[Bugfix][ROCm] Create linear layer biases with requires_grad=False (#56161)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
This commit is contained in:
Micah Williamson
2026-09-10 13:54:52 -05:00
committed by GitHub
parent 48cb12c184
commit 7de70fa7ae
2 changed files with 9 additions and 4 deletions
+1 -1
View File
@@ -4022,7 +4022,7 @@ steps:
#------------------------------------------------------ mi355 · weight_loading -------------------------------------------------------#
- label: ":amd: (MI355) Weight Loading Multi-GPU" # TBD
timeout_in_minutes: 30
timeout_in_minutes: 40
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
dind: false
agent_pool: mi355_2
+8 -3
View File
@@ -380,7 +380,8 @@ class ReplicatedLinear(LinearBase):
if bias:
self.bias = Parameter(
torch.empty(self.output_size, dtype=self.params_dtype)
torch.empty(self.output_size, dtype=self.params_dtype),
requires_grad=False,
)
set_weight_attrs(
self.bias,
@@ -525,7 +526,8 @@ class ColumnParallelLinear(LinearBase):
if bias:
self.bias = Parameter(
torch.empty(self.output_size_per_partition, dtype=params_dtype)
torch.empty(self.output_size_per_partition, dtype=params_dtype),
requires_grad=False,
)
set_weight_attrs(
self.bias,
@@ -1708,7 +1710,10 @@ class RowParallelLinear(LinearBase):
)
if bias:
self.bias = Parameter(torch.empty(self.output_size, dtype=params_dtype))
self.bias = Parameter(
torch.empty(self.output_size, dtype=params_dtype),
requires_grad=False,
)
set_weight_attrs(
self.bias,
{