Merge pull request vllm-project#8 from Starmys/dev/chengzhang/phi3moe128k

xiaoxiawu-microsoft · web-flow · commit dfaba7ce6bcf · 2024-05-28T11:28:30.000-07:00
Support Phi3SuScaledRotaryEmbedding for 128k model
diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py
@@ -452,6 +452,7 @@ def __init__(self,
             max_position=max_position,
             base=int(self.rope_theta),
             is_neox_style=True,
+            rope_scaling=getattr(config, 'rope_scaling', None),
         )
 
         self.attn = Attention(

Original file line number	Diff line number	Diff line change
`@@ -452,6 +452,7 @@ def __init__(self,`
`452`	`452`	`max_position=max_position,`
`453`	`453`	`base=int(self.rope_theta),`
`454`	`454`	`is_neox_style=True,`
	`455`	`+ rope_scaling=getattr(config, 'rope_scaling', None),`
`455`	`456`	`)`
`456`	`457`
`457`	`458`	`self.attn = Attention(`