kohya-ss / kohya-ss/sd-scripts
Cannot make multi node training work
- Dominant language
- Python
- Stars
- 7.2k
- Forks
- 1.2k
- Avg merge
- 11m
- Merged PRs (30d)
- 2
Description
We ran into various problems, such as `NOTE: Redirects are currently not supported in Windows or MacOs.`, `The client socket has failed to connect to [ip: port we specify]`, and also the nccl backend issue (which can be solved).
We run two windows 10 machines in the same network, each has one 3090. The accelerate version we tried are `0.15.0` and `0.16.0`, torch version is `2.0.1+cu118`, our configuration is mainly like below, only the `machine_rank` field is different for the two machines. We tried some slight variations of the config, feels more wrong and also failed.
```
compute_environment: LOCAL_MACHINE
deepspeed_config: {}
distributed_type: MULTI_GPU
downcast_bf16: 'no'
dynamo_backend: 'NO'
fsdp_config: {}
gpu_ids: all
machine_rank: 0
main_process_ip: "127.0.0.1"
main_process_port: 8000
main_training_function: main
megatron_lm_config: {}
mixed_precision: fp16
num_machines: 2
num_processes: 1
rdzv_backend: static
same_network: true
use_cpu: false
```
What did we do wrong? Did anyone run multi node training successfully?
Contributor guide
No contributing guide indexed for this repository
Research direction
Start by reproducing the two-machine setup described in the issue with the provided configuration, comparing the machine_rank values and the main_process_ip setting. Review the reported redirect, client socket, and NCCL errors together. Done means multi-node training starts successfully on both Windows 10 machines.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- distributed-systems, machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100