Upload HDLM model with complete HF integration
Browse files- config.json +102 -0
config.json
ADDED
|
@@ -0,0 +1,102 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"ngpus": 4,
|
| 3 |
+
"gradient_accumulation_steps": 8,
|
| 4 |
+
"pretrain_autoregressive_path": "/home/toolkit/research-diffcodegen/exp_local/openwebtext/mdlm-autoregressive/org-DiTAR-absorb-v2/checkpoints-meta/checkpoint.pth",
|
| 5 |
+
"tokenizer": {
|
| 6 |
+
"tokens": 50257,
|
| 7 |
+
"model": "gpt2"
|
| 8 |
+
},
|
| 9 |
+
"training": {
|
| 10 |
+
"batch_size": 512,
|
| 11 |
+
"accum": 8,
|
| 12 |
+
"n_iters": 1000000,
|
| 13 |
+
"snapshot_freq": 100,
|
| 14 |
+
"log_freq": 10,
|
| 15 |
+
"eval_freq": 100,
|
| 16 |
+
"snapshot_freq_for_preemption": 3000,
|
| 17 |
+
"weight": "standard",
|
| 18 |
+
"snapshot_sampling": true,
|
| 19 |
+
"ema": 0.9999,
|
| 20 |
+
"warmup_iter": -1
|
| 21 |
+
},
|
| 22 |
+
"data": {
|
| 23 |
+
"train": "openwebtext-train",
|
| 24 |
+
"valid": "wikitext103",
|
| 25 |
+
"cache_dir": "/home/toolkit/research-diffcodegen/data",
|
| 26 |
+
"debug": false
|
| 27 |
+
},
|
| 28 |
+
"graph": {
|
| 29 |
+
"type": "QGamma",
|
| 30 |
+
"gamma": 0.01,
|
| 31 |
+
"file": "/home/toolkit/research-diffcodegen/data",
|
| 32 |
+
"report_all": false,
|
| 33 |
+
"expanded_sigma": true
|
| 34 |
+
},
|
| 35 |
+
"noise": {
|
| 36 |
+
"type": "loglinear",
|
| 37 |
+
"sigma_min": 0.0001,
|
| 38 |
+
"sigma_max": 2.0,
|
| 39 |
+
"ar_diffusion": false,
|
| 40 |
+
"expanded_sigma": true
|
| 41 |
+
},
|
| 42 |
+
"sampling": {
|
| 43 |
+
"predictor": "analytic",
|
| 44 |
+
"steps_per_level": 1,
|
| 45 |
+
"noise_removal": true,
|
| 46 |
+
"strategy": "direct",
|
| 47 |
+
"strategy_param": 0.9
|
| 48 |
+
},
|
| 49 |
+
"annealing": {
|
| 50 |
+
"type": "block",
|
| 51 |
+
"efficient": false,
|
| 52 |
+
"width": 1024,
|
| 53 |
+
"tau": 2048,
|
| 54 |
+
"eval_tau": 512,
|
| 55 |
+
"steps_per_level": 1,
|
| 56 |
+
"sampling_method": "SAR",
|
| 57 |
+
"diffusion_loss_weight": 1.0,
|
| 58 |
+
"ce_loss_weight": 4.0,
|
| 59 |
+
"sampling_eps": 0.0001,
|
| 60 |
+
"attention": {
|
| 61 |
+
"context_type": "block_causal",
|
| 62 |
+
"block_type": "full"
|
| 63 |
+
},
|
| 64 |
+
"match_inference": true
|
| 65 |
+
},
|
| 66 |
+
"eval": {
|
| 67 |
+
"batch_size": 32,
|
| 68 |
+
"perplexity": true,
|
| 69 |
+
"perplexity_batch_size": 16
|
| 70 |
+
},
|
| 71 |
+
"optim": {
|
| 72 |
+
"weight_decay": 0.0,
|
| 73 |
+
"optimizer": "AdamW",
|
| 74 |
+
"lr": 0.0003,
|
| 75 |
+
"beta1": 0.9,
|
| 76 |
+
"beta2": 0.999,
|
| 77 |
+
"eps": 1e-08,
|
| 78 |
+
"warmup": 10000,
|
| 79 |
+
"grad_clip": 1.0,
|
| 80 |
+
"scheduler": "lambda"
|
| 81 |
+
},
|
| 82 |
+
"experiment": {
|
| 83 |
+
"name": "QGamma0.01-v2",
|
| 84 |
+
"wandb_project": "debug-QGamma"
|
| 85 |
+
},
|
| 86 |
+
"model": {
|
| 87 |
+
"name": "gamma_hdlm",
|
| 88 |
+
"type": "ddit",
|
| 89 |
+
"hidden_size": 768,
|
| 90 |
+
"cond_dim": 128,
|
| 91 |
+
"length": 1024,
|
| 92 |
+
"n_blocks": 12,
|
| 93 |
+
"n_heads": 12,
|
| 94 |
+
"scale_by_sigma": false,
|
| 95 |
+
"dropout": 0.1,
|
| 96 |
+
"transformer_sigma_conditioning": true,
|
| 97 |
+
"hybrid_sigma_embedding": true,
|
| 98 |
+
"post_process_logits": true,
|
| 99 |
+
"use_timestep_embedding": true
|
| 100 |
+
},
|
| 101 |
+
"model_type": "gamma_hybrid"
|
| 102 |
+
}
|