biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Extend collect_children and no_init_weights contexts (#32757)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-22 16:20:27 +08:00
committed by GitHub
parent 1bf1a34b19
commit 2b8a38b6d6
20 changed files with 444 additions and 257 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/models/whisper.py
View File

@@ -909,7 +909,12 @@ class WhisperForConditionalGeneration(
self.config = config
self.dtype = vllm_config.model_config.dtype
self.model = WhisperModel(vllm_config=vllm_config, prefix=prefix)
with self._mark_composite_model(
vllm_config,
language_targets=WhisperDecoder,
tower_targets={"audio": WhisperEncoder},
):
self.model = WhisperModel(vllm_config=vllm_config, prefix=prefix)
self.proj_out = ParallelLMHead(
config.vocab_size,
@@ -937,9 +942,6 @@ class WhisperForConditionalGeneration(
)
return decoder_outputs
def get_language_model(self) -> torch.nn.Module:
return self.model.decoder
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
# Required as part of SupportsMultiModal interface.
audio_input = self._parse_and_validate_audio_input(**kwargs)