2022-12-22 02:15:34 +00:00
|
|
|
import gzip
|
|
|
|
import random
|
2023-01-10 08:18:56 +00:00
|
|
|
from functools import partial
|
2023-01-16 06:44:29 +00:00
|
|
|
from time import time
|
|
|
|
|
2022-12-22 02:15:34 +00:00
|
|
|
import numpy as np
|
|
|
|
import torch
|
2023-01-10 08:18:56 +00:00
|
|
|
import torch.nn as nn
|
2023-01-16 06:44:29 +00:00
|
|
|
import torch.optim as optim
|
2022-12-22 02:15:34 +00:00
|
|
|
import tqdm
|
2022-12-29 06:28:31 +00:00
|
|
|
from packaging import version
|
2022-12-22 02:15:34 +00:00
|
|
|
from palm_pytorch import PaLM
|
|
|
|
from palm_pytorch.autoregressive_wrapper import AutoregressiveWrapper
|
|
|
|
from torch.utils.data import DataLoader, Dataset
|
2022-12-29 06:01:09 +00:00
|
|
|
|
|
|
|
import colossalai
|
2022-12-29 06:28:31 +00:00
|
|
|
from colossalai.logging import disable_existing_loggers, get_dist_logger
|
|
|
|
from colossalai.nn.optimizer.gemini_optimizer import GeminiAdamOptimizer
|
2023-01-16 07:55:41 +00:00
|
|
|
from colossalai.nn.parallel import ZeroDDP
|
2022-12-29 06:01:09 +00:00
|
|
|
from colossalai.tensor import ColoParameter, ComputePattern, ComputeSpec, ProcessGroup, ReplicaSpec, ShardSpec
|
|
|
|
from colossalai.utils import MultiTimer, get_current_device
|
2022-12-29 06:28:31 +00:00
|
|
|
from colossalai.utils.model.colo_init_context import ColoInitContext
|
2022-12-22 02:15:34 +00:00
|
|
|
|
|
|
|
# constants
|
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
NUM_BATCHES = int(10)
|
2023-01-10 08:18:56 +00:00
|
|
|
WARMUP_BATCHES = 1
|
2022-12-29 06:28:31 +00:00
|
|
|
GRADIENT_ACCUMULATE_EVERY = 1
|
2022-12-22 02:15:34 +00:00
|
|
|
LEARNING_RATE = 2e-4
|
|
|
|
VALIDATE_EVERY = 100
|
|
|
|
GENERATE_EVERY = 500
|
|
|
|
GENERATE_LENGTH = 512
|
|
|
|
SEQ_LEN = 1024
|
|
|
|
|
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
def parse_args():
|
|
|
|
parser = colossalai.get_default_parser()
|
|
|
|
parser.add_argument(
|
|
|
|
"--distplan",
|
|
|
|
type=str,
|
|
|
|
default='colossalai',
|
|
|
|
help="The distributed plan [colossalai, pytorch].",
|
|
|
|
)
|
|
|
|
parser.add_argument(
|
|
|
|
"--tp_degree",
|
|
|
|
type=int,
|
|
|
|
default=1,
|
|
|
|
help="Tensor Parallelism Degree. Valid when using colossalai as dist plan.",
|
|
|
|
)
|
|
|
|
parser.add_argument(
|
|
|
|
"--placement",
|
|
|
|
type=str,
|
|
|
|
default='cpu',
|
|
|
|
help="Placement Policy for Gemini. Valid when using colossalai as dist plan.",
|
|
|
|
)
|
|
|
|
parser.add_argument(
|
|
|
|
"--shardinit",
|
|
|
|
type=bool,
|
|
|
|
default=False,
|
|
|
|
help=
|
|
|
|
"Shard the tensors when init the model to shrink peak memory size on the assigned device. Valid when using colossalai as dist plan.",
|
|
|
|
)
|
|
|
|
parser.add_argument(
|
|
|
|
"--batch_size",
|
|
|
|
type=int,
|
|
|
|
default=8,
|
|
|
|
help="batch size per DP group of training.",
|
|
|
|
)
|
2023-01-16 06:44:29 +00:00
|
|
|
parser.add_argument(
|
|
|
|
"--dummy_data",
|
|
|
|
type=bool,
|
|
|
|
default=False,
|
|
|
|
help="use dummy dataset.",
|
|
|
|
)
|
2023-01-03 09:49:00 +00:00
|
|
|
args = parser.parse_args()
|
|
|
|
return args
|
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2022-12-29 06:28:31 +00:00
|
|
|
# helpers
|
2022-12-22 02:15:34 +00:00
|
|
|
def cycle(loader):
|
|
|
|
while True:
|
|
|
|
for data in loader:
|
|
|
|
yield data
|
|
|
|
|
|
|
|
|
|
|
|
def decode_token(token):
|
|
|
|
return str(chr(max(32, token)))
|
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2023-01-10 08:18:56 +00:00
|
|
|
def get_tflops(model_numel, batch_size, seq_len, step_time):
|
|
|
|
return model_numel * batch_size * seq_len * 8 / 1e12 / (step_time + 1e-12)
|
2022-12-22 02:15:34 +00:00
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2022-12-22 02:15:34 +00:00
|
|
|
def decode_tokens(tokens):
|
|
|
|
return "".join(list(map(decode_token, tokens)))
|
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2023-01-10 08:18:56 +00:00
|
|
|
def get_model_size(model: nn.Module):
|
|
|
|
total_numel = 0
|
|
|
|
for module in model.modules():
|
|
|
|
for p in module.parameters(recurse=False):
|
|
|
|
total_numel += p.numel()
|
|
|
|
return total_numel
|
2022-12-29 06:28:31 +00:00
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2022-12-29 06:01:09 +00:00
|
|
|
# Gemini + ZeRO DDP
|
|
|
|
def gemini_zero_dpp(model: torch.nn.Module, pg: ProcessGroup, placememt_policy: str = "auto"):
|
|
|
|
cai_version = colossalai.__version__
|
|
|
|
if version.parse(cai_version) > version.parse("0.1.10"):
|
|
|
|
from colossalai.nn.parallel import GeminiDDP
|
|
|
|
model = GeminiDDP(model,
|
|
|
|
device=get_current_device(),
|
|
|
|
placement_policy=placememt_policy,
|
|
|
|
pin_memory=True,
|
|
|
|
search_range_mb=32)
|
|
|
|
elif version.parse(cai_version) <= version.parse("0.1.10") and version.parse(cai_version) >= version.parse("0.1.9"):
|
|
|
|
from colossalai.gemini import ChunkManager, GeminiManager
|
|
|
|
chunk_size = ChunkManager.search_chunk_size(model, 64 * 1024**2, 32)
|
|
|
|
gemini_manager = GeminiManager(placememt_policy, chunk_manager)
|
|
|
|
chunk_manager = ChunkManager(chunk_size,
|
|
|
|
pg,
|
|
|
|
enable_distributed_storage=True,
|
|
|
|
init_device=GeminiManager.get_default_device(placememt_policy))
|
|
|
|
model = ZeroDDP(model, gemini_manager)
|
|
|
|
else:
|
|
|
|
raise NotImplemented(f"CAI version {cai_version} is not supported")
|
|
|
|
return model
|
2022-12-29 06:28:31 +00:00
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2023-01-05 09:57:50 +00:00
|
|
|
## Parameter Sharding Strategies for Tensor Parallelism
|
|
|
|
def split_param_single_dim_tp1d(dim: int, param: ColoParameter, pg: ProcessGroup):
|
|
|
|
spec = (ShardSpec([dim], [pg.tp_world_size()]), ComputeSpec(ComputePattern.TP1D))
|
|
|
|
param.set_tensor_spec(*spec)
|
|
|
|
|
|
|
|
|
|
|
|
def split_param_row_tp1d(param: ColoParameter, pg: ProcessGroup):
|
|
|
|
split_param_single_dim_tp1d(0, param, pg)
|
|
|
|
|
|
|
|
|
|
|
|
def split_param_col_tp1d(param: ColoParameter, pg: ProcessGroup):
|
|
|
|
split_param_single_dim_tp1d(-1, param, pg)
|
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2023-01-05 09:57:50 +00:00
|
|
|
# Tensor Parallel
|
|
|
|
def tensor_parallelize(model: torch.nn.Module, pg: ProcessGroup):
|
|
|
|
"""tensor_parallelize
|
|
|
|
Sharding the Model Parameters.
