2022-03-01 10:17:01 +00:00
|
|
|
from functools import partial
|
2022-03-02 10:28:29 +00:00
|
|
|
|
2022-03-01 10:17:01 +00:00
|
|
|
import torch
|
2022-03-02 10:28:29 +00:00
|
|
|
import torch.distributed as dist
|
2022-03-04 05:40:48 +00:00
|
|
|
from colossalai.logging import get_dist_logger
|
2022-03-02 10:28:29 +00:00
|
|
|
from colossalai.utils import checkpoint
|
2022-03-18 07:44:47 +00:00
|
|
|
from colossalai.zero.shard_utils import TensorShardStrategy
|
2022-03-10 01:57:26 +00:00
|
|
|
from colossalai.zero.sharded_model import ShardedModelV2
|
2022-03-01 10:17:01 +00:00
|
|
|
|
2022-03-17 05:16:22 +00:00
|
|
|
LOGGER = get_dist_logger('zero_test')
|
|
|
|
|
|
|
|
MP_PARALLEL_CONFIG = dict(fp16=dict(mode=None,), parallel=dict(pipeline=dict(size=1), tensor=dict(size=2, mode=None)))
|
|
|
|
|
|
|
|
_ZERO_MODEL_CONFIG = dict(reduce_scatter_bucket_size_mb=25,
|
|
|
|
fp32_reduce_scatter=False,
|
|
|
|
offload_config=None,
|
|
|
|
gradient_predivide_factor=1.0,
|
2022-03-18 07:44:47 +00:00
|
|
|
use_memory_tracer=False,
|
2022-03-23 06:59:59 +00:00
|
|
|
shard_strategy=TensorShardStrategy(),
|
|
|
|
reuse_fp16_shard=False)
|
2022-03-18 07:44:47 +00:00
|
|
|
|
|
|
|
_ZERO_OPTIMIZER_CONFIG = dict(cpu_offload=False,
|
|
|
|
initial_scale=2**5,
|
|
|
|
min_scale=1,
|
|
|
|
growth_factor=2,
|
|
|
|
backoff_factor=0.5,
|
|
|
|
growth_interval=1000,
|
|
|
|
hysteresis=2,
|
2022-03-18 08:18:31 +00:00
|
|
|
max_scale=2**32)
|
2022-03-17 02:20:53 +00:00
|
|
|
|
|
|
|
ZERO_PARALLEL_CONFIG = dict(fp16=dict(mode=None,),
|
|
|
|
zero=dict(
|
2022-03-17 05:16:22 +00:00
|
|
|
model_config=_ZERO_MODEL_CONFIG,
|
|
|
|
optimizer_config=_ZERO_OPTIMIZER_CONFIG,
|
2022-03-18 05:57:20 +00:00
|
|
|
),
|
|
|
|
parallel=dict(pipeline=dict(size=1), tensor=dict(size=1, mode=None)))
|
2022-03-17 02:20:53 +00:00
|
|
|
|
2022-03-04 05:40:48 +00:00
|
|
|
CONFIG = dict(fp16=dict(mode=None,),
|
|
|
|
zero=dict(level=3,
|
|
|
|
verbose=False,
|
|
|
|
offload_optimizer_config=dict(device='cpu', pin_memory=True, buffer_count=5, fast_init=False),
|
|
|
|
offload_param_config=dict(device='cpu',
|
|
|
|
pin_memory=True,
|
|
|
|
buffer_count=5,
|
|
|
|
buffer_size=1e8,
|
|
|
|
max_in_cpu=1e9)),
|
|
|
|
parallel=dict(pipeline=dict(size=1), tensor=dict(size=1, mode=None)))
|
2022-03-01 10:17:01 +00:00
|
|
|
|
2022-03-02 10:28:29 +00:00
|
|
|
|
2022-03-10 01:57:26 +00:00
|
|
|
def run_fwd_bwd(model, data, label, criterion, enable_autocast=False):
|
|
|
|
model.train()
|
|
|
|
with torch.cuda.amp.autocast(enabled=enable_autocast):
|
|
|
|
if criterion:
|
|
|
|
y = model(data)
|
|
|
|
loss = criterion(y, label)
|
|
|
|
else:
|
|
|
|
loss = model(data, label)
|
|
|
|
loss = loss.float()
|
|
|
|
if isinstance(model, ShardedModelV2):
|
|
|
|
model.backward(loss)
|
|
|
|
else:
|
|
|
|
loss.backward()
|
|
|
|
|
|
|
|
|
2022-03-01 10:17:01 +00:00
|
|
|
def checkpoint_wrapper(module, enable=True):
|
|
|
|
if enable:
|
|
|
|
module.forward = partial(checkpoint, module.forward)
|
|
|
|
return module
|
|
|
|
|
|
|
|
|
|
|
|
def allclose(tensor_a: torch.Tensor, tensor_b: torch.Tensor, loose=False) -> bool:
|
|
|
|
if loose:
|
2022-03-08 10:18:06 +00:00
|
|
|
return torch.allclose(tensor_a, tensor_b, atol=1e-2, rtol=1e-3)
|
2022-03-01 10:17:01 +00:00
|
|
|
return torch.allclose(tensor_a, tensor_b)
|
|
|
|
|
|
|
|
|
|
|
|
def check_grads(model, zero_model, loose=False):
|
|
|
|
for p, zero_p in zip(model.parameters(), zero_model.parameters()):
|
|
|
|
zero_grad = zero_p.grad.clone().to(p.device)
|
2022-03-08 10:18:06 +00:00
|
|
|
grad = p.grad.float()
|
|
|
|
assert grad.dtype == zero_grad.dtype
|
|
|
|
assert allclose(grad, zero_grad, loose=loose)
|
2022-03-02 10:28:29 +00:00
|
|
|
|
2022-03-01 10:17:01 +00:00
|
|
|
|
|
|
|
def check_params(model, zero_model, loose=False):
|
|
|
|
for p, zero_p in zip(model.parameters(), zero_model.parameters()):
|
|
|
|
zero_p = zero_p.clone().to(p.device)
|
2022-03-17 05:16:22 +00:00
|
|
|
# assert p.dtype == zero_p.dtype
|
|
|
|
assert allclose(p.float(), zero_p.float(), loose=loose), f"diff {p.float() - zero_p.float()}"
|
2022-03-01 10:17:01 +00:00
|
|
|
|
2022-03-02 10:28:29 +00:00
|
|
|
|
|
|
|
def check_grads_padding(model, zero_model, loose=False):
|
|
|
|
rank = dist.get_rank()
|
|
|
|
for p, zero_p in zip(model.parameters(), zero_model.parameters()):
|
|
|
|
zero_grad = zero_p.grad.clone().to(p.device)
|
|
|
|
chunks = torch.flatten(p.grad).chunk(dist.get_world_size())
|
|
|
|
if rank >= len(chunks):
|
|
|
|
continue
|
2022-03-08 10:18:06 +00:00
|
|
|
grad = chunks[rank].float()
|
2022-03-02 10:28:29 +00:00
|
|
|
if zero_grad.size(0) > grad.size(0):
|
|
|
|
zero_grad = zero_grad[:grad.size(