def testAddSubFallback(self):
torch._lazy.config.set_force_fallback("aten::sub")
device = get_test_device()
x = torch.randn(2, 3, 4, device=device)
y = torch.randn(2, 3, 4, device=device)
z = torch.zeros(2, 3, 4, device=device)
device = 'lazy'
x_lazy = x.detach().clone().to(device=device)
y_lazy = y.detach().clone().to(device=device)
z_lazy = z.detach().clone().to(device=device)
for i in range(10):
if i < 5:
z += (x + y)
else:
z += (x - y)
for i in range(10):
if i < 5:
z_lazy += (x_lazy + y_lazy)
else:
z_lazy += (x_lazy - y_lazy)
torch._lazy.mark_step()
torch.testing.assert_close(z.cpu(), z_lazy.cpu())
assert metrics.counter_value(
"IrNodeReused_torch::lazy::AddTensor") >= 8
metrics.reset()
torch._lazy.ir_cache.reset()
torch._lazy.config.set_force_fallback("")
def testBatchNorm(self):
device = get_test_device()
x = torch.randn(16, 3, 224, 224, device=device)
weight = torch.randn(3, device=device)
bias = torch.randn(3, device=device)
for i in range(10):
# BatchNorm2d does extra checks on dimensions which SymInts don't support yet
# so we call `torch.ops.aten.native_batch_norm` to bypass the checks.
z, _, _ = torch.ops.aten.native_batch_norm(x, weight, bias, None,
None, True, 0.1, 1e-5)
device = "lazy"
x_lazy = x.detach().clone().to(device=device)
weight_lazy = weight.detach().clone().to(device=device)
bias_lazy = bias.detach().clone().to(device=device)
for i in range(10):
z_lazy, _, _ = torch.ops.aten.native_batch_norm(
x_lazy, weight_lazy, bias_lazy, None, None, True, 0.1, 1e-5)
torch._lazy.mark_step()
torch.testing.assert_close(z.cpu(), z_lazy.cpu())
assert metrics.counter_value(
"IrNodeReused_torch::lazy::TSNativeBatchNormForward") >= 7
metrics.reset()
torch._lazy.ir_cache.reset()
def get_fallback_ops():
fallback_ops = []
for opname in metrics.counter_names():
if "aten::" not in opname:
continue
val = int(metrics.counter_value(opname))
if val > 0:
fallback_ops.append(f"{opname}={val}")
return fallback_ops
def testBatchNorm(self):
device = get_test_device()
x = torch.randn(16, 3, 224, 224, device=device)
bn = torch.nn.BatchNorm2d(3).to(device=device)
for i in range(10):
z = bn(x)
device = "lazy"
x_lazy = x.detach().clone().to(device=device)
bn = bn.to(device=device)
for i in range(10):
z_lazy = bn(x_lazy)
torch._lazy.mark_step()
torch.testing.assert_close(z.cpu(), z_lazy.cpu())
assert metrics.counter_value(
"IrNodeReused_torch::lazy::TSNativeBatchNormForward") >= 7
metrics.reset()
torch._lazy.ir_cache.reset()
def testAdd(self):
device = get_test_device()
x = torch.randn(2, 3, 4, device=device)
y = torch.randn(2, 3, 4, device=device)
z = torch.zeros(2, 3, 4, device=device)
device = 'lazy'
x_lazy = x.detach().clone().to(device=device)
y_lazy = y.detach().clone().to(device=device)
z_lazy = z.detach().clone().to(device=device)
for i in range(10):
z += (x + y)
for i in range(10):
z_lazy += (x_lazy + y_lazy)
torch._lazy.mark_step()
torch.testing.assert_close(z.cpu(), z_lazy.cpu())
assert metrics.counter_value(
"IrNodeReused_torch::lazy::AddTensor") >= 14
metrics.reset()
torch._lazy.ir_cache.reset()