zedong

cs master @sjtu1896

[zedong @zedongpeng1]

Open and reproducible in one command:

Code: https://t.co/S1TWxfu4ad
Weights: https://t.co/a7yxv5Sb8x

[zedong @zedongpeng1]

Qwen3.8-27B on a single RTX 4090: 322 tok/s decode.

Same card, same benchmark (Spec-Bench, batch 1, greedy), same session:
vLLM 200, llama.cpp 102.

245 tok/s on the official weights; better 3/4-bit weights do the rest.

Write-up, interactive figures:
https://t.co/YRPcdwwZID

[zedong @zedongpeng1]

Most late wins were 0.3–1%, smaller than run-to-run noise. Measuring them took forced-text A/Bs (every config decodes the same tokens) and drift-balanced run orders.

An AI coding agent (Claude Code) ran the experiments for four days.

[zedong @zedongpeng1]

@ferstar_org @grok 和nvidia的openshell区别

[zedong @zedongpeng1]

@ScarletKc 用jev做一个

[zedong @zedongpeng1]

@theo @grok 这什么软件界面

[zedong @zedongpeng1]

@zesenhuang 是吗 有段时间ultra有问题一直没用 之前一直用max开发 手动安排几个线程做事 看来可以试试让他call subagent自动多线程

[Nathan Barry @nathanrs]

BERT is just a Single Text Diffusion Step! (1/n)
When I first read about language diffusion models, I was surprised to find that their training objective was just a generalization of masked language modeling (MLM), something we’ve been doing since BERT from 2018.
The first thought I had was, “can we finetune a BERT-like model to do text generation?”

[Hardware Architecture Papers @WWVY]

ForgeHLS: A Large-Scale, Open-Source Dataset for High-Level Synthesis. https://t.co/quIbaE3xj6