frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

Avoid multiple instantiations of the RoPE class (#1828)

Browse Source
This commit is contained in:
Jee Li
2023-11-30 15:06:27 +08:00
committed by GitHub
parent 27feead2f8
commit 63b2206ad0
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/layers/rotary_embedding.py
View File

@ -272,6 +272,9 @@ class YaRNScalingRotaryEmbedding(RotaryEmbedding):
return cache
_ROPE_DICT: Dict[Tuple, RotaryEmbedding] = {}
def get_rope(
head_size: int,
rotary_dim: int,
@ -280,6 +283,10 @@ def get_rope(
is_neox_style: bool = True,
rope_scaling: Optional[Dict[str, Any]] = None,
) -> RotaryEmbedding:
key = (head_size, rotary_dim, max_position, base, is_neox_style,
rope_scaling)
if key in _ROPE_DICT:
return _ROPE_DICT[key]
if rope_scaling is None:
rotary_emb = RotaryEmbedding(head_size, rotary_dim, max_position, base,
is_neox_style)
@ -312,4 +319,5 @@ def get_rope(
**extra_kwargs)
else:
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
_ROPE_DICT[key] = rotary_emb
return rotary_emb