[Bugfix][ROCm] Create linear layer biases with `requires_grad=False` (#56161)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Sep 10, 2026 at 13:54 UTC
7de70fa7ae9fdda688257eed13639b56cc677fde
2 files changed
+9
-4
.buildkite/test-amd.yaml
+1
-1
@@ -4022,7 +4022,7 @@ steps:
4022
#------------------------------------------------------ mi355 · weight_loading -------------------------------------------------------#
4023
4024
- label: ":amd: (MI355) Weight Loading Multi-GPU" # TBD
4025
- timeout_in_minutes: 30
4025
+ timeout_in_minutes: 40
4026
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4027
dind: false
4028
agent_pool: mi355_2
vllm/model_executor/layers/linear.py
+8
-3
@@ -380,7 +380,8 @@ class ReplicatedLinear(LinearBase):
380
381
if bias:
382
self.bias = Parameter(
383
- torch.empty(self.output_size, dtype=self.params_dtype)
383
+ torch.empty(self.output_size, dtype=self.params_dtype),
384
+ requires_grad=False,
385
)
386
set_weight_attrs(
387
self.bias,
@@ -525,7 +526,8 @@ class ColumnParallelLinear(LinearBase):
526
527
if bias:
528
self.bias = Parameter(
528
- torch.empty(self.output_size_per_partition, dtype=params_dtype)
529
+ torch.empty(self.output_size_per_partition, dtype=params_dtype),
530
+ requires_grad=False,
531
)
532
set_weight_attrs(
533
self.bias,
@@ -1708,7 +1710,10 @@ class RowParallelLinear(LinearBase):
1710
)
1711
1712
if bias:
1711
- self.bias = Parameter(torch.empty(self.output_size, dtype=params_dtype))
1713
+ self.bias = Parameter(
1714
+ torch.empty(self.output_size, dtype=params_dtype),
1715
+ requires_grad=False,
1716
+ )
1717
set_weight_attrs(
1718
self.bias,
1719
{