[ROCm][Bugfix] Run the AITER AR+RMSNorm fallback test on the platform… (#61350)

Signed-off-by: Stefan Koncarevic <stefan.koncarevic@amd.com>
Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
This commit is contained in:
stefankoncarevic
2026-10-13 08:35:57 +00:00
committed by GitHub
parent ff305934f1
commit 24aa1191f4
+8 -2
View File
@@ -11,6 +11,7 @@ import torch
import vllm._aiter_ops as aiter_ops
import vllm.distributed as vllm_distributed
import vllm.kernels # noqa: F401
from vllm.platforms import current_platform
def test_rocm_aiter_gemma_rmsnorm_falls_back_after_allreduce(
@@ -72,9 +73,14 @@ def test_rocm_aiter_gemma_rmsnorm_falls_back_after_allreduce(
fake_aiter.rmsnorm2d_fwd_with_add = fake_rmsnorm2d_fwd_with_add
monkeypatch.setitem(sys.modules, "aiter", fake_aiter)
inp = torch.randn(8192, 16, dtype=torch.bfloat16)
# On the platform's device, not the CPU. The fallback leaves Python through
# `torch.ops.vllm_aiter.fused_add_rms_norm`, which is registered for the
# platform's dispatch key alone, so CPU tensors have nothing to dispatch to
# on a GPU build. The kernel behind it is still the fake above.
device = current_platform.device_type
inp = torch.randn(8192, 16, dtype=torch.bfloat16, device=device)
residual = torch.randn_like(inp)
weight = torch.randn(16, dtype=torch.bfloat16)
weight = torch.randn(16, dtype=torch.bfloat16, device=device)
out, residual_out = aiter_ops._rocm_aiter_fused_allreduce_rmsnorm_impl(
inp, residual, weight, 1e-6, gemma_norm=True