三年未合：开发者吐槽 Hugging Face 拒并 T5 lm_head 初始化一行改动
Birchlabs · x · 2026-10-21
Developer Birch-san notes Hugging Face never merged his one-line transformers PR (#26441) from 2023. It changes T5 lmhead weight init variance from 1 to the reciprocal of hiddendim: before, logits had variance hiddendim and initial cross-entropy loss of 110; after, variance 1 and initial loss 11, in line with expectations. Related to issue #16749; the PR remains closed and unmerged.
「Fun」频道最新
- Anthropic 涉足蝎毒抗毒血清研究，需挤数千只蝎子毒液免疫马匹 — JasonBotterill · 2026-10-21
- 洗稿被举报后封号，Alchain 称 X 这方面处理很到位 — AlchainHust · 2026-10-21
- Jerry Seinfeld 谈创作:「正确的方式就是难的方式」 — soumitrashukla9 · 2026-10-21
- Modal 在旧金山街头竖广告牌：Runway 实时视频推理跑在 Modal 上 — charles_irl · 2026-10-21
- PyTorch CUDA Graph 报错印上 T 恤，AI 圈程序员梗图走红 — AlpinDale · 2026-10-21
- TBPN 走红 17 个月后,其报道对象 Abe Brown 升任 The Information 专题编辑 — johncoogan · 2026-10-21