Deepspeed not offloading to CPU

Question

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.stack.imgur.com/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?