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 15, 09:45 UTC
[Bugfix] Inconsistent token calculation compared to HF in llava family (#18479)
Signed-off-by: jaycha <jaycha@ncsoft.com>
youngrok cha committed
May 22, 2025 at 12:21 UTC
d022115cc603ed5985996f85460c3be5c7ae749e
2 files changed
+8
-4
vllm/model_executor/models/llava_next.py
+4
-2
@@ -135,11 +135,13 @@ class LlavaNextProcessingInfo(BaseLlavaProcessingInfo):
135
current_aspect_ratio = current_width / current_height
136
137
if aspect_ratio > current_aspect_ratio:
138
- new_height = (original_height * current_width) // original_width
138
+ new_height = int(
139
+ round(original_height * (current_width / original_width), 7))
140
padding = (current_height - new_height) // 2
141
current_height = current_height - (2 * padding)
142
else:
142
- new_width = (original_width * current_height) // original_height
143
+ new_width = int(
144
+ round(original_width * (current_height / original_height), 7))
145
padding = (current_width - new_width) // 2
146
current_width = current_width - (2 * padding)
147
vllm/model_executor/models/llava_onevision.py
+4
-2
@@ -116,11 +116,13 @@ class LlavaOnevisionProcessingInfo(LlavaNextProcessingInfo):
116
current_aspect_ratio = current_width / current_height
117
118
if aspect_ratio > current_aspect_ratio:
119
- new_height = (original_height * current_width) // original_width
119
+ new_height = int(
120
+ round(original_height * (current_width / original_width), 7))
121
padding = (current_height - new_height) // 2
122
current_height = current_height - (2 * padding)
123
else:
123
- new_width = (original_width * current_height) // original_height
124
+ new_width = int(
125
+ round(original_width * (current_height / original_height), 7))
126
padding = (current_width - new_width) // 2
127
current_width = current_width - (2 * padding)
128