[AMD] diffusion: expect the residual-gate fast path to be off on ROCm in test_modulate (#40660)

This commit is contained in:
Michael
2026-10-13 23:34:31 -07:00
committed by GitHub
parent 68c18a5701
commit 6f646563a2
@@ -40,6 +40,10 @@ register_cuda_ci(est_time=50, stage="nightly", runner_config="1-gpu-large")
register_amd_ci(est_time=38, suite="nightly-amd-kernel-1-gpu", nightly=True)
DEVICE = "cuda"
CUDA_NATIVE_ONLY = pytest.mark.skipif(
torch.version.hip is not None,
reason="the bit-exact residual-gate fast path is CUDA-only",
)
@pytest.fixture(autouse=True)
@@ -170,9 +174,12 @@ def test_residual_gate_add_matches_torch(residual_shape, gate_shape):
update = torch.randn_like(residual)
gate = torch.randn(gate_shape, device=DEVICE, dtype=torch.bfloat16)
assert can_use_residual_gate_add_cuda(residual, update, gate)
ref = residual + update * gate
_assert_gate_add(residual_gate_add_cuda(residual, update, gate), ref)
if torch.version.hip is None:
assert can_use_residual_gate_add_cuda(residual, update, gate)
_assert_gate_add(residual_gate_add_cuda(residual, update, gate), ref)
else:
assert not can_use_residual_gate_add_cuda(residual, update, gate)
assert torch.equal(residual_gate_add(residual, update, gate), ref)
@@ -181,6 +188,7 @@ def test_residual_gate_add_matches_torch(residual_shape, gate_shape):
"shape,gate_shape",
[((1, 9, 64), (1, 1, 64)), ((1, 9, 64), (1, 9, 64)), PER_TOKEN_GATE_CASES[0]],
)
@CUDA_NATIVE_ONLY
def test_residual_gate_add_dtypes(dtype, shape, gate_shape):
residual = torch.randn(shape, device=DEVICE, dtype=dtype)
update = torch.randn_like(residual)
@@ -192,6 +200,7 @@ def test_residual_gate_add_dtypes(dtype, shape, gate_shape):
@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32])
@pytest.mark.parametrize("shape", [(1, 17, 65), (1, 7800, 2240), (2, 33, 128)])
@CUDA_NATIVE_ONLY
def test_residual_gate_add_transposed_residual(dtype, shape):
batch, tokens, hidden_size = shape
residual = torch.randn(
@@ -208,6 +217,7 @@ def test_residual_gate_add_transposed_residual(dtype, shape):
assert out.stride() == ref.stride() == residual.stride()
@CUDA_NATIVE_ONLY
def test_residual_gate_add_transposed_storage_offsets():
tokens, hidden_size = 33, 128
residual = (
@@ -231,6 +241,7 @@ def test_residual_gate_add_transposed_storage_offsets():
@pytest.mark.parametrize("transposed", [False, True])
@CUDA_NATIVE_ONLY
def test_residual_gate_add_torch_compile_fullgraph(transposed):
shape = (1, 128, 32) if transposed else (1, 32, 128)
residual = torch.randn(shape, device=DEVICE, dtype=torch.bfloat16)
@@ -244,6 +255,7 @@ def test_residual_gate_add_torch_compile_fullgraph(transposed):
assert out.stride() == residual.stride()
@CUDA_NATIVE_ONLY
def test_residual_gate_add_transposed_cuda_graph():
residual = torch.randn((1, 128, 32), device=DEVICE, dtype=torch.bfloat16).transpose(
1, 2
@@ -269,7 +281,9 @@ def test_residual_gate_add_guards_and_eager_fallback():
residual = torch.randn((1, 8, 64), device=DEVICE, dtype=torch.bfloat16)
update = torch.randn_like(residual)
gate = torch.randn((1, 1, 64), device=DEVICE, dtype=torch.bfloat16)
assert can_use_residual_gate_add_cuda(residual, update, gate)
assert can_use_residual_gate_add_cuda(residual, update, gate) == (
torch.version.hip is None
)
rejected = [
(residual.cpu(), update, gate), # not on device