[1/N] Initialize MM components in context managers (A-D) (#32632)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 14:12:42 +08:00
committed by GitHub
parent 4753f3bf69
commit b75e85dede
11 changed files with 240 additions and 268 deletions
+13 -15
View File
@@ -460,20 +460,21 @@ class AudioFlamingo3ForConditionalGeneration(
multimodal_config = vllm_config.model_config.multimodal_config
self.config = config
self.multimodal_config = multimodal_config
self.audio_tower = AudioFlamingo3Encoder(
config.audio_config,
)
self.multi_modal_projector = AudioFlamingo3MultiModalProjector(config)
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 = AudioFlamingo3Encoder(
config.audio_config,
)
self.multi_modal_projector = AudioFlamingo3MultiModalProjector(config)
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
@@ -599,9 +600,6 @@ class AudioFlamingo3ForConditionalGeneration(
current_idx += count
return tuple(grouped_embeddings)
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: