@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 0f961b3ce9
Fork 0

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 16, 15:45 UTC

[Bugfix] Fix incorrect llava next feature size calculation (#8496)

zyddnys committed Sep 20, 2024 at 18:48 UTC 0f961b3ce9ac3d3fd13e201c4358884bc094905e
1 file changed +8 -6
vllm/model_executor/models/llava_next.py
+8 -6
@@ -87,17 +87,19 @@ def _get_llava_next_num_unpadded_features(
87 current_height = npatches * num_patch_height
88 current_width = npatches * num_patch_width
89
90 - aspect_ratio = original_width / original_height
90 + original_aspect_ratio = original_width / original_height
91 current_aspect_ratio = current_width / current_height
92
93 - if aspect_ratio > current_aspect_ratio:
94 - new_height = (original_height * current_width) // original_width
93 + if original_aspect_ratio > current_aspect_ratio:
94 + scale_factor = current_width / original_width
95 + new_height = int(original_height * scale_factor)
96 padding = (current_height - new_height) // 2
96 - current_height -= padding * 2
97 + current_height -= 2 * padding
98 else:
98 - new_width = (original_width * current_height) // original_height
99 + scale_factor = current_height / original_height
100 + new_width = int(original_width * scale_factor)
101 padding = (current_width - new_width) // 2
100 - current_width -= padding * 2
102 + current_width -= 2 * padding
103
104 unpadded_features = current_height * current_width
105 newline_features = current_height