[3/N] Initialize MM components in context managers (I-L) (#32650)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 10:21:56 +00:00
committed by GitHub
parent 8be263c3fb
commit 7f1bcd18ff
10 changed files with 203 additions and 389 deletions
+23 -25
View File
@@ -268,27 +268,30 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsP
self.config = config
self.multimodal_config = multimodal_config
# TODO: Optionally initializes this for supporting embeddings.
self.vision_tower = init_vision_tower_for_llava(
config,
quant_config=quant_config,
multimodal_config=multimodal_config,
require_post_norm=False,
prefix=maybe_prefix(prefix, "vision_tower"),
)
self.image_newline = nn.Parameter(torch.empty(config.text_config.hidden_size))
self.multi_modal_projector = LlavaMultiModalProjector(
vision_hidden_size=vision_hidden_size,
text_hidden_size=config.text_config.hidden_size,
projector_hidden_act=config.projector_hidden_act,
multimodal_projector_bias=config.multimodal_projector_bias,
)
with self._mark_tower_model(vllm_config, "image"):
self.vision_tower = init_vision_tower_for_llava(
config,
quant_config=quant_config,
multimodal_config=multimodal_config,
require_post_norm=False,
prefix=maybe_prefix(prefix, "vision_tower"),
)
self.image_newline = nn.Parameter(
torch.empty(config.text_config.hidden_size)
)
self.multi_modal_projector = LlavaMultiModalProjector(
vision_hidden_size=vision_hidden_size,
text_hidden_size=config.text_config.hidden_size,
projector_hidden_act=config.projector_hidden_act,
multimodal_projector_bias=config.multimodal_projector_bias,
)
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):
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config.text_config,
prefix=maybe_prefix(prefix, "language_model"),
)
self.make_empty_intermediate_tensors = (
self.language_model.make_empty_intermediate_tensors
@@ -427,8 +430,6 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsP
self,
inputs: LlavaNextImagePixelInputs,
) -> torch.Tensor | tuple[torch.Tensor, ...]:
assert self.vision_tower is not None
pixel_values = inputs["pixel_values"]
if isinstance(pixel_values, torch.Tensor):
@@ -480,9 +481,6 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsP
for i, patch_features_batch in enumerate(patch_embeddings)
]
def get_language_model(self) -> torch.nn.Module:
return self.language_model
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
image_input = self._parse_and_validate_image_input(**kwargs)
if image_input is None: