File size: 8,825 Bytes
4947683 | 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 | from typing import Optional, List
import torch.distributed as dist
from torch.utils.data import Sampler
import numpy as np
import numba
@numba.njit
def ffd_check(a: np.ndarray, c: int, n: int):
# First-fit-decreasing bin packing
# Check if a[] could fit in n bins with capacity c
# https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
a = np.sort(a)[::-1]
bins = np.full((n, ), c, dtype=a.dtype)
for size in a:
not_found = True
for idx in range(n):
if bins[idx] >= size:
bins[idx] -= size
not_found = False
break
if not_found:
return False
return True
@numba.njit
def ffd_with_result(a: np.ndarray, c: int, start_index: int):
# First-fit-decreasing bin packing (with result return)
indices = np.argsort(a)[::-1]
a = a[indices]
bins = []
bins_result = []
for a_id, size in enumerate(a):
add_new = True
for idx in range(len(bins)):
if bins[idx] >= size:
bins[idx] -= size
bins_result[idx].append(indices[a_id] + start_index)
add_new = False
break
if add_new:
bins.append(c - size)
bins_result.append([indices[a_id] + start_index])
return bins_result
@numba.njit
def allocate(lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int):
# Dynamic batch allocator, similar to Multifit
# https://en.wikipedia.org/wiki/Multifit_algorithm
# ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
s = 0
start_index = 0
result = []
while True:
# binary search [l, r)
l = 1
r = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
while r - l > 1:
m = (l + r) // 2
if ffd_check(lengths[start_index: start_index + m], c, n):
l = m
else:
r = m
# use length l
batch = ffd_with_result(lengths[start_index: start_index + l], c, start_index)
assert len(batch) <= n
if len(batch) < n:
break
start_index += l
s = lengths_cumsum[start_index - 1]
# add local rank
result.append(batch[rank])
return result, s, len(result) * c * n
class MultipackDistributedBatchSampler(Sampler):
"""Unpadded length sampling using Multipack.
Approximate (at most ~1.22x) the optimal solution of the identical-machines scheduling problem, which is NP-hard."""
def __init__(
self,
batch_max_length: int,
lengths: List[int],
num_replicas: Optional[int] = None,
rank: Optional[int] = None,
seed: int = 0,
):
# Get rank
if num_replicas is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
num_replicas = dist.get_world_size()
if rank is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
rank = dist.get_rank()
self.num_replicas = num_replicas
self.rank = rank
self.seed = seed
self.batch_max_length = batch_max_length
self.lengths = lengths
assert isinstance(self.lengths, np.ndarray)
self.epoch = 0
# statistics
self.eff_total_used = 0
self.eff_total_slots = 0
def set_epoch(self, epoch: int):
self.epoch = epoch
def generate_batches(self, set_stats=False):
indices = np.random.default_rng(seed=self.seed + self.epoch).permutation(len(self.lengths))
lengths = self.lengths[indices]
lengths_cumsum = np.cumsum(lengths)
batches, total_used, total_slots = allocate(lengths=lengths,
lengths_cumsum=lengths_cumsum,
rank=self.rank,
c=self.batch_max_length,
n=self.num_replicas)
batches = [indices[batch] for batch in batches]
# statistics
if set_stats:
self.eff_total_used += total_used
self.eff_total_slots += total_slots
return batches
def __iter__(self):
batches = self.generate_batches(set_stats=True)
return iter(batches)
def __len__(self):
return self.num_batches()
def num_batches(self):
batches = self.generate_batches()
return len(batches)
def efficiency(self):
return self.eff_total_used / self.eff_total_slots
class MultipackDistributedBatchSamplerHybrid(Sampler):
"""Unpadded length sampling using Multipack.
Approximate (at most ~1.22x) the optimal solution of the identical-machines scheduling problem, which is NP-hard."""
def __init__(
self,
batch_max_length: int,
lengths: List[int],
num_replicas: Optional[int] = None,
rank: Optional[int] = None,
seed: int = 0,
min_max_ratio: int = 200
):
# Get rank
if num_replicas is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
num_replicas = dist.get_world_size()
if rank is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
rank = dist.get_rank()
self.num_replicas = num_replicas
self.rank = rank
self.seed = seed
self.batch_max_length = batch_max_length
self.lengths = lengths
assert isinstance(self.lengths, np.ndarray)
self.epoch = 0
# statistics
self.eff_total_used = 0
self.eff_total_slots = 0
self.min_max_ratio = min_max_ratio
def set_epoch(self, epoch: int):
self.epoch = epoch
def generate_batches(self, set_stats=False):
small = np.where(self.lengths <= self.min_max_ratio)[0]
large = np.where(self.lengths > self.min_max_ratio)[0]
# indices = np.random.default_rng(seed=self.seed + self.epoch).permutation(len(self.lengths))
indices_small = np.random.default_rng(seed=self.seed + self.epoch).permutation(len(small))
indices_large = np.random.default_rng(seed=self.seed + self.epoch).permutation(len(large))
indices_small = small[indices_small]
indices_large = large[indices_large]
lengths_small = self.lengths[indices_small]
lengths_cumsum_small = np.cumsum(lengths_small)
batches_small, total_used_small, total_slots_small = allocate(lengths=lengths_small,
lengths_cumsum=lengths_cumsum_small,
rank=self.rank,
c=self.batch_max_length,
n=self.num_replicas)
batches_small = [indices_small[batch] for batch in batches_small]
lengths_large = self.lengths[indices_large]
lengths_cumsum_large = np.cumsum(lengths_large)
batches_large, total_used_large, total_slots_large = allocate(lengths=lengths_large,
lengths_cumsum=lengths_cumsum_large,
rank=self.rank,
c=self.batch_max_length,
n=self.num_replicas)
batches_large = [indices_large[batch] for batch in batches_large]
batches = batches_small + batches_large
indices = np.random.default_rng(seed=self.seed + self.epoch).permutation(len(batches))
batches = [batches[i] for i in indices]
# statistics
if set_stats:
self.eff_total_used += total_used_large
self.eff_total_slots += total_slots_large
self.eff_total_used += total_used_small
self.eff_total_slots += total_slots_small
return batches
def __iter__(self):
batches = self.generate_batches(set_stats=True)
return iter(batches)
def __len__(self):
return self.num_batches()
def num_batches(self):
batches = self.generate_batches()
return len(batches)
def efficiency(self):
return self.eff_total_used / self.eff_total_slots |