Shen Zhuoran

icon indicating a website link https://cmsflash.github.io icon indicating an email [email protected]

icon company Augment icon location San Jose icon location Code LLM/agent training at @augmentcode . Ex-@google brain / @cruise-automation / Pony.ai. HKU alum. 💎 Terran in StarCraft II.

Results 11 issues of Shen Zhuoran

[6-line repro] Cloning input `x` to `stk.ops.sdd()` leads to drastically different outputs

2 comment

I was debugging a data-parallel forward mismatch when using `megablocks` (DP and non-DP give different forward results). During debugging, I tried to reproduce such difference minimally, and found that simply...