| { |
| "type": "slither.SlitherModel", |
| "pretrained_url": null, |
| "pretrained_step": null, |
| "pretrained_strict": null, |
| "torch_dtype": "float32", |
| "vocab_size": 49153, |
| "bos_token_id": 0, |
| "eos_token_id": 0, |
| "pad_token_id": 49152, |
| "hidden_size": 1024, |
| "num_hidden_layers": 16, |
| "num_attention_heads": 8, |
| "num_key_value_heads": 4, |
| "intermediate_size": 1664, |
| "hidden_act": "silu", |
| "max_position_embeddings": 4096, |
| "rope_theta": 10000, |
| "rms_norm_eps": 1e-05, |
| "attention_kernel": "flash_attention", |
| "pure_modules": [ |
| "SlitherStateWriter" |
| ], |
| "position_embedding_dim": 256, |
| "num_output_layers": 4, |
| "num_memory_layers": 4, |
| "state_size": 512, |
| "num_state_heads": 16, |
| "init_state_out_scale": 0.1, |
| "write_mixer_kernel_size": 5, |
| "write_mixer_init_scale": 0.1, |
| "chunk_length": 1024, |
| "sharding": { |
| "replicate_default": true, |
| "backbone_layers.layers.*": [ |
| [ |
| "data", |
| "fsdp" |
| ], |
| null, |
| null |
| ], |
| "output_layers.layers.*": [ |
| [ |
| "data", |
| "fsdp" |
| ], |
| null, |
| null |
| ], |
| "memory_layers.layers.*": [ |
| [ |
| "data", |
| "fsdp" |
| ], |
| null, |
| null |
| ], |
| "lm_head": [ |
| [ |
| "data", |
| "fsdp" |
| ], |
| null, |
| null |
| ] |
| }, |
| "remat": { |
| "advanced": [ |
| { |
| "name": "backbone_layers", |
| "settings": { |
| "activation_checkpoint_layers": [ |
| "SlitherBackboneLayer" |
| ], |
| "optimization_barrier_layers": [ |
| "SlitherBackboneLayer" |
| ], |
| "scan_layers": "layers", |
| "is_layer_pure": true |
| } |
| }, |
| { |
| "name": "output_layers", |
| "settings": { |
| "activation_checkpoint_layers": [ |
| "SlitherOutputLayer" |
| ], |
| "optimization_barrier_layers": [ |
| "SlitherOutputLayer" |
| ], |
| "scan_layers": "layers", |
| "is_layer_pure": true |
| } |
| }, |
| { |
| "name": "memory_layers", |
| "settings": { |
| "activation_checkpoint_layers": [ |
| "SlitherMemoryLayer" |
| ], |
| "optimization_barrier_layers": [ |
| "SlitherMemoryLayer" |
| ], |
| "scan_layers": "layers", |
| "is_layer_pure": true |
| } |
| } |
| ] |
| } |
| } |