zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-08-29 01:46:09 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge 04a7c07a9e into 4b2a0df237

Browse Source
This commit is contained in:
Awni Hannun 2025-08-25 09:45:17 -07:00 committed by GitHub
commit d57815c51c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
whisper/mlx_whisper/_version.py
View File

@ -1,3 +1,3 @@
# Copyright © 2023-2024 Apple Inc.
__version__ = "0.4.1"
__version__ = "0.4.2"

9
whisper/mlx_whisper/decoding.py
View File

@ -265,7 +265,7 @@ class GreedyDecoder(TokenDecoder):
else:
next_tokens = categorical(logits, self.temperature)
logprobs = logits - mx.logsumexp(logits, axis=-1)
logprobs = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
current_logprobs = logprobs[mx.arange(logprobs.shape[0]), next_tokens]
sum_logprobs += current_logprobs * (tokens[:, -1] != self.eot)
@ -380,7 +380,7 @@ class ApplyTimestampRules(LogitFilter):
# if sum of probability over timestamps is above any other token, sample timestamp
mask = mx.array(mask)
logprobs = logits - mx.logsumexp(logits, axis=-1)
logprobs = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
timestamp_logprob = logprobs[:, self.tokenizer.timestamp_begin :].logsumexp(
axis=-1, keepdims=True
)
@ -603,6 +603,7 @@ class DecodingTask:
inputs = tokens[:, -1:]
if tokens.shape[-1] > self.n_ctx:
break
next_tokens, next_completed, next_sum_logprobs, _ = _step(
inputs, audio_features, tokens, sum_logprobs
)
@ -643,9 +644,7 @@ class DecodingTask:
tokens = mx.broadcast_to(
tokens, [n_audio, self.n_group, len(self.initial_tokens)]
)
tokens = tokens.reshape(
tokens, (n_audio * self.n_group, len(self.initial_tokens))
)
tokens = tokens.reshape((n_audio * self.n_group, len(self.initial_tokens)))
# call the main sampling loop
tokens, sum_logprobs, no_speech_probs = self._main_loop(audio_features, tokens)