Upload config.json with huggingface_hub
Browse files- config.json +63 -1
config.json
CHANGED
|
@@ -1 +1,63 @@
|
|
| 1 |
-
{
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "parcae-small-140m",
|
| 3 |
+
"hf_config": {
|
| 4 |
+
"org": "SandyResearch",
|
| 5 |
+
"name": "parcae-small-140m"
|
| 6 |
+
},
|
| 7 |
+
"block_size": 2048,
|
| 8 |
+
"n_embd": 768,
|
| 9 |
+
"intermediate_size": 3072,
|
| 10 |
+
"num_attention_heads": 6,
|
| 11 |
+
"num_key_value_heads": 6,
|
| 12 |
+
"vocab_size": 32768,
|
| 13 |
+
"padding_multiple": 64,
|
| 14 |
+
"padded_vocab_size": 32768,
|
| 15 |
+
"rope_settings": {
|
| 16 |
+
"use_rope": true,
|
| 17 |
+
"rope_condense_ratio": 1,
|
| 18 |
+
"rope_base": 50000
|
| 19 |
+
},
|
| 20 |
+
"use_abacus": false,
|
| 21 |
+
"randomize_positions_from": null,
|
| 22 |
+
"block_class_name": "TransformerPreNormBlock",
|
| 23 |
+
"norm_class_name": "RMSNorm",
|
| 24 |
+
"attn_impl": "flash",
|
| 25 |
+
"norm_eps": 1e-05,
|
| 26 |
+
"mlp_class_name": "BaseMLP",
|
| 27 |
+
"nonlin_name": "ReLU2",
|
| 28 |
+
"bias": false,
|
| 29 |
+
"qk_bias": false,
|
| 30 |
+
"init_strategy": "scaled-zero",
|
| 31 |
+
"init_orthogonal": true,
|
| 32 |
+
"skip_initialization": false,
|
| 33 |
+
"mup_model_scaling_factor": 1,
|
| 34 |
+
"use_fused_head": "pytorch",
|
| 35 |
+
"debias_attention": false,
|
| 36 |
+
"center_attention": false,
|
| 37 |
+
"clip_qkv": null,
|
| 38 |
+
"qk_norm": true,
|
| 39 |
+
"logit_softcap": null,
|
| 40 |
+
"activation_checkpoint_impl": "per-iteration",
|
| 41 |
+
"simple_ops": false,
|
| 42 |
+
"strategy": "single",
|
| 43 |
+
"injection_type": "diagonal",
|
| 44 |
+
"n_layers_in_recurrent_block": 2,
|
| 45 |
+
"n_layers_in_prelude": 2,
|
| 46 |
+
"n_layers_in_coda": 2,
|
| 47 |
+
"state_init": "like-init",
|
| 48 |
+
"recurrent_embedding_dimension": 768,
|
| 49 |
+
"recurrent_intermediation_embedding_dimension": 3072,
|
| 50 |
+
"recurrent_num_attention_heads": null,
|
| 51 |
+
"prelude_norm": true,
|
| 52 |
+
"sampling_scheme": "poisson-truncated-full",
|
| 53 |
+
"mean_recurrence": 8,
|
| 54 |
+
"mean_backprop_depth": 4,
|
| 55 |
+
"lockstep_n": false,
|
| 56 |
+
"lockstep_k": false,
|
| 57 |
+
"curriculum_target": "forward",
|
| 58 |
+
"recurrent_iteration_method": "per-sequence",
|
| 59 |
+
"tie_embeddings": true,
|
| 60 |
+
"model_class_name": "Parcae",
|
| 61 |
+
"_is_recurrent_block_config": false,
|
| 62 |
+
"_class_name": "ParcaeConfig"
|
| 63 |
+
}
|