Mirroristas/whisper
2
0
Fork 0
mirror of https://github.com/openai/whisper.git synced 2025-11-24 06:26:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge a0936816d5624f975a039ffeb1ca25d473e26d5b into c0d2f624c09dc18e709e37c2ad90c039a4eb72a2

Browse Source
This commit is contained in:
Pramod Pai 2025-06-27 02:27:40 +00:00 committed by GitHub
commit b00382ac7b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
whisper/decoding.py
View File

@ -112,7 +112,13 @@ class DecodingOptions:
# implementation details # implementation details
fp16: bool = True # use fp16 for most of the calculation fp16: bool = True # use fp16 for most of the calculation
bf16: bool = False # use bf16 for most of the calculation
def __post_init__(self):
if self.fp16 and self.bf16:
raise ValueError("Both fp16 and bf16 cannot be True at the same time")
if self.bf16:
object.__setattr__(self, "fp16", False)
@dataclass(frozen=True) @dataclass(frozen=True)
class DecodingResult: class DecodingResult:
@ -655,7 +661,9 @@ class DecodingTask:
audio_features = self.model.encoder(mel) audio_features = self.model.encoder(mel)
if audio_features.dtype != ( if audio_features.dtype != (
torch.float16 if self.options.fp16 else torch.float32 torch.float16 if self.options.fp16 else
torch.bfloat16 if self.options.bf16 else
torch.float32
): ):
return TypeError( return TypeError(
f"audio_features has an incorrect dtype: {audio_features.dtype}" f"audio_features has an incorrect dtype: {audio_features.dtype}"

2
whisper/transcribe.py
View File

@ -132,7 +132,7 @@ def transcribe(
warnings.warn("FP16 is not supported on CPU; using FP32 instead") warnings.warn("FP16 is not supported on CPU; using FP32 instead")
dtype = torch.float32 dtype = torch.float32
if dtype == torch.float32: if dtype == torch.float32 or dtype == torch.bfloat16:
decode_options["fp16"] = False decode_options["fp16"] = False
# Pad 30-seconds of silence to the input audio, for slicing # Pad 30-seconds of silence to the input audio, for slicing