fix TEGroupedLinear (#94)

tastelikefeet · web-flow · commit 6b48d61fde6e · 2026-03-02T14:08:07.000+08:00
diff --git a/src/twinkle/model/megatron/tuners/lora.py b/src/twinkle/model/megatron/tuners/lora.py
@@ -136,8 +136,6 @@ def update_layer(self, adapter_name: str, r: int, *, lora_alpha: int, lora_dropo
             'config': self.config,
             'is_expert': self.is_expert,
         }
-        if exists('megatron_core>=0.13'):
-            kwargs['tp_group'] = self.base_layer.tp_group
 
         if isinstance(self.base_layer, TopKRouter):
             # Router layer - no parallelism needed

Original file line number	Diff line number	Diff line change
`@@ -136,8 +136,6 @@ def update_layer(self, adapter_name: str, r: int, *, lora_alpha: int, lora_dropo`
`136`	`136`	`'config': self.config,`
`137`	`137`	`'is_expert': self.is_expert,`
`138`	`138`	`}`
`139`		`- if exists('megatron_core>=0.13'):`
`140`		`- kwargs['tp_group'] = self.base_layer.tp_group`
`141`	`139`
`142`	`140`	`if isinstance(self.base_layer, TopKRouter):`
`143`	`141`	`# Router layer - no parallelism needed`