AutoCast for mixed precision/fp16 fails? about x-transformers

[LarsHill]

I actually encountered a similar scenario.
The standard Huggingface bert-base-cased model trained with 16 bit mixed precision (using pytorch-lightning), a vocab size of 100K and a seq len of 1024 uses around 34GB of Memory with a batch size=8 (on Nvidia A100). If I switch to full 32 bit precision the RAM usage almost doubles to 67GB of Memory which is expected.

However, if I use the x-transformers "Bert-like" implementation (mimicking the Huggingface config):

self.bert = TransformerWrapper(
            num_tokens=100_000,
            max_seq_len=1024,
            emb_dropout=0.1,
            tie_embedding=True,
            attn_layers=Encoder(
                dim=768,
                depth=12,
                heads=12,
                attn_flash=False,
                layer_dropout=0.1,  # stochastic depth - dropout entire layer
                attn_dropout=0.1,  # dropout post-attention
                ff_dropout=0.1,  # feedforward dropout
                use_abs_pos_emb=True,
            ),
        )

the memory usage does not change if I switch between 16-mixed and 32 precision. The overall usage (same batch size and hardware) remains at a constant 52GB, which is substantially higher than the HF model with 34GB.
Why does the precision setting of the lightning trainer not affect the x-transformers implementation?

I would love to use the x-transformer implementation due to the large amount of new features. However, I am wondering where these significant GPU RAM differences come from? And why does torch.autocast, which I think is used by lightning under the hood show no effect?