wehub-resource-sync
156 lines
4.9 KiB
Python
156 lines
4.9 KiB
Python
# Copyright (c) 2025 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 os
|
|
import random
|
|
|
|
import numpy as np
|
|
|
|
import paddle
|
|
import paddle.distributed as dist
|
|
from paddle.distributed.auto_parallel.ring_attention import (
|
|
shard_seq_load_balance,
|
|
unshard_seq_load_balance,
|
|
)
|
|
|
|
dist.init_parallel_env()
|
|
|
|
|
|
class TestContextParallel:
|
|
def __init__(self):
|
|
self.rank = dist.get_rank()
|
|
self.world_size = dist.get_world_size()
|
|
self._sep_mesh = dist.ProcessMesh(
|
|
list(range(self.world_size)), dim_names=["sep"]
|
|
)
|
|
|
|
def set_seed(self, seed):
|
|
paddle.seed(seed)
|
|
np.random.seed(seed)
|
|
random.seed(seed)
|
|
|
|
def _test_cp_base(
|
|
self,
|
|
is_causal=True,
|
|
):
|
|
mesh = dist.ProcessMesh(list(range(self.world_size)), dim_names=['sep'])
|
|
dist.auto_parallel.set_mesh(mesh)
|
|
self.set_seed(1024)
|
|
bs = 2
|
|
seq_len = 256 # flash_attn seq_len/card > 128
|
|
dim = 16
|
|
nheads = 2
|
|
dtype = paddle.bfloat16
|
|
q = paddle.rand(
|
|
(bs, seq_len, nheads, dim),
|
|
dtype=dtype,
|
|
)
|
|
k = paddle.rand(
|
|
(bs, seq_len, nheads, dim),
|
|
dtype=dtype,
|
|
)
|
|
v = paddle.rand(
|
|
(bs, seq_len, nheads, dim),
|
|
dtype=dtype,
|
|
)
|
|
q.stop_gradient = False
|
|
k.stop_gradient = False
|
|
v.stop_gradient = False
|
|
|
|
with paddle.no_grad():
|
|
dist.broadcast(q, src=0)
|
|
dist.broadcast(k, src=0)
|
|
dist.broadcast(v, src=0)
|
|
# base compute
|
|
output_ref = paddle.nn.functional.scaled_dot_product_attention(
|
|
q, k, v, is_causal=is_causal
|
|
)
|
|
loss_ref = output_ref.mean()
|
|
loss_ref.backward()
|
|
|
|
cp_q = q.detach().clone()
|
|
cp_k = k.detach().clone()
|
|
cp_v = v.detach().clone()
|
|
placements = [dist.Replicate() for _ in range(len(mesh.dim_names))]
|
|
|
|
# shard compute
|
|
sharded_q = dist.shard_tensor(cp_q, mesh, placements)
|
|
sharded_k = dist.shard_tensor(cp_k, mesh, placements)
|
|
sharded_v = dist.shard_tensor(cp_v, mesh, placements)
|
|
sharded_q = shard_seq_load_balance(sharded_q, 1)
|
|
sharded_k = shard_seq_load_balance(sharded_k, 1)
|
|
sharded_v = shard_seq_load_balance(sharded_v, 1)
|
|
sharded_q.stop_gradient = False
|
|
sharded_k.stop_gradient = False
|
|
sharded_v.stop_gradient = False
|
|
|
|
output_sharded = paddle.nn.functional.scaled_dot_product_attention(
|
|
sharded_q, sharded_k, sharded_v, is_causal=is_causal, backend='p2p'
|
|
)
|
|
loss_sharded = paddle.mean(output_sharded)
|
|
loss_sharded.backward()
|
|
|
|
with paddle.no_grad():
|
|
reorder_t = unshard_seq_load_balance(output_sharded, 1)
|
|
np.testing.assert_allclose(
|
|
loss_ref.numpy(), loss_sharded.numpy(), rtol=5e-06, atol=5e-06
|
|
)
|
|
np.testing.assert_allclose(
|
|
output_ref.to("float32").numpy(),
|
|
reorder_t.to("float32").numpy(),
|
|
rtol=2e-01,
|
|
atol=6e-02,
|
|
)
|
|
|
|
with paddle.no_grad():
|
|
reorder_q_grad = unshard_seq_load_balance(sharded_q.grad, 1)
|
|
reorder_k_grad = unshard_seq_load_balance(sharded_k.grad, 1)
|
|
reorder_v_grad = unshard_seq_load_balance(sharded_v.grad, 1)
|
|
|
|
rtol = 3e-05
|
|
atol = 3e-05
|
|
np.testing.assert_allclose(
|
|
q.grad.to("float32").numpy(),
|
|
reorder_q_grad.to("float32").numpy(),
|
|
rtol=rtol,
|
|
atol=atol,
|
|
)
|
|
np.testing.assert_allclose(
|
|
k.grad.to("float32").numpy(),
|
|
reorder_k_grad.to("float32").numpy(),
|
|
rtol=rtol,
|
|
atol=atol,
|
|
)
|
|
np.testing.assert_allclose(
|
|
v.grad.to("float32").numpy(),
|
|
reorder_v_grad.to("float32").numpy(),
|
|
rtol=rtol,
|
|
atol=atol,
|
|
)
|
|
|
|
def run_test_cases(self):
|
|
# flash attention is not supported yet for cpu
|
|
if os.getenv("backend") == "gpu":
|
|
cuda_version_main = int(paddle.version.cuda().split(".")[0])
|
|
device_prop_main = paddle.device.cuda.get_device_capability()[0]
|
|
if cuda_version_main >= 11 and device_prop_main >= 8:
|
|
self._test_cp_base()
|
|
self._test_cp_base(is_causal=False)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
tester = TestContextParallel()
|
|
tester.run_test_cases()
|