File size: 9,376 Bytes
d7b3a74 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | import logging
import os
from contextlib import contextmanager
import torch
import torch.distributed as dist
from slime.utils.memory_utils import available_memory, clear_memory, print_memory
logger = logging.getLogger(__name__)
old_new_group_dict = {}
def monkey_patch_torch_dist():
pid = os.getpid()
if pid in old_new_group_dict:
assert dist.old_new_group == old_new_group_dict[pid]
return
logger.info("Applying monkey patch to torch.distributed")
old_new_group = dist.new_group
old_new_group_dict[pid] = old_new_group
dist.old_new_group = old_new_group
def new_group(*args, **kwargs):
group = old_new_group(*args, **kwargs)
# skip none nccl group.
if len(args) >= 3 and args[2] == "gloo" or "backend" in kwargs and kwargs["backend"] == "gloo":
return group
# Get ranks from arguments
if len(args) >= 1 and args[0] is not None:
ranks = args[0]
elif "ranks" in kwargs and kwargs["ranks"] is not None:
ranks = kwargs["ranks"]
else:
# If no ranks specified, use all ranks in world
ranks = list(range(dist.get_world_size()))
if len(ranks) == 1:
return group
group = ReloadableProcessGroup(group, ranks)
return group
dist.new_group = new_group
def get_new_function(func):
def new_function(*args, **kwargs):
args = tuple([arg.group if isinstance(arg, ReloadableProcessGroup) else arg for arg in args])
kwargs = {k: (v.group if isinstance(v, ReloadableProcessGroup) else v) for k, v in kwargs.items()}
with _wrap_low_level_call():
return func(*args, **kwargs)
return new_function
dist.get_rank = get_new_function(dist.get_rank)
dist.get_world_size = get_new_function(dist.get_world_size)
dist.get_backend = get_new_function(dist.get_backend)
dist.get_global_rank = get_new_function(dist.get_global_rank)
dist.get_group_rank = get_new_function(dist.get_group_rank)
dist.get_process_group_ranks = get_new_function(dist.get_process_group_ranks)
dist.all_reduce = get_new_function(dist.all_reduce)
dist.all_gather = get_new_function(dist.all_gather)
dist.all_gather_into_tensor = get_new_function(dist.all_gather_into_tensor)
dist.all_gather_object = get_new_function(dist.all_gather_object)
dist.all_to_all = get_new_function(dist.all_to_all)
dist.all_to_all_single = get_new_function(dist.all_to_all_single)
dist.broadcast = get_new_function(dist.broadcast)
dist.reduce = get_new_function(dist.reduce)
dist.reduce_scatter = get_new_function(dist.reduce_scatter)
dist.reduce_scatter_tensor = get_new_function(dist.reduce_scatter_tensor)
dist.scatter = get_new_function(dist.scatter)
dist.gather = get_new_function(dist.gather)
dist.barrier = get_new_function(dist.barrier)
dist.send = get_new_function(dist.send)
dist.recv = get_new_function(dist.recv)
dist._coalescing_manager = get_new_function(dist._coalescing_manager)
# p2p
old_isend = dist.isend
old_irecv = dist.irecv
dist.isend = get_new_function(dist.isend)
dist.irecv = get_new_function(dist.irecv)
def get_new_p2pop_function(func):
def new_function(*args, **kwargs):
def convert(arg):
if isinstance(arg, ReloadableProcessGroup):
return arg.group
elif arg == dist.isend:
arg = old_isend
elif arg == dist.irecv:
arg = old_irecv
return arg
args = (convert(arg) for arg in args)
kwargs = {k: convert(v) for k, v in kwargs.items()}
return func(*args, **kwargs)
return new_function
dist.P2POp.__new__ = get_new_p2pop_function(dist.P2POp.__new__)
dist.P2POp.__init__ = get_new_p2pop_function(dist.P2POp.__init__)
class ReloadableProcessGroup(torch.distributed.ProcessGroup):
GROUPS = {}
def __init__(self, group, ranks):
super().__init__(
rank=dist.get_rank(group),
size=dist.get_world_size(group),
)
self.group = group
self.group_info = {
"ranks": ranks,
}
pid = os.getpid()
if pid not in ReloadableProcessGroup.GROUPS:
ReloadableProcessGroup.GROUPS[pid] = []
ReloadableProcessGroup.GROUPS[pid].append(self)
def __getattr__(self, name):
return getattr(self.group, name)
@staticmethod
def destroy_process_groups():
pid = os.getpid()
for reloadable_group in ReloadableProcessGroup.GROUPS.get(pid, []):
if reloadable_group.group is None:
continue
try:
dist.destroy_process_group(reloadable_group.group)
except ValueError as e:
logger.warning(
f"Process group already invalid/destroyed; skipping cleanup. Exception: {e}",
exc_info=True,
)
del reloadable_group.group
reloadable_group.group = None
@staticmethod
def reload_process_groups():
pid = os.getpid()
reloadable_groups = ReloadableProcessGroup.GROUPS.get(pid, [])
logger.info(f"Reloading {len(reloadable_groups)} process groups in pid {pid}")
old_new_group = old_new_group_dict.get(pid)
for reloadable_group in reloadable_groups:
if reloadable_group.group is not None:
continue
group = old_new_group(ranks=reloadable_group.group_info["ranks"], backend="nccl")
reloadable_group.group = group
def rank(self) -> int:
return self.group.rank()
def size(self) -> int:
return self.group.size()
def name(self) -> str:
return self.group.name()
def shutdown(self) -> None:
if self.group is not None:
self.group.shutdown()
def abort(self) -> None:
if self.group is not None:
self.group.abort()
def _fwd(self, method, *args, **kwargs):
inner = self.group
if inner is None:
raise RuntimeError("ReloadableProcessGroup: inner PG is None, call reload() first.")
with _wrap_low_level_call():
return getattr(inner, method)(*args, **kwargs)
def barrier(self, *a, **kw):
return self._fwd("barrier", *a, **kw)
def broadcast(self, *a, **kw):
return self._fwd("broadcast", *a, **kw)
def allreduce(self, *a, **kw):
return self._fwd("allreduce", *a, **kw)
def allreduce_coalesced(self, *a, **kw):
return self._fwd("allreduce_coalesced", *a, **kw)
def reduce(self, *a, **kw):
return self._fwd("reduce", *a, **kw)
def allgather(self, *a, **kw):
return self._fwd("allgather", *a, **kw)
def _allgather_base(self, *a, **kw):
return self._fwd("_allgather_base", *a, **kw)
def allgather_coalesced(self, *a, **kw):
return self._fwd("allgather_coalesced", *a, **kw)
def allgather_into_tensor_coalesced(self, *a, **kw):
return self._fwd("allgather_into_tensor_coalesced", *a, **kw)
def gather(self, *a, **kw):
return self._fwd("gather", *a, **kw)
def scatter(self, *a, **kw):
return self._fwd("scatter", *a, **kw)
def reduce_scatter(self, *a, **kw):
return self._fwd("reduce_scatter", *a, **kw)
def _reduce_scatter_base(self, *a, **kw):
return self._fwd("_reduce_scatter_base", *a, **kw)
def reduce_scatter_tensor_coalesced(self, *a, **kw):
return self._fwd("reduce_scatter_tensor_coalesced", *a, **kw)
def alltoall_base(self, *a, **kw):
return self._fwd("alltoall_base", *a, **kw)
def alltoall(self, *a, **kw):
return self._fwd("alltoall", *a, **kw)
def send(self, *a, **kw):
return self._fwd("send", *a, **kw)
def recv(self, *a, **kw):
return self._fwd("recv", *a, **kw)
def recv_anysource(self, *a, **kw):
return self._fwd("recv_anysource", *a, **kw)
def _start_coalescing(self, *a, **kw):
return self._fwd("_start_coalescing", *a, **kw)
def _end_coalescing(self, *a, **kw):
return self._fwd("_end_coalescing", *a, **kw)
def _get_backend_name(self):
return self._fwd("_get_backend_name")
def _get_backend(self, *a, **kw):
return self._fwd("_get_backend", *a, **kw)
def _set_default_backend(self, *a, **kw):
return self._fwd("_set_default_backend", *a, **kw)
@property
def bound_device_id(self):
return self.group.bound_device_id
@bound_device_id.setter
def bound_device_id(self, dev):
self.group.bound_device_id = dev
def destroy_process_groups():
"""Destroy all reloadable process groups."""
ReloadableProcessGroup.destroy_process_groups()
def reload_process_groups():
"""Reload all reloadable process groups."""
ReloadableProcessGroup.reload_process_groups()
@contextmanager
def _wrap_low_level_call():
try:
mem_info = available_memory()
if mem_info["free_GB"] < 5:
clear_memory()
yield
except Exception as e:
mem_info = print_memory("after torch distributed error")
e.add_note(f"{mem_info=}")
raise
|