[4/N] Initialize MM components in context managers (M-P) (#32663)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 14:06:32 +00:00
committed by GitHub
parent bb9172030e
commit fda3f03eb2
24 changed files with 290 additions and 353 deletions
+11 -12
View File
@@ -820,16 +820,18 @@ class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal):
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
self.encoder = RadioWithNeck(
config=config, quant_config=quant_config, prefix=f"{prefix}.encoder"
)
with self._mark_tower_model(vllm_config, "image"):
self.encoder = RadioWithNeck(
config=config, quant_config=quant_config, prefix=f"{prefix}.encoder"
)
self.decoder = MBartDecoderNoPos(
config.decoder,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.decoder",
)
with self._mark_language_model(vllm_config):
self.decoder = MBartDecoderNoPos(
config.decoder,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.decoder",
)
self.vocab_size = config.decoder.vocab_size
self.lm_head = ParallelLMHead(
@@ -883,9 +885,6 @@ class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal):
pixel_values = pixel_values.to(dtype)
return self.encoder(pixel_values)
def get_language_model(self) -> torch.nn.Module:
return self.decoder
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None:
image_input = self._parse_and_validate_image_input(**kwargs)
if image_input is None: