[Bugfix][ROCm] Create linear layer biases with requires_grad=False

[micah-wil]

Purpose

After upgrading to AITER v0.1.21.post2 on ROCm in #55968, we are seeing the following test failure:

pytest -s -v evals/gpt_oss/test_gpqa_correctness.py::test_gpqa_correctness[gpt-oss-20b-rocm-quark-mxfp4-fp8-triton] --config-list-file=configs/models-gfx950.txt

(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]   File "/usr/local/lib/python3.12/dist-packages/flydsl/compiler/jit_argument.py", line 450, in __init__
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]     dl = dltensor.__dlpack__()
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]          ^^^^^^^^^^^^^^^^^^^^^
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]   File "/usr/local/lib/python3.12/dist-packages/torch/_tensor.py", line 1760, in __dlpack__
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]     raise BufferError(
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055] BufferError: Can't export tensors that require gradient, use tensor.detach()
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055] Traceback (most recent call last):
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]   File "/usr/local/lib/python3.12/dist-packages/flydsl/compiler/jit_argument.py", line 446, in __init__
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]     dl = dltensor.__dlpack__(stream=-1)
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]   File "/usr/local/lib/python3.12/dist-packages/torch/_tensor.py", line 1760, in __dlpack__
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055]     raise BufferError(
(Worker_TP1 pid=2684) ERROR 09-09 10:03:01 [multiproc_executor.py:1055] BufferError: Can't export tensors that require gradient, use tensor.detach()

e.g. https://buildkite.com/vllm/amd-ci/builds/12759/list?jid=01a08458-ac70-43e6-89e7-7c9668624b86&tab=output

This happened after ROCm/aiter#5145, which replaced splitk_hgemm / small_m_hgemm FlyDSL kernels with gemm_a16w16_gfx950. This kernel uses DLPack, so the biases are rejected because they use the default requires_grad=True (pytorch prohibits sending tensors to dlpack w/ requires_grad=True):

     dl = dltensor.__dlpack__(stream=-1)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   File "/usr/local/lib/python3.12/dist-packages/torch/_tensor.py", line 1760, in __dlpack__
     raise BufferError(
 BufferError: Can't export tensors that require gradient, use tensor.detach()

Weights explicitly set requires_grad=False so they export fine, but linear biases do not (ReplicatedLinear, ColumnParallelLinear, RowParallelLinear). Here, I explicitly set requires_grad=False on the bias tensors to make them compatible with DLPack. This should have no unintended side effects as nothing in vLLM reads a bias's requires_grad.

Test Plan

pytest -s -v evals/gpt_oss/test_gpqa_correctness.py::test_gpqa_correctness[gpt-oss-20b-rocm-quark-mxfp4-fp8-triton] --config-list-file=configs/models-gfx950.txt

Test Result

pytest -s -v evals/gpt_oss/test_gpqa_correctness.py::test_gpqa_correctness[gpt-oss-20b-rocm-quark-mxfp4-fp8-triton] --config-list-file=configs/models-gfx950.txt

============= 1 passed, 14 warnings in 233.59s (0:03:53) ===============

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[github-actions]

✅ @micah-wil, CI is now available for this PR.

• /ci run starts upstream CI; /amd-ci run starts AMD CI only.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
• /ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

[micah-wil]

/ci run

[github-actions]

✅ Triggered Buildkite CI #88000 for commit 60e422a9d371.

[micah-wil]

/ci run

[github-actions]

✅ Triggered Buildkite CI #88140 for commit 0a23accab080.