFINITY;
+    lprobs[lprob_start + token_to_be_banned] = -std::numeric_limits<float>::infinity();
   }
 }
 
diff --git a/onnxruntime/contrib_ops/cuda/transformers/beam_search_topk.cu b/onnxruntime/contrib_ops/cuda/transformers/beam_search_topk.cu
index 5ac10f6321e63..44be2ef2375ee 100644
--- a/onnxruntime/contrib_ops/cuda/transformers/beam_search_topk.cu
+++ b/onnxruntime/contrib_ops/cuda/transformers/beam_search_topk.cu
@@ -60,7 +60,7 @@ struct TopK {
   __device__ __forceinline__ void Init() {
     for (int i = 0; i < max_k; i++) {
       key[i] = -1;
-      value[i] = NumericLimits<T>::Min();
+      value[i] = NumericLimits<T>::Lowest();
     }
   }
 };
diff --git a/onnxruntime/contrib_ops/cuda/transformers/greedy_search_top_one.cu b/onnxruntime/contrib_ops/cuda/transformers/greedy_search_top_one.cu
index 68a2e16482af9..b2969194ff400 100644
--- a/onnxruntime/contrib_ops/cuda/transformers/greedy_search_top_one.cu
+++ b/onnxruntime/contrib_ops/cuda/transformers/greedy_search_top_one.cu
@@ -5,6 +5,7 @@
 
 #include <cub/cub.cuh>
 
+
 #include "core/providers/cuda/shared_inc/cuda_utils.h"
 #include "core/providers/cuda/cu_inc/common.cuh"
 
@@ -19,7 +20,10 @@ struct TopOne {
   int32_t key;
   T value;
 
-  __device__ __host__ __forceinline__ TopOne(int32_t key = -1, T value = NumericLimits<T>::Min()) : key(key), value(value) {
+  __device__ __host__ __forceinline__ TopOne() : key(-1), value(NumericLimits<T>::Lowest()) {
+  }
+
+  __device__ __host__ __forceinline__ TopOne(int32_t key, T value) : key(key), value(value) {
   }
 
   __device__ __forceinline__ void Reduce(int32_t k, T v) {
diff --git a/onnxruntime/core/optimizer/attention_fusion_helper.h b/onnxruntime/core/optimizer/attention_fusion_helper.h
index 267a82b72670c..935114c40d1a7 100644
--- a/onnxruntime/core/optimizer/attention_fusion_helper.h
+++ b/onnxruntime/core/optimizer/attention_fusion_helper.h
@@ -1,5 +1,6 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
+#include <limits>
 #include "onnx/defs/shape_inference.h"
 #include "onnx/defs/tensor_proto_util.h"
 #include "core/framework/tensorprotoutils.h"
@@ -767,7 +768,8 @@ bool MatchInputMaskSubgraph(const Graph& graph, const Node& layer_norm, const No
   }
 
   // check where has X=-Infinity
-  if (!optimizer_utils::IsInitializerWithExpectedValue(graph, *(where.InputDefs()[1]), -INFINITY, true)) {
+  if (!optimizer_utils::IsInitializerWithExpectedValue(graph, *(where.InputDefs()[1]),
+                                                       -std::numeric_limits<float>::infinity(), true)) {
     DEBUG_LOG("where const not matched.");
     return false;
   }
diff --git a/onnxruntime/core/providers/cuda/math/topk_impl.cuh b/onnxruntime/core/providers/cuda/math/topk_impl.cuh
index cbde6da457fdb..112566e54bbba 100644
--- a/onnxruntime/core/providers/cuda/math/topk_impl.cuh
+++ b/onnxruntime/core/providers/cuda/math/topk_impl.cuh
@@ -412,7 +412,7 @@ Status TopKImpl(const CudaKernel* kernel, bool use_deterministic_compute,
   if (aligned_dimension <= GridDim::maxThreadsPerBlock) {
     BitonicTopK<CudaT><<<N, GridDim::maxThreadsPerBlock, aligned_dimension * sizeof(KV<CudaT>), stream>>>(
         input_x_ptr, output_v_ptr, output_i, elem_nums, size, axis, K, aligned_K, largest, sorted, dimension,
-        aligned_dimension, NumericLimits<T>::Min(), NumericLimits<T>::Max());
+        aligned_dimension, NumericLimits<CudaT>::Lowest(), NumericLimits<CudaT>::Max());
   } else if (K <= BT * 16 || 0 == sorted) {
     if (use_deterministic_compute) {
       static std::once_flag log_warning;
@@ -425,19 +425,19 @@ Status TopKImpl(const CudaKernel* kernel, bool use_deterministic_compute,
     if (BT * 2 >= K || 0 == sorted) {
       RadixTopK<CudaT, BT, 2><<<N, BT, 256 * sizeof(uint32_t), stream>>>(
           input_x_ptr, output_v_ptr, output_i, elem_nums, size, axis, K, largest, sorted, dimension, XPT,
-          NumericLimits<T>::Min(), NumericLimits<T>::Max());
+          NumericLimits<CudaT>::Lowest(), NumericLimits<CudaT>::Max());
     } else if (BT * 4 >= K) {
       RadixTopK<CudaT, BT, 4><<<N, BT, 256 * sizeof(uint32_t), stream>>>(
           input_x_ptr, output_v_ptr, output_i, elem_nums, size, axis, K, largest, sorted, dimension, XPT,
-          NumericLimits<T>::Min(), NumericLimits<T>::Max());
+          NumericLimits<CudaT>::Lowest(), NumericLimits<CudaT>::Max());
     } else if (BT * 8 >= K) {
       RadixTopK<CudaT, BT, 8><<<N, BT, 256 * sizeof(uint32_t), stream>>>(
           input_x_ptr, output_v_ptr, output_i, elem_nums, size, axis, K, largest, sorted, dimension, XPT,
-          NumericLimits<T>::Min(), NumericLimits<T>::Max());
+          NumericLimits<CudaT>::Lowest(), NumericLimits<CudaT>::Max());
     } else {
       RadixTopK<CudaT, BT, 16><<<N, BT, 256 * sizeof(uint32_t), stream>>>(
           input_x_ptr, output_v_ptr, output_i, elem_nums, size, axis, K, largest, sorted, dimension, XPT,
-          NumericLimits<T>::Min(), NumericLimits<T>::Max());
+          NumericLimits<CudaT>::Lowest(), NumericLimits<CudaT>::Max());
     }
   } else {
     auto input_key_buffer = kernel->GetScratchBuffer<CudaT>(dimension, ort_stream);
diff --git a/onnxruntime/core/providers/cuda/shared_inc/cuda_utils.h b/onnxruntime/core/providers/cuda/shared_inc/cuda_utils.h
index ed642754af3ba..f9433642f0857 100644
--- a/onnxruntime/core/providers/cuda/shared_inc/cuda_utils.h
+++ b/onnxruntime/core/providers/cuda/shared_inc/cuda_utils.h
@@ -10,6 +10,7 @@
 #include <memory>
 #include <type_traits>
 #include <vector>
+#include <limits>
 
 #include <gsl/gsl>
 #include "core/framework/float16.h"
@@ -120,7 +121,7 @@ constexpr int kNumBitsPerBitmaskElement = std::numeric_limits<BitmaskElementType
 
 template <typename T>
 struct NumericLimits {
-  __inline__ __host__ __device__ static T Min() {
+  __inline__ __host__ __device__ static T Lowest() {
     return std::numeric_limits<T>::lowest();
   }
   __inline__ __host__ __device__ static T Max() {
@@ -128,43 +129,18 @@ struct NumericLimits {
   }
 };
 
-template <>
-struct NumericLimits<MLFloat16> {
-  __inline__ __host__ __device__ static half Min() {
-    return -65504.0;
-  }
-  __inline__ __host__ __device__ static half Max() {
-    return 65504.0;
-  }
-};
-
 template <>
 struct NumericLimits<half> {
-  __inline__ __host__ __device__ static half Min() {
-    return -65504.0;
-  }
-  __inline__ __host__ __device__ static half Max() {
-    return 65504.0;
+  __inline__ __host__ __device__ static half Lowest() {
+    return -65504.0f;
   }
-};
 
-template <>
-struct NumericLimits<float> {
-  __inline__ __host__ __device__ static float Min() {
-    return -INFINITY;
-  }
-  __inline__ __host__ __device__ static float Max() {
-    return INFINITY;
-  }
-};
-
-template <>
-struct NumericLimits<double> {
-  __inline__ __host__ __device__ static double Min() {
-    return -HUGE_VAL;
-  }
-  __inline__ __host__ __device__ static double Max() {
-    return HUGE_VAL;
+  __inline__ __host__ __device__ static half Max() {
+#ifdef CUDART_MAX_NORMAL_FP16  // defined in cuda 12.3 or later
+    return CUDART_MAX_NORMAL_FP16;
+#else
+    return 65504.0f;
+#endif
   }
 };
 
diff --git a/onnxruntime/core/providers/qnn/qnn_execution_provider.cc b/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
index ed193904fe7a8..a7a91d107b142 100644
--- a/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
+++ b/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
@@ -377,13 +377,15 @@ QNNExecutionProvider::QNNExecutionProvider(const ProviderOptions& provider_optio
     LOGS_DEFAULT(VERBOSE) << "User specified enable_htp_weight_sharing: " << enable_htp_weight_sharing_;
   }
 
-  model_settings_.offload_graph_io_quantization = ParseBoolOption("offload_graph_io_quantization", false,
+  model_settings_.offload_graph_io_quantization = ParseBoolOption("offload_graph_io_quantization", true,
                                                                   provider_options_map);
 
   if (disable_cpu_ep_fallback_ && model_settings_.offload_graph_io_quantization) {
-    LOGS_DEFAULT(WARNING) << "Fallback to CPU EP is disabled, but user configured QNN EP to offload graph I/O "
-                          << "quantization/dequantization to another EP. Session creation will fail if the CPU EP "
-                          << "handles the graph I/O quantization/dequantization.";
+    LOGS_DEFAULT(INFO) << "Fallback to CPU EP is disabled, but user tried to configure QNN EP to offload graph I/O "
+                       << "quantization/dequantization to another EP. These are conflicting options. Fallback to CPU "
+                       << "EP will remain disabled and graph I/O quantization/dequantization will not be offloaded "
+                       << "to another EP.";
+    model_settings_.offload_graph_io_quantization = false;
   }
 
   qnn_backend_manager_ = std::make_unique<qnn::QnnBackendManager>(
diff --git a/onnxruntime/core/providers/xnnpack/detail/utils.cc b/onnxruntime/core/providers/xnnpack/detail/utils.cc
index 4eef14dddecd3..2adf8339b4b66 100644
--- a/onnxruntime/core/providers/xnnpack/detail/utils.cc
+++ b/onnxruntime/core/providers/xnnpack/detail/utils.cc
@@ -5,6 +5,7 @@
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
+#include <limits>
 
 #include "core/common/common.h"
 #include "core/common/safeint.h"
@@ -239,8 +240,8 @@ std::unique_ptr<IndexedSubGraph::MetaDef> FuseActivation(const NodeUnit& node_un
   def.attributes = node_unit.GetNode().GetAttributes();
 
   // use infinity as the default as that's what xnnpack uses if min/max are not set
-  float min = -INFINITY;
-  float max = INFINITY;
+  float min = -std::numeric_limits<float>::infinity();
+  float max = std::numeric_limits<float>::infinity();
 
   const auto& activation_type = activation.OpType();
   if (activation_type == "Clip") {
diff --git a/onnxruntime/core/providers/xnnpack/math/gemm.cc b/onnxruntime/core/providers/xnnpack/math/gemm.cc
index 35a06cb7eb89f..a3ff3b585ae45 100644
--- a/onnxruntime/core/providers/xnnpack/math/gemm.cc
+++ b/onnxruntime/core/providers/xnnpack/math/gemm.cc
@@ -2,6 +2,9 @@
 // Licensed under the MIT License.
 
 #include "gemm.h"
+
+#include <limits>
+
 #include "core/framework/transpose_helper.h"
 #include "core/providers/utils.h"
 #include "core/providers/xnnpack/xnnpack_init.h"
@@ -140,8 +143,8 @@ Status Gemm::PrePack(const Tensor& tensor, int input_idx, AllocatorPtr,
   auto weights_cache = GetWeightsCache();
   xnn_status status = xnn_status::xnn_status_uninitialized;
   struct xnn_operator* p = nullptr;
-  float foutput_min = clip_min_max_ ? clip_min_max_->first : -INFINITY;
-  float foutput_max = clip_min_max_ ? clip_min_max_->second : INFINITY;
+  float foutput_min = clip_min_max_ ? clip_min_max_->first : -std::numeric_limits<float>::infinity();
+  float foutput_max = clip_min_max_ ? clip_min_max_->second : std::numeric_limits<float>::infinity();
   if (op_compute_type_ == OpComputeType::op_compute_type_fp32) {
     const float* bias_data = nullptr;
     if (C_matrix_exists_) {
diff --git a/onnxruntime/core/providers/xnnpack/math/matmul.cc b/onnxruntime/core/providers/xnnpack/math/matmul.cc
index 44a6fb4ee835a..f574238195ffd 100644
--- a/onnxruntime/core/providers/xnnpack/math/matmul.cc
+++ b/onnxruntime/core/providers/xnnpack/math/matmul.cc
@@ -2,6 +2,7 @@
 // Licensed under the MIT License.
 
 #include "matmul.h"
+#include <limits>
 #include "core/providers/cpu/math/matmul_helper.h"
 #include "core/providers/xnnpack/xnnpack_init.h"
 
@@ -109,8 +110,8 @@ Status MatMul::PrePack(const Tensor& tensor, int input_idx, AllocatorPtr alloc,
   xnn_weights_cache_t weight_cache = nullptr;
 #endif
 
-  float foutput_min = -INFINITY;
-  float foutput_max = INFINITY;
+  float foutput_min = -std::numeric_limits<float>::infinity();
+  float foutput_max = std::numeric_limits<float>::infinity();
   if (op_type_ == OpComputeType::op_compute_type_fp32) {
     status = xnn_create_fully_connected_nc_f32(
         shape_broadcast[0],    // size_t input_channels,
diff --git a/onnxruntime/core/providers/xnnpack/nn/average_pool.cc b/onnxruntime/core/providers/xnnpack/nn/average_pool.cc
index 1c8ed556e90d7..1fc941d9f52f6 100644
--- a/onnxruntime/core/providers/xnnpack/nn/average_pool.cc
+++ b/onnxruntime/core/providers/xnnpack/nn/average_pool.cc
@@ -33,8 +33,8 @@ Status CreateXnnpackKernel(const PoolAttributes& pool_attrs,
   if (pool_attrs.auto_pad == AutoPadType::SAME_UPPER) {
     flags |= XNN_FLAG_TENSORFLOW_SAME_PADDING;
   }
-  float foutput_min = clip_min_max ? clip_min_max->first : -INFINITY;
-  float foutput_max = clip_min_max ? clip_min_max->second : INFINITY;
+  float foutput_min = clip_min_max ? clip_min_max->first : -std::numeric_limits<float>::infinity();
+  float foutput_max = clip_min_max ? clip_min_max->second : std::numeric_limits<float>::infinity();
   xnn_status status = xnn_status_unsupported_parameter;
   if (avgpool_type == OpComputeType::op_compute_type_fp32) {
     status = xnn_create_average_pooling2d_nhwc_f32(input_padding_top, input_padding_right,
diff --git a/onnxruntime/core/providers/xnnpack/nn/conv_base.cc b/onnxruntime/core/providers/xnnpack/nn/conv_base.cc
index e0723c0e7690e..458e6000c8d70 100644
--- a/onnxruntime/core/providers/xnnpack/nn/conv_base.cc
+++ b/onnxruntime/core/providers/xnnpack/nn/conv_base.cc
@@ -54,8 +54,8 @@ Status CreateXnnpackKernel(const ConvAttributes& conv_attrs,
 
   xnn_status status = xnn_status::xnn_status_uninitialized;
   p = nullptr;
-  float foutput_min = clip_min_max ? clip_min_max->first : -INFINITY;
-  float foutput_max = clip_min_max ? clip_min_max->second : INFINITY;
+  float foutput_min = clip_min_max ? clip_min_max->first : -std::numeric_limits<float>::infinity();
+  float foutput_max = clip_min_max ? clip_min_max->second : std::numeric_limits<float>::infinity();
   // with the following IC and OC number, we can cover depthwise and regular conv at the same time
   // the equation 'IC (group_input_channels) == C ' set up when group_count==1 (regular convolution)
   // and OC (group_output_channels) follows the same rule.
diff --git a/onnxruntime/core/providers/xnnpack/nn/max_pool.cc b/onnxruntime/core/providers/xnnpack/nn/max_pool.cc
index 6742e51e55082..c828ae9400174 100644
--- a/onnxruntime/core/providers/xnnpack/nn/max_pool.cc
+++ b/onnxruntime/core/providers/xnnpack/nn/max_pool.cc
@@ -3,6 +3,8 @@
 
 #include "max_pool.h"
 
+#include <limits>
+
 #include "core/graph/graph.h"
 #include "core/providers/utils.h"
 #include "core/providers/xnnpack/xnnpack_init.h"
@@ -168,8 +170,8 @@ MaxPool::MaxPool(const OpKernelInfo& info)
   auto input_dtype = X_arg.TypeAsProto()->tensor_type().elem_type();
   xnn_status status = xnn_status_invalid_state;
   struct xnn_operator* p = nullptr;
-  float foutput_min = clip_min_max_ ? clip_min_max_->first : -INFINITY;
-  float foutput_max = clip_min_max_ ? clip_min_max_->second : INFINITY;
+  float foutput_min = clip_min_max_ ? clip_min_max_->first : -std::numeric_limits<float>::infinity();
+  float foutput_max = clip_min_max_ ? clip_min_max_->second : std::numeric_limits<float>::infinity();
   if (input_dtype == ONNX_NAMESPACE::TensorProto_DataType_FLOAT) {
     maxpool_type_ = OpComputeType::op_compute_type_fp32;
     status = xnn_create_max_pooling2d_nhwc_f32(input_padding_top, input_padding_right,
diff --git a/onnxruntime/core/session/onnxruntime_c_api.cc b/onnxruntime/core/session/onnxruntime_c_api.cc
index e4c58ba51c3df..01991710986cf 100644
--- a/onnxruntime/core/session/onnxruntime_c_api.cc
+++ b/onnxruntime/core/session/onnxruntime_c_api.cc
@@ -2842,7 +2842,7 @@ static_assert(offsetof(OrtApi, SessionOptionsAppendExecutionProvider_OpenVINO_V2
 static_assert(offsetof(OrtApi, AddExternalInitializersFromFilesInMemory) / sizeof(void*) == 279, "Size of version 18 API cannot change");
 
 // So that nobody forgets to finish an API version, this check will serve as a reminder:
-static_assert(std::string_view(ORT_VERSION) == "1.20.1",
+static_assert(std::string_view(ORT_VERSION) == "1.20.2",
               "ORT_Version change detected, please follow below steps to ensure OrtApi is updated properly");
 // 1. Update the hardcoded version string in above static_assert to silence it
 // 2. If there were any APIs added to ort_api_1_to_20 above:
diff --git a/onnxruntime/test/platform/apple/apple_package_test/Podfile.template b/onnxruntime/test/platform/apple/apple_package_test/Podfile.template
index dd99e5104853f..ec598748c4d8b 100644
--- a/onnxruntime/test/platform/apple/apple_package_test/Podfile.template
+++ b/onnxruntime/test/platform/apple/apple_package_test/Podfile.template
@@ -15,7 +15,7 @@ if ENV['SKIP_MACOS_TEST'] != 'true'
       # Comment the next line if you don't want to use dynamic frameworks
       use_frameworks!
 
-      platform :osx, '11.0'
+      platform :osx, '13.3'
 
       target 'macos_package_testUITests' do
         inherit! :search_paths
diff --git a/onnxruntime/test/providers/qnn/argmaxmin_op_test.cc b/onnxruntime/test/providers/qnn/argmaxmin_op_test.cc
index c514cf16b2f3c..41de81fd47585 100644
--- a/onnxruntime/test/providers/qnn/argmaxmin_op_test.cc
+++ b/onnxruntime/test/providers/qnn/argmaxmin_op_test.cc
@@ -70,6 +70,7 @@ static void RunQDQArgMxxOpTest(const std::string& op_type, TestInputDef<float> i
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>(op_type, {input_def}, {}, attrs),   // baseline float32 model
                        BuildQDQArgMxxTestCase<QType>(op_type, input_def, attrs),  // QDQ model
diff --git a/onnxruntime/test/providers/qnn/average_pool_test.cc b/onnxruntime/test/providers/qnn/average_pool_test.cc
index 1a0f9bfcbae97..45d9a4fd2d45a 100644
--- a/onnxruntime/test/providers/qnn/average_pool_test.cc
+++ b/onnxruntime/test/providers/qnn/average_pool_test.cc
@@ -31,6 +31,7 @@ static void RunAveragePoolOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<float>(op_type, input_defs, {}, attrs),
                   provider_options,
@@ -53,6 +54,7 @@ static void RunQDQAveragePoolOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>(op_type, input_defs, {}, attrs),
                        BuildQDQOpTestCase<QuantType>(op_type, input_defs, {}, attrs),
diff --git a/onnxruntime/test/providers/qnn/batch_norm_htp_test.cc b/onnxruntime/test/providers/qnn/batch_norm_htp_test.cc
index 0a39413a4ec1b..7471b44fafc63 100644
--- a/onnxruntime/test/providers/qnn/batch_norm_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/batch_norm_htp_test.cc
@@ -160,6 +160,7 @@ static void RunBatchNormQDQTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Runs model with DQ-> InstanceNorm -> Q and compares the outputs of the CPU and QNN EPs.
   TestQDQModelAccuracy(BuildBatchNormTestCase(input_def, scale_def, bias_def),
@@ -180,6 +181,7 @@ static void RunBatchNormFP16Test(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestInputDef<MLFloat16> input_fp16_def = ConvertToFP16InputDef(input_def);
   TestInputDef<MLFloat16> scale_fp16_def = ConvertToFP16InputDef(scale_def);
diff --git a/onnxruntime/test/providers/qnn/cast_test.cc b/onnxruntime/test/providers/qnn/cast_test.cc
index 9b83dd281a56d..713baab0b77b2 100644
--- a/onnxruntime/test/providers/qnn/cast_test.cc
+++ b/onnxruntime/test/providers/qnn/cast_test.cc
@@ -57,6 +57,7 @@ static void RunCastOpTest(const std::vector<int64_t>& shape, ONNX_NAMESPACE::Ten
 #else
   provider_options["backend_path"] = use_htp ? "libQnnHtp.so" : "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   if (use_htp && enable_fp16_precision) {
     provider_options["enable_htp_fp16_precision"] = "1";
diff --git a/onnxruntime/test/providers/qnn/clip_op_test.cc b/onnxruntime/test/providers/qnn/clip_op_test.cc
index cfa77a46210b3..ae169f7b255f0 100644
--- a/onnxruntime/test/providers/qnn/clip_op_test.cc
+++ b/onnxruntime/test/providers/qnn/clip_op_test.cc
@@ -117,6 +117,7 @@ static void RunQDQClipTestOnHTP(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float, float>("Clip", {input_def}, {min_max_defs}, {});
   auto qdq_model_builder = BuildQDQOpTestCase<QType, float>("Clip", {input_def}, {min_max_defs}, {},
@@ -205,6 +206,7 @@ TEST_F(QnnHTPBackendTests, Clip_U8_Rank5) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(model_fn,
                   provider_options,
diff --git a/onnxruntime/test/providers/qnn/conv_test.cc b/onnxruntime/test/providers/qnn/conv_test.cc
index cf37fc00335d3..91677781e8d98 100644
--- a/onnxruntime/test/providers/qnn/conv_test.cc
+++ b/onnxruntime/test/providers/qnn/conv_test.cc
@@ -93,6 +93,8 @@ static void RunCPUConvOpTest(const std::string& conv_op_type, const TestInputDef
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
+
   auto build_fn = BuildF32ConvTestCase(conv_op_type, input_def, weights_def, bias_def, strides, pads,
                                        dilations, group, auto_pad);
   RunQnnModelTest(build_fn,
@@ -317,6 +319,7 @@ static void RunHTPConvOpTest(const std::string& conv_op_type, const TestInputDef
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildF32ConvTestCase(conv_op_type, input_def, weights_def, bias_def, strides, pads, dilations,
                                             group, auto_pad, output_activation),
@@ -354,6 +357,7 @@ static void RunHTPConvOpPerChannelTest(const std::string& conv_op_type, const Te
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_fn = BuildF32ConvTestCase(conv_op_type, input_def, weights_def, bias_def, strides, pads, dilations,
                                      group, auto_pad, output_activation);
@@ -665,6 +669,7 @@ TEST_F(QnnHTPBackendTests, Test_QDQConvWithDynamicWeightsFromMul) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto BuildConvMulGraph = [](ModelTestBuilder& builder) {
     // DQ node for Conv input
diff --git a/onnxruntime/test/providers/qnn/flatten_op_test.cc b/onnxruntime/test/providers/qnn/flatten_op_test.cc
index 637d3257ddea7..6370d4a4f16ed 100644
--- a/onnxruntime/test/providers/qnn/flatten_op_test.cc
+++ b/onnxruntime/test/providers/qnn/flatten_op_test.cc
@@ -101,6 +101,7 @@ static void RunQDQFlattenTestOnHTP(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float>("Flatten", {input_def}, {}, attrs);
   auto qdq_model_builder = BuildQDQOpTestCase<QType>("Flatten", {input_def}, {}, attrs, kOnnxDomain, use_contrib_qdq);
@@ -172,6 +173,7 @@ TEST_F(QnnHTPBackendTests, Flatten_QDQ8bit_Rank5) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(model_fn,
                   provider_options,
diff --git a/onnxruntime/test/providers/qnn/gather_elems_op_test.cc b/onnxruntime/test/providers/qnn/gather_elems_op_test.cc
index 81c08873064c8..68c9867617a83 100644
--- a/onnxruntime/test/providers/qnn/gather_elems_op_test.cc
+++ b/onnxruntime/test/providers/qnn/gather_elems_op_test.cc
@@ -67,6 +67,7 @@ static void RunCPUGatherElemsOpTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<DataType, IndexType>("GatherElements", {input_def}, {indices_def}, attrs),
                   provider_options,
@@ -91,6 +92,7 @@ static void RunHTPQDQGatherElemsOpTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float, IndexType>("GatherElements", {input_def}, {indices_def}, attrs);
   auto qdq_model_builder = BuildQDQGatherElemsTestCase<QuantType, IndexType>(input_def, indices_def, attrs,
@@ -119,6 +121,7 @@ static void RunHTPGatherElemsOpTest(const TestInputDef<DataType>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<DataType, IndexType>("GatherElements", {input_def}, {indices_def}, attrs),
                   provider_options,
diff --git a/onnxruntime/test/providers/qnn/gather_op_htp_test.cc b/onnxruntime/test/providers/qnn/gather_op_htp_test.cc
index 55177cc7ed131..4478d36ebfe28 100644
--- a/onnxruntime/test/providers/qnn/gather_op_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/gather_op_htp_test.cc
@@ -63,6 +63,7 @@ static void RunQDQGatherOpTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float, IndicesType>("Gather", {input_def}, {indices_def}, attrs);
   auto qdq_model_builder = BuildQDQGatherTestCase<QuantType, IndicesType>(input_def, indices_def, attrs,
@@ -152,4 +153,4 @@ TEST_F(QnnHTPBackendTests, DISABLED_GatherOp_IndicesStaticInt32_Axis1) {
 }  // namespace test
 }  // namespace onnxruntime
 
-#endif
\ No newline at end of file
+#endif
diff --git a/onnxruntime/test/providers/qnn/gemm_op_test.cc b/onnxruntime/test/providers/qnn/gemm_op_test.cc
index 33c868694c9c0..b2aa6280ef282 100644
--- a/onnxruntime/test/providers/qnn/gemm_op_test.cc
+++ b/onnxruntime/test/providers/qnn/gemm_op_test.cc
@@ -29,6 +29,7 @@ static void RunGemmTestOnCPU(const std::vector<TestInputDef<DataType>>& input_de
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<float>("Gemm", input_defs, {}, attrs),
                   provider_options,
@@ -246,6 +247,8 @@ static void RunQDQGemmTestOnHTP(const std::vector<TestInputDef<float>>& input_de
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
+
   auto f32_model_builder = BuildOpTestCase<float>("Gemm", input_defs, {}, attrs);
   auto qdq_model_builder = BuildQDQGemmTestCase<InputAQType, InputBQType>(input_defs, attrs, use_contrib_qdq);
   TestQDQModelAccuracy<InputAQType>(f32_model_builder,
diff --git a/onnxruntime/test/providers/qnn/instance_norm_htp_test.cc b/onnxruntime/test/providers/qnn/instance_norm_htp_test.cc
index 3598ba1ac8851..d4f66b72e0a4c 100644
--- a/onnxruntime/test/providers/qnn/instance_norm_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/instance_norm_htp_test.cc
@@ -79,6 +79,7 @@ static void RunInstanceNormQDQTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Runs model with DQ-> InstanceNorm -> Q and compares the outputs of the CPU and QNN EPs.
   TestQDQModelAccuracy(BuildOpTestCase<float>("InstanceNormalization", {input_def, scale_def, bias_def}, {}, attrs),
diff --git a/onnxruntime/test/providers/qnn/layer_norm_test.cc b/onnxruntime/test/providers/qnn/layer_norm_test.cc
index 947ac19be40a8..b2997c627827d 100644
--- a/onnxruntime/test/providers/qnn/layer_norm_test.cc
+++ b/onnxruntime/test/providers/qnn/layer_norm_test.cc
@@ -28,6 +28,7 @@ static void RunLayerNormCpuTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<float>("LayerNormalization", {input_def, scale_def}, {}, attrs),
                   provider_options,
@@ -152,6 +153,7 @@ static void RunLayerNormQDQTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>("LayerNormalization", {input_def, scale_def}, {}, attrs),
                        BuildQDQLayerNormTestCase<InputQType, ScaleQType>(input_def, scale_def, bias_def, attrs,
diff --git a/onnxruntime/test/providers/qnn/leakyrelu_op_htp_test.cc b/onnxruntime/test/providers/qnn/leakyrelu_op_htp_test.cc
index ece8d91d53648..77d96b56d23ef 100644
--- a/onnxruntime/test/providers/qnn/leakyrelu_op_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/leakyrelu_op_htp_test.cc
@@ -28,6 +28,7 @@ static void RunLeakyReluOpQDQTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>("LeakyRelu", {input_def}, {}, attrs),
                        BuildQDQOpTestCase<QuantType>("LeakyRelu", {input_def}, {}, attrs),
@@ -66,6 +67,7 @@ TEST_F(QnnHTPBackendTests, LeakyReluFP16OpSet16) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto input_def = TestInputDef<float>({1, 2, 3}, false, {-40.0f, -20.0f, 1.0f, 10.0f, 30.0f, 40.0f});
   TestInputDef<MLFloat16> input_fp16_def = ConvertToFP16InputDef(input_def);
diff --git a/onnxruntime/test/providers/qnn/logical_comp_ops_test.cc b/onnxruntime/test/providers/qnn/logical_comp_ops_test.cc
index 59105136781f4..a49fa7c5fcf61 100644
--- a/onnxruntime/test/providers/qnn/logical_comp_ops_test.cc
+++ b/onnxruntime/test/providers/qnn/logical_comp_ops_test.cc
@@ -73,6 +73,7 @@ static void RunCPULogicalOpTest(const std::string& op_type, const std::vector<in
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildLogicalOpTestCase(op_type, shape),
                   provider_options,
@@ -92,6 +93,7 @@ static void RunQDQLogicalOpTest(const std::string& op_type, const std::vector<in
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildQDQLogicalOpTestCase<QuantType>(op_type, shape),
                   provider_options,
@@ -157,6 +159,7 @@ TEST_F(QnnHTPBackendTests, EqualToCast4D) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Model building function that creates a QDQ graph with an Equal node followed by
   // a Cast to float32.
diff --git a/onnxruntime/test/providers/qnn/lrn_op_test.cc b/onnxruntime/test/providers/qnn/lrn_op_test.cc
index a99cba66bf167..8327849a175ba 100644
--- a/onnxruntime/test/providers/qnn/lrn_op_test.cc
+++ b/onnxruntime/test/providers/qnn/lrn_op_test.cc
@@ -70,6 +70,7 @@ static void RunCPULRNOpTest(const TestInputDef<float>& input_def, int64_t size,
   provider_options["backend_path"] = "libQnnCpu.so";
   fp32_abs_err = 1.5e-5f;  // On linux we need slightly larger tolerance.
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildLRNTestCase(input_def, size, alpha, beta, bias),
                   provider_options,
@@ -91,6 +92,7 @@ static void RunQDQLRNOpTest(const TestInputDef<float>& input_def, int64_t size,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildLRNTestCase(input_def, size, alpha, beta, bias),
                        BuildQDQLRNTestCase<QuantType>(input_def, size, alpha, beta, bias),
diff --git a/onnxruntime/test/providers/qnn/matmul_test.cpp b/onnxruntime/test/providers/qnn/matmul_test.cpp
index 5c6967761b1db..24a74b9c4cd3d 100644
--- a/onnxruntime/test/providers/qnn/matmul_test.cpp
+++ b/onnxruntime/test/providers/qnn/matmul_test.cpp
@@ -125,6 +125,7 @@ static void RunQDQPerChannelMatMulOpOpTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   if (enable_fp16_precision) {
     provider_options["enable_htp_fp16_precision"] = "1";
@@ -178,6 +179,7 @@ static void RunQDQMatMulOpOpTest(const TestInputDef<float>& input1_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildMatMulOpTestCase(input1_def, input2_def),
                        BuildMatMulOpQDQTestCase<Input0QType, Input1QType, OutputQType>(input1_def, input2_def,
diff --git a/onnxruntime/test/providers/qnn/max_min_op_test.cc b/onnxruntime/test/providers/qnn/max_min_op_test.cc
index 3deff121f3c72..4db1fdcec4b59 100644
--- a/onnxruntime/test/providers/qnn/max_min_op_test.cc
+++ b/onnxruntime/test/providers/qnn/max_min_op_test.cc
@@ -26,6 +26,7 @@ static void RunCPUMinOrMaxOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<float>(op_type, input_defs, {}, {}, kOnnxDomain),
                   provider_options,
@@ -47,6 +48,7 @@ static void RunQDQMinOrMaxOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>(op_type, input_defs, {}, {}, kOnnxDomain),     // baseline float32 model
                        BuildQDQOpTestCase<QType>(op_type, input_defs, {}, {}, kOnnxDomain),  // QDQ model
diff --git a/onnxruntime/test/providers/qnn/pad_op_test.cpp b/onnxruntime/test/providers/qnn/pad_op_test.cpp
index a6b8664c6c0c9..17629b92188a4 100644
--- a/onnxruntime/test/providers/qnn/pad_op_test.cpp
+++ b/onnxruntime/test/providers/qnn/pad_op_test.cpp
@@ -116,6 +116,7 @@ static void RunPadOpTest(const TestInputDef<float>& data_def,
     provider_options["backend_path"] = "libQnnCpu.so";
 #endif
   }
+  provider_options["offload_graph_io_quantization"] = "0";
 
   if (enable_fp16_precision) {
     provider_options["enable_htp_fp16_precision"] = "1";
@@ -144,6 +145,7 @@ static void RunQDQPadOpTest(const TestInputDef<float>& data_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildPadTestCase(data_def, pads_def, constant_value_def, attrs),
                        BuildPadQDQTestCase<QuantType>(data_def, pads_def, constant_value_def, attrs,
diff --git a/onnxruntime/test/providers/qnn/pool_op_test.cpp b/onnxruntime/test/providers/qnn/pool_op_test.cpp
index 5dd3a6aaa3620..8905a64ab4d7f 100644
--- a/onnxruntime/test/providers/qnn/pool_op_test.cpp
+++ b/onnxruntime/test/providers/qnn/pool_op_test.cpp
@@ -60,6 +60,7 @@ static void RunPoolOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<float>(op_type, {input_def}, {}, attrs),
                   provider_options,
@@ -83,6 +84,7 @@ static void RunQDQPoolOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>(op_type, {input_def}, {}, attrs),
                        BuildPoolQDQTestCase<QuantType>(op_type, input_def, attrs, use_contrib_qdq_ops),
diff --git a/onnxruntime/test/providers/qnn/qnn_basic_test.cc b/onnxruntime/test/providers/qnn/qnn_basic_test.cc
index e8282dbad9f72..8335bf0b15796 100644
--- a/onnxruntime/test/providers/qnn/qnn_basic_test.cc
+++ b/onnxruntime/test/providers/qnn/qnn_basic_test.cc
@@ -117,6 +117,7 @@ TEST(QnnEP, TestDisableCPUFallback_ModelNotFullySupported) {
 #else
     options["backend_path"] = "libQnnCpu.so";
 #endif
+    options["offload_graph_io_quantization"] = "0";
 
     so.AppendExecutionProvider("QNN", options);
 
@@ -148,6 +149,7 @@ TEST(QnnEP, TestDisableCPUFallback_TryingToRunOnQnnCPU) {
 #else
   options["backend_path"] = "libQnnCpu.so";
 #endif
+  options["offload_graph_io_quantization"] = "0";
 
   auto input_defs = {TestInputDef<float>({1, 2, 2, 2}, false, -10.0f, 10.0f),
                      TestInputDef<float>({1, 2, 2, 2}, false, -10.0f, 10.0f)};
@@ -196,6 +198,7 @@ TEST(QnnEP, TestDisableCPUFallback_ConflictingConfig) {
 #else
     options["backend_path"] = "libQnnCpu.so";
 #endif
+    options["offload_graph_io_quantization"] = "0";
 
     so.AppendExecutionProvider("QNN", options);
 
@@ -226,6 +229,7 @@ TEST_F(QnnHTPBackendTests, TestConvWithExternalData) {
 #else
   options["backend_path"] = "libQnnHtp.so";
 #endif
+  options["offload_graph_io_quantization"] = "0";
 
   so.AppendExecutionProvider("QNN", options);
 
@@ -301,6 +305,7 @@ static void RunNHWCResizeModel(const ORTCHAR_T* ort_model_path, bool use_htp, bo
   so.SetGraphOptimizationLevel(ORT_ENABLE_ALL);
 
   onnxruntime::ProviderOptions options;
+  options["offload_graph_io_quantization"] = "0";
 
 #if defined(_WIN32)
   options["backend_path"] = use_htp ? "QnnHtp.dll" : "QnnCpu.dll";
@@ -591,6 +596,7 @@ TEST_F(QnnHTPBackendTests, MultithreadSessionRun) {
 #else
   options["backend_path"] = "libQnnHtp.so";
 #endif
+  options["offload_graph_io_quantization"] = "0";
 
   auto qnn_ep = QnnExecutionProviderWithOptions(options, &session_opts);
   EXPECT_TRUE(session_obj.RegisterExecutionProvider(std::move(qnn_ep)).IsOK());
@@ -640,6 +646,7 @@ TEST_F(QnnHTPBackendTests, MultithreadHtpPowerCfgSessionRunOption) {
 #else
   options["backend_path"] = "libQnnHtp.so";
 #endif
+  options["offload_graph_io_quantization"] = "0";
 
   auto qnn_ep = QnnExecutionProviderWithOptions(options, &session_opts);
   EXPECT_TRUE(session_obj.RegisterExecutionProvider(std::move(qnn_ep)).IsOK());
@@ -705,6 +712,7 @@ TEST_F(QnnHTPBackendTests, MultithreadDefaultHtpPowerCfgFromEpOption) {
 #else
   options["backend_path"] = "libQnnHtp.so";
 #endif
+  options["offload_graph_io_quantization"] = "0";
   options["htp_performance_mode"] = "burst";
 
   auto qnn_ep = QnnExecutionProviderWithOptions(options, &session_opts);
@@ -756,6 +764,7 @@ TEST_F(QnnHTPBackendTests, MultithreadHtpPowerCfgDefaultAndRunOption) {
 #else
   options["backend_path"] = "libQnnHtp.so";
 #endif
+  options["offload_graph_io_quantization"] = "0";
   options["htp_performance_mode"] = "burst";
 
   auto qnn_ep = QnnExecutionProviderWithOptions(options, &session_opts);
@@ -920,6 +929,7 @@ TEST_F(QnnHTPBackendTests, ProfilingTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   provider_options["enable_htp_fp16_precision"] = "1";
   provider_options["profiling_level"] = "detailed";
   provider_options["profiling_file_path"] = "detailed_profile.csv";
@@ -940,6 +950,7 @@ TEST_F(QnnHTPBackendTests, CastAddHTPAccuracyTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildCastAddTestCase(),
                   provider_options,
@@ -1010,6 +1021,7 @@ TEST_F(QnnHTPBackendTests, EPRejectsDynamicShapesF32) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   provider_options["enable_htp_fp16_precision"] = "1";  // QNN EP will use fp16 precision.
                                                         // CPU EP will use fp32, so we can relax accuracy requirements.
 
diff --git a/onnxruntime/test/providers/qnn/qnn_ep_context_test.cc b/onnxruntime/test/providers/qnn/qnn_ep_context_test.cc
index a3f0ed55b83f2..dda03829dcff5 100644
--- a/onnxruntime/test/providers/qnn/qnn_ep_context_test.cc
+++ b/onnxruntime/test/providers/qnn/qnn_ep_context_test.cc
@@ -83,6 +83,7 @@ void QnnContextBinaryMultiPartitionTestBody(bool single_ep_node = true) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   const std::unordered_map<std::string, int> domain_to_version = {{"", 13}, {kMSDomain, 1}};
 
@@ -225,6 +226,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryGeneration2InputTypes) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   const std::unordered_map<std::string, int> domain_to_version = {{"", 13}, {kMSDomain, 1}};
 
@@ -275,6 +277,7 @@ TEST_F(QnnHTPBackendTests, QnnContextGeneration2InputsOrderIssue) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Add kMSDomain to cover contrib op like Gelu
   const std::unordered_map<std::string, int> domain_to_version = {{"", 13}, {kMSDomain, 1}};
@@ -311,6 +314,7 @@ TEST_F(QnnHTPBackendTests, QnnContextGenerationNodeNamePrefix) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   std::string node_name_prefix = "node_name_prefix_test";
 
   // Add kMSDomain to cover contrib op like Gelu
@@ -353,6 +357,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryCacheEmbedModeTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   const std::string context_binary_file = "./qnn_context_binary_test.onnx";
   std::remove(context_binary_file.c_str());
 
@@ -401,6 +406,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryCacheNonEmbedModeTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   const std::string context_binary_file = "./testdata/qnn_context_cache_non_embed.onnx";
   std::string qnn_ctx_bin = "./testdata/qnn_context_cache_non_embed.onnx_QNNExecutionProvider_QNN_8283143575221199085_1_0.bin";
 
@@ -482,6 +488,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryCache_InvalidGraph) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   const std::string context_binary_file = "./qnn_context_cache_non_embed.onnx";
   std::filesystem::path context_bin = "qnn_context_cache_non_embed.onnx_QNNExecutionProvider_QNN_8283143575221199085_1_0.bin";
   std::remove(context_binary_file.c_str());
@@ -579,6 +586,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryRelativePathTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   ASSERT_STATUS_OK(session_object.RegisterExecutionProvider(QnnExecutionProviderWithOptions(provider_options)));
   ASSERT_STATUS_OK(session_object.Load(model_data.data(), static_cast<int>(model_data.size())));
@@ -609,6 +617,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryAbsolutePathTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   ASSERT_STATUS_OK(session_object.RegisterExecutionProvider(QnnExecutionProviderWithOptions(provider_options)));
   ASSERT_STATUS_OK(session_object.Load(model_data.data(), static_cast<int>(model_data.size())));
@@ -634,6 +643,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryFileNotExistTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   ASSERT_STATUS_OK(session_object.RegisterExecutionProvider(QnnExecutionProviderWithOptions(provider_options)));
   ASSERT_STATUS_OK(session_object.Load(model_data.data(), static_cast<int>(model_data.size())));
@@ -659,6 +669,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryFileEmptyStringTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   ASSERT_STATUS_OK(session_object.RegisterExecutionProvider(QnnExecutionProviderWithOptions(provider_options)));
   ASSERT_STATUS_OK(session_object.Load(model_data.data(), static_cast<int>(model_data.size())));
@@ -676,6 +687,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinary2InputsTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   const std::string context_binary_file = "./qnn_context_binary_2inputs_test.onnx";
   std::remove(context_binary_file.c_str());
 
@@ -727,6 +739,7 @@ TEST_F(QnnHTPBackendTests, QnnContextBinaryCache_SingleNodeNameNotMatchGraphName
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
   const std::string context_binary_file = "./qnn_context_cache_non_embed.onnx";
   std::filesystem::path context_bin = "qnn_context_cache_non_embed.onnx_QNNExecutionProvider_QNN_8283143575221199085_1_0.bin";
   std::remove(context_binary_file.c_str());
@@ -804,6 +817,7 @@ TEST_F(QnnHTPBackendTests, QnnMultiContextEmbeded) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   Ort::SessionOptions so;
   so.AppendExecutionProvider("QNN", provider_options);
@@ -819,6 +833,7 @@ TEST_F(QnnHTPBackendTests, QnnMultiContextExternal) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   Ort::SessionOptions so;
   so.AppendExecutionProvider("QNN", provider_options);
@@ -950,6 +965,7 @@ TEST_F(QnnHTPBackendTests, QnnContextShareAcrossSessions1) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Create QDQ models
   std::vector<std::string> onnx_model_paths{"./weight_share1.onnx", "./weight_share2.onnx"};
@@ -1047,6 +1063,7 @@ TEST_F(QnnHTPBackendTests, QnnContextShareAcrossSessions2) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Create QDQ models
   std::vector<std::string> onnx_model_paths{"./weight_share21.onnx", "./weight_share22.onnx"};
diff --git a/onnxruntime/test/providers/qnn/qnn_test_utils.cc b/onnxruntime/test/providers/qnn/qnn_test_utils.cc
index 79e7d39e85518..3f6efc453cfe4 100644
--- a/onnxruntime/test/providers/qnn/qnn_test_utils.cc
+++ b/onnxruntime/test/providers/qnn/qnn_test_utils.cc
@@ -278,7 +278,7 @@ static BackendSupport GetHTPSupport(const onnxruntime::logging::Logger& logger)
   MockKernelLookup kernel_lookup;
   onnxruntime::GraphViewer graph_viewer(graph);
   std::unique_ptr<onnxruntime::IExecutionProvider> qnn_ep = QnnExecutionProviderWithOptions(
-      {{"backend_path", "QnnHtp.dll"}});
+      {{"backend_path", "QnnHtp.dll"}, {"offload_graph_io_quantization", "0"}});
 
   qnn_ep->SetLogger(&logger);
   auto result = qnn_ep->GetCapability(graph_viewer, kernel_lookup);
@@ -341,7 +341,7 @@ static BackendSupport GetCPUSupport(const onnxruntime::logging::Logger& logger)
   MockKernelLookup kernel_lookup;
   onnxruntime::GraphViewer graph_viewer(graph);
   std::unique_ptr<onnxruntime::IExecutionProvider> qnn_ep = QnnExecutionProviderWithOptions(
-      {{"backend_path", "QnnCpu.dll"}});
+      {{"backend_path", "QnnCpu.dll"}, {"offload_graph_io_quantization", "0"}});
 
   qnn_ep->SetLogger(&logger);
   auto result = qnn_ep->GetCapability(graph_viewer, kernel_lookup);
diff --git a/onnxruntime/test/providers/qnn/reduce_op_test.cc b/onnxruntime/test/providers/qnn/reduce_op_test.cc
index 13173d9a87f55..69df89ebcfbaa 100644
--- a/onnxruntime/test/providers/qnn/reduce_op_test.cc
+++ b/onnxruntime/test/providers/qnn/reduce_op_test.cc
@@ -82,6 +82,7 @@ static void RunReduceTest(const std::string& op_type,
                           float fp32_abs_err = 1e-5f,
                           bool enable_fp16 = false) {
   ProviderOptions provider_options;
+  provider_options["offload_graph_io_quantization"] = "0";
   if (enable_fp16) {
 #if defined(_WIN32)
     provider_options["backend_path"] = "QnnHtp.dll";
@@ -401,6 +402,7 @@ static void RunReduceOpQDQTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   constexpr bool noop_with_empty_axes = false;
   const bool axes_as_input = ReduceOpHasAxesInput(op_type, opset);  // Later opsets have "axes" as an input.
diff --git a/onnxruntime/test/providers/qnn/reshape_expand_op_test.cc b/onnxruntime/test/providers/qnn/reshape_expand_op_test.cc
index 3964edc11461b..21abc66e679a8 100644
--- a/onnxruntime/test/providers/qnn/reshape_expand_op_test.cc
+++ b/onnxruntime/test/providers/qnn/reshape_expand_op_test.cc
@@ -30,6 +30,7 @@ static void RunReshapeExpandTestOnCPU(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<DataType, int64_t>(op_type, {input_def}, {shape_def}, attrs),
                   provider_options,
@@ -161,6 +162,7 @@ static void RunReshapeExpandTestOnHTP(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<DataType, int64_t>(op_type, {input_def}, {shape_def}, attrs),
                   provider_options,
@@ -185,6 +187,7 @@ static void RunQDQReshapeExpandTestOnHTP(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float, int64_t>(op_type, {input_def}, {shape_def}, attrs);
   auto qdq_model_builder = BuildQDQReshapeExpandTestCase<QType>(op_type, input_def, shape_def, attrs, use_contrib_qdq);
diff --git a/onnxruntime/test/providers/qnn/resize_test.cc b/onnxruntime/test/providers/qnn/resize_test.cc
index 15612e3267a75..15fcd039a05e9 100644
--- a/onnxruntime/test/providers/qnn/resize_test.cc
+++ b/onnxruntime/test/providers/qnn/resize_test.cc
@@ -127,6 +127,7 @@ static void RunCPUResizeOpTest(const TestInputDef<float>& input_def, const std::
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(GetResizeModelBuilder(input_def, sizes_data, mode, coordinate_transformation_mode, nearest_mode),
                   provider_options,
@@ -145,6 +146,7 @@ static void RunCPUResizeOpTestWithScales(const TestInputDef<float>& input_def, c
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(GetResizeModelBuilderWithScales(input_def, scales_data, mode, coordinate_transformation_mode, nearest_mode),
                   provider_options,
@@ -166,6 +168,7 @@ static void RunQDQResizeOpTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(GetResizeModelBuilder(input_def, sizes_data, mode, coordinate_transformation_mode, nearest_mode),
                        GetQDQResizeModelBuilder<QuantType>(input_def, sizes_data, mode, coordinate_transformation_mode,
diff --git a/onnxruntime/test/providers/qnn/simple_op_htp_test.cc b/onnxruntime/test/providers/qnn/simple_op_htp_test.cc
index 7541d94bac0c6..c0e567b07d798 100644
--- a/onnxruntime/test/providers/qnn/simple_op_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/simple_op_htp_test.cc
@@ -32,6 +32,7 @@ static void RunOpTestOnCPU(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnCpu.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(BuildOpTestCase<InputType>(op_type, input_defs, {}, attrs, op_domain),
                   provider_options,
@@ -129,6 +130,7 @@ static void RunQDQOpTest(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>(op_type, input_defs, {}, attrs, op_domain),
                        BuildQDQOpTestCase<InputQType>(op_type, input_defs, {}, attrs, op_domain, use_contrib_qdq),
@@ -780,6 +782,7 @@ TEST_F(QnnHTPBackendTests, QuantAccuracyTest) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Note: a graph input -> Q -> DQ -> is optimized by Qnn to have a perfectly accurate output.
   // ORT's CPU EP, on the otherhand, actually quantizes and dequantizes the input, which leads to different outputs.
@@ -1206,6 +1209,7 @@ TEST_F(QnnHTPBackendTests, Add_U8_U16_Convert) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   TestQDQModelAccuracy(BuildOpTestCase<float>("Add", {input0_def, input1_def}, {}, {}, kOnnxDomain),
                        BuildQDQConvertAddTestCase(input0_def, input1_def),
@@ -1271,6 +1275,7 @@ TEST_F(QnnHTPBackendTests, DQ_Q_ConvertFusion_SameType) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   QuantParams<uint8_t> out_qparams_u8 = {1.0f, 128};
   QuantParams<uint16_t> out_qparams_u16 = {1.0f, 32768};
diff --git a/onnxruntime/test/providers/qnn/slice_htp_test.cc b/onnxruntime/test/providers/qnn/slice_htp_test.cc
index 07c97d2d7b1fa..dc16192188e0b 100644
--- a/onnxruntime/test/providers/qnn/slice_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/slice_htp_test.cc
@@ -84,6 +84,7 @@ static void RunSliceQDQTest(const TestInputDef<float>& data_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   const std::vector<TestInputDef<float>> f32_inputs = {data_def};
   const std::vector<TestInputDef<int64_t>> int64_inputs = {starts_def, ends_def, axes_def, steps_def};
diff --git a/onnxruntime/test/providers/qnn/split_op_test.cc b/onnxruntime/test/providers/qnn/split_op_test.cc
index 6dc721edb421e..2b8a913ba4056 100644
--- a/onnxruntime/test/providers/qnn/split_op_test.cc
+++ b/onnxruntime/test/providers/qnn/split_op_test.cc
@@ -276,6 +276,7 @@ static void RunQDQSplitOpTestOnHTP(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   const bool split_is_input = opset >= 13;
   auto f32_model_builder = BuildSplitTestCase<float>(input_def, split, split_is_input, axis, num_outputs);
diff --git a/onnxruntime/test/providers/qnn/squeeze_unsqueeze_op_test.cc b/onnxruntime/test/providers/qnn/squeeze_unsqueeze_op_test.cc
index 33d2f64c0315e..249a5da733ae9 100644
--- a/onnxruntime/test/providers/qnn/squeeze_unsqueeze_op_test.cc
+++ b/onnxruntime/test/providers/qnn/squeeze_unsqueeze_op_test.cc
@@ -161,6 +161,7 @@ static void RunQDQSqueezeTestOnHTP(const std::string& op_type,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float, int64_t>(op_type, {input_def}, {axes_def}, {});
   auto qdq_model_builder = BuildQDQSqueezeTestCase<QType>(op_type, input_def, axes_def, use_contrib_qdq);
@@ -219,6 +220,7 @@ TEST_F(QnnHTPBackendTests, Squeeze_Rank5_Rank2_f32) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(model_fn,
                   provider_options,
@@ -275,6 +277,7 @@ TEST_F(QnnHTPBackendTests, Unsqueeze_Rank3_Rank5_f32) {
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   RunQnnModelTest(model_fn,
                   provider_options,
diff --git a/onnxruntime/test/providers/qnn/tile_op_test.cc b/onnxruntime/test/providers/qnn/tile_op_test.cc
index 2b35c730ee5fe..20fb2801176e6 100644
--- a/onnxruntime/test/providers/qnn/tile_op_test.cc
+++ b/onnxruntime/test/providers/qnn/tile_op_test.cc
@@ -98,6 +98,7 @@ static void RunQDQTileTestOnHTP(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildOpTestCase<float, int64_t>("Tile", {input_def}, {repeats_def}, {});
   auto qdq_model_builder = BuildQDQTileTestCase<QType>(input_def, repeats_def, use_contrib_qdq);
diff --git a/onnxruntime/test/providers/qnn/topk_op_test.cc b/onnxruntime/test/providers/qnn/topk_op_test.cc
index 5a9351b9366ec..aee86428b1b5d 100644
--- a/onnxruntime/test/providers/qnn/topk_op_test.cc
+++ b/onnxruntime/test/providers/qnn/topk_op_test.cc
@@ -154,6 +154,7 @@ static void RunQDQTopKTestOnHTP(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   auto f32_model_builder = BuildTopKTestCase<float>(input_def, k_def, attrs);
   auto qdq_model_builder = BuildQDQTopKTestCase<QType>(input_def, k_def, attrs, use_contrib_qdq);
diff --git a/onnxruntime/test/providers/qnn/transpose_htp_test.cc b/onnxruntime/test/providers/qnn/transpose_htp_test.cc
index 63746e22d214d..b7bec34f7dcbd 100644
--- a/onnxruntime/test/providers/qnn/transpose_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/transpose_htp_test.cc
@@ -70,6 +70,7 @@ static void RunTransposeQDQTest(const TestInputDef<float>& input_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Runs model with DQ-> Transpose -> Q and compares the outputs of the CPU and QNN EPs.
   TestQDQModelAccuracy(BuildTransposeTestCase<float>(input_def, attrs),
diff --git a/onnxruntime/test/providers/qnn/where_htp_test.cc b/onnxruntime/test/providers/qnn/where_htp_test.cc
index ec525ef4eb3cc..e1b0604b314e1 100644
--- a/onnxruntime/test/providers/qnn/where_htp_test.cc
+++ b/onnxruntime/test/providers/qnn/where_htp_test.cc
@@ -79,6 +79,7 @@ static void RunWhereQDQTest(const TestInputDef<bool>& condition_def,
 #else
   provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  provider_options["offload_graph_io_quantization"] = "0";
 
