更新配置文件
Browse files- config.json +4 -6
- modeling_telechat.py +1 -2
config.json
CHANGED
|
@@ -4,7 +4,7 @@
|
|
| 4 |
"architectures": [
|
| 5 |
"TelechatForCausalLM"
|
| 6 |
],
|
| 7 |
-
"attention_dropout": 0.
|
| 8 |
"attention_softmax_in_fp32": true,
|
| 9 |
"auto_map": {
|
| 10 |
"AutoConfig": "configuration_telechat.TelechatConfig",
|
|
@@ -16,25 +16,23 @@
|
|
| 16 |
"eos_token_id": 2,
|
| 17 |
"ffn_hidden_size": 12288,
|
| 18 |
"flash_attn": true,
|
| 19 |
-
"hidden_dropout": 0.
|
| 20 |
"hidden_size": 4096,
|
| 21 |
"initializer_range": 0.02,
|
| 22 |
"layer_norm_epsilon": 1e-05,
|
| 23 |
"logn": false,
|
| 24 |
"masked_softmax_fusion": true,
|
| 25 |
"model_type": "telechat",
|
|
|
|
| 26 |
"n_head": 32,
|
| 27 |
"n_inner": null,
|
| 28 |
"n_layer": 30,
|
| 29 |
-
"offset_alibi": 100,
|
| 30 |
"pad_token_id": 3,
|
| 31 |
-
"pretraining_tp": 2,
|
| 32 |
-
"seq_length": 8192,
|
| 33 |
"skip_bias_add": true,
|
| 34 |
"skip_bias_add_qkv": false,
|
| 35 |
"slow_but_exact": false,
|
| 36 |
"torch_dtype": "float16",
|
| 37 |
-
"training_seqlen":
|
| 38 |
"transformers_version": "4.30.0",
|
| 39 |
"unk_token_id": 0,
|
| 40 |
"use_cache": true,
|
|
|
|
| 4 |
"architectures": [
|
| 5 |
"TelechatForCausalLM"
|
| 6 |
],
|
| 7 |
+
"attention_dropout": 0.1,
|
| 8 |
"attention_softmax_in_fp32": true,
|
| 9 |
"auto_map": {
|
| 10 |
"AutoConfig": "configuration_telechat.TelechatConfig",
|
|
|
|
| 16 |
"eos_token_id": 2,
|
| 17 |
"ffn_hidden_size": 12288,
|
| 18 |
"flash_attn": true,
|
| 19 |
+
"hidden_dropout": 0.1,
|
| 20 |
"hidden_size": 4096,
|
| 21 |
"initializer_range": 0.02,
|
| 22 |
"layer_norm_epsilon": 1e-05,
|
| 23 |
"logn": false,
|
| 24 |
"masked_softmax_fusion": true,
|
| 25 |
"model_type": "telechat",
|
| 26 |
+
"seq_length": 8192,
|
| 27 |
"n_head": 32,
|
| 28 |
"n_inner": null,
|
| 29 |
"n_layer": 30,
|
|
|
|
| 30 |
"pad_token_id": 3,
|
|
|
|
|
|
|
| 31 |
"skip_bias_add": true,
|
| 32 |
"skip_bias_add_qkv": false,
|
| 33 |
"slow_but_exact": false,
|
| 34 |
"torch_dtype": "float16",
|
| 35 |
+
"training_seqlen": 8192,
|
| 36 |
"transformers_version": "4.30.0",
|
| 37 |
"unk_token_id": 0,
|
| 38 |
"use_cache": true,
|
modeling_telechat.py
CHANGED
|
@@ -105,8 +105,7 @@ class RotaryEmbedding(torch.nn.Module):
|
|
| 105 |
return ntk_alpha
|
| 106 |
|
| 107 |
def forward(self, x, seq_dim=0, seq_len=None):
|
| 108 |
-
|
| 109 |
-
seq_len = x.shape[seq_dim]
|
| 110 |
seq_len = max(seq_len, self.config.training_seqlen)
|
| 111 |
ntk_alpha = self.get_ntk_alpha(seq_len)
|
| 112 |
self.mscale = float(self.get_mscale(seq_len / self.config.training_seqlen))
|
|
|
|
| 105 |
return ntk_alpha
|
| 106 |
|
| 107 |
def forward(self, x, seq_dim=0, seq_len=None):
|
| 108 |
+
seq_len = x.shape[seq_dim]
|
|
|
|
| 109 |
seq_len = max(seq_len, self.config.training_seqlen)
|
| 110 |
ntk_alpha = self.get_ntk_alpha(seq_len)
|
| 111 |
self.mscale = float(self.get_mscale(seq_len / self.config.training_seqlen))
|