[Model] Use context managers for encoder- and LM-only mode (#32605)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 11:43:38 +08:00
committed by GitHub
parent 6c01ffb897
commit 4753f3bf69
21 changed files with 290 additions and 353 deletions
+8 -18
View File
@@ -457,8 +457,7 @@ class Mistral3ForConditionalGeneration(
):
config.projector_hidden_act = "gelu"
# TODO: Optionally initializes this for supporting embeddings.
if multimodal_config.get_limit_per_prompt("image"):
with self._mark_tower_model(vllm_config, "image"):
self.vision_tower = init_vision_tower_for_llava(
config,
quant_config=quant_config,
@@ -476,15 +475,13 @@ class Mistral3ForConditionalGeneration(
quant_config=quant_config,
prefix=maybe_prefix(prefix, "multi_modal_projector"),
)
else:
self.vision_tower = None
self.multi_modal_projector = None
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
@@ -534,9 +531,6 @@ class Mistral3ForConditionalGeneration(
image_embeds = (image_embeds,)
return image_embeds
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:
@@ -607,11 +601,7 @@ class Mistral3ForConditionalGeneration(
return self.language_model.compute_logits(hidden_states)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
skip_prefixes = []
if self.vision_tower is None and self.multi_modal_projector is None:
skip_prefixes = ["vision_tower.", "multi_modal_projector."]
loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
loader = AutoWeightsLoader(self)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
def get_mm_mapping(self) -> MultiModelKeys: