A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 11:00 UTC
[Bugfix] Fix HunyuanVL XD-RoPE (#29593)
Signed-off-by: Roger Wang <hey@rogerw.io> Co-authored by: grider-transwithai <grider@transwith.ai>
Roger Wang committed
Nov 27, 2025 at 04:36 UTC
cf348c8d27c34247f5976a86ebe6f4a3b4f9e888
2 files changed
+4
-4
vllm/model_executor/models/hunyuan_vision.py
+1
-1
@@ -847,7 +847,7 @@ class HunYuanVLForConditionalGeneration(
847
.expand(-1, llm_grid_w + 1)
848
.reshape(-1)
849
)
850
- h_index[pos : pos + token_num] = 0
850
+ t_index[pos : pos + token_num] = image_index
851
852
if xd_num == 4:
853
llm_positions = torch.stack([p_index, w_index, h_index, t_index])
vllm/transformers_utils/processors/hunyuan_vl_image.py
+3
-3
@@ -195,9 +195,9 @@ class HunYuanVLImageProcessor(BaseImageProcessor):
195
processed_images = []
196
for image in images:
197
if do_resize:
198
- resized_width, resized_height = smart_resize(
199
- width,
200
- height,
198
+ resized_height, resized_width = smart_resize(
199
+ height=height,
200
+ width=width,
201
factor=patch_size * merge_size,
202
min_pixels=self.min_pixels,
203
max_pixels=self.max_pixels,