Mirroristas/whisper
2
0
Fork 0
mirror of https://github.com/openai/whisper.git synced 2025-11-23 22:15:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

Avoid rearranging all caches (#1483)

Browse Source 
* avoid rearranging all kv_caches

* avoid calculating the same kv_cache from cross attn

* Update decoding.py

* linter fix

---------

Co-authored-by: Jong Wook Kim <jongwook@openai.com>
This commit is contained in:
WangChou Lu 2023-07-07 03:48:08 +08:00 committed by GitHub
parent f572f2161b
commit b91c907694
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 9 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
whisper/decoding.py
View File

@ -146,6 +146,10 @@ class PyTorchInference(Inference):
self.kv_cache = {} self.kv_cache = {}
self.hooks = [] self.hooks = []
key_modules = [block.attn.key for block in self.model.decoder.blocks]
value_modules = [block.attn.value for block in self.model.decoder.blocks]
self.kv_modules = key_modules + value_modules
def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor: def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
if not self.kv_cache: if not self.kv_cache:
self.kv_cache, self.hooks = self.model.install_kv_cache_hooks() self.kv_cache, self.hooks = self.model.install_kv_cache_hooks()
@ -164,9 +168,10 @@ class PyTorchInference(Inference):
self.hooks = [] self.hooks = []
def rearrange_kv_cache(self, source_indices): def rearrange_kv_cache(self, source_indices):
for module, tensor in self.kv_cache.items(): if source_indices != list(range(len(source_indices))):
# update the key/value cache to contain the selected sequences for module in self.kv_modules:
self.kv_cache[module] = tensor[source_indices].detach() # update the key/value cache to contain the selected sequences
self.kv_cache[module] = self.kv_cache[module][source_indices].detach()
class SequenceRanker: class SequenceRanker:
@ -668,7 +673,6 @@ class DecodingTask:
return languages, lang_probs return languages, lang_probs
def _main_loop(self, audio_features: Tensor, tokens: Tensor): def _main_loop(self, audio_features: Tensor, tokens: Tensor):
assert audio_features.shape[0] == tokens.shape[0]
n_batch = tokens.shape[0] n_batch = tokens.shape[0]
sum_logprobs: Tensor = torch.zeros(n_batch, device=audio_features.device) sum_logprobs: Tensor = torch.zeros(n_batch, device=audio_features.device)
no_speech_probs = [np.nan] * n_batch no_speech_probs = [np.nan] * n_batch
@ -721,8 +725,7 @@ class DecodingTask:
) )
] ]
# repeat the audio & text tensors by the group size, for beam search or best-of-n sampling # repeat text tensors by the group size, for beam search or best-of-n sampling
audio_features = audio_features.repeat_interleave(self.n_group, dim=0)
tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device) tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device)
# call the main sampling loop # call the main sampling loop