[LoRA] Gate all_gather on fully_sharded_loras inside _mcp_apply; rewrite regression test (#45715)
Signed-off-by: lcheng <lcheng321@gatech.edu> Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai> Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
co-authored by
Jee Jee Li
Jee Jee Li
parent
f59db63732
commit
e51e700470
@@ -22,10 +22,10 @@ from .utils import _fully_sharded_can_replace, _not_fully_sharded_can_replace
|
|||||||
|
|
||||||
|
|
||||||
def _mcp_apply(x, bias, layer: "ColumnParallelLinearWithLoRA"):
|
def _mcp_apply(x, bias, layer: "ColumnParallelLinearWithLoRA"):
|
||||||
"""
|
"""Fully-sharded (S-LoRA) apply path for column-parallel LoRA layers."""
|
||||||
For `ColumnParallelLinearWithLoRA` or classes that inherit from
|
assert layer.lora_config.fully_sharded_loras, (
|
||||||
`ColumnParallelLinearWithLoRA`, they share the same `apply` logic.
|
"_mcp_apply is only used for fully sharded LoRA"
|
||||||
"""
|
)
|
||||||
assert (
|
assert (
|
||||||
layer.n_slices
|
layer.n_slices
|
||||||
== len(layer.lora_a_stacked)
|
== len(layer.lora_a_stacked)
|
||||||
@@ -341,7 +341,7 @@ class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
|
|||||||
and base_forward is not merged_forward
|
and base_forward is not merged_forward
|
||||||
):
|
):
|
||||||
return self._apply_base_forward(x)
|
return self._apply_base_forward(x)
|
||||||
return _mcp_apply(x, bias, self)
|
return super().apply(x, bias)
|
||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
def can_replace_layer(
|
def can_replace_layer(
|
||||||
|
|||||||
Reference in New Issue
Block a user