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

[VLM RLHF] Take Image input for verl vlm rollout (#4915)

Browse Source 
Signed-off-by: Xinyuan Tong <justinning0323@outlook.com>
Co-authored-by: GeLee <leege233@gmail.com>
This commit is contained in:
XinyuanTong
2025-04-01 20:03:17 -07:00
committed by GitHub
parent 12047f5e94
commit 9eb49e878b
4 changed files with 18 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/sglang/srt/managers/multimodal_processors/base_processor.py
View File

@@ -139,8 +139,6 @@ class BaseMultimodalProcessor(ABC):
else:
multimodal_tokens.image_token = multimodal_tokens.image_token
assert isinstance(prompt, str)
if isinstance(prompt, list) and return_text:
assert len(prompt) and isinstance(prompt[0], int)
prompt = self._processor.tokenizer.decode(prompt)
@@ -204,7 +202,16 @@ class BaseMultimodalProcessor(ABC):
continue
image_sizes += frames[0].size * len(frames)
hashes += [hash(image_file)] * len(frames)
# Generate a hashable value for the image file
if isinstance(image_file, Image.Image):
# For PIL.Image objects, use the ID as a hashable value
hash_value = hash(id(image_file))
else:
# For other types (strings, etc.), use the regular hash
hash_value = hash(image_file)
hashes += [hash_value] * len(frames)
images += frames
image_index += 1
if frames_to_process != 0:

2
python/sglang/srt/managers/multimodal_processors/qwen_vl.py
View File

@@ -5,7 +5,7 @@ from typing import List, Union
import torch
from PIL import Image
from sglang.srt.managers.multimodal_processor import (
from sglang.srt.managers.multimodal_processors.base_processor import (
BaseMultimodalProcessor as SGLangBaseProcessor,
)
from sglang.srt.managers.multimodal_processors.base_processor import (