diff --git a/.github/scripts/calculate_best_perf.py b/.github/scripts/calculate_best_perf.py
index c19194385..3d2f53233 100644
--- a/.github/scripts/calculate_best_perf.py
+++ b/.github/scripts/calculate_best_perf.py
@@ -21,8 +21,8 @@
 
 parser = argparse.ArgumentParser(description="Get Best Performance",
                                  formatter_class=argparse.ArgumentDefaultsHelpFormatter)
-parser.add_argument("--best", default=None, required=True, help="Saved best performance file")
-parser.add_argument("--new", default=None, required=True, help="New round launch")
+parser.add_argument("--best", required=True, help="Saved best performance file")
+parser.add_argument("--new", required=True, help="New round launch")
 parser.add_argument("--device", default=None, type=str, help="Device name, such as PVC1100")
 parser.add_argument("--os", default=None, type=str, help="OS version, such as Ubuntu 22.04")
 parser.add_argument("--driver", default=None, type=str, help="Driver version, such as 25.05.32567")
@@ -63,8 +63,8 @@ def find_files(pattern, path):
     new_data = pd.read_csv(new_file)
     new_data = new_data.reset_index(drop=True)
     for index, row in new_data.iterrows():
-        new_eager = row["abs_latency"] * row["speedup"]
         # eager
+        new_eager = row["abs_latency"] * row["speedup"]
         eager_row = next(([i, line] for i, line in best_data.iterrows()
                           if (line["Category"] == category
                               and line["Model"] == row["name"]
@@ -74,15 +74,15 @@ def find_files(pattern, path):
             (best_data["Model"] == row["name"]) &
             (best_data["Value Type"] == "eager")]
         if eager_row != "N/A":
-            if new_eager < best_eager_value["Value"].values:
-                best_data.loc[eager_row[0]] = category, row["name"], "eager", new_eager, \
-                    args.device, args.os, args.driver, args.oneapi, args.gcc, args.python, \
-                    args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')
+            if new_eager < best_eager_value["Value"].values[0]:
+                best_data.loc[eager_row[0]] = [category, row["name"], "eager", new_eager,
+                    args.device, args.os, args.driver, args.oneapi, args.gcc, args.python,
+                    args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')]
         else:
             best_data.loc[len(best_data), :] = None
-            best_data.loc[len(best_data) - 1] = category, row["name"], "eager", new_eager, \
-                args.device, args.os, args.driver, args.oneapi, args.gcc, args.python, \
-                args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')
+            best_data.loc[len(best_data) - 1] = [category, row["name"], "eager", new_eager,
+                args.device, args.os, args.driver, args.oneapi, args.gcc, args.python,
+                args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')]
         # inductor
         inductor_row = next(([i, line] for i, line in best_data.iterrows()
                              if (line["Category"] == category
@@ -93,14 +93,14 @@ def find_files(pattern, path):
             (best_data["Model"] == row["name"]) &
             (best_data["Value Type"] == "inductor")]
         if inductor_row != "N/A":
-            if row["abs_latency"] < best_inductor_value["Value"].values:
-                best_data.at[inductor_row[0]] = category, row["name"], "inductor", row["abs_latency"], \
-                    args.device, args.os, args.driver, args.oneapi, args.gcc, args.python, \
-                    args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')
+            if row["abs_latency"] < best_inductor_value["Value"].values[0]:
+                best_data.loc[inductor_row[0]] = [category, row["name"], "inductor", row["abs_latency"],
+                    args.device, args.os, args.driver, args.oneapi, args.gcc, args.python,
+                    args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')]
         else:
             best_data.loc[len(best_data), :] = None
-            best_data.loc[len(best_data) - 1] = category, row["name"], "inductor", row["abs_latency"], \
-                args.device, args.os, args.driver, args.oneapi, args.gcc, args.python, \
-                args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')
+            best_data.loc[len(best_data) - 1] = [category, row["name"], "inductor", row["abs_latency"],
+                args.device, args.os, args.driver, args.oneapi, args.gcc, args.python,
+                args.pytorch, args.torch_xpu_ops, date.today().strftime('%F')]
 
 best_data.to_csv(args.best, sep=',', encoding='utf-8', index=False)
diff --git a/.github/scripts/e2e_summary.sh b/.github/scripts/e2e_summary.sh
index 2b5bf97b8..5894bb4b5 100644
--- a/.github/scripts/e2e_summary.sh
+++ b/.github/scripts/e2e_summary.sh
@@ -1,7 +1,7 @@
 #!/bin/bash
 
 results_dir="$1"
-artifact_type="$2"
+reference_dir="$2"
 check_file="$(dirname "$0")/../ci_expected_accuracy/check_expected.py"
 rm -rf /tmp/tmp-*.txt
 
@@ -147,32 +147,13 @@ fi
 performance=$(find "${results_dir}" -name "*.csv" |grep -E "_xpu_performance.csv" -c)
 if [ "${performance}" -gt 0 ];then
     echo "### Performance"
-    pip install jq > /dev/null 2>&1
-    if [ "${artifact_type}" != "" ];then
-        gh api \
-            --method GET -F per_page=100 -F page=10 \
-            -H "Accept: application/vnd.github+json" -H "X-GitHub-Api-Version: 2022-11-28" \
-            /repos/${GITHUB_REPOSITORY:-"intel/torch-xpu-ops"}/actions/artifacts \
-            > ${GITHUB_WORKSPACE:-"/tmp"}/refer.json
-        artifact_id="$(eval "jq -r \
-                '[.artifacts[] | \
-                select(.name|test(\"${artifact_type}.*\")) | \
-                select(.workflow_run.head_branch|test(\"main\"))][0].id' \
-            ${GITHUB_WORKSPACE:-"/tmp"}/refer.json")"
-        if [ "$artifact_id" -gt 1 ];then
-            gh api \
-                -H "Accept: application/vnd.github+json" \
-                -H "X-GitHub-Api-Version: 2022-11-28" \
-                /repos/${GITHUB_REPOSITORY:-"intel/torch-xpu-ops"}/actions/artifacts/${artifact_id}/zip > reference.zip
-        fi
+    unzip ${reference_dir}/*.zip -d ${reference_dir} > /dev/null 2>&1
+    if [ "${IS_PR}" == "1" ];then
+        python "$(dirname "$0")/perf_comparison.py" --xpu ${results_dir} --refer ${reference_dir} --pr
+    else
+        python "$(dirname "$0")/perf_comparison.py" --xpu ${results_dir} --refer ${reference_dir}
     fi
-    rm -rf ${GITHUB_WORKSPACE:-"/tmp"}/reference
-    mkdir ${GITHUB_WORKSPACE:-"/tmp"}/reference
-    mv reference.zip ${GITHUB_WORKSPACE:-"/tmp"}/reference
-    unzip ${GITHUB_WORKSPACE:-"/tmp"}/reference/reference.zip -d ${GITHUB_WORKSPACE:-"/tmp"}/reference > /dev/null 2>&1
-    reference_dir="${GITHUB_WORKSPACE:-"/tmp"}/reference"
-    python "$(dirname "$0")/perf_comparison.py" -xpu ${results_dir} -refer ${reference_dir}
-    cp ${GITHUB_WORKSPACE:-"/tmp"}/reference/best.csv ${results_dir}/best.csv > /dev/null 2>&1 || true
+    cp ${reference_dir}/best.csv ${results_dir}/best.csv > /dev/null 2>&1 || true
     python "$(dirname "$0")/calculate_best_perf.py" \
         --new ${results_dir} \
         --best ${results_dir}/best.csv \
diff --git a/.github/scripts/perf_comparison.py b/.github/scripts/perf_comparison.py
index 92ac6c2c3..9a314d533 100644
--- a/.github/scripts/perf_comparison.py
+++ b/.github/scripts/perf_comparison.py
@@ -1,6 +1,6 @@
 # To compare the performance diff
 # Usage:
-#   python perf_comparison.py -xpu /path/to/xpu/performance/result/dir -refer /path/to/reference/dir
+#   python perf_comparison.py --xpu /path/to/xpu/performance/result/dir --refer /path/to/reference/dir
 
 import re
 import os
@@ -10,8 +10,9 @@
 from statistics import geometric_mean
 
 parser = argparse.ArgumentParser(description="Analysis", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
-parser.add_argument("-xpu", default=None, help="XPU performance result csv files dir")
-parser.add_argument("-refer", default=None, help="XPU refrerence result csv files dir")
+parser.add_argument("--xpu", default=None, help="XPU performance result csv files dir")
+parser.add_argument("--refer", default=None, help="XPU refrerence result csv files dir")
+parser.add_argument("--pr", action="store_true", help="Only show results xpu has")
 args = parser.parse_args()
 
 
@@ -46,7 +47,7 @@ def find_files(pattern, path):
     if os.path.isfile(refer_file):
         refer_data= pd.read_csv(refer_file)
         refer_names = [row["name"] for index, row in refer_data.iterrows()]
-        names = xpu_names + refer_names
+        names = xpu_names if args.pr else xpu_names + refer_names
         names = set(names)
         names = sorted(names)
         for name in names:
@@ -72,18 +73,19 @@ def find_files(pattern, path):
             xpu_value = next((row for index, row in xpu_data.iterrows() if row["name"] == name), "")
             xpu_eager_latency = xpu_value["speedup"] * xpu_value["abs_latency"]
             output_data.append([multiple_replace(xpu_file), name, xpu_eager_latency, xpu_value["abs_latency"], xpu_value["speedup"], -1, -1, -1, -1, -1])
-refer_files = find_files("*_xpu_performance.csv", args.refer)
-for refer_file in refer_files:
-    refer_data = pd.read_csv(refer_file)
-    refer_names = [row["name"] for index, row in refer_data.iterrows()]
-    xpu_file = re.sub(args.refer, args.xpu + "/", refer_file, flags=re.IGNORECASE)
-    if not os.path.isfile(xpu_file):
-        names = set(refer_names)
-        names = sorted(names)
-        for name in names:
-            refer_value = next((row for index, row in refer_data.iterrows() if row["name"] == name), "")
-            refer_eager_latency = refer_value["speedup"] * refer_value["abs_latency"]
-            output_data.append([multiple_replace(refer_file), name, -1, -1, -1, refer_eager_latency, refer_value["abs_latency"], refer_value["speedup"], -1, -1])
+if not args.pr:
+    refer_files = find_files("*_xpu_performance.csv", args.refer)
+    for refer_file in refer_files:
+        refer_data = pd.read_csv(refer_file)
+        refer_names = [row["name"] for index, row in refer_data.iterrows()]
+        xpu_file = re.sub(args.refer, args.xpu + "/", refer_file, flags=re.IGNORECASE)
+        if not os.path.isfile(xpu_file):
+            names = set(refer_names)
+            names = sorted(names)
+            for name in names:
+                refer_value = next((row for index, row in refer_data.iterrows() if row["name"] == name), "")
+                refer_eager_latency = refer_value["speedup"] * refer_value["abs_latency"]
+                output_data.append([multiple_replace(refer_file), name, -1, -1, -1, refer_eager_latency, refer_value["abs_latency"], refer_value["speedup"], -1, -1])
 
 # summary
 output_data = pd.DataFrame(output_data, columns=output_header)
diff --git a/.github/workflows/_performance_comparison.yml b/.github/workflows/_performance_comparison.yml
index 76f216cee..21fe79692 100644
--- a/.github/workflows/_performance_comparison.yml
+++ b/.github/workflows/_performance_comparison.yml
@@ -7,17 +7,19 @@ on:
         required: true
         type: string
         default: ''
-        description: Target artifact id
+        description: Target run id
       baseline:
         required: true
         type: string
         default: ''
-        description: Baseline artifact id
+        description: Baseline run id
 
 permissions: read-all
 
 jobs:
   Performance-Comparison:
+    env:
+      GH_TOKEN: ${{ github.token }}
     runs-on: ubuntu-latest
     steps:
       - name: Cleanup workspace
@@ -29,23 +31,15 @@ jobs:
           python-version: '3.10' 
       - name: Checkout torch-xpu-ops
         uses: actions/checkout@v4
-      - name: Downlaod target artifacts
-        uses: actions/download-artifact@v4
-        with:
-          run-id: ${{ inputs.target }}
-          pattern: XPU-E2E
-          path: target
-      - name: Downlaod baseline artifacts
-        uses: actions/download-artifact@v4
-        with:
-          run-id: ${{ inputs.baseline }}
-          pattern: XPU-E2E
-          path: baseline
-      - name: Get comparison artifacts
+      - name: Downlaod artifacts
         run: |
-          cd target && unzip *.zip
-          cd ../baseline && unzip *.zip
+          rm -rf target && mkdir target && cd target
+          gh --repo ${GITHUB_REPOSITORY} run download ${{ inputs.target }} -p "Inductor-*-XPU-E2E-*"
+          mv Inductor-*-XPU-E2E-*/* . && cd ..
+          rm -rf baseline && mkdir baseline && cd baseline
+          gh --repo ${GITHUB_REPOSITORY} run download ${{ inputs.baseline }} -p "Inductor-*-XPU-E2E-*"
+          mv Inductor-*-XPU-E2E-*/* . && cd ..
       - name: Comparison result
         run: |
           python -m pip install pandas
-          python .github/scripts/perf_comparison.py -xpu target -refer baseline >> ${GITHUB_STEP_SUMMARY}
+          python .github/scripts/perf_comparison.py --xpu target --refer baseline >> ${GITHUB_STEP_SUMMARY}
diff --git a/.github/workflows/nightly_ondemand.yml b/.github/workflows/nightly_ondemand.yml
index 48c594293..9c1baed45 100644
--- a/.github/workflows/nightly_ondemand.yml
+++ b/.github/workflows/nightly_ondemand.yml
@@ -59,7 +59,7 @@ on:
         default: '3.10'
         description: Python version
 
-permissions: read-all
+permissions: write-all
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.sha }}-${{ github.event_name }}-${{ inputs.pytorch }}-${{ inputs.keep_torch_xpu_ops }}-${{ inputs.ut }}-${{ inputs.triton }}-${{ inputs.suite }}-${{ inputs.dt }}-${{ inputs.mode }}-${{ inputs.scenario }}-${{ inputs.model }}-${{ inputs.python }}
@@ -102,6 +102,7 @@ jobs:
     timeout-minutes: 3600
     env:
       GH_TOKEN: ${{ github.token }}
+      reference_issue: 1645
       pytorch: ${{ needs.Linux-Nightly-Ondemand-Build.outputs.torch_commit_id }}
       keep_torch_xpu_ops: ${{ github.event_name == 'schedule' && 'false' || inputs.keep_torch_xpu_ops }}
       python: ${{ github.event_name == 'schedule' && '3.10' || inputs.python }}
@@ -300,10 +301,26 @@ jobs:
           scenario: ${{ inputs.scenario }}
           hf_token: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
 
+      - name: Download Reference Artifact
+        id: reference_id
+        run: |
+          set -xe
+          source activate e2e_ci
+          conda install gh --channel conda-forge -y
+          if [ "${{ env.run_type }}" == "on-demand" ];then
+            artifact_type="weekly"
+          else
+            artifact_type="${{ env.run_type }}"
+          fi
+          REFERENCE_RUN_ID="$(gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} \
+            --json body -q .body |grep "Inductor-${artifact_type}-LTS-XPU-E2E" |sed 's/.*: *//')"
+          gh --repo ${GITHUB_REPOSITORY} run download ${REFERENCE_RUN_ID} -p "Inductor-*-XPU-E2E-*"
+          rm -rf reference && mv Inductor-*-XPU-E2E-* reference
       - name: Summarize archieve files
         id: summary
         if: ${{ ! cancelled() }}
         run: |
+          set -x -e -o pipefail
           rm -rf ${{ github.workspace }}/upload_files
           cp -r ${{ github.workspace }}/../pytorch/inductor_log ${{ github.workspace }}/upload_files
           mkdir -p ${{ github.workspace }}/../../_backup/ && cd ${{ github.workspace }}/../../_backup/
@@ -314,7 +331,7 @@ jobs:
             source activate e2e_ci
             bash ${{ github.workspace }}/.github/scripts/e2e_summary.sh \
               ${{ github.workspace }}/upload_files \
-              Inductor-${{ env.run_type }}-LTS-XPU-E2E \
+              ${{ github.workspace }}/reference \
             >> ${GITHUB_STEP_SUMMARY}
             exit_label=$(awk 'BEGIN{sum=0}{if($2>0){sum++}}END{print sum}' /tmp/tmp-result.txt)
             if [ ${exit_label} -ne 0 ];then
@@ -335,6 +352,12 @@ jobs:
         with:
           name: Inductor-${{ env.run_type }}-LTS-XPU-E2E-Data-${{ github.event.pull_request.number || github.sha }}
           path: ${{ github.workspace }}/upload_files
+      - name: Upload Reference Run ID
+        if: ${{ env.run_type != 'on-demand' }}
+        run: |
+          gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} --json body -q .body | \
+            sed "s/Inductor-${{ env.run_type }}-Rolling-XPU-E2E:.*/Inductor-${{ env.run_type }}-Rolling-XPU-E2E: ${GITHUB_RUN_ID}/" | sed '/^$/d' > new_body.txt
+          gh --repo ${GITHUB_REPOSITORY} issue edit ${reference_issue} --body-file new_body.txt
 
   Windows-Nightly-Ondemand-UT-Tests:
     if: ${{ github.event_name == 'schedule' || inputs.ut != '' }}
diff --git a/.github/workflows/nightly_ondemand_rolling.yml b/.github/workflows/nightly_ondemand_rolling.yml
index a985c9b78..8355f236c 100644
--- a/.github/workflows/nightly_ondemand_rolling.yml
+++ b/.github/workflows/nightly_ondemand_rolling.yml
@@ -59,7 +59,7 @@ on:
         default: '3.10'
         description: Python version
 
-permissions: read-all
+permissions: write-all
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.sha }}-${{ github.event_name }}-${{ inputs.pytorch }}-${{ inputs.keep_torch_xpu_ops }}-${{ inputs.ut }}-${{ inputs.triton }}-${{ inputs.suite }}-${{ inputs.dt }}-${{ inputs.mode }}-${{ inputs.scenario }}-${{ inputs.model }}-${{ inputs.python }}
@@ -117,6 +117,7 @@ jobs:
     timeout-minutes: 3600
     env:
       GH_TOKEN: ${{ github.token }}
+      reference_issue: 1645
       pytorch: ${{ needs.Linux-Nightly-Ondemand-Build-Rolling.outputs.torch_commit_id }}
       keep_torch_xpu_ops: ${{ github.event_name == 'schedule' && 'false' || inputs.keep_torch_xpu_ops }}
       python: ${{ github.event_name == 'schedule' && '3.10' || inputs.python }}
@@ -326,11 +327,26 @@ jobs:
           scenario: ${{ inputs.scenario }}
           hf_token: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
           driver: rolling
-
+      - name: Download Reference Artifact
+        id: reference_id
+        run: |
+          set -xe
+          source activate e2e_ci
+          conda install gh --channel conda-forge -y
+          if [ "${{ env.run_type }}" == "on-demand" ];then
+            artifact_type="weekly"
+          else
+            artifact_type="${{ env.run_type }}"
+          fi
+          REFERENCE_RUN_ID="$(gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} \
+            --json body -q .body |grep "Inductor-${artifact_type}-LTS-XPU-E2E" |sed 's/.*: *//')"
+          gh --repo ${GITHUB_REPOSITORY} run download ${REFERENCE_RUN_ID} -p "Inductor-*-XPU-E2E-*"
+          rm -rf reference && mv Inductor-*-XPU-E2E-* reference
       - name: Summarize archieve files
         id: summary
         if: ${{ ! cancelled() }}
         run: |
+          set -x -e -o pipefail
           rm -rf ${{ github.workspace }}/upload_files
           cp -r ${{ github.workspace }}/../pytorch/inductor_log ${{ github.workspace }}/upload_files
           mkdir -p ${{ github.workspace }}/../../_backup/ && cd ${{ github.workspace }}/../../_backup/
@@ -341,7 +357,7 @@ jobs:
             source activate e2e_ci
             bash ${{ github.workspace }}/.github/scripts/e2e_summary.sh \
                 ${{ github.workspace }}/upload_files \
-                Inductor-${{ env.run_type }}-Rolling-XPU-E2E \
+                ${{ github.workspace }}/reference \
             >> ${GITHUB_STEP_SUMMARY}
             exit_label=$(awk 'BEGIN{sum=0}{if($2>0){sum++}}END{print sum}' /tmp/tmp-result.txt)
             if [ ${exit_label} -ne 0 ];then
@@ -362,6 +378,12 @@ jobs:
         with:
           name: Inductor-${{ env.run_type }}-Rolling-XPU-E2E-Data-${{ github.event.pull_request.number || github.sha }}
           path: ${{ github.workspace }}/upload_files
+      - name: Upload Reference Run ID
+        if: ${{ env.run_type != 'on-demand' }}
+        run: |
+          gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} --json body -q .body | \
+            sed "s/Inductor-${{ env.run_type }}-Rolling-XPU-E2E:.*/Inductor-${{ env.run_type }}-Rolling-XPU-E2E: ${GITHUB_RUN_ID}/" | sed '/^$/d' > new_body.txt
+          gh --repo ${GITHUB_REPOSITORY} issue edit ${reference_issue} --body-file new_body.txt
 
   Tests-Failure-And-Report:
     if: ${{ ! cancelled() }}
diff --git a/.github/workflows/nightly_ondemand_whl.yml b/.github/workflows/nightly_ondemand_whl.yml
index b8585a451..fbee5dac3 100644
--- a/.github/workflows/nightly_ondemand_whl.yml
+++ b/.github/workflows/nightly_ondemand_whl.yml
@@ -49,7 +49,7 @@ on:
         default: '3.10'
         description: Python version
 
-permissions: read-all
+permissions: write-all
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.sha }}-${{ github.event_name }}-${{ inputs.pytorch }}-${{ inputs.ut }}-${{ inputs.suite }}-${{ inputs.dt }}-${{ inputs.mode }}-${{ inputs.scenario }}-${{ inputs.model }}-${{ inputs.python }}
@@ -72,6 +72,7 @@ jobs:
     timeout-minutes: 3600
     env:
       GH_TOKEN: ${{ github.token }}
+      reference_issue: 1645
       pytorch: ${{ github.event_name == 'schedule' && 'nightly' || inputs.pytorch }}
       python: ${{ github.event_name == 'schedule' && '3.10' || inputs.python }}
       run_type: ${{ (github.event_name == 'schedule' && (github.event.schedule == '0 17 * * 5' && 'weekly' || 'nightly')) || 'on-demand' }}
@@ -262,11 +263,26 @@ jobs:
           scenario: ${{ inputs.scenario }}
           pytorch: nightly_wheel
           hf_token: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
-
+      - name: Download Reference Artifact
+        id: reference_id
+        run: |
+          set -xe
+          source activate e2e_ci
+          conda install gh --channel conda-forge -y
+          if [ "${{ env.run_type }}" == "on-demand" ];then
+            artifact_type="weekly"
+          else
+            artifact_type="${{ env.run_type }}"
+          fi
+          REFERENCE_RUN_ID="$(gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} \
+            --json body -q .body |grep "Inductor-${artifact_type}-LTS-XPU-E2E" |sed 's/.*: *//')"
+          gh --repo ${GITHUB_REPOSITORY} run download ${REFERENCE_RUN_ID} -p "Inductor-*-XPU-E2E-*"
+          rm -rf reference && mv Inductor-*-XPU-E2E-* reference
       - name: Summarize archieve files
         id: summary
         if: ${{ ! cancelled() }}
         run: |
+          set -x -e -o pipefail
           rm -rf ${{ github.workspace }}/upload_files
           cp -r ${{ github.workspace }}/../pytorch/inductor_log ${{ github.workspace }}/upload_files
           mkdir -p ${{ github.workspace }}/../../_backup/ && cd ${{ github.workspace }}/../../_backup/
@@ -277,7 +293,7 @@ jobs:
             source activate e2e_ci
             bash ${{ github.workspace }}/.github/scripts/e2e_summary.sh \
                 ${{ github.workspace }}/upload_files \
-                Inductor-${{ env.run_type }}-Pre-XPU-E2E \
+                ${{ github.workspace }}/reference \
             >> ${GITHUB_STEP_SUMMARY}
             exit_label=$(awk 'BEGIN{sum=0}{if($2>0){sum++}}END{print sum}' /tmp/tmp-result.txt)
             if [ ${exit_label} -ne 0 ];then
@@ -298,6 +314,12 @@ jobs:
         with:
           name: Inductor-${{ env.run_type }}-Pre-XPU-E2E-Data-${{ github.event.pull_request.number || github.sha }}
           path: ${{ github.workspace }}/upload_files
+      - name: Upload Reference Run ID
+        if: ${{ env.run_type != 'on-demand' }}
+        run: |
+          gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} --json body -q .body | \
+            sed "s/Inductor-${{ env.run_type }}-Rolling-XPU-E2E:.*/Inductor-${{ env.run_type }}-Rolling-XPU-E2E: ${GITHUB_RUN_ID}/" | sed '/^$/d' > new_body.txt
+          gh --repo ${GITHUB_REPOSITORY} issue edit ${reference_issue} --body-file new_body.txt
 
