OC
OC
I tried in ubuntu 1804+ kernel 4.15. There are 3 combinations: 1. `iptables -t mangle -A INPUT -m dscp --dscp 1 -j DADDR --set-daddr=192.168.0.2` works; 2. `tc filter add dev...
I have tested item 3 with ICMP. iptables NAT does not work either. ``` PING 192.168.0.2 (192.168.0.2) 56(84) bytes of data. 64 bytes from 192.168.0.3: icmp_seq=1 ttl=64 time=0.575 ms ```...
Please submit it as PR
1. enable launch_reward_fn_async and give compute_reward_async more cpus in ``` @ray.remote(num_cpus=1) def compute_reward_async ``` 2. use an external sandbox: https://verl.readthedocs.io/en/latest/examples/sandbox_fusion_example.html
related to https://github.com/volcengine/verl/issues/658
blocked by https://github.com/sgl-project/sglang/issues/6723
@ocss884 I have to add lots of patch function to sglang in order to get it works. You can see those code in verl/workers/rollout/sglang_rollout/async_sglang_rollout.py ``` # patch to avoid issue...
Found 2 bugs, 1. eos token has no effect in async_generate 2. ray actor is incorrect in multi node env I am working on those 2 errors.
> Found 2 bugs, > > 1. eos token has no effect in async_generate > 2. ray actor is incorrect in multi node env > > I am working on...
> The motivation and the experiment look good to me! I can help review if needed. Yes，please review.