diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
index 083517a6..c7fde9a3 100644
--- a/.github/workflows/benchmark.yml
+++ b/.github/workflows/benchmark.yml
@@ -16,13 +16,14 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        build: [cuda11]
+        build: [cuda11, cuda12]
         include:
           - build: cuda11
             image: nvidia/cuda:11.7.1-devel-ubuntu22.04
             modeldir: /llamasharp_ci/models_benchmark
-          # - build: cuda12
-          #   image: nvidia/cuda:12.1.1-runtime-ubuntu22.04
+          - build: cuda12
+            image: nvidia/cuda:12.1.1-devel-ubuntu22.04
+            modeldir: /llamasharp_ci/models_benchmark
 
     container:
       image: ${{ matrix.image }}
@@ -72,3 +73,45 @@ jobs:
       with:
         name: Benchmark_Results
         path: BenchmarkDotNet.Artifacts/results/*
+
+  windows-benchmark-cuda:
+    if: contains(github.event.pull_request.labels.*.name, 'benchmark')
+    runs-on: [self-hosted, windows, gpu]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        build: [cuda11]
+        include:
+          - build: cuda11
+            modeldir: F:\Models\LLamaSharpBenchmark
+
+    env:
+      AGENT_TOOLSDIRECTORY: D:\Libs\github\runner-cache
+      BENCHMARK_MODEL_DIR: ${{ matrix.modeldir }}
+
+    steps:
+    - name: Settings
+      run: |
+        set http_proxy=127.0.0.1:7891
+        set https_proxy=127.0.0.1:7891
+
+    - uses: actions/checkout@v4
+
+    - name: Clear package cache
+      run: dotnet clean LLamaSharp.sln && dotnet nuget locals all --clear
+    - name: Restore packages
+      run: dotnet restore LLamaSharp.sln
+    - name: Build
+      run: | 
+        dotnet clean
+        dotnet build LLama/LLamaSharp.csproj -c Release --no-restore
+        dotnet build LLama.Benchmark/LLama.Benchmark.csproj -c Release --no-restore
+    - name: Run benchmark test
+      run: dotnet run --project LLama.Benchmark/LLama.Benchmark.csproj -c Release --anyCategories LLama
+    - name: Upload artifacts
+      if: always()
+      uses: actions/upload-artifact@v3
+      with:
+        name: Benchmark_Results
+        path: BenchmarkDotNet.Artifacts/results/*
diff --git a/LLama.Benchmark/LLamaExecutorBenchmark/Prefill.cs b/LLama.Benchmark/LLamaExecutorBenchmark/Prefill.cs
index 7c540d08..d7475c6e 100644
--- a/LLama.Benchmark/LLamaExecutorBenchmark/Prefill.cs
+++ b/LLama.Benchmark/LLamaExecutorBenchmark/Prefill.cs
@@ -6,6 +6,7 @@ using BenchmarkDotNet.Engines;
 using BenchmarkDotNet.Jobs;
 using LLama.Abstractions;
 using LLama.Common;
+using LLama.Native;
 
 namespace LLama.Benchmark.LLamaExecutorBenchmark
 {
@@ -100,6 +101,17 @@ namespace LLama.Benchmark.LLamaExecutorBenchmark
         [GlobalSetup(Targets = [nameof(Basic)])]
         public void GlobalSetup()
         {
+            var showLLamaCppLogs = true;
+            NativeLibraryConfig
+               .Instance
+               .WithLogCallback((level, message) =>
+               {
+                   if (showLLamaCppLogs)
+                       Console.WriteLine($"[llama {level}]: {message.TrimEnd('\n')}");
+               }).WithCuda().SkipCheck().WithAutoFallback(false);
+
+            // Calling this method forces loading to occur now.
+            NativeApi.llama_empty_call();
             InitializeParamsAndModel();
         }