[Model] Remove Mono-InternVL (InternLM2VEForCausalLM) (#45129)

Signed-off-by: Xianbao QIAN <xianbao.qian@gmail.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Tiezhen WANG
2026-06-12 10:41:09 +00:00
committed by GitHub
co-authored by Claude Isotr0py
parent 88ed636218
commit f1e13f7df9
9 changed files with 53 additions and 262 deletions
+12 -39
View File
@@ -23,7 +23,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.quantization.awq import AWQConfig
from vllm.model_executor.models.intern_vit import (
InternVisionModel,
InternVisionPatchModel,
)
from vllm.model_executor.models.module_mapping import MultiModelKeys
from vllm.multimodal import MULTIMODAL_REGISTRY
@@ -582,14 +581,10 @@ class InternVLChatModel(
self.downsample_ratio = config.downsample_ratio
self.ps_version = config.ps_version
llm_arch_name = config.text_config.architectures[0]
self.is_mono = llm_arch_name == "InternLM2VEForCausalLM"
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.vision_model = self._init_vision_model(
config,
quant_config=quant_config,
is_mono=self.is_mono,
prefix=maybe_prefix(prefix, "vision_model"),
)
self.mlp1 = self._init_mlp1(config)
@@ -604,7 +599,6 @@ class InternVLChatModel(
self.img_context_token_id = None
self.video_context_token_id = None
self.visual_token_mask = None
self.make_empty_intermediate_tensors = (
self.language_model.make_empty_intermediate_tensors
)
@@ -627,26 +621,22 @@ class InternVLChatModel(
config: PretrainedConfig,
quant_config: QuantizationConfig | None,
*,
is_mono: bool,
prefix: str,
):
if not is_mono:
vision_feature_layer = config.select_layer
if vision_feature_layer < 0:
num_hidden_layers = (
config.vision_config.num_hidden_layers + vision_feature_layer + 1
)
else:
num_hidden_layers = vision_feature_layer + 1
return InternVisionModel(
config.vision_config,
quant_config=quant_config,
num_hidden_layers_override=num_hidden_layers,
prefix=prefix,
vision_feature_layer = config.select_layer
if vision_feature_layer < 0:
num_hidden_layers = (
config.vision_config.num_hidden_layers + vision_feature_layer + 1
)
else:
return InternVisionPatchModel(config.vision_config)
num_hidden_layers = vision_feature_layer + 1
return InternVisionModel(
config.vision_config,
quant_config=quant_config,
num_hidden_layers_override=num_hidden_layers,
prefix=prefix,
)
def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:
vit_hidden_size = config.vision_config.hidden_size
@@ -805,15 +795,6 @@ class InternVLChatModel(
return modalities
def _set_visual_token_mask(self, input_ids: torch.Tensor) -> None:
if self.is_mono:
assert self.img_context_token_id is not None
self.visual_token_mask = (input_ids == self.img_context_token_id).reshape(
-1, 1
)
else:
self.visual_token_mask = None
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
if not modalities:
@@ -844,9 +825,6 @@ class InternVLChatModel(
*,
is_multimodal: torch.Tensor | None = None,
) -> torch.Tensor:
if multimodal_embeddings is not None and len(multimodal_embeddings) > 0:
self._set_visual_token_mask(input_ids)
# This is to satisfy the type checker for each overload
if multimodal_embeddings is None or is_multimodal is None:
return super().embed_input_ids(input_ids)
@@ -875,11 +853,6 @@ class InternVLChatModel(
"inputs_embeds": inputs_embeds,
}
# Only required if the model is mono-architecture
if self.visual_token_mask is not None:
forward_kwargs.update({"visual_token_mask": self.visual_token_mask})
self.visual_token_mask = None
hidden_states = self.language_model.model(**forward_kwargs)
return hidden_states