snie
snie
Hi, I checked again based on your comment. It looks like my correction is wrong. Inside the `computeSignatureSize` function, `r` is the number of rows and `s` is the number...
Hi! I am almost done. Will create a pull request tmr. Proofreading would be great! I am not sure with many parts of the translation.
The pull request is here https://github.com/ncase/remember/pull/14#issue-231055698 @chinchihyu
> Could you help run an end-to-end example and show the accuracy matches? Which example do you recommend running? grpo.sh?
@vermouth1992 @eric-haibin-lin Had a run comparing pp=1 and pp=2 for grpo.sh with 2 H100 nodes. The rewards are comparable: But in my run pp=2 is slower than pp=1 (red line...
> @snie2012 that's expected. I'd recommend users to switch from TP16 to TP8PP2 who wants to run large models but does not have RDMA/IB across nodes sounds good, lmk any...
> pls fix pre-commit error. i guess we have to wait for [vllm-project/vllm#21786](https://github.com/vllm-project/vllm/pull/21786) thanks, updated
@eric-haibin-lin appreciate if you can take a look to see if we can close this, thanks
> > I hit this error when running with: > > > > 1. vlllm 0.9.1 > > 2. rollout_mode="async" > > 3. VLLM_USE_V1=1 > > > > VLLM_USE_V1=0 also...
rebase and fix format