What if attention wasn't about matching tokens, but operating in function space?
Glad to share our #ICML2026 paper:
📄 Functional Attention: From Pairwise Affinities to Functional Correspondences
w/ @Jiefang_Xiao@GaoMaolin @stevenygd Daniel Cremers
📄 https://t.co/rhn9NtwrBm
For over a decade, we’ve accepted that end-to-end backprop is the only way to train deep networks. But holding the entire network in memory all at once is why AI training is hitting a resource wall.
We found a new way to break the network into blocks and train them independently. The trick? Treating the network’s forward pass like a diffusion model denoising a signal.
This reinterpretation slashes the memory needed to train deep models. In our #ICLR2026 paper (https://t.co/PK5h0mqQSo), we matched end-to-end performance across ViTs, DiTs, and LLMs. We did this while training just one isolated block at a time.
Qwopus3.5-27b really rocks.
My best setup of llama-server on a single RTX3090:
hf = Jackrong/Qwopus3.5-27B-v3-GGUF:Q4_K_M
c = 262144
jinja = on
reasoning = off
fa = on
cache-type-k = q4_0
cache-type-v = q4_0
spec-default = on
262k, 160-200 TPS, almost perfect tool calling in Hermes.
I don't know the negative effect of the q4_0 kv cache.
Haven't tried vLLM, but tool parsing could be better.