   // Runs model with DQ-> Where -> Q and compares the outputs of the CPU and QNN EPs.
   TestQDQModelAccuracy(BuildWhereTestCase(condition_def, x_def, y_def),
diff --git a/tools/ci_build/github/apple/default_full_apple_framework_build_settings.json b/tools/ci_build/github/apple/default_full_apple_framework_build_settings.json
index 84d7e355ed5b4..f2d034dbfc000 100644
--- a/tools/ci_build/github/apple/default_full_apple_framework_build_settings.json
+++ b/tools/ci_build/github/apple/default_full_apple_framework_build_settings.json
@@ -24,7 +24,7 @@
         ],
         "macosx": [
             "--macos=MacOSX",
-            "--apple_deploy_target=11.0"
+            "--apple_deploy_target=13.3"
         ],
         "iphoneos": [
             "--ios",
diff --git a/tools/ci_build/github/apple/default_training_ios_framework_build_settings.json b/tools/ci_build/github/apple/default_training_ios_framework_build_settings.json
index 1d4a8c038c07b..41f11ebae7d54 100644
--- a/tools/ci_build/github/apple/default_training_ios_framework_build_settings.json
+++ b/tools/ci_build/github/apple/default_training_ios_framework_build_settings.json
@@ -33,7 +33,7 @@
         ],
         "macosx": [
             "--macos=MacOSX",
-            "--apple_deploy_target=11.0"
+            "--apple_deploy_target=13.3"
         ]
     }
 }
diff --git a/tools/ci_build/github/azure-pipelines/npm-packaging-pipeline.yml b/tools/ci_build/github/azure-pipelines/npm-packaging-pipeline.yml
index e7034b563a0fc..f2bd7f2b57d5a 100644
--- a/tools/ci_build/github/azure-pipelines/npm-packaging-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/npm-packaging-pipeline.yml
@@ -47,8 +47,8 @@ stages:
     UseWebPoolName: true
     RunWebGpuTestsForDebugBuild: false
     RunWebGpuTestsForReleaseBuild: true
-    WebGpuPoolName: 'onnxruntime-Win2022-webgpu-A10'
-    WebCpuPoolName: 'Azure-Pipelines-EO-Windows2022-aiinfra'
+    WebGpuPoolName: 'onnxruntime-Win2022-VS2022-webgpu-A10'
+    WebCpuPoolName: 'onnxruntime-Win2022-VS2022-webgpu-A10'
 
 - template: templates/react-native-ci.yml
   parameters:
diff --git a/tools/ci_build/github/azure-pipelines/orttraining-linux-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/orttraining-linux-ci-pipeline.yml
deleted file mode 100644
index 5c3273f79bd30..0000000000000
--- a/tools/ci_build/github/azure-pipelines/orttraining-linux-ci-pipeline.yml
+++ /dev/null
@@ -1,95 +0,0 @@
-##### start trigger Don't edit it manually, Please do edit set-trigger-rules.py ####
-### please do rerun set-trigger-rules.py ###
-trigger:
-  branches:
-    include:
-    - main
-    - rel-*
-  paths:
-    exclude:
-    - docs/**
-    - README.md
-    - CONTRIBUTING.md
-    - BUILD.md
-    - 'js/web'
-    - 'onnxruntime/core/providers/js'
-pr:
-  branches:
-    include:
-    - main
-    - rel-*
-  paths:
-    exclude:
-    - docs/**
-    - README.md
-    - CONTRIBUTING.md
-    - BUILD.md
-    - 'js/web'
-    - 'onnxruntime/core/providers/js'
-#### end trigger ####
-
-jobs:
-- job: Linux_Build
-  timeoutInMinutes: 180
-  workspace:
-    clean: all
-  variables:
-    skipComponentGovernanceDetection: true
-    CCACHE_DIR: $(Pipeline.Workspace)/ccache
-    TODAY: $[format('{0:dd}{0:MM}{0:yyyy}', pipeline.startTime)]
-  pool: onnxruntime-Ubuntu-2204-Training-CPU
-  steps:
-  - task: mspremier.PostBuildCleanup.PostBuildCleanup-task.PostBuildCleanup@3
-    displayName: 'Clean Agent Directories'
-    condition: always()
-
-  - checkout: self
-    clean: true
-    submodules: none
-
-  - template: templates/get-docker-image-steps.yml
-    parameters:
-      Dockerfile: tools/ci_build/github/linux/docker/inference/x86_64/default/cpu/Dockerfile
-      Context: tools/ci_build/github/linux/docker/inference/x86_64/default/cpu
-      DockerBuildArgs: "--build-arg BUILD_UID=$( id -u ) --build-arg BASEIMAGE=registry.access.redhat.com/ubi8/ubi"
-      Repository: onnxruntimecpubuildcentos8x64_packaging
-
-  - task: Cache@2
-    inputs:
-      key: '"$(TODAY)" | "$(Build.SourceBranch)" | "$(Build.SourceVersion)"'
-      path: $(CCACHE_DIR)
-      cacheHitVar: CACHE_RESTORED
-      restoreKeys: |
-        "$(TODAY)" | "$(Build.SourceBranch)"
-        "$(TODAY)" |
-    displayName: Cach Task
-
-  - task: CmdLine@2
-    displayName: 'build'
-    inputs:
-      script: |
-        set -e -x
-        mkdir -p $HOME/.onnx
-        mkdir -p $(Pipeline.Workspace)/ccache
-        docker run --rm \
-          --volume /data/onnx:/data/onnx:ro \
-          --volume /data/models:/build/models:ro \
-          --volume $(Build.SourcesDirectory):/onnxruntime_src \
-          --volume $(Build.BinariesDirectory):/build \
-          --volume $HOME/.onnx:/home/onnxruntimedev/.onnx \
-          --volume $(Pipeline.Workspace)/ccache:/cache \
-          -e ALLOW_RELEASED_ONNX_OPSET_ONLY=0 \
-          -e NIGHTLY_BUILD \
-          -e BUILD_BUILDNUMBER \
-          -e CCACHE_DIR=/cache \
-          onnxruntimecpubuildcentos8x64_packaging \
-          /onnxruntime_src/tools/ci_build/github/linux/build_training_ci.sh
-      workingDirectory: $(Build.SourcesDirectory)
-
-  - task: PublishTestResults@2
-    displayName: 'Publish unit test results'
-    inputs:
-      testResultsFiles: '**/*.results.xml'
-      searchFolder: '$(Build.BinariesDirectory)'
-      testRunTitle: 'Unit Test Run'
-    condition: succeededOrFailed()
diff --git a/tools/ci_build/github/azure-pipelines/orttraining-linux-gpu-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/orttraining-linux-gpu-ci-pipeline.yml
deleted file mode 100644
index 494035637a79d..0000000000000
--- a/tools/ci_build/github/azure-pipelines/orttraining-linux-gpu-ci-pipeline.yml
+++ /dev/null
@@ -1,55 +0,0 @@
-##### start trigger Don't edit it manually, Please do edit set-trigger-rules.py ####
-### please do rerun set-trigger-rules.py ###
-trigger:
-  branches:
-    include:
-    - main
-    - rel-*
-  paths:
-    exclude:
-    - docs/**
-    - README.md
-    - CONTRIBUTING.md
-    - BUILD.md
-    - 'js/web'
-    - 'onnxruntime/core/providers/js'
-pr:
-  branches:
-    include:
-    - main
-    - rel-*
-  paths:
-    exclude:
-    - docs/**
-    - README.md
-    - CONTRIBUTING.md
-    - BUILD.md
-    - 'js/web'
-    - 'onnxruntime/core/providers/js'
-#### end trigger ####
-
-jobs:
-- template: templates/linux-ci.yml
-  parameters:
-    AgentPool : 'Onnxruntime-Linux-GPU-NC6sv3'
-    JobName: 'Onnxruntime_Linux_GPU_Training'
-    RunDockerBuildArgs: >
-      -o ubuntu20.04 -d gpu
-      -t onnxruntime_orttraining_ortmodule_tests_image
-      -u
-      -e
-      -x "
-      --enable_training
-      --config Release
-      --use_cuda --cuda_version=11.8 --cuda_home=/usr/local/cuda-11.8 --cudnn_home=/usr/local/cuda-11.8
-      --build_wheel
-      --enable_nvtx_profile
-      --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=70
-      "
-    RunInjectedPipeline: 'true'
-    InjectedPipeline: 'orttraining-linux-gpu-test-ci-pipeline.yml'
-    DockerImageTag: 'onnxruntime_orttraining_ortmodule_tests_image'
-    TimeoutInMinutes: 190
-    # Enable unreleased onnx opsets in CI builds
-    # This facilitates testing the implementation for the new opsets
-    AllowReleasedOpsetOnly: '0'
diff --git a/tools/ci_build/github/azure-pipelines/post-merge-jobs.yml b/tools/ci_build/github/azure-pipelines/post-merge-jobs.yml
index 833e97b437c33..497de2368b4bc 100644
--- a/tools/ci_build/github/azure-pipelines/post-merge-jobs.yml
+++ b/tools/ci_build/github/azure-pipelines/post-merge-jobs.yml
@@ -8,7 +8,7 @@ stages:
       BuildStaticLib: true
       ExtraBuildArgs: ''
       UseWebPoolName: true
-      WebCpuPoolName: 'Onnxruntime-Win-CPU-2022'
+      WebCpuPoolName: 'onnxruntime-Win2022-VS2022-webgpu-A10'
 
 # The follow section has 15 different build jobs that can be divided into 3 groups:
 # 1. Default CPU build with normal win32 linking, without ORT extension
diff --git a/tools/ci_build/github/azure-pipelines/templates/mac-cpu-packaging-pipeline.yml b/tools/ci_build/github/azure-pipelines/templates/mac-cpu-packaging-pipeline.yml
index 080079388a76c..ab31e592d7d71 100644
--- a/tools/ci_build/github/azure-pipelines/templates/mac-cpu-packaging-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/templates/mac-cpu-packaging-pipeline.yml
@@ -68,9 +68,6 @@ stages:
   jobs:
     - job: MacOS_C_API_Package_Publish
       pool:
-        ${{ if eq(parameters.DoESRP, true)}}:
-          vmImage: 'macOS-12'
-        ${{ else }}:
           vmImage: 'macOS-13'
       steps:
       - checkout: none
diff --git a/tools/ci_build/github/azure-pipelines/templates/react-native-ci.yml b/tools/ci_build/github/azure-pipelines/templates/react-native-ci.yml
index 8593aa2d821fa..ce40de7b67a51 100644
--- a/tools/ci_build/github/azure-pipelines/templates/react-native-ci.yml
+++ b/tools/ci_build/github/azure-pipelines/templates/react-native-ci.yml
@@ -201,6 +201,30 @@ stages:
       workingDirectory: '$(Build.SourcesDirectory)/js/react_native/e2e'
       displayName: Bootstrap Android and iOS e2e tests
 
+# TODO: remove this once we upgrade the RN to post 0.70.15 or 0.73.3+ versions this is just a temporary workaround
+    - script: |
+        if [[ "$OSTYPE" == "darwin"* ]]; then
+          sed -i .bak "/boostorg.jfrog.io/c\\
+        spec.source = { :http => 'https://archives.boost.io/release/1.76.0/source/boost_1_76_0.tar.bz2', " boost.podspec
+        else
+          sed -i .bak "/boostorg.jfrog.io/c\spec.source = { :http => 'https://archives.boost.io/release/1.76.0/source/boost_1_76_0.tar.bz2', " boost.podspec
+        fi
+        rm -f boost.podspec.bak
+      workingDirectory: '$(Build.SourcesDirectory)/js/react_native/node_modules/react-native/third-party-podspecs'
+      displayName: Path the boost 1.76.0 source URL in boost.podspec js/react_native/node_modules/react-native/third-party-podspecs/boost.podspec
+
+# TODO: remove this once we upgrade the RN to post 0.70.15 or 0.73.3+ versions this is just a temporary workaround
+    - script: |
+        if [[ "$OSTYPE" == "darwin"* ]]; then
+          sed -i .bak "/boostorg.jfrog.io/c\\
+        spec.source = { :http => 'https://archives.boost.io/release/1.76.0/source/boost_1_76_0.tar.bz2', " boost.podspec
+        else
+          sed -i .bak "/boostorg.jfrog.io/c\spec.source = { :http => 'https://archives.boost.io/release/1.76.0/source/boost_1_76_0.tar.bz2', " boost.podspec
+        fi
+        rm -f boost.podspec.bak
+      workingDirectory: '$(Build.SourcesDirectory)/js/react_native/e2e/node_modules/react-native/third-party-podspecs'
+      displayName: Path the boost 1.76.0 source URL in boost.podspec in js/react_native/e2e/node_modules/react-native/third-party-podspecs/boost.podspec
+
     - script: |
         ORT_C_LOCAL_POD_PATH=$(Build.BinariesDirectory)/ios-full-pod/onnxruntime-c \
         pod install
diff --git a/tools/ci_build/github/azure-pipelines/templates/web-ci.yml b/tools/ci_build/github/azure-pipelines/templates/web-ci.yml
index b2e1833156657..053fa80dff417 100644
--- a/tools/ci_build/github/azure-pipelines/templates/web-ci.yml
+++ b/tools/ci_build/github/azure-pipelines/templates/web-ci.yml
@@ -39,10 +39,10 @@ parameters:
   default: false
 - name: WebGpuPoolName
   type: string
-  default: 'onnxruntime-Win2022-webgpu-A10'
+  default: 'onnxruntime-Win2022-VS2022-webgpu-A10'
 - name: WebCpuPoolName
   type: string
-  default: 'onnxruntime-Win-CPU-2022-web'
+  default: 'onnxruntime-Win2022-VS2022-webgpu-A10'
 
 - name: ExtraBuildArgs
   displayName: 'Extra build command line arguments'
diff --git a/tools/ci_build/github/azure-pipelines/web-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/web-ci-pipeline.yml
index 713ad4d1a6b1c..01920ad1f7fbb 100644
--- a/tools/ci_build/github/azure-pipelines/web-ci-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/web-ci-pipeline.yml
@@ -61,6 +61,6 @@ stages:
     UseWebPoolName: true
     RunWebGpuTestsForDebugBuild: false
     RunWebGpuTestsForReleaseBuild: true
-    WebGpuPoolName: 'onnxruntime-Win2022-webgpu-A10'
-    WebCpuPoolName: 'onnxruntime-Win-CPU-2022-web'
+    WebGpuPoolName: 'onnxruntime-Win2022-VS2022-webgpu-A10'
+    WebCpuPoolName: 'onnxruntime-Win2022-VS2022-webgpu-A10'
     WithCache: false
diff --git a/tools/ci_build/github/linux/build_training_ci.sh b/tools/ci_build/github/linux/build_training_ci.sh
deleted file mode 100755
index 82f75a5cbbc50..0000000000000
--- a/tools/ci_build/github/linux/build_training_ci.sh
+++ /dev/null
@@ -1,4 +0,0 @@
-#!/bin/bash
-set -e -x
-python3.12 -m pip install -r /onnxruntime_src/tools/ci_build/github/linux/python/requirements.txt
-python3.12 /onnxruntime_src/tools/ci_build/build.py --build_dir /build --config Release --enable_training --skip_submodule_sync --parallel
diff --git a/tools/ci_build/github/linux/docker/Dockerfile.ubuntu_gpu_training b/tools/ci_build/github/linux/docker/Dockerfile.ubuntu_gpu_training
deleted file mode 100644
index 4d11cbbde3354..0000000000000
--- a/tools/ci_build/github/linux/docker/Dockerfile.ubuntu_gpu_training
+++ /dev/null
@@ -1,60 +0,0 @@
-ARG BASEIMAGE=nvcr.io/nvidia/cuda:11.8.0-cudnn8-devel-ubuntu18.04
-
-FROM $BASEIMAGE
-
-ARG PYTHON_VERSION=3.9
-ARG INSTALL_DEPS_EXTRA_ARGS
-ARG USE_CONDA=false
-
-ADD scripts /tmp/scripts
-RUN /tmp/scripts/install_ubuntu.sh -p $PYTHON_VERSION && \
-    /tmp/scripts/install_os_deps.sh -d gpu $INSTALL_DEPS_EXTRA_ARGS
-
-# If USE_CONDA is false, use root to install python dependencies.
-RUN if [ "$USE_CONDA" = false ] ; \
-    then /tmp/scripts/install_python_deps.sh -p $PYTHON_VERSION -d gpu $INSTALL_DEPS_EXTRA_ARGS ; \
-    fi
-
-WORKDIR /root
-
-# Allow configure to pick up GDK and CuDNN where it expects it.
-# (Note: $CUDNN_VERSION is defined by NVidia's base image)
-RUN _CUDNN_VERSION=$(echo $CUDNN_VERSION | cut -d. -f1-2) && \
-    mkdir -p /usr/local/cudnn-$_CUDNN_VERSION/cuda/include && \
-    ln -s /usr/include/cudnn.h /usr/local/cudnn-$_CUDNN_VERSION/cuda/include/cudnn.h && \
-    mkdir -p /usr/local/cudnn-$_CUDNN_VERSION/cuda/lib64 && \
-    ln -s /etc/alternatives/libcudnn_so /usr/local/cudnn-$_CUDNN_VERSION/cuda/lib64/libcudnn.so && \
-    ln -s /usr/local/cudnn{-$_CUDNN_VERSION,}
-
-ENV LD_LIBRARY_PATH /usr/local/openblas/lib:$LD_LIBRARY_PATH
-
-ARG BUILD_USER=onnxruntimedev
-ARG BUILD_UID=1000
-RUN adduser --gecos 'onnxruntime Build User' --disabled-password $BUILD_USER --uid $BUILD_UID
-WORKDIR /home/$BUILD_USER
-USER $BUILD_USER
-
-ARG MINICONDA_PREFIX=/home/$BUILD_USER/miniconda3
-RUN if [ "$USE_CONDA" = true ] ; \
-    then MINICONDA=miniconda.sh && \
-    wget --no-verbose https://repo.anaconda.com/miniconda/Miniconda3-py37_4.9.2-Linux-x86_64.sh -O $MINICONDA && \
-    chmod a+x $MINICONDA && \
-    ./$MINICONDA -b -p $MINICONDA_PREFIX && \
-    rm ./$MINICONDA && \
-    $MINICONDA_PREFIX/bin/conda clean --yes --all && \
-    $MINICONDA_PREFIX/bin/conda install -y python=$PYTHON_VERSION ; \
-    fi
-
-ENV PATH /home/$BUILD_USER/miniconda3/bin:$PATH
-
-# If USE_CONDA is true, use onnxruntimedev user to install python dependencies
-RUN if [ "$USE_CONDA" = true ] ; \
-    then /tmp/scripts/install_python_deps.sh -p $PYTHON_VERSION -d gpu $INSTALL_DEPS_EXTRA_ARGS -c ; \
-    fi
-
-WORKDIR /root
-USER root
-RUN rm -rf /tmp/scripts
-
-WORKDIR /home/$BUILD_USER
-USER $BUILD_USER