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

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

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 18:21:56 +08:00
committed by GitHub
parent 8be263c3fb
commit 7f1bcd18ff
10 changed files with 203 additions and 389 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
vllm/model_executor/models/interns1.py
View File

@@ -547,20 +547,20 @@ class InternS1ForConditionalGeneration(
)
self.downsample_ratio = config.downsample_ratio
self.llm_arch_name = config.text_config.architectures[0]
self.vision_tower = self._init_vision_model(
config,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "vision_tower"),
)
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.vision_tower = self._init_vision_model(
config,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "vision_tower"),
)
self.multi_modal_projector = self._init_mlp1(config)
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config.text_config,
prefix=maybe_prefix(prefix, "language_model"),
)
self.multi_modal_projector = self._init_mlp1(config)
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.img_context_token_id = None
self.video_context_token_id = None
@@ -699,8 +699,6 @@ class InternS1ForConditionalGeneration(
):
return image_input["data"]
assert self.vision_tower is not None
image_embeds = self.extract_feature(image_input["pixel_values"])
num_patches = image_input["num_patches"]
@@ -737,9 +735,6 @@ class InternS1ForConditionalGeneration(
def _set_visual_token_mask(self, input_ids: torch.Tensor) -> None:
self.visual_token_mask = None
def get_language_model(self) -> torch.nn.Module:
return self.language_model
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
if not modalities: