| { | |
| "architectures": [ | |
| "DFlashDraftModel" | |
| ], | |
| "model_type": "qwen3", | |
| "hidden_size": 4096, | |
| "intermediate_size": 13312, | |
| "num_hidden_layers": 5, | |
| "num_attention_heads": 64, | |
| "num_key_value_heads": 8, | |
| "head_dim": 128, | |
| "vocab_size": 120832, | |
| "rms_norm_eps": 1e-05, | |
| "max_position_embeddings": 262144, | |
| "rope_theta": 11158840.0, | |
| "num_target_layers": 5, | |
| "target_hidden_size": 4096, | |
| "target_num_hidden_layers": 80, | |
| "target_layer_ids": [ | |
| 1, | |
| 20, | |
| 39, | |
| 58, | |
| 77 | |
| ], | |
| "mask_token_id": 120023, | |
| "tie_word_embeddings": false, | |
| "draft_vocab_size": 120832, | |
| "_torchspec_version": "0.1.0", | |
| "torch_dtype": "bfloat16", | |
| "dflash_config": { | |
| "target_layer_ids": [ | |
| 1, | |
| 20, | |
| 39, | |
| 58, | |
| 77 | |
| ], | |
| "mask_token_id": 120023 | |
| } | |
| } |