Scale the default Whisper batch size to the detected GPU memory

README.md

@@ -261,7 +261,7 @@ _Not all options working now!_
                             For video URLs, the video language is used automatically.
                             WARNING: setting this for non-matching songs will degrade
                             alignment quality (e.g. --language en for German songs).
-    --whisper_batch_size    Segments processed in parallel. Lower if low on GPU mem: slower, but transcription is UNCHANGED (the safe lever) >> ((default) is 16)
+    --whisper_batch_size    Segments processed in parallel. 'auto' (default) scales to GPU VRAM: 16 (>=8GB) / 8 (6GB) / 4 (4GB). Lower manually if still low on GPU mem: slower, but transcription is UNCHANGED (the safe lever)
     --whisper_compute_type  Change to "int8" to save more GPU mem at a small accuracy cost; use only if lowering the batch size is not enough >> ((default) is "float16" for cuda devices, "int8" for cpu)
     --keep_numbers          Numbers will be transcribed as numerics instead of as words
     --vad_onset             VAD (Voice Activity Detection) speech activation threshold (0.0-1.0). Lower
@@ -854,12 +854,15 @@ uv pip install --index-url https://download.pytorch.org/whl/cu121 torch torchvis
 
 #### Crashes due to low VRAM
 
-The default Whisper model (`large-v2`) needs more than 8 GB VRAM. On a smaller
-GPU there are two independent ways to fit it, in order of preference:
+At full batch size (16) the default Whisper model (`large-v2`) needs more than
+8 GB VRAM. UltraSinger scales the batch size to the detected GPU memory
+automatically (16 on 8+ GB, 8 on 6 GB, 4 on 4 GB cards), so most GPUs fit
+as-is. If it still runs out of memory, there are two independent levers, in
+order of preference:
 
-1. **Lower the batch size** (`--whisper_batch_size 4`, or 2/1). This processes
-   fewer audio segments in parallel: it is slower, but the transcription is
-   **unchanged** — so this is the safe lever, try it first.
+1. **Lower the batch size further** (`--whisper_batch_size 2`, or 1). This
+   processes fewer audio segments in parallel: it is slower, but the
+   transcription is **unchanged** — so this is the safe lever, try it first.
 2. **Switch to int8** (`--whisper_compute_type int8`). This halves the model's
    memory at a small accuracy cost — add it only if lowering the batch size
    alone is not enough.

install/auto_install.bat

@@ -220,13 +220,13 @@ if "!BUILD!"=="cpu" (
 ) else (
     if not defined GPU_VRAM (
         echo GPU VRAM could not be verified ^(forced CUDA build^).
-        echo If your GPU has less than 8 GB VRAM, the default Whisper model
-        echo ^(large-v2^) may run out of memory. Two independent ways to fit it
-        echo ^(not set automatically^):
-        echo   --whisper_batch_size 4    Fewer segments in parallel: slower,
-        echo                             but the transcription is UNCHANGED
-        echo                             ^(the safe lever; lower to 2 or 1 if it
-        echo                             still runs out^).
+        echo UltraSinger scales the Whisper batch size automatically to the
+        echo GPU memory ^(16 on 8+ GB, 8 on 6 GB, 4 on 4 GB cards^), so the
+        echo default Whisper model ^(large-v2^) should fit as-is. If it still
+        echo runs out of memory:
+        echo   --whisper_batch_size 2    ^(or 1^) Fewer segments in parallel:
+        echo                             slower, but the transcription is
+        echo                             UNCHANGED ^(the safe lever^).
         echo   --whisper_compute_type int8    Halves the model's memory at a
         echo                             small accuracy cost - add this only if
         echo                             lowering the batch size is not enough.
@@ -238,13 +238,15 @@ if "!BUILD!"=="cpu" (
         echo GUI users: enable this under Settings -^> 'Remote Speech-to-Text'.
     ) else (
         if !GPU_VRAM! LSS 8192 (
-            echo Your GPU has less than 8 GB VRAM. The default Whisper model
-            echo ^(large-v2^) may run out of memory. Two independent ways to fit
-            echo it ^(not set automatically^):
-            echo   --whisper_batch_size 4    Fewer segments in parallel: slower,
-            echo                             but the transcription is UNCHANGED
-            echo                             ^(the safe lever; lower to 2 or 1 if
-            echo                             it still runs out^).
+            echo Your GPU has less than 8 GB VRAM. UltraSinger scales the
+            echo Whisper batch size automatically to the GPU memory ^(16 on
+            echo 8+ GB, 8 on 6 GB, 4 on 4 GB cards^), so the default Whisper
+            echo model ^(large-v2^) should fit as-is. If it still runs out of
+            echo memory:
+            echo   --whisper_batch_size 2    ^(or 1^) Fewer segments in
+            echo                             parallel: slower, but the
+            echo                             transcription is UNCHANGED
+            echo                             ^(the safe lever^).
             echo   --whisper_compute_type int8    Halves the model's memory at a
             echo                             small accuracy cost - add this only
             echo                             if lowering the batch size is not

install/auto_install.sh

@@ -246,13 +246,13 @@ case "$BUILD" in
     cuda)
         if [ -z "$GPU_VRAM" ]; then
             echo "GPU VRAM could not be verified (forced CUDA build)."
-            echo "If your GPU has less than 8 GB VRAM, the default Whisper model"
-            echo "(large-v2) may run out of memory. Two independent ways to fit it"
-            echo "(not set automatically):"
-            echo "  --whisper_batch_size 4    Fewer segments in parallel: slower,"
-            echo "                            but the transcription is UNCHANGED"
-            echo "                            (the safe lever; lower to 2 or 1 if"
-            echo "                            it still runs out)."
+            echo "UltraSinger scales the Whisper batch size automatically to the"
+            echo "GPU memory (16 on 8+ GB, 8 on 6 GB, 4 on 4 GB cards), so the"
+            echo "default Whisper model (large-v2) should fit as-is. If it still"
+            echo "runs out of memory:"
+            echo "  --whisper_batch_size 2    (or 1) Fewer segments in parallel:"
+            echo "                            slower, but the transcription is"
+            echo "                            UNCHANGED (the safe lever)."
             echo "  --whisper_compute_type int8   Halves the model's memory at a"
             echo "                            small accuracy cost - add this only"
             echo "                            if lowering the batch size is not enough."
@@ -263,13 +263,14 @@ case "$BUILD" in
             echo "runs transcription in the cloud instead of on your GPU."
             echo "GUI users: enable this under Settings -> 'Remote Speech-to-Text'."
         elif [ "$GPU_VRAM" -lt 8192 ]; then
-            echo "Your GPU has less than 8 GB VRAM. The default Whisper model"
-            echo "(large-v2) may run out of memory. Two independent ways to fit it"
-            echo "(not set automatically):"
-            echo "  --whisper_batch_size 4    Fewer segments in parallel: slower,"
-            echo "                            but the transcription is UNCHANGED"
-            echo "                            (the safe lever; lower to 2 or 1 if"
-            echo "                            it still runs out)."
+            echo "Your GPU has less than 8 GB VRAM. UltraSinger scales the"
+            echo "Whisper batch size automatically to the GPU memory (16 on"
+            echo "8+ GB, 8 on 6 GB, 4 on 4 GB cards), so the default Whisper"
+            echo "model (large-v2) should fit as-is. If it still runs out of"
+            echo "memory:"
+            echo "  --whisper_batch_size 2    (or 1) Fewer segments in parallel:"
+            echo "                            slower, but the transcription is"
+            echo "                            UNCHANGED (the safe lever)."
             echo "  --whisper_compute_type int8   Halves the model's memory at a"
             echo "                            small accuracy cost - add this only"
             echo "                            if lowering the batch size is not enough."

pytest/gui/test_freestyle_threshold_widgets.py

@@ -204,3 +204,24 @@ def test_default_thresholds_round_trip_to_settings_defaults(self):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class TestBatchSizeAutoArgs(unittest.TestCase):
+    """Batch size 0 means Auto: the flag must be omitted so the CLI can
+    scale it to the detected GPU memory; explicit values are forwarded."""
+
+    def setUp(self):
+        self.runner = UltraSingerRunner()
+
+    def test_auto_omits_batch_size_flag(self):
+        config = dict(_DEFAULTS)
+        config["whisper_batch_size"] = 0
+        args = self.runner.build_args(config, "test.mp3")
+        self.assertNotIn("--whisper_batch_size", args)
+
+    def test_explicit_batch_size_is_forwarded(self):
+        config = dict(_DEFAULTS)
+        config["whisper_batch_size"] = 8
+        args = self.runner.build_args(config, "test.mp3")
+        self.assertIn("--whisper_batch_size", args)
+        self.assertEqual(args[args.index("--whisper_batch_size") + 1], "8")

pytest/modules/DeviceDetection/test_device_detection.py

@@ -0,0 +1,52 @@
+"""Tests for the device detection module (auto Whisper batch size)."""
+
+from unittest.mock import MagicMock, patch
+
+from modules.DeviceDetection.device_detection import auto_whisper_batch_size
+
+
+def _gpu_props(vram_gb: float):
+    props = MagicMock()
+    props.total_memory = int(vram_gb * 1024 ** 3)
+    return props
+
+
+class TestAutoWhisperBatchSize:
+    """VRAM-based scaling: 16 on 8+ GB, 8 on 6 GB, 4 on 4 GB cards."""
+
+    def test_cpu_keeps_historic_default(self):
+        assert auto_whisper_batch_size("cpu") == 16
+
+    def test_cuda_unavailable_keeps_historic_default(self):
+        with patch("torch.cuda.is_available", return_value=False):
+            assert auto_whisper_batch_size("cuda") == 16
+
+    def test_4gb_card_gets_4(self):
+        with patch("torch.cuda.is_available", return_value=True), \
+             patch("torch.cuda.get_device_properties",
+                   return_value=_gpu_props(4.0)):
+            assert auto_whisper_batch_size("cuda") == 4
+
+    def test_6gb_card_gets_8(self):
+        with patch("torch.cuda.is_available", return_value=True), \
+             patch("torch.cuda.get_device_properties",
+                   return_value=_gpu_props(6.0)):
+            assert auto_whisper_batch_size("cuda") == 8
+
+    def test_8gb_card_gets_16(self):
+        with patch("torch.cuda.is_available", return_value=True), \
+             patch("torch.cuda.get_device_properties",
+                   return_value=_gpu_props(8.0)):
+            assert auto_whisper_batch_size("cuda") == 16
+
+    def test_12gb_card_gets_16(self):
+        with patch("torch.cuda.is_available", return_value=True), \
+             patch("torch.cuda.get_device_properties",
+                   return_value=_gpu_props(12.0)):
+            assert auto_whisper_batch_size("cuda") == 16
+
+    def test_query_failure_keeps_historic_default(self):
+        with patch("torch.cuda.is_available", return_value=True), \
+             patch("torch.cuda.get_device_properties",
+                   side_effect=RuntimeError("driver in a bad state")):
+            assert auto_whisper_batch_size("cuda") == 16

src/Settings.py

@@ -90,7 +90,7 @@ class Settings:
     whisper_model = WhisperModel.LARGE_V2  # Multilingual model tiny|base|small|medium|large-v1|large-v2|large-v3
     # English-only model tiny.en|base.en|small.en|medium.en
     whisper_align_model = None   # Model for other languages from huggingface.co e.g -> "gigant/romanian-wav2vec2"
-    whisper_batch_size = 16   # reduce if low on GPU mem
+    whisper_batch_size = None  # None = auto: scaled to GPU VRAM (16 / 8 / 4); set an int to override
     whisper_compute_type = None   # change to "int8" if low on GPU mem (may reduce accuracy)
     keep_numbers = False
 

src/UltraSinger.py

@@ -97,7 +97,10 @@
 from modules.musicbrainz_client import search_musicbrainz
 from modules.sheet import create_sheet
 from modules.ProcessData import ProcessData, ProcessDataPaths, MediaInfo
-from modules.DeviceDetection.device_detection import check_gpu_support
+from modules.DeviceDetection.device_detection import (
+    auto_whisper_batch_size,
+    check_gpu_support,
+)
 from modules.Image.image_helper import save_image
 from modules.ffmpeg_helper import (
     is_ffmpeg_available,
@@ -1923,6 +1926,19 @@ def main(argv: list[str]) -> None:
 def check_requirements() -> None:
     if not settings.force_cpu:
         settings.pytorch_device = check_gpu_support()
+
+    # Resolve the automatic Whisper batch size now that the device (and
+    # its VRAM) is known. An explicit --whisper_batch_size wins.
+    if settings.whisper_batch_size is None:
+        whisper_device = (
+            "cpu" if settings.force_whisper_cpu else settings.pytorch_device
+        )
+        settings.whisper_batch_size = auto_whisper_batch_size(whisper_device)
+        print(
+            f"{ULTRASINGER_HEAD} Whisper batch size: "
+            f"{blue_highlighted(str(settings.whisper_batch_size))} "
+            f"(auto, scaled to GPU memory)"
+        )
     print(f"{ULTRASINGER_HEAD} ----------------------")
 
     if not is_ffmpeg_available(settings.user_ffmpeg_path):
@@ -1988,7 +2004,9 @@ def init_settings(argv: list[str]) -> Settings:
         elif opt in ("--whisper_align_model"):
             settings.whisper_align_model = arg
         elif opt in ("--whisper_batch_size"):
-            settings.whisper_batch_size = int(arg)
+            settings.whisper_batch_size = (
+                None if arg.strip().lower() == "auto" else int(arg)
+            )
         elif opt in ("--whisper_compute_type"):
             settings.whisper_compute_type = arg
         elif opt in ("--keep_numbers"):

src/gui/config.py

@@ -26,7 +26,7 @@
     "output_folder": "",
     # Whisper
     "whisper_model": "large-v2",
-    "whisper_batch_size": 16,
+    "whisper_batch_size": 0,  # 0 = Auto: UltraSinger scales it to GPU VRAM
     "whisper_compute_type": "",
     "whisper_align_model": "",
     # VAD / ASR (tuned for singing)

src/gui/settings_tab.py

@@ -133,15 +133,20 @@ def _build_transcription_section(self):
                          _DEFAULTS["whisper_model"]))
 
         self._whisper_batch_size = _NoScrollSpinBox()
-        self._whisper_batch_size.setRange(1, 64)
-        self._whisper_batch_size.setValue(self._config.get("whisper_batch_size", 16))
+        self._whisper_batch_size.setRange(0, 64)
+        self._whisper_batch_size.setSpecialValueText("Auto")
+        self._whisper_batch_size.setValue(
+            self._config.get("whisper_batch_size", 0))
         card.add_row("Batch Size", self._whisper_batch_size,
                      "How many audio segments Whisper processes at once. "
+                     "'Auto' (recommended) scales to your GPU memory: "
+                     "16 on 8+ GB, 8 on 6 GB, 4 on 4 GB cards. "
                      "Higher values are faster but use more GPU memory. "
-                     "Lower it (4, or 2/1) if you get out-of-memory errors: "
-                     "this only makes transcription slower — the result is "
-                     "unchanged, so it is the safe lever to try first "
-                     "(before switching Compute Type to int8).",
+                     "Set a low value (4, or 2/1) if you still get "
+                     "out-of-memory errors: this only makes transcription "
+                     "slower — the result is unchanged, so it is the safe "
+                     "lever to try first (before switching Compute Type to "
+                     "int8).",
                      reset_callback=lambda: self._whisper_batch_size.setValue(
                          _DEFAULTS["whisper_batch_size"]))
 

src/modules/DeviceDetection/device_detection.py

@@ -92,6 +92,28 @@ def _enable_deterministic_mode():
     )
 
 
+def auto_whisper_batch_size(device: str) -> int:
+    """Pick a Whisper batch size that fits the GPU memory.
+
+    Field-tested scaling: the historic default of 16 needs roughly 8 GB
+    of VRAM with the large-v2 model; 8 runs stable on 6 GB cards and 4
+    on 4 GB cards. A lower batch size only slows transcription down -
+    the result is unchanged. On CPU there is no VRAM cliff, so the
+    historic default is kept.
+    """
+    if device != "cuda" or not torch.cuda.is_available():
+        return 16
+    try:
+        vram_gb = torch.cuda.get_device_properties(0).total_memory / 1024 ** 3
+    except RuntimeError:
+        return 16
+    if vram_gb < 5:
+        return 4
+    if vram_gb < 7:
+        return 8
+    return 16
+
+
 def __check_pytorch_support():
     pytorch_gpu_supported = torch.cuda.is_available()
     if not pytorch_gpu_supported:
@@ -105,6 +127,8 @@ def __check_pytorch_support():
         print(f"{ULTRASINGER_HEAD} Found GPU: {blue_highlighted(gpu_name)} VRAM: {blue_highlighted(gpu_vram)} GB.")
         if gpu_vram < 6:
             print(
-                f"{ULTRASINGER_HEAD} {red_highlighted('GPU VRAM is less than 6GB. Program may crash due to insufficient memory.')}")
+                f"{ULTRASINGER_HEAD} GPU VRAM is less than 6GB - the Whisper "
+                f"batch size is scaled down automatically to fit (a manual "
+                f"--whisper_batch_size overrides this).")
         print(f"{ULTRASINGER_HEAD} {blue_highlighted('pytorch')} - using {red_highlighted('cuda')} gpu.")
     return pytorch_gpu_supported

src/modules/common_print.py

@@ -65,7 +65,7 @@ def print_help() -> None:
                             via Whisper tiny (~2-3s) instead of full transcription.
                             WARNING: setting this for non-matching songs will degrade
                             alignment quality (e.g. --language en for German songs).
-    --whisper_batch_size    Segments processed in parallel. Lower if low on GPU mem: slower, but transcription is UNCHANGED (the safe lever) >> ((default) is 16)
+    --whisper_batch_size    Segments processed in parallel. 'auto' (default) scales to GPU VRAM: 16 (>=8GB) / 8 (6GB) / 4 (4GB). Lower manually if still low on GPU mem: slower, but transcription is UNCHANGED (the safe lever)
     --whisper_compute_type  Change to "int8" to save more GPU mem at a small accuracy cost; use only if lowering the batch size is not enough >> ((default) is "float16" for cuda devices, "int8" for cpu)
     --keep_numbers          Numbers will be transcribed as numerics instead of as words >> True|False >> ((default) is False)
     --vad_onset             VAD (Voice Activity Detection) speech activation threshold (0.0-1.0). Lower

src/modules/init_interactive_mode.py

@@ -59,11 +59,11 @@ def configure_additional_options(console, settings, header):
     if additional_options_input == 'y':  
         console.print(f"\n{header} [bold underline]Additional options:[/bold underline]\n")  
 
-        # Whisper Batch Size  
-        whipser_batch_size_response = console.input(  
-            f"{header} Enter the [green]Whisper batch size[/green] (default [cyan]16[/cyan]): "  
-        ).strip()  
-        settings.whisper_batch_size = int(whipser_batch_size_response) if whipser_batch_size_response.isdigit() else 16  
+        # Whisper Batch Size
+        whipser_batch_size_response = console.input(
+            f"{header} Enter the [green]Whisper batch size[/green] (default [cyan]auto[/cyan] - scaled to GPU memory): "
+        ).strip()
+        settings.whisper_batch_size = int(whipser_batch_size_response) if whipser_batch_size_response.isdigit() else None
 
         # Whisper Compute Type  
         whisper_compute_choice = console.input(