EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

model: Minicpmo (#3023)

Browse Source
This commit is contained in:
Mick
2025-03-25 11:08:40 +08:00
committed by GitHub
parent 64129fa632
commit 1e86457c90
40 changed files with 2906 additions and 493 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
python/sglang/srt/models/qwen2_5_vl.py
View File

@@ -57,7 +57,7 @@ from sglang.srt.managers.mm_utils import (
MultiModalityDataPaddingPatternTokenPairs,
general_mm_embed_routine,
)
from sglang.srt.managers.schedule_batch import ImageInputs
from sglang.srt.managers.schedule_batch import MultimodalInputs
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.models.qwen2 import Qwen2Model
@@ -513,7 +513,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
self.logits_processor = LogitsProcessor(config)
self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
def pad_input_ids(self, input_ids: List[int], image_inputs: ImageInputs):
def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
# Get all special token IDs
im_start_id: int = image_inputs.im_start_id
im_end_id: int = image_inputs.im_end_id
@@ -523,7 +523,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
return pattern.pad_input_tokens(input_ids, image_inputs)
def get_image_feature(self, image_input: ImageInputs) -> torch.Tensor:
def get_image_feature(self, image_input: MultimodalInputs) -> torch.Tensor:
pixel_values = image_input.pixel_values.type(self.visual.dtype)
image_embeds = self.visual(pixel_values, grid_thw=image_input.image_grid_thws)
return image_embeds
@@ -572,10 +572,9 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
inputs_embeds = general_mm_embed_routine(
input_ids=input_ids,
positions=positions,
forward_batch=forward_batch,
embed_tokens=self.get_input_embeddings(),
image_embedding_func=self.get_image_feature,
mm_data_embedding_func=self.get_image_feature,
)
hidden_states = self.model(