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

[5/N] Initialize MM components in context managers (Q-Z) (#32695)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-21 03:10:23 +08:00
committed by GitHub
parent f0feb1cf81
commit 193069d129
9 changed files with 178 additions and 168 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
vllm/model_executor/models/qwen2_audio.py
View File

@@ -334,20 +334,21 @@ class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal, Supports
multimodal_config = vllm_config.model_config.multimodal_config
self.config = config
self.multimodal_config = multimodal_config
self.audio_tower = Qwen2AudioEncoder(config.audio_config)
self.multi_modal_projector = Qwen2AudioMultiModalProjector(
config.audio_config.d_model, config.text_config.hidden_size
)
self.quant_config = quant_config
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config.text_config,
prefix=maybe_prefix(prefix, "language_model"),
architectures=["Qwen2ForCausalLM"],
)
with self._mark_tower_model(vllm_config, "audio"):
self.audio_tower = Qwen2AudioEncoder(config.audio_config)
self.multi_modal_projector = Qwen2AudioMultiModalProjector(
config.audio_config.d_model, config.text_config.hidden_size
)
with self._mark_language_model(vllm_config):
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config.text_config,
prefix=maybe_prefix(prefix, "language_model"),
architectures=["Qwen2ForCausalLM"],
)
self.make_empty_intermediate_tensors = (
self.language_model.make_empty_intermediate_tensors
@@ -441,9 +442,6 @@ class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal, Supports
masked_audio_features, audio_output_lengths.flatten().tolist()
)
def get_language_model(self) -> torch.nn.Module:
return self.language_model
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
audio_input = self._parse_and_validate_audio_input(**kwargs)
if audio_input is None: