DeepSpeed with ZeRO3 strategy cannot build 'fused_adam' #6892
Comments
LeonardoZini
changed the title
|
@LeonardoZini, can you please share the log showing the
With zero stage 3 model sharding, special handling is required to access parameters. See following links for more details.
|
|
The log are this one and this Thank you for the references! |
|
@LeonardoZini, I noticed a compiler error in your build log Can you try the following to see if it would reproduce the compile-time error? >>> import torch, deepspeed
>>> from deepspeed.ops.adam.fused_adam import FusedAdam
>>> x = FusedAdam([torch.empty(100)])
`` |
|
@LeonardoZini, were you able to confirm the compile-time error? |
|
Sorry for being late. I attach you the logs. STDERR STDOUT GCC version |
|
I know that the error comes from the unsupported compiler version. I don't have access to an older compiler on the cluster I am working on, so this is now a relatively low priority for me. I don't know if this is expected behavior, but I suggest extending compatibility with newer compilers. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe the bug
I am using Deepspeed with the huggingface trainer to fine-tune an llm. While with ZeRO2 strategy I don't have any problem I need to shard also the parameters since i'm working on long-context sequences.
When using ZeRO3 the trainer at the beginning of the training, raise me an excpetion
RuntimeError: Error building extension 'fused_adam'I installed DeepSpeed with the command
TORCH_CUDA_ARCH_LIST="8.0;8.6;9.0" DS_BUILD_OPS=1 pip install deepspeed --global-option="build_ext"(i tried also the 0.15.4 version)I tried also
TORCH_CUDA_ARCH_LIST="8.0;8.6;9.0" DS_BUILD_FUSED_ADAM=1 pip install deepspeed --global-option="build_ext".What changes is that if i specify in the deepspeed json config file the
"offload_optimizer"and"offload_param"it doesn't throws me any error, but i lose any reference to the model parameters (the weights are void tensors).I am using a SLURM scheduler, and one thing i noticed is that the ds_report output are different. Outisde the SLURM
fused_adamseems installed, while inside SLURM no.pip env
ds_report output
output
ds_report output in SLURM
output
output log
Ass enlighten in this logs, the number of parameter goes from 4568002560 before the training loop, to 266240 after the training loop (the voice Parameter Offload makes me thinking..).
System info :
Launcher context
I am launching with torchrun,
srun torchrun --nnodes=1 --nproc-per-node=2 --rdzv-endpoint=$MASTER_ADDR:$MASTER_PORT --rdzv-id=$SLURM_JOB_NAME --rdzv-backend="c10d" --max_restarts=$MAX_RESTARTS trainer.pyThe text was updated successfully, but these errors were encountered: