[model] DBRX-instruct support (#337)

This commit is contained in:
Liangsheng Yin
2024-03-29 01:05:19 +08:00
committed by GitHub
parent 3842eba5fa
commit 2af565b3bb
5 changed files with 726 additions and 4 deletions

View File

@@ -29,6 +29,13 @@ class ModelConfig:
)
self.num_attention_heads = self.hf_config.num_attention_heads
self.num_key_value_heads = getattr(self.hf_config, "num_key_value_heads", None)
# for Dbrx and MPT models
if self.hf_config.model_type in ["dbrx", "mpt"]:
self.num_key_value_heads = getattr(
self.hf_config.attn_config, "kv_n_heads", None
)
if self.num_key_value_heads is None:
self.num_key_value_heads = self.num_attention_heads
self.hidden_size = self.hf_config.hidden_size