|
|
|
|
Args:
|
|
|
|
model (torch.nn.Module): a torch module to be sharded
|
|
|
|
"""
|
|
|
|
for mn, module in model.named_modules():
|
|
|
|
for pn, param in module.named_parameters(recurse=False):
|
|
|
|
if hasattr(param, 'visited'):
|
|
|
|
continue
|
|
|
|
param.set_dist_spec(ReplicaSpec())
|
|
|
|
if 'net.0' in mn:
|
|
|
|
split_param_col_tp1d(param, pg) # colmn slice
|
|
|
|
elif 'to_q' in mn:
|
|
|
|
split_param_col_tp1d(param, pg) # colmn slice
|
|
|
|
elif 'to_kv' in mn:
|
|
|
|
split_param_row_tp1d(param, pg) # row slice
|
|
|
|
elif 'to_out' in mn:
|
|
|
|
split_param_row_tp1d(param, pg) # row slice
|
|
|
|
elif '1.1' in mn:
|
|
|
|
split_param_col_tp1d(param, pg) # colmn slice
|
|
|
|
elif '1.2' in mn:
|
|
|
|
split_param_row_tp1d(param, pg) # row slice
|
|
|
|
else:
|
|
|
|
param.set_dist_spec(ReplicaSpec())
|
|
|
|
param.visited = True
|
|
|
|
|
2022-12-29 06:28:31 +00:00
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
args = parse_args()
|
|
|
|
if args.distplan not in ["colossalai", "pytorch"]:
|
2023-01-16 06:44:29 +00:00
|
|
|
raise TypeError(f"{args.distplan} is error")
|
2022-12-29 06:01:09 +00:00
|
|
|
disable_existing_loggers()
|
2023-01-03 09:49:00 +00:00
|
|
|
colossalai.launch_from_torch(config={})
|
2023-01-10 08:18:56 +00:00
|
|
|
logger = get_dist_logger()
|
2022-12-29 06:01:09 +00:00
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
|
|
|
|
def generate_dataset(dummy_data: bool = False):
|
|
|
|
if not dummy_data:
|
|
|
|
with gzip.open("./data/enwik8.gz") as file:
|
|
|
|
X = np.fromstring(file.read(int(95e6)), dtype=np.uint8)
|
|
|
|
trX, vaX = np.split(X, [int(90e6)])
|
|
|
|
data_train, data_val = torch.from_numpy(trX), torch.from_numpy(vaX)
|
|
|
|
# print(f"data_train {data_train.shape} {data_train.dtype} {max(data_train)} {min(data_train)}")
|
|
|
|
# print(f"data_val {data_val.shape} {data_val.dtype} {max(data_val)} {min(data_val)}")
|
|
|
|
return data_train, data_val
|
|
|
|
else:
|
|
|
|
return torch.randint(0, 100, (90000000,)), torch.randint(0, 100, (5000000,))
|
|
|
|
|
|
|
|
|
|
|
|
data_train, data_val = generate_dataset(args.dummy_data)
|
|
|
|
|
|
|
|
print("generate dataset ready!")
|
2022-12-22 02:15:34 +00:00
|
|
|
|
|
|
|
|
|
|
|
class TextSamplerDataset(Dataset):
|
|
|
|
|
|
|
|
def __init__(self, data, seq_len):
|
|
|
|
super().__init__()
|
|
|
|
self.data = data
|
|
|
|
self.seq_len = seq_len
|
|
|
|
|
|
|
|
def __getitem__(self, index):
|
|
|
|
rand_start = torch.randint(0, self.data.size(0) - self.seq_len, (1,))
|
|
|
|
full_seq = self.data[rand_start:rand_start + self.seq_len + 1].long()
|
|
|
|
return full_seq.cuda()
|
|
|
|
|
|
|
|
def __len__(self):
|
|
|
|
return self.data.size(0) // self.seq_len
|
|
|
|
|
|
|
|
|
|
|
|
train_dataset = TextSamplerDataset(data_train, SEQ_LEN)
|
|
|
|
val_dataset = TextSamplerDataset(data_val, SEQ_LEN)
|
2023-01-03 09:49:00 +00:00
|
|
|
train_loader = cycle(DataLoader(train_dataset, batch_size=args.batch_size))
|
|
|
|
val_loader = cycle(DataLoader(val_dataset, batch_size=args.batch_size))
|
|
|
|
|
|
|
|
if args.distplan == "colossalai":
|
|
|
|
# instantiate GPT-like decoder model
|
2022-12-22 02:15:34 +00:00
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
default_pg = ProcessGroup(tp_degree=args.tp_degree)
|
|
|
|
default_dist_spec = ShardSpec([-1], [args.tp_degree]) if args.shardinit else None
|
|
|
|
ctx = ColoInitContext(device='cpu', default_dist_spec=default_dist_spec, default_pg=default_pg)
|
2022-12-29 06:01:09 +00:00
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
with ctx:
|
2023-01-10 08:18:56 +00:00
|
|
|
model = PaLM(num_tokens=50304, dim=4096, depth=64)
|
2023-01-03 09:49:00 +00:00
|
|
|
model = AutoregressiveWrapper(model, max_seq_len=SEQ_LEN)
|
|
|
|
|
|
|
|
pg = default_pg
|
2023-01-05 09:57:50 +00:00
|
|
|
tensor_parallelize(model, pg)
|
2023-01-03 09:49:00 +00:00
|
|
|
model = gemini_zero_dpp(model, pg, args.placement)
|
|
|
|
|
|
|
|
#optimizer
|
|
|
|
|
|
|
|
#optimizer = GeminiAdamOptimizer(model, lr=1e-7, initial_scale=2**5)
|
|
|
|
optimizer = GeminiAdamOptimizer(model, lr=LEARNING_RATE, initial_scale=2**5)
|
|
|
|
else:
|
|
|
|
model = PaLM(num_tokens=256, dim=512, depth=8)
|
|
|
|
model = AutoregressiveWrapper(model, max_seq_len=2048)
|
|
|
|
model.cuda()
|
|
|
|
optim = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
|
2022-12-29 06:01:09 +00:00
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
# model is shared after TP
|
2023-01-10 08:18:56 +00:00
|
|
|
numel = get_model_size(model)
|
|
|
|
get_tflops_func = partial(get_tflops, numel, args.batch_size, SEQ_LEN)
|
2022-12-22 02:15:34 +00:00
|
|
|
|
|
|
|
# training
|
2022-12-29 06:28:31 +00:00
|
|
|
model.train()
|
2023-01-10 08:18:56 +00:00
|
|
|
tflops_list = []
|
2022-12-22 02:15:34 +00:00
|
|
|
for i in tqdm.tqdm(range(NUM_BATCHES), mininterval=10.0, desc="training"):
|
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
if args.distplan == "colossalai":
|
|
|
|
optimizer.zero_grad()
|
2023-01-10 08:18:56 +00:00
|
|
|
start = time()
|
2023-01-03 09:49:00 +00:00
|
|
|
loss = model(next(train_loader))
|
2023-01-10 08:18:56 +00:00
|
|
|
fwd_end = time()
|
|
|
|
fwd_time = fwd_end - start
|
2023-01-03 09:49:00 +00:00
|
|
|
# loss.backward()
|
|
|
|
optimizer.backward(loss)
|
2023-01-10 08:18:56 +00:00
|
|
|
bwd_end = time()
|
|
|
|
bwd_time = bwd_end - fwd_end
|
2022-12-29 06:28:31 +00:00
|
|
|
|
2023-01-10 08:18:56 +00:00
|
|
|
# print(f"training loss: {loss.item()}")
|
2023-01-03 09:49:00 +00:00
|
|
|
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
|
|
|
|
# optim.step()
|
|
|
|
# optim.zero_grad()
|
|
|
|
optimizer.step()
|
2023-01-10 08:18:56 +00:00
|
|
|
optim_time = time() - bwd_end
|
|
|
|
step_time = time() - start
|
|
|
|
|
|
|
|
step_tflops = get_tflops_func(step_time)
|
|
|
|
logger.info(
|
|
|
|
f"[{i + 1}/{NUM_BATCHES}] Loss:{loss.item():.3f}, Step time: {step_time:.3f}s, TFLOPS: {get_tflops_func(step_time):.3f}, FWD time: {fwd_time:.3f}s, BWD time: {bwd_time:.3f}s, OPTIM time: {optim_time:.3f}s",
|
|
|
|
ranks=[0],
|
|
|
|
)
|
|
|
|
if i >= WARMUP_BATCHES:
|
|
|
|
tflops_list.append(step_tflops)
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
else:
|
|
|
|
for __ in range(GRADIENT_ACCUMULATE_EVERY):
|
|
|
|
loss = model(next(train_loader))
|
|
|
|
loss.backward()
|
2022-12-22 02:15:34 +00:00
|
|
|
|
2023-01-03 09:49:00 +00:00
|
|
|
print(f"training loss: {loss.item()}")
|
|
|
|
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
|
|
|
|
optim.step()
|
|
|
|
optim.zero_grad()
|
2023-01-16 06:44:29 +00:00
|
|
|
|
2023-01-10 08:18:56 +00:00
|
|
|
tflops_list.sort()
|
|
|
|
median_index = ((NUM_BATCHES - WARMUP_BATCHES) >> 1) + WARMUP_BATCHES
|
|
|
|
logger.info(f"Median TFLOPS is {tflops_list[median_index]:.3f}")
|
|
|
|
|
2023-01-16 06:44:29 +00:00
|
|
|
# TODO
|
|
|
|
# if i % VALIDATE_EVERY == 0:
|
|
|
|
# model.eval()
|
|
|
|
# with torch.no_grad():
|
|
|
|
# loss = model(next(val_loader))
|
|
|
|
# print(f"validation loss: {loss.item()}")
|
2022-12-29 06:28:31 +00:00
|
|
|
|
|
|
|
# if i % GENERATE_EVERY == 0:
|
|
|
|
# model.eval()
|
|
|
|
# inp = random.choice(val_dataset)[:-1]
|
|
|
|
# prime = decode_tokens(inp)
|
|
|
|
# print(f"%s \n\n %s", (prime, "*" * 100))
|
|
|
|
|
|
|
|
# sample = model.generate(inp[None, ...], GENERATE_LENGTH)
|
|
|
|
# output_str = decode_tokens(sample[0])
|
2023-01-16 06:44:29 +00:00
|
|
|
# print(output_str)
|