Upload V2: stackformer GPT-2 + sparse cross-attention vision model (128 visual tokens)
4666d47 verified | { | |
| "vocab_size": 50258, | |
| "embed_dim": 768, | |
| "num_layers": 12, | |
| "num_heads": 12, | |
| "hidden_dim": 3072, | |
| "context_length": 128, | |
| "dropout": 0.1, | |
| "qkv_bias": true, | |
| "cross_attention_pos": [ | |
| 3, | |
| 6, | |
| 9 | |
| ], | |
| "num_visual_tokens": 128, | |
| "vision_dim": 768, | |
| "perceiver_depth": 2, | |
| "perceiver_heads": 8, | |
| "batch_size": 16, | |
| "grad_accum_steps": 2, | |
| "learning_rate": 0.0002, | |
| "weight_decay": 0.01, | |
| "time_budget_minutes": 270, | |
| "amp_dtype": "bf16" | |
| } |