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

[2/N] Initialize MM components in context managers (E-H) (#32641)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 16:12:56 +08:00
committed by GitHub
parent 148117ea2e
commit e1a34c3a5d
12 changed files with 161 additions and 189 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
vllm/model_executor/models/granite_speech.py
View File

@@ -597,27 +597,29 @@ class GraniteSpeechForConditionalGeneration(
self.quant_config = quant_config
self.cache_config = cache_config
# The language model is typically a Granite LLM
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config.text_config,
prefix=maybe_prefix(prefix, "language_model"),
)
with self._mark_language_model(vllm_config):
# The language model is typically a Granite LLM
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config.text_config,
prefix=maybe_prefix(prefix, "language_model"),
)
# Conformer encoder
self.encoder = GraniteSpeechCTCEncoder(
config=config.encoder_config,
quant_config=quant_config,
prefix=f"{prefix}.encoder",
)
with self._mark_tower_model(vllm_config, "audio"):
# Conformer encoder
self.encoder = GraniteSpeechCTCEncoder(
config=config.encoder_config,
quant_config=quant_config,
prefix=f"{prefix}.encoder",
)
# Blip2 QFormer
self.projector = GraniteSpeechEncoderProjector(
config=config,
quant_config=quant_config,
cache_config=cache_config,
prefix=f"{prefix}.projector",
)
# Blip2 QFormer
self.projector = GraniteSpeechEncoderProjector(
config=config,
quant_config=quant_config,
cache_config=cache_config,
prefix=f"{prefix}.projector",
)
self.make_empty_intermediate_tensors = (
self.language_model.make_empty_intermediate_tensors
@@ -770,9 +772,6 @@ class GraniteSpeechForConditionalGeneration(
# Split variable length features into a tuple
return torch.split(masked_embeds, audio_input["audio_embed_sizes"])
def get_language_model(self) -> torch.nn.Module:
return self.language_model
def embed_multimodal(
self,
**kwargs: object,