Update config.json
Browse filesUse the correct activation function : related discussion: https://huggingface.co/google/flan-t5-xxl/discussions/11
The previous config file was using gelu function instead of gated-gelu that is automatically set when forcing is_gated_actto True, more specifically here
This is not a breaking change since it fixes only for inference. Users that trained a model with gelu instead of gated-gelu should not be affected by this change.
- config.json +1 -3
config.json
CHANGED
@@ -6,13 +6,11 @@
|
|
6 |
"d_kv": 64,
|
7 |
"d_model": 4096,
|
8 |
"decoder_start_token_id": 0,
|
9 |
-
"dense_act_fn": "gelu",
|
10 |
"dropout_rate": 0.1,
|
11 |
"eos_token_id": 1,
|
12 |
-
"feed_forward_proj": "gelu",
|
13 |
"initializer_factor": 1.0,
|
14 |
"is_encoder_decoder": true,
|
15 |
-
"is_gated_act": true,
|
16 |
"layer_norm_epsilon": 1e-06,
|
17 |
"model_type": "t5",
|
18 |
"num_decoder_layers": 24,
|
|
|
6 |
"d_kv": 64,
|
7 |
"d_model": 4096,
|
8 |
"decoder_start_token_id": 0,
|
|
|
9 |
"dropout_rate": 0.1,
|
10 |
"eos_token_id": 1,
|
11 |
+
"feed_forward_proj": "gated-gelu",
|
12 |
"initializer_factor": 1.0,
|
13 |
"is_encoder_decoder": true,
|
|
|
14 |
"layer_norm_epsilon": 1e-06,
|
15 |
"model_type": "t5",
|
16 |
"num_decoder_layers": 24,
|