chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Upgrade transformers==5.3.0 (#17784)

Browse Source 
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>
Co-authored-by: Alison Shao <alisonshao@mac.lan>
Co-authored-by: Mick <mickjagger19@icloud.com>
This commit is contained in:
Xinyuan Tong
2026-03-18 20:50:43 +00:00
committed by GitHub
parent e5750a572c
commit d1e95af282
95 changed files with 1134 additions and 341 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/multimodal/mm_utils.py
View File

@@ -48,6 +48,11 @@ from sglang.srt.distributed.communication_op import tensor_model_parallel_all_ga
from sglang.srt.utils import flatten_nested_list
def ensure_numpy(x):
"""Convert torch.Tensor to numpy array if needed (v5 compat)."""
return x.numpy() if isinstance(x, torch.Tensor) else x
def has_valid_data(data) -> bool:
if data is None:
return False
@@ -237,10 +242,11 @@ def process_anyres_image(image, processor, grid_pinpoints):
best_resolution = select_best_resolution(image.size, possible_resolutions)
image_padded = resize_and_pad_image(image, best_resolution)
# For Siglip processor, only have size but no crop size
# For Siglip processor, only have size but no crop size.
# In transformers v5, crop_size may exist but be None.
crop_size = (
processor.crop_size["height"]
if "crop_size" in processor.__dict__
if getattr(processor, "crop_size", None) is not None
else processor.size["height"]
)
shortest_edge = (
@@ -257,6 +263,8 @@ def process_anyres_image(image, processor, grid_pinpoints):
processor.preprocess(image_patch.convert("RGB"))["pixel_values"][0]
for image_patch in image_patches
]
# In transformers v5, image processors may return torch.Tensor instead of numpy arrays
image_patches = [ensure_numpy(p) for p in image_patches]
return np.stack(image_patches, axis=0)

11
python/sglang/srt/multimodal/processors/llava.py
View File

@@ -16,7 +16,11 @@ from sglang.srt.models.llava import (
)
from sglang.srt.models.llavavid import LlavaVidForCausalLM
from sglang.srt.models.mistral import Mistral3ForConditionalGeneration
from sglang.srt.multimodal.mm_utils import expand2square, process_anyres_image
from sglang.srt.multimodal.mm_utils import (
ensure_numpy,
expand2square,
process_anyres_image,
)
from sglang.srt.multimodal.processors.base_processor import BaseMultimodalProcessor
from sglang.srt.utils import ImageData, load_image, logger
from sglang.utils import get_exception_traceback
@@ -50,8 +54,8 @@ class LlavaImageProcessor(BaseMultimodalProcessor):
# It is a video with multiple images
image_hash = hash(url)
pixel_values = image_processor(image)["pixel_values"]
for _ in range(len(pixel_values)):
pixel_values[_] = pixel_values[_].astype(np.float16)
for i in range(len(pixel_values)):
pixel_values[i] = ensure_numpy(pixel_values[i]).astype(np.float16)
pixel_values = np.stack(pixel_values, axis=0)
return pixel_values, image_hash, image_size
else:
@@ -75,6 +79,7 @@ class LlavaImageProcessor(BaseMultimodalProcessor):
else:
pixel_values = image_processor(image)["pixel_values"][0]
pixel_values = ensure_numpy(pixel_values)
if isinstance(pixel_values, np.ndarray):
pixel_values = pixel_values.astype(np.float16)