[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
+23 -24
View File
@@ -853,28 +853,30 @@ class CLIPEmbeddingModel(nn.Module, SupportsMultiModal, SupportsQuant):
self.text_embed_dim = text_config.hidden_size
self.vision_embed_dim = vision_config.hidden_size
self.text_model = CLIPTextTransformer(
text_config,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "text_model"),
)
self.vision_model = CLIPVisionTransformer(
vision_config,
quant_config=quant_config,
multimodal_config=multimodal_config,
prefix=maybe_prefix(prefix, "vision_model"),
)
with self._mark_language_model(vllm_config):
self.text_model = CLIPTextTransformer(
text_config,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "text_model"),
)
self.text_projection = nn.Linear(
self.text_embed_dim,
self.projection_dim,
bias=False,
)
self.visual_projection = nn.Linear(
self.vision_embed_dim,
self.projection_dim,
bias=False,
)
self.text_projection = nn.Linear(
self.text_embed_dim,
self.projection_dim,
bias=False,
)
with self._mark_tower_model(vllm_config, "image"):
self.vision_model = CLIPVisionTransformer(
vision_config,
quant_config=quant_config,
multimodal_config=multimodal_config,
prefix=maybe_prefix(prefix, "vision_model"),
)
self.visual_projection = nn.Linear(
self.vision_embed_dim,
self.projection_dim,
bias=False,
)
pooler_config = vllm_config.model_config.pooler_config
assert pooler_config is not None
@@ -940,9 +942,6 @@ class CLIPEmbeddingModel(nn.Module, SupportsMultiModal, SupportsQuant):
return self.get_image_features(pixel_values)
def get_language_model(self) -> torch.nn.Module:
return self.text_model
def _embed_text_input_ids(
self,
input_ids: torch.Tensor,