Upload config.json with huggingface_hub
Browse files- config.json +71 -0
config.json
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"data_dir": "/dev/shm/data/fineweb-edu-25B",
|
| 3 |
+
"batch_size": 64,
|
| 4 |
+
"seq_len": 1024,
|
| 5 |
+
"max_iters": 5723,
|
| 6 |
+
"gradient_accumulation_steps": 8,
|
| 7 |
+
"seed": 1337,
|
| 8 |
+
"optimizer": "muon",
|
| 9 |
+
"learning_rate": 0.0006,
|
| 10 |
+
"min_lr": 6e-05,
|
| 11 |
+
"weight_decay": 0.1,
|
| 12 |
+
"beta1": 0.9,
|
| 13 |
+
"beta2": 0.95,
|
| 14 |
+
"grad_clip": 1.0,
|
| 15 |
+
"warmup_iters": 200,
|
| 16 |
+
"lr_decay_iters": 5723,
|
| 17 |
+
"muon_lr": 0.008,
|
| 18 |
+
"muon_min_lr": 0.0001,
|
| 19 |
+
"adamw_lr": 0.0003,
|
| 20 |
+
"adamw_min_lr": 1e-05,
|
| 21 |
+
"muon_momentum": 0.95,
|
| 22 |
+
"muon_nesterov": true,
|
| 23 |
+
"muon_ns_steps": 5,
|
| 24 |
+
"muon_weight_decay": 0.0,
|
| 25 |
+
"lr_schedule": "linear_warmdown",
|
| 26 |
+
"warmdown_frac": 0.5,
|
| 27 |
+
"parity_target": 0.0,
|
| 28 |
+
"parity_margin": 0.03,
|
| 29 |
+
"parity_warmdown_iters": 1500,
|
| 30 |
+
"parity_trigger_iter": -1,
|
| 31 |
+
"retain_from_iter": -1,
|
| 32 |
+
"device": "cuda",
|
| 33 |
+
"dtype": "bfloat16",
|
| 34 |
+
"compile": true,
|
| 35 |
+
"n_layer": 24,
|
| 36 |
+
"n_head": 16,
|
| 37 |
+
"n_embd": 2048,
|
| 38 |
+
"block_size": 1024,
|
| 39 |
+
"vocab_size": 50304,
|
| 40 |
+
"pos_encoding": "learned",
|
| 41 |
+
"rope_theta": 10000.0,
|
| 42 |
+
"sparsity_mode": "none",
|
| 43 |
+
"rmsnorm_affine": true,
|
| 44 |
+
"cayley_levels": [],
|
| 45 |
+
"cayley_per_parent_budget": false,
|
| 46 |
+
"cayley_seed": 42,
|
| 47 |
+
"cayley_backend": "auto",
|
| 48 |
+
"cayley_locations": [
|
| 49 |
+
"resid_mid"
|
| 50 |
+
],
|
| 51 |
+
"cayley_score_standardize": false,
|
| 52 |
+
"cayley_forward_standardized": false,
|
| 53 |
+
"cayley_score_std_ema_beta": 0.99,
|
| 54 |
+
"cayley_replaces_pre_rmsnorm": false,
|
| 55 |
+
"cayley_richardson_refine": false,
|
| 56 |
+
"cayley_matching_pursuit": false,
|
| 57 |
+
"tied_block_init": false,
|
| 58 |
+
"dead_threshold_c": 0.1,
|
| 59 |
+
"resume": false,
|
| 60 |
+
"out_dir": "/dev/shm/out/probe-vanilla-vanilla-large-parity-3b",
|
| 61 |
+
"eval_interval": 100,
|
| 62 |
+
"eval_iters": 50,
|
| 63 |
+
"log_interval": 10,
|
| 64 |
+
"wandb_project": "sparse-nanogpt",
|
| 65 |
+
"wandb_run_name": "probe-vanilla-vanilla-large-parity-3b",
|
| 66 |
+
"wandb_step_offset": 0,
|
| 67 |
+
"quick_eval_interval": 0,
|
| 68 |
+
"quick_eval_timeout": 900,
|
| 69 |
+
"hierarchy_eval_enabled": false,
|
| 70 |
+
"hierarchy_eval_iters": null
|
| 71 |
+
}
|