transformers
Optimize Qwen2VL vision model by precomputing cos/sin embeds before ViT blocks
#35837

Merged

Optimize Qwen2VL vision model by precomputing cos/sin embeds before ViT blocks #35837

ArthurZucker merged 4 commits into huggingface:main from li-plus:fast-qwen2vl-vision-rope

qubvel added optimization

qubvel added Multimodal

zucchini-nlp commented on 2025-01-22

li-plus force pushed from 2e912f3f to 343e47ed 1 year ago

zucchini-nlp approved these changes on 2025-01-24

li-plus force pushed from 343e47ed to 64c28270 1 year ago

li-plus force pushed from 64c28270 to 8c220896 1 year ago

ArthurZucker commented on 2025-02-12

Optimize Qwen2VL vision model by precomputing cos/sin embeds before V…

2357261a

Make rotary_pos_emb optional & fix type

984c2ffb

Adapt pre-computed cos/sin to Qwen2.5VL

79907108

li-plus force pushed from 8c220896 to 79907108 1 year ago

More concise

65b620da

li-plus force pushed from 40d625d9 to 65b620da 1 year ago

ArthurZucker approved these changes on 2025-02-13

ArthurZucker merged 5f0fd118 into main 1 year ago

li-plus deleted the fast-qwen2vl-vision-rope branch 1 year ago

Reviewers

ArthurZucker

zucchini-nlp

qubvel

Assignees

No one assigned

Labels

optimization Multimodal

Milestone

No milestone