sunzj

请教如何丢包的问题?

7

https://github.com/skywind3000/kcp/issues/230#issuecomment-541292962 "KCP 任何情况下都不会丢包，自动重传协议本身就是干这个的。" 怎样在ikcp_waitsnd 中累积到一定数据的时候，或重传次数到达一定的阈值时，直接丢包？还是kcp内部不会丢包，只能在上层发现ikcp_waitsnd中累积过多，直接将要送入kcp的数据包丢掉？

如何安全的清理KCP中未发送的数据

1

使用KCP传送PCM的音频，当一首歌结束后，因为网络延时原因，还有部分数据缓存在KCP中，如何安全的将未发送的数据清理掉？以下实现可以不？ void ikcp_clearsnd(ikcpcb *kcp) { if(kcp) { IKCPSEG *seg; while (!iqueue_is_empty(&kcp->snd_buf)) { seg = iqueue_entry(kcp->snd_buf.next, IKCPSEG, node); iqueue_del(&seg->node); ikcp_segment_delete(kcp, seg); } while (!iqueue_is_empty(&kcp->snd_queue)) { seg = iqueue_entry(kcp->snd_queue.next, IKCPSEG, node); iqueue_del(&seg->node);...

Add limitation for local_planner parameters

5

**Describe problem solved by the proposed feature** After PR281 https://github.com/PX4/avoidance/pull/281, local planner use the parameter from PX4 firmware. However, there isn't limitation for these parameters, as i test, the MPC_XY_CRUISE...

[Bug] llama2 radix_tree->MatchPrefix fail cause lots of redunction prefilling.

5

I found a interesting issue on llama2, radix_tree->MatchPrefix fail, cause lots of redunction prefilling. The issue is: 1. Enter prompt S1, llama2 tokenize to "S1+29871". 2. Decode result is S2....

[Bug] llama3: same text tokenize to different token cause prefix cache match fail.

2

When run llama3, sometime prefilled tokens are much more than input. After doing some investigation, i found there is issue on prefix cache match, the log is: 03-17 10:22:55.165 18761...