From 7de70fa7ae9fdda688257eed13639b56cc677fde Mon Sep 17 00:00:00 2001 From: Micah Williamson Date: Thu, 10 Sep 2026 13:54:52 -0500 Subject: [PATCH] [Bugfix][ROCm] Create linear layer biases with `requires_grad=False` (#56161) Signed-off-by: Micah Williamson --- .buildkite/test-amd.yaml | 2 +- vllm/model_executor/layers/linear.py | 11 ++++++++--- 2 files changed, 9 insertions(+), 4 deletions(-) diff --git a/.buildkite/test-amd.yaml b/.buildkite/test-amd.yaml index 9561fec5e0..cca01e45ba 100644 --- a/.buildkite/test-amd.yaml +++ b/.buildkite/test-amd.yaml @@ -4022,7 +4022,7 @@ steps: #------------------------------------------------------ mi355 ยท 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 diff --git a/vllm/model_executor/layers/linear.py b/vllm/model_executor/layers/linear.py index cd1c0e69d2..2e77fac06c 100644 --- a/vllm/model_executor/layers/linear.py +++ b/vllm/model_executor/layers/linear.py @@ -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, {