@qualcomm.bsky.social @machinelearning.bsky.social

Excited to announce that our Paper from #Qualcomm Canada has been accepted at #NeurIPS2025 OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding.
Looking forward to sharing our work at NeurIPS 2025. @neuripsconf.bsky.social
Preprint - arxiv.org/abs/2507.02659

Thanks everyone for the wonderful and engaging discussions during our poster sessions and live demo of our work - Stepping Forward On The Last Mile. Thank you 😊.
arxiv.org/html/2411.04...

@neuripsconf.bsky.social
@qualcomm.bsky.social

Gosh. Thanks I got confused if this was another pack. Apologies:)

I'd like to be added. Thanks

Alright thank you for the clarification.

I did work on structured pruning on weights a few years ago and as we were focused on deployment to edge devices , it was critical. But this approach on the activation/attention head is interesting although the inference graph wouldn't really change on the base model. Would love to further discuss.

I do have a few points of clarification. Maybe I will drop by on the poster session.
But the stage 1 looks to me like it's structured pruning on the activation. What I am curious about is does this approach help improving inference ? So I presume we won't need to do compute for certain heads.

Neurips 2024 poster link
neurips.cc/virtual/2024...

@neuripsconf.bsky.social

Very interesting work on LoFiT!

I'd like to be added to the pack as I would be at NeurIPS 2024 as well. Thanks

@neuripsconf.bsky.social @qualcomm.bsky.social

Qualcomm at NeurIPS 2024: Our groundbreaking innovations and cutting-edge advancements in AI | Qualcomm 2024 is notable for the remarkable advancements in generative artificial intelligence (GenAI), and Qualcomm Technologies is at the forefront of bringing these capabilities to edge devices. Our Qualcom...

Qualcomm at NeurIPS 2024: Our groundbreaking innovations and cutting-edge advancements in AI.
www.qualcomm.com/news/onq/202...

If you're attending NeurIPS 2024 in Vancouver, be sure to visit us at Qualcomm's booth #533 and our poster(#6102) on Wednesday - Stepping forward on the last mile.

@neuripsconf.bsky.social

Our work at Qualcomm AI research, Qualcomm Canada was accepted at NeurIPS 2024. This work presents on-device training for any network (transformer, convolutional or rnn architecture) using fixed points(quantized) forward gradients (no back-prop).
Paper link: arxiv.org/html/2411.04...

#Neurips2024