feat: 新增NTK-Aware RoPE缩放支持

- RotaryEmbedding接受rope_scaling配置,自动计算scaled base
- AutoRegressiveLMConfig和EncoderConfig新增rope_scaling字段
This commit is contained in:
2026-05-25 21:22:07 +08:00
parent a4688021bf
commit 737585a32a
4 changed files with 28 additions and 4 deletions
+2
View File
@@ -49,6 +49,7 @@ class AutoRegressiveLMConfig(BaseModelConfig):
max_len: Optional[int] = None
rope_theta: Optional[float] = None
rope_scaling: Optional[dict] = None
attn_type: str = "gqa"
n_heads: Optional[int] = None
@@ -80,6 +81,7 @@ class EncoderConfig(BaseModelConfig):
max_len: Optional[int] = None
rope_theta: Optional[float] = None
rope_scaling: Optional[dict] = None
n_heads: Optional[int] = None
n_kv_heads: Optional[int] = None