docs(voxtral_realtime): document CUDA Windows workflow

Add CUDA-Windows streaming export and run examples to README.
Add streaming mode support to Windows CI test script with
meaningful output validation.

Generated with Claude Code.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: larryliu0820

.ci/scripts/test_model_e2e_windows.ps1

@@ -13,7 +13,8 @@ param(
     [Parameter(Mandatory = $true)]
     [string]$QuantName,
     [string]$ModelDir = ".",
-    [string]$ExpectedCudaVersion = ""
+    [string]$ExpectedCudaVersion = "",
+    [string]$Mode = ""
 )
 
 Set-StrictMode -Version Latest
@@ -25,6 +26,15 @@ if ($Device -ne "cuda-windows") {
     throw "Unsupported device '$Device'. Expected 'cuda-windows'."
 }
 
+if ($Mode -ne "") {
+    if ($Mode -notin @("vr-streaming", "vr-offline")) {
+        throw "Unsupported mode '$Mode'. Supported modes: vr-streaming, vr-offline"
+    }
+    if ($HfModel -ne "mistralai/Voxtral-Mini-4B-Realtime-2602") {
+        throw "Mode '$Mode' can only be used with Voxtral Realtime model"
+    }
+}
+
 Write-Host "Testing model: $HfModel (quantization: $QuantName)"
 
 $resolvedModelDir = (Resolve-Path -Path $ModelDir).Path
@@ -79,7 +89,7 @@ switch ($HfModel) {
         $runnerTarget = "voxtral_realtime_runner"
         $runnerPath = "voxtral_realtime"
         $runnerPreset = "voxtral-realtime-cuda"
-        $expectedOutput = "Loading audio from"
+        $expectedOutput = "Quilter"
         $preprocessor = "preprocessor.pte"
         $tokenizerUrl = ""
         $tokenizerFile = "tekken.json"
@@ -207,6 +217,9 @@ try {
                 "--audio_path", (Join-Path -Path $resolvedModelDir -ChildPath $audioFile),
                 "--preprocessor_path", (Join-Path -Path $resolvedModelDir -ChildPath $preprocessor)
             )
+            if ($Mode -ne "vr-offline") {
+                $runnerArgs += "--streaming"
+            }
         }
     }
 

examples/models/voxtral_realtime/CMakePresets.json

@@ -14,12 +14,16 @@
         {
             "name": "voxtral-realtime-cpu",
             "displayName": "Voxtral Realtime runner (CPU)",
-            "inherits": ["voxtral-realtime-base"]
+            "inherits": [
+                "voxtral-realtime-base"
+            ]
         },
         {
             "name": "voxtral-realtime-metal",
             "displayName": "Voxtral Realtime runner (Metal)",
-            "inherits": ["voxtral-realtime-base"],
+            "inherits": [
+                "voxtral-realtime-base"
+            ],
             "cacheVariables": {
                 "EXECUTORCH_BUILD_METAL": "ON"
             },
@@ -32,14 +36,19 @@
         {
             "name": "voxtral-realtime-cuda",
             "displayName": "Voxtral Realtime runner (CUDA)",
-            "inherits": ["voxtral-realtime-base"],
+            "inherits": [
+                "voxtral-realtime-base"
+            ],
             "cacheVariables": {
                 "EXECUTORCH_BUILD_CUDA": "ON"
             },
             "condition": {
                 "type": "inList",
                 "string": "${hostSystemName}",
-                "list": ["Linux", "Windows"]
+                "list": [
+                    "Linux",
+                    "Windows"
+                ]
             }
         }
     ],
@@ -48,20 +57,26 @@
             "name": "voxtral-realtime-cpu",
             "displayName": "Build Voxtral Realtime runner (CPU)",
             "configurePreset": "voxtral-realtime-cpu",
-            "targets": ["voxtral_realtime_runner"]
+            "targets": [
+                "voxtral_realtime_runner"
+            ]
         },
         {
             "name": "voxtral-realtime-metal",
             "displayName": "Build Voxtral Realtime runner (Metal)",
             "configurePreset": "voxtral-realtime-metal",
             "configuration": "Release",
-            "targets": ["voxtral_realtime_runner"]
+            "targets": [
+                "voxtral_realtime_runner"
+            ]
         },
         {
             "name": "voxtral-realtime-cuda",
             "displayName": "Build Voxtral Realtime runner (CUDA)",
             "configurePreset": "voxtral-realtime-cuda",
-            "targets": ["voxtral_realtime_runner"]
+            "targets": [
+                "voxtral_realtime_runner"
+            ]
         }
     ],
     "workflowPresets": [
@@ -108,4 +123,4 @@
             ]
         }
     ]
-}
+}

examples/models/voxtral_realtime/README.md

@@ -97,6 +97,7 @@ python export_voxtral_rt.py \
 | `xnnpack` | ✓ | ✓ | `4w`, `8w`, `8da4w`, `8da8w` |
 | `metal` | ✓ | ✓ | none (fp32) or `fpa4w` (Metal-specific 4-bit) |
 | `cuda` | ✓ | ✓ | `4w`, `8w` |
+| `cuda-windows` | ✓ | ✓ | `4w`, `8w` |
 
 Metal backend provides Apple GPU acceleration. CUDA backend provides NVIDIA GPU
 acceleration via AOTInductor.
@@ -171,6 +172,54 @@ Alternatively, you can build torchao with Metal support while installing ExecuTo
 EXECUTORCH_BUILD_KERNELS_TORCHAO=1 TORCHAO_BUILD_EXPERIMENTAL_MPS=1 ./install_executorch.sh
 ```
 
+### CUDA-Windows Export
+
+Before running `cuda-windows` export, make sure these requirements are set up:
+- `x86_64-w64-mingw32-g++` is installed and on `PATH` (mingw-w64 cross-compiler).
+- `WINDOWS_CUDA_HOME` points to the extracted Windows CUDA package directory.
+
+Example setup on Ubuntu (refer to [Parakeet README](../parakeet/README.md) for detailed extraction steps):
+
+```bash
+# Ensure the WINDOWS_CUDA_HOME environment variable is set
+export WINDOWS_CUDA_HOME=/opt/cuda-windows/extracted/cuda_cudart/cudart
+```
+
+Export the model for Windows CUDA (example with int4 quantization):
+
+```bash
+python export_voxtral_rt.py \
+    --model-path ~/models/Voxtral-Mini-4B-Realtime-2602 \
+    --backend cuda-windows \
+    --dtype bf16 \
+    --output-dir ./voxtral_rt_exports \
+    --qlinear-encoder 4w \
+    --qlinear-encoder-packing-format tile_packed_to_4d \
+    --qlinear 4w \
+    --qlinear-packing-format tile_packed_to_4d \
+    --qembedding 8w
+```
+
+For streaming, add `--streaming`:
+
+```bash
+python export_voxtral_rt.py \
+    --model-path ~/models/Voxtral-Mini-4B-Realtime-2602 \
+    --backend cuda-windows \
+    --dtype bf16 \
+    --streaming \
+    --output-dir ./voxtral_rt_exports \
+    --qlinear-encoder 4w \
+    --qlinear-encoder-packing-format tile_packed_to_4d \
+    --qlinear 4w \
+    --qlinear-packing-format tile_packed_to_4d \
+    --qembedding 8w
+```
+
+Both offline and streaming exports generate:
+- `model.pte`
+- `aoti_cuda_blob.ptd`
+
 ### Options
 
 | Flag | Default | Description |
@@ -228,6 +277,18 @@ make voxtral_realtime-metal
 This builds ExecuTorch with Metal backend support. The runner binary is at
 the same path as above. Metal exports can only run on macOS with Apple Silicon.
 
+### CUDA-Windows
+
+On Windows (PowerShell), use CMake workflow presets directly from the executorch root directory. Note that if you exported the model with 4-bit quantization, you may need to specify your GPU's compute capability (e.g., `80;86;89;90;120` for Ampere, Lovelace, Hopper, and Blackwell) to avoid "invalid device function" errors at runtime, as the `int4mm` kernels require SM 80 or newer.
+
+```powershell
+$env:CMAKE_CUDA_ARCHITECTURES="80;86;89;90;120"
+cmake --workflow --preset llm-release-cuda
+Push-Location examples/models/voxtral_realtime
+cmake --workflow --preset voxtral-realtime-cuda
+Pop-Location
+```
+
 ## Run
 
 The runner requires:
@@ -237,35 +298,49 @@ The runner requires:
 - A 16kHz mono WAV audio file (or live audio via `--mic`)
 - For CUDA: `aoti_cuda_blob.ptd` — delegate data file (pass via `--data_path`)
 
-```bash
-cmake-out/examples/models/voxtral_realtime/voxtral_realtime_runner \
-    --model_path voxtral_rt_exports/model.pte \
-    --tokenizer_path ~/models/Voxtral-Mini-4B-Realtime-2602/tekken.json \
-    --preprocessor_path voxtral_rt_exports/preprocessor.pte \
+### Windows (PowerShell)
+
+```powershell
+.\cmake-out\examples\models\voxtral_realtime\Release\voxtral_realtime_runner.exe `
+    --model_path voxtral_rt_exports\model.pte `
+    --tokenizer_path C:\path\to\tekken.json `
+    --preprocessor_path voxtral_rt_exports\preprocessor.pte `
     --audio_path input.wav
 ```
 
 For CUDA, include the `.ptd` data file:
 
-```bash
-cmake-out/examples/models/voxtral_realtime/voxtral_realtime_runner \
-    --model_path voxtral_rt_exports/model.pte \
-    --data_path voxtral_rt_exports/aoti_cuda_blob.ptd \
-    --tokenizer_path ~/models/Voxtral-Mini-4B-Realtime-2602/tekken.json \
-    --preprocessor_path voxtral_rt_exports/preprocessor.pte \
+```powershell
+.\cmake-out\examples\models\voxtral_realtime\Release\voxtral_realtime_runner.exe `
+    --model_path voxtral_rt_exports\model.pte `
+    --data_path voxtral_rt_exports\aoti_cuda_blob.ptd `
+    --tokenizer_path C:\path\to\tekken.json `
+    --preprocessor_path voxtral_rt_exports\preprocessor.pte `
     --audio_path input.wav
 ```
 
 For streaming, add `--streaming`. This requires a model exported with
 `--streaming`. The runner processes audio in 80ms steps, computing mel
 and running the encoder+decoder incrementally.
 
-```bash
-cmake-out/examples/models/voxtral_realtime/voxtral_realtime_runner \
-    --model_path voxtral_rt_exports/model.pte \
-    --tokenizer_path ~/models/Voxtral-Mini-4B-Realtime-2602/tekken.json \
-    --preprocessor_path voxtral_rt_exports/preprocessor.pte \
-    --audio_path input.wav \
+```powershell
+.\cmake-out\examples\models\voxtral_realtime\Release\voxtral_realtime_runner.exe `
+    --model_path voxtral_rt_exports\model.pte `
+    --tokenizer_path C:\path\to\tekken.json `
+    --preprocessor_path voxtral_rt_exports\preprocessor.pte `
+    --audio_path input.wav `
+    --streaming
+```
+
+For CUDA streaming, include the `.ptd` data file:
+
+```powershell
+.\cmake-out\examples\models\voxtral_realtime\Release\voxtral_realtime_runner.exe `
+    --model_path voxtral_rt_exports\model.pte `
+    --data_path voxtral_rt_exports\aoti_cuda_blob.ptd `
+    --tokenizer_path C:\path\to\tekken.json `
+    --preprocessor_path voxtral_rt_exports\preprocessor.pte `
+    --audio_path input.wav `
     --streaming
 ```
 
@@ -285,6 +360,17 @@ ffmpeg -f avfoundation -i ":0" -ar 16000 -ac 1 -f f32le -nostats -loglevel error
 
 Ctrl+C stops recording and flushes remaining text.
 
+**Windows (PowerShell):**
+
+```powershell
+.\cmake-out\examples\models\voxtral_realtime\Release\voxtral_realtime_runner.exe `
+  --model_path C:\path\to\voxtral_rt_exports\model.pte `
+  --data_path C:\path\to\voxtral_rt_exports\aoti_cuda_blob.ptd `
+  --tokenizer_path C:\path\to\tekken.json `
+  --preprocessor_path C:\path\to\voxtral_rt_exports\preprocessor.pte `
+  --audio_path C:\path\to\input.wav
+```
+
 **CUDA:** Add `--data_path voxtral_rt_exports/aoti_cuda_blob.ptd` to all
 run commands above when using the CUDA backend.