# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest import numpy as np from op_test import ( OpTest, check_cudnn_version_and_compute_capability, get_device, is_custom_device, ) import paddle import paddle.distributed as dist import paddle.nn.functional as F from paddle import _C_ops from paddle.base import core from paddle.distributed.auto_parallel.static.dist_attribute import ( DistTensorSpec, TensorDistAttr, ) from paddle.nn.functional import swiglu as fused_swiglu_impl def swiglu(x, y, out_grad): if isinstance(x, np.ndarray): x = paddle.to_tensor(x) y = paddle.to_tensor(y) out_grad = paddle.to_tensor(out_grad) origin_x = x.detach().clone() origin_x.stop_gradient = False x = origin_x origin_y = y.detach().clone() origin_y.stop_gradient = False y = origin_y dtype = x.dtype need_convert = False assert dtype == y.dtype output_dtype = dtype if paddle.is_compiled_with_cuda() or is_custom_device(): if dtype in [paddle.float16, paddle.bfloat16]: output_dtype = paddle.float32 x = x.astype(output_dtype) y = y.astype(output_dtype) need_convert = True out = F.silu(x) * y if need_convert: out = out.astype(dtype) out.backward(out_grad) ret = [ out.astype(output_dtype), origin_x.grad.astype(output_dtype), origin_y.grad.astype(output_dtype), ] return ret def fused_swiglu(x, y, out_grad, swiglu_func=fused_swiglu_impl): x = x.detach().clone() x.stop_gradient = False if y is not None: y = y.detach().clone() y.stop_gradient = False out = swiglu_func(x, y) out.backward(out_grad) output_dtype = x.dtype if paddle.is_compiled_with_cuda() or is_custom_device(): if x.dtype in [paddle.float16, paddle.bfloat16]: output_dtype = paddle.float32 ret = [ out.astype(output_dtype), ] if y is not None: x_grad, y_grad = x.grad, y.grad else: x_grad, y_grad = paddle.split(x.grad, 2, axis=-1) ret.append(x_grad.astype(output_dtype)) ret.append(y_grad.astype(output_dtype)) return ret tol_map = { paddle.float64: [1e-8, 1e-8], paddle.float32: [1e-6, 1e-6], paddle.float16: [1e-3, 1e-3], paddle.bfloat16: [1e-3, 1e-3], } class TestSwiGLUDygraph(unittest.TestCase): def fused_swiglu(self, x, y, out_grad): return fused_swiglu(x, y, out_grad) def fused_swiglu_impl(self, x, y=None): return fused_swiglu_impl(x, y) def check_dygraph_impl(self, device, shape, dtype): x = paddle.randn(shape, dtype=dtype) y = paddle.randn(shape, dtype=dtype) out_grad = paddle.randn(shape, dtype=dtype) ret1 = swiglu(x, y, out_grad) ret2 = self.fused_swiglu(x, y, out_grad) ret3 = self.fused_swiglu(paddle.concat([x, y], axis=-1), None, out_grad) atol, rtol = tol_map[dtype] err_msg = ( f"Failed when device = {device}, dtype = {dtype}, shape = {shape}" ) for t1, t2, t3 in zip(ret1, ret2, ret3): t1, t2, t3 = t1.numpy(), t2.numpy(), t3.numpy() np.testing.assert_allclose( t1, t2, atol=atol, rtol=rtol, err_msg=err_msg ) np.testing.assert_equal(t2, t3, err_msg=err_msg) def check_dygraph(self, shape): metas = [('cpu', paddle.float32), ('cpu', paddle.float64)] if paddle.is_compiled_with_cuda() or is_custom_device(): metas.append((get_device(), paddle.float32)) metas.append((get_device(), paddle.float64)) metas.append((get_device(), paddle.float16)) if check_cudnn_version_and_compute_capability( min_device_capability=8 ): metas.append((get_device(), paddle.bfloat16)) for device, dtype in metas: origin_device = paddle.get_device() paddle.set_device(device) for with_split in [True]: self.check_dygraph_impl(device, shape, dtype) paddle.set_device(origin_device) def check_static_graph(self, shape, dtype="float32"): x = paddle.static.data(name='x', shape=shape, dtype=dtype) y = paddle.static.data(name='y', shape=shape, dtype=dtype) concated_x = paddle.static.data( name='concated_x', shape=[*shape[:-1], shape[-1] * 2], dtype=dtype, ) out1 = self.fused_swiglu_impl(x, y) out2 = self.fused_swiglu_impl(concated_x) concated_x_np = np.random.random(concated_x.shape).astype(dtype) x_np, y_np = np.split(concated_x_np, 2, axis=-1) exe = paddle.static.Executor() t1, t2 = exe.run( feed={'x': x_np, 'y': y_np, 'concated_x': concated_x_np}, fetch_list=[out1, out2], ) np.testing.assert_equal(t1, t2) def check_main(self, shape): self.check_dygraph(shape) paddle.enable_static() with paddle.static.program_guard( paddle.static.Program(), paddle.static.Program() ): self.check_static_graph(shape) paddle.disable_static() def test_main(self): self.check_main([8, 100]) self.check_main([4, 101]) class TestSwigluOp(OpTest): def config(self): self.x_shape = (8, 128) self.check_auto_parallel = True def setUp(self): self.config() self.op_type = "swiglu" self.prim_op_type = "comp" self.python_api = fused_swiglu_impl self.public_python_api = fused_swiglu_impl x = np.random.uniform(-1, 1, self.x_shape).astype("float64") y = np.random.uniform(-1, 1, self.x_shape).astype("float64") out_grad = np.random.uniform(-1, 1, self.x_shape).astype("float64") res = swiglu(x, y, out_grad) self.inputs = {'x': x, 'y': y} self.outputs = {'out': res[0].numpy()} self.placements = { 'x': [dist.Shard(1)], 'y': [dist.Shard(1)], 'out': [dist.Shard(1)], } def test_check_output(self): self.check_output(check_prim_pir=True) def test_check_grad(self): self.check_grad( ['x', 'y'], 'out', check_auto_parallel=self.check_auto_parallel, check_dygraph=1, check_prim_pir=True, ) class TestSwigluOp2(TestSwigluOp): def setUp(self): self.config() self.op_type = "swiglu" self.prim_op_type = "comp" self.python_api = fused_swiglu_impl self.public_python_api = fused_swiglu_impl x = np.random.uniform(-1, 1, self.x_shape).astype("float64") tmp_inputs = np.split(x, 2, axis=-1) x = tmp_inputs[0] y = tmp_inputs[1] out_grad = np.random.uniform(-1, 1, x.shape).astype("float64") res = swiglu(x, y, out_grad) self.inputs = {'x': x, 'y': y} self.outputs = {'out': res[0].numpy()} self.placements = { 'x': [dist.Shard(1)], 'y': [dist.Shard(1)], 'out': [dist.Shard(1)], } @unittest.skipIf( not (paddle.base.core.is_compiled_with_dist() or is_custom_device()), "The spmd rule is should be tested with distributed=ON", ) class TestSwigluSpmd(unittest.TestCase): def setUp(self): self.kernel = 'swiglu' self.rule = paddle.base.core.get_phi_spmd_rule(self.kernel) x_shape = [64, 32] process_mesh = dist.ProcessMesh(mesh=[0, 1, 2, 3]) x_tensor_dist_attr = TensorDistAttr() x_tensor_dist_attr.dims_mapping = [-1, 0] x_tensor_dist_attr.process_mesh = process_mesh self.x_dist_tensor_spec = DistTensorSpec(x_shape, x_tensor_dist_attr) self.y_dist_tensor_spec = DistTensorSpec(x_shape, x_tensor_dist_attr) self.out_dist_tensor_spec = DistTensorSpec(self.x_dist_tensor_spec) def test_input_x_y(self): result_dist_attrs = self.rule.infer_forward( self.x_dist_tensor_spec, self.y_dist_tensor_spec ) inferred_input_dist_attrs = result_dist_attrs[0] inferred_output_dist_attrs = result_dist_attrs[1] self.assertEqual(len(result_dist_attrs), 2) self.assertEqual(len(inferred_input_dist_attrs), 2) self.assertEqual(len(inferred_output_dist_attrs), 1) self.assertEqual(inferred_output_dist_attrs[0].dims_mapping, [-1, 0]) def test_input_x_unshard_last_dim(self): x_shape = [64, 32] process_mesh = dist.ProcessMesh(mesh=[0, 1, 2, 3]) x_tensor_dist_attr = TensorDistAttr() x_tensor_dist_attr.dims_mapping = [0, -1] x_tensor_dist_attr.process_mesh = process_mesh self.x_dist_tensor_spec = DistTensorSpec(x_shape, x_tensor_dist_attr) result_dist_attrs = self.rule.infer_forward( self.x_dist_tensor_spec, DistTensorSpec() ) inferred_input_dist_attrs = result_dist_attrs[0] inferred_output_dist_attrs = result_dist_attrs[1] self.assertEqual(len(result_dist_attrs), 2) self.assertEqual(len(inferred_input_dist_attrs), 2) self.assertEqual(len(inferred_output_dist_attrs), 1) self.assertEqual(inferred_output_dist_attrs[0].dims_mapping, [0, -1]) @unittest.skipIf( not (core.is_compiled_with_cuda() or is_custom_device()), "mamtul 0 size only with in cuda", ) class TestSwiglu0SizeDygraph(unittest.TestCase): def test_swiglu(self): x = paddle.ones([0, 128], dtype="float32") y = paddle.ones([0, 128], dtype="float32") x.stop_gradient = False y.stop_gradient = False out = fused_swiglu_impl(x, y) dz = paddle.ones([0, 128], dtype="float32") out = _C_ops.swiglu_grad(x, y, dz) self.assertEqual(out[0].shape, x.shape) self.assertEqual(out[1].shape, y.shape) class TestSwigluOp_ZeroSize(OpTest): def config(self): self.x_shape = (0, 128) self.y_shape = (1, 128) self.out_shape = (0, 128) def setUp(self): self.config() self.op_type = "swiglu" self.python_api = fused_swiglu_impl self.public_python_api = fused_swiglu_impl x = np.random.uniform(-1, 1, self.x_shape).astype("float64") y = np.random.uniform(-1, 1, self.y_shape).astype("float64") out_grad = np.random.uniform(-1, 1, self.out_shape).astype("float64") res = swiglu(x, y, out_grad) self.inputs = {'x': x, 'y': y} self.outputs = {'out': res[0].numpy()} def test_check_output(self): self.check_output() def test_check_grad(self): self.check_grad( ['x', 'y'], 'out', ) class TestSwigluOp_ZeroSize2(TestSwigluOp_ZeroSize): def config(self): self.x_shape = (1, 128) self.y_shape = (0, 128) self.out_shape = (0, 128) class TestSwigluOp_ZeroSize3(TestSwigluOp_ZeroSize): def config(self): self.x_shape = (0, 128) self.y_shape = (0, 128) self.out_shape = (0, 128) class TestSwigluGradOp(unittest.TestCase): def test_swiglu_grad(self): x = paddle.randn([10, 2]).astype("float32") out_grad = paddle.randn([10, 1]).astype("float32") x_grad, y_grad = paddle._C_ops.swiglu_grad(x, None, out_grad) self.assertFalse( paddle.all(x_grad == 0).item(), "x_grad should not be all zero" ) # y_grad is not initialized when y is none if __name__ == "__main__": unittest.main()