[Bug]: NCCL timed out during inference

[enkiid]

Your current environment

Using:

• vllm 0.4.1
• nccl 2.18.1
• pytorch 2.2.1

🐛 Describe the bug

During inference I sometimes get this error:

(RayWorkerWrapper pid=2376582) [rank1]:[E ProcessGroupNCCL.cpp:523] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=50404, OpType=GATHER, NumelIn=8000, NumelOut=0, Timeout(ms)=600000) ran for 600327 milliseconds before timing out.

Havn't seen it in earlier versions of vllm, any thoughts?

[Ch3ngY1]

The same issue, which occurs randomly on my dataset.
vllm 0.4.1
torch 2.2.0+cu118

[DefTruth]

i have encountered the same issue, try --disable-custom-all-reduce and --enforce-eager is worked for me.

[changyuanzhangchina]

Please refer to #4430

1. --disable-custom-all-reduce = True
2. --enforce-eager = True (may be unnecessary)
3. update to the [Core] Ignore infeasible swap requests. #4557

This three can solve the watchdog problem for me
before this, nccl watchdog error happens several times per day,
and now, it works well

[yunfeng-scale]

we're seeing this on 0.4.2 as well with mixtral 8x22b. --disable-custom-all-reduce resolves the problem.

[yunfeng-scale]

can we again disable custom all reduce by default?

[syr-cn]

i have encountered the same issue, try --disable-custom-all-reduce and --enforce-eager is worked for me.

Works for me! thanks a lot!

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!

[fake-warrior8]

i have encountered the same issue, try --disable-custom-all-reduce and --enforce-eager is worked for me.

My code uses vllm for evaluation, how to set these parameters to vllm.LLM?

My running command is python eval.py
the eval.py use the following vllm code

from vllm import LLM, SamplingParams
import vllm.envs as envs

envs.VLLM_HOST_IP="0.0.0.0" or "127.0.0.1"
llm = LLM(model=model_name_or_path, 
              tensor_parallel_size=len(available_gpus), 
              trust_remote_code=True, 
            #   swap_space=60,
              gpu_memory_utilization=0.96,
              )