What is the biggest LLM model tinybox can pretrain from scratch

[banyan-god]

With 144GB of Vram what is biggest model we can train from scratch using tinybox. Will it be possible to train model similar to lets say llama 3 8B from scratch using this ? Here are some model params

    llama3(
        vocab_size=128_256,
        num_layers=32,
        num_heads=32,
        num_kv_heads=8,
        embed_dim=4096,
        max_seq_len=8192,
        intermediate_dim=14336,
        attn_dropout=0.0,
        norm_eps=1e-5,
        rope_base=500000.0,
    )

Now we need to also consider block size, batch size etc