| { | |
| "model_type": "FlashSTU", | |
| "_name_or_path": "FlashSTU-340M-0408", | |
| "architectures": ["STUForCausalLM"], | |
| "dim": 1024, | |
| "num_heads": 4, | |
| "num_layers": 12, | |
| "seq_len": 4096, | |
| "window_size": 512, | |
| "vocab_size": 200064, | |
| "weight_tying": true, | |
| "inter_dim": 4096, | |
| "mlp_scale": 12, | |
| "bias": false, | |
| "num_eigh": 24, | |
| "r": 896, | |
| "use_hankel_L": false, | |
| "num_epochs": 1, | |
| "global_bsz": 524288, | |
| "bsz": 8, | |
| "warmup_steps": 1907, | |
| "eval_period": 50, | |
| "save_period": 1000, | |
| "max_lr": 4.0e-4, | |
| "min_lr": 4.0e-5, | |
| "max_norm": 1.0, | |
| "fsdp": true, | |
| "ddp": false, | |
| "reshard_after_forward_policy": "default", | |
| "mixed_precision": true, | |
| "torch_dtype": "bfloat16", | |
| "cpu_offload": false, | |
| "sharding_strategy": "full_shard", | |
| "state_dict_type": "full", | |
| "auto_wrap_policy": "partial", | |
| "backward_prefetch": "backward_pre", | |
| "forward_prefetch": false, | |
| "sync_module_states": true, | |
| "use_orig_params": true, | |
| "device_id": null, | |
| "precision": { | |
| "param": "bfloat16", | |
| "reduce": "bfloat16", | |
| "buffer": "bfloat16" | |
| }, | |
| "fsdp_modules": [ | |
| "STULayer", | |
| "AttentionLayer" | |
| ], | |
| "num_workers": 0, | |
| "snapshot_every_n_steps": 50, | |
| "use_activation_checkpointing": true, | |
| "use_flash_fft": false, | |
| "use_tensordot": true, | |
| "use_attn": true, | |
| "use_alibi": false, | |
| "softcap": 50.0, | |
| "rope_theta": 10000.0, | |
| "torch_compile": true, | |
| "torch_compile_kwargs": { | |
| "mode": "default", | |
| "fullgraph": false | |
| }, | |
| "enable_compiled_autograd": false | |
| } | |