   Tests-Failure-And-Report:
     if: ${{ ! cancelled() }}
diff --git a/.github/workflows/pull.yml b/.github/workflows/pull.yml
index 02cd96b5b..0d18d3ae5 100644
--- a/.github/workflows/pull.yml
+++ b/.github/workflows/pull.yml
@@ -73,6 +73,7 @@ jobs:
     runs-on: pvc_e2e
     env:
       GH_TOKEN: ${{ github.token }}
+      reference_issue: 1645
     # Don't run on forked repos and draft PRs
     if: ${{ (github.repository_owner == 'intel') && (github.event.pull_request.draft == false) }}
     timeout-minutes: 900
@@ -183,16 +184,28 @@ jobs:
           scenario: accuracy,performance
           env_prepare: true
           hf_token: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+      - name: Download Reference Artifact
+        id: reference_id
+        run: |
+          set -xe
+          source activate e2e_ci
+          conda install gh --channel conda-forge -y
+          REFERENCE_RUN_ID="$(gh --repo ${GITHUB_REPOSITORY} issue view ${reference_issue} \
+            --json body -q .body |grep "Inductor-weekly-LTS-XPU-E2E" |sed 's/.*: *//')"
+          gh --repo ${GITHUB_REPOSITORY} run download ${REFERENCE_RUN_ID} -p "Inductor-*-XPU-E2E-*"
+          rm -rf reference && mv Inductor-*-XPU-E2E-* reference
       - name: Summarize archieve files
         if: ${{ ! cancelled() }}
         run: |
+          set -x -e -o pipefail
           rm -rf ${{ github.workspace }}/upload_files
           cp -r ${{ github.workspace }}/../pytorch/inductor_log ${{ github.workspace }}/upload_files
           # Print summary
           source activate e2e_ci
+          export IS_PR=1
           bash ${{ github.workspace }}/.github/scripts/e2e_summary.sh \
               ${{ github.workspace }}/upload_files \
-              Inductor-weekly-LTS-XPU-E2E \
+              ${{ github.workspace }}/reference \
           >> ${GITHUB_STEP_SUMMARY}
           exit_label=$(awk 'BEGIN{sum=0}{if($2>0){sum++}}END{print sum}' /tmp/tmp-result.txt)
           if [ ${exit_label} -ne 0 ];then