Deepspeed fails to offload operations to the CPU, like I thought it should do when it runs out of GPU memory. I guess I have some setting wrong. When the batch size is increased it gives an error like
(https://i.sstatic.net/StcTz.png) torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 7.04 GiB (GPU 1; 79.15 GiB total capacity; 68.07 GiB already allocated; 5.90 GiB free; 72.14 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
(doesnt happen for smaller batch sizes).
Using Adam optimizer, and an AMD EPYC 7V13 64-Core Processor (an Azure VM).
The DeepSpeed config is -
{
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"allgather_partitions": true,
"allgather_bucket_size": 2e8,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 2e8,
"contiguous_gradients": true
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto"
}
Training is done by HuggingFace Trainer, and the DeepSpeed config is used by adding the config dict to TrainingArguments
with open("./Multi_Modal_Model/zero_config/stage_2_config.json") as f:
z_optimiser = json.load(f)
training_args = TrainingArguments(
...
deepspeed=z_optimiser,
...
)
Using PyTorch 1.13, trying to train a HuggingFace CLIP model.
Anyone know what I'm doing wrong?
