[ROCm][CI] Fix ROCm LoRA Transformers fallback with full CUDA graphs (#41577)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-05-23 04:31:32 +00:00
committed by GitHub
parent 84e351555a
commit d28bdf9344
4 changed files with 88 additions and 19 deletions
+51
View File
@@ -141,6 +141,29 @@ def test_aiter_rejects_unsupported_dtypes():
assert not impl.supports_args(*args), f"aiter should reject dtype={dtype}"
@pytest.mark.skipif(
not current_platform.is_rocm(),
reason="ROCm vllm_c RMSNorm needs explicit ND input handling",
)
def test_vllm_c_rms_norm_accepts_nd_input():
torch.set_default_device(current_platform.device_type)
impl = ir.ops.rms_norm.impls["vllm_c"]
if not impl.supported:
pytest.skip("vllm_c impl not supported on this platform")
base = torch.randn(3, 8, 192, dtype=torch.float16)
x = base.split(64, dim=-1)[0].view(3, 8, 4, 16)
assert not x.is_contiguous()
weight = torch.randn(16, dtype=torch.float16)
epsilon = 1e-5
output = impl.impl_fn(x, weight, epsilon)
ref_output = rms_norm_native(x, weight, epsilon)
assert output.shape == x.shape
assert_close(ir.ops.rms_norm, output, ref_output)
fused_add_rms_norm_native = ir.ops.fused_add_rms_norm.impls["native"].impl_fn
@@ -167,6 +190,34 @@ def test_fused_add_rms_norm_registration():
assert actual == expected
@pytest.mark.skipif(
not current_platform.is_rocm(),
reason="ROCm vllm_c fused_add_rms_norm needs explicit ND input handling",
)
def test_vllm_c_fused_add_rms_norm_accepts_nd_input():
torch.set_default_device(current_platform.device_type)
impl = ir.ops.fused_add_rms_norm.impls["vllm_c"]
if not impl.supported:
pytest.skip("vllm_c impl not supported on this platform")
base = torch.randn(3, 8, 192, dtype=torch.float16)
residual_base = torch.randn(3, 8, 192, dtype=torch.float16)
x = base.split(64, dim=-1)[0].view(3, 8, 4, 16)
x_residual = residual_base.split(64, dim=-1)[0].view(3, 8, 4, 16)
assert not x.is_contiguous()
assert not x_residual.is_contiguous()
weight = torch.randn(16, dtype=torch.float16)
epsilon = 1e-5
output, residual = impl.impl_fn(x.clone(), x_residual.clone(), weight, epsilon)
ref_output, ref_residual = fused_add_rms_norm_native(x, x_residual, weight, epsilon)
assert output.shape == x.shape
assert residual.shape == x_residual.shape
assert_close(ir.ops.fused_add_rms_norm, output, ref_output)
assert_close(ir.ops.fused_add_rms_norm, residual, ref_residual)
@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32])
@pytest.mark.parametrize("n_tokens", NUM_TOKENS)
@pytest.mark.parametrize("hidden_size", COMMON_HIDDEN_SIZES)