Reject beam search options up front instead of inconsistently

tests/test_beam_search.py

@@ -0,0 +1,41 @@
+"""Beam search is not implemented: say so instead of failing inconsistently."""
+
+import pytest
+
+from tests.conftest import StubModel
+from whisper_mlx import LightningWhisperMLX
+from whisper_mlx.cli import build_parser
+
+
+@pytest.fixture
+def stub(tokenizer, ts):
+    return StubModel(lambda o, w: [ts(0), *tokenizer.encode(" hello"), ts(2)])
+
+
+@pytest.mark.parametrize("batch_size", [1, 4])
+@pytest.mark.parametrize("option", [{"beam_size": 5}, {"patience": 1.0}])
+def test_transcribe_rejects_beam_search_options(
+    stub, run_transcribe, batch_size, option
+):
+    with pytest.raises(NotImplementedError, match="beam search"):
+        run_transcribe(stub, seconds=30, batch_size=batch_size, **option)
+
+
+def test_unset_beam_search_options_are_fine(stub, run_transcribe):
+    result = run_transcribe(stub, seconds=30, beam_size=None, patience=None)
+
+    assert result["text"] == " hello"
+
+
+def test_lightning_wrapper_rejects_beam_size():
+    whisper = LightningWhisperMLX(model="tiny")
+
+    with pytest.raises(NotImplementedError, match="beam search"):
+        whisper.transcribe("audio.mp3", beam_size=5)
+
+
+def test_cli_has_no_patience_option(capsys):
+    with pytest.raises(SystemExit):
+        build_parser().parse_args(["audio.mp3", "--patience", "1.0"])
+
+    assert "--patience" in capsys.readouterr().err

whisper_mlx/cli.py

@@ -118,12 +118,6 @@ def build_parser():
         default=5,
         help="Number of candidates when sampling with non-zero temperature",
     )
-    parser.add_argument(
-        "--patience",
-        type=float,
-        default=None,
-        help="Optional patience value to use in beam decoding, as in https://arxiv.org/abs/2204.05424, the default (1.0) is equivalent to conventional beam search",
-    )
     parser.add_argument(
         "--length-penalty",
         type=float,

whisper_mlx/lightning.py

@@ -104,7 +104,9 @@ def transcribe(
             - temperature, compression_ratio_threshold, logprob_threshold
             - no_speech_threshold, condition_on_previous_text, initial_prompt
             - prepend_punctuations, append_punctuations, clip_timestamps
-            - hallucination_silence_threshold, fp16, beam_size, patience, etc.
+            - hallucination_silence_threshold (batch_size=1 only), fp16
+            - suppress_tokens, best_of (batch_size=1 only), etc.
+            Beam search (beam_size, patience) is not implemented.
 
         Returns
         -------

whisper_mlx/transcribe.py

@@ -111,7 +111,9 @@ def transcribe(
         to make it more likely to predict those word correctly.
 
     decode_options: dict
-        Keyword arguments to construct `DecodingOptions` instances
+        Keyword arguments to construct `DecodingOptions` instances. Beam search
+        (`beam_size`, `patience`) is not implemented and raises NotImplementedError;
+        `best_of` only applies when batch_size is 1.
 
     clip_timestamps: Union[str, List[float]]
         Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to process.
@@ -134,6 +136,11 @@ def transcribe(
             f"batch_size={batch_size} may cause out-of-memory errors. "
             "Consider using batch_size <= 64."
         )
+    for option in ("beam_size", "patience"):
+        if decode_options.get(option) is not None:
+            raise NotImplementedError(
+                f"{option} is not supported: beam search is not implemented yet"
+            )
     if batch_size > 1 and hallucination_silence_threshold is not None:
         # skipping silence needs the window-by-window seeking of batch_size=1
         warnings.warn(