Latest

Every published item, in order · 7652 items
Oct 20
Chinese contentPAA-Bench 提出重新评估 VLM 视觉 token 剪枝:准确率和 token 数都不是好指标

论文发现现行的 VLM 剪枝评测方式有漏洞,模型可能靠文本捷径蒙混过关。作者做了个 PAA-Bench 框架和 PCS 指标,8 种方法实测后发现压缩更多 token 未必更快,想给剪枝方法做评测的可以看看。

arXiv cs.AI@Geng Li et al. (4)Source
Oct 19
PapersBuildersMulti-sourceSelected12:18
FuTransformer: FP8 prefill and INT4 decode kernels for Hopper GPUs

The paper shows the quality-price of aggressive quantization is now measured, not guessed: INT4 weights plus a Hadamard-rotated INT4 cache cost under one perplexity point while roughly tripling decode speed on H100. Engineers serving Llama-3.1-class models on Hopper GPUs can use its fragment-layout dequantization and KV-cache placement techniques to cut latency without falling back to padded prefill. Compared with PyTorch's INT4 kernel, which loses to BF16 at batch 32, this design stays fast up to batch 32.

Event
arXiv cs.AI@Wei WangSource
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.

Only the latest 2000 items are listed; for older ones see the daily archive