1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
| import traceback
import os
import socket
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
import ray
from ray.util.placement_group import placement_group, remove_placement_group
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
def find_free_port():
"""Find a free port for master communication"""
import socket
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.bind(("", 0))
s.listen(1)
port = s.getsockname()[1]
return port
def set_random_seed(seed=42):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
set_random_seed(42)
ray.init()
class DummyAttn(nn.Module):
"""
定义一个简单的 model，完成 forward，backward，梯度同步等操作。
"""
def __init__(
self,
num_heads: int = 16,
head_dim: int = 64,
hidden_size: int = 1024,
):
super().__init__()
self.num_heads = num_heads
self.head_dim = head_dim
self.q_proj = nn.Linear(hidden_size, head_dim * num_heads, bias=False)
self.k_proj = nn.Linear(hidden_size, head_dim * num_heads, bias=False)
self.v_proj = nn.Linear(hidden_size, head_dim * num_heads, bias=False)
self.out_proj = nn.Linear(head_dim * num_heads, hidden_size, bias=False)
self._init_weights()
self.register_backward_hook(self._allreduce_grads)
def _init_weights(self):
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
if m.bias is not None:
nn.init.zeros_(m.bias)
def register_backward_hook(self, hook):
for p in self.parameters():
if p.requires_grad is True:
p.register_hook(hook)
def _allreduce_grads(self, grad):
"""
在当前 group (default group) 中执行梯度 all_reduce 同步操作。
"""
dist.all_reduce(grad, op=dist.ReduceOp.SUM)
return grad
def forward(self, x):
"""
Args:
x: (batch_size, seq_len, hidden_size)
"""
bs, seq_len, hidden_size = x.size()
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x)
q = q.view(bs, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(bs, seq_len, self.num_heads, self.head_dim).permute(0, 2, 3, 1)
v = v.view(bs, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
attn = q @ k / math.sqrt(self.head_dim)
attn_scores = F.softmax(attn, dim=-1)
y = attn_scores @ v
y = y.transpose(1, 2).contiguous().view(bs, seq_len, -1)
y = self.out_proj(y)
return y
class WorkerBase:
""" 用作 Actors 的类
可以直接使用装饰器定义 @ray.remote 或者 ray.remote(WorkerBase)
"""
def __init__(self, temp_init: bool = False):
self._node_id = ray.get_runtime_context().get_node_id()
self._actor_id = ray.get_runtime_context().get_actor_id()
self._task_id = ray.get_runtime_context().get_task_id()
self._job_id = ray.get_runtime_context().get_job_id()
self._hostname = socket.gethostname()
self._ip_address = socket.gethostbyname(socket.gethostname())
if temp_init:
return
self._set_seed(42)
if not dist.is_initialized():
dist.init_process_group(
backend="cpu:gloo,cuda:nccl",
world_size=int(os.getenv("WORLD_SIZE", "1")),
rank=int(os.getenv("RANK", "0")),
)
self._rank = dist.get_rank()
self._world_size = dist.get_world_size()
self.model = DummyAttn()
self.model.to("cuda")
print(f"=> Rank {self._rank} init model")
def get_actor_info(self):
return {
"node_id": self._node_id,
"actor_id": self._actor_id,
"task_id": self._task_id,
"job_id": self._job_id,
"hostname": self._hostname,
"ip_address": self._ip_address,
}
def shutdown(self):
dist.destroy_process_group()
def _set_seed(self, seed: int = 42):
set_random_seed(seed)
def train_step(self, data):
self.model.train()
x = data.to("cuda")
y = self.model(x)
loss = y.sum()
loss.backward()
return loss.cpu()
def sample_grads(self):
for name, p in self.model.named_parameters():
if p.requires_grad is True:
return name, p.grad.cpu()
def main():
num_devices = get_num_devices()
pg = placement_group([
{"CPU": 1, "GPU": 1} for _ in range(num_devices)
], strategy="STRICT_PACK", name="ray_actor_communication")
ray.get(pg.ready())
print(f"=> Placement group {pg.id} is ready, num_devices: {num_devices}")
worker_cls = WorkerBase
Worker = ray.remote(worker_cls)
print("========= Get network info ==========")
temp_worker = Worker.options(
scheduling_strategy=PlacementGroupSchedulingStrategy(
placement_group=pg,
placement_group_bundle_index=0,
),
runtime_env={"env_vars": {"WORLD_SIZE": "1", "RANK": "0", "MASTER_ADDR": "127.0.0.1", "MASTER_PORT": str(find_free_port())}},
num_gpus=1,
num_cpus=1,
).remote(temp_init=True)
network_info = ray.get(temp_worker.get_actor_info.remote())
master_addr = network_info["ip_address"]
master_port = str(find_free_port())
print(f"Using master: {master_addr}:{master_port}")
ray.kill(temp_worker)
print("========= Terminate temporary worker ==========\n")
workers = []
for i in range(num_devices):
env_vars = {
"WORLD_SIZE": str(num_devices),
"RANK": str(i),
"MASTER_ADDR": master_addr,
"MASTER_PORT": master_port,
}
workers.append(
Worker.options(
scheduling_strategy=PlacementGroupSchedulingStrategy(
placement_group=pg,
placement_group_bundle_index=i,
),
runtime_env={
"env_vars": env_vars,
},
num_gpus=1,
num_cpus=1,
).remote()
)
datas = torch.randn(4, 128, 1024)
for i, worker in enumerate(workers):
data = datas.chunk(num_devices)[i]
worker.train_step.remote(data)
grads = ray.get([worker.sample_grads.remote() for worker in workers])
for i, (name, grad) in enumerate(grads):
print(f"=> Rank {i} grad: {name}\n{grad}")
if __name__ == "__main__":
try:
main()
except KeyboardInterrupt:
pass
except Exception as e:
print(e)
traceback.print_exc()
finally:
ray.shutdown()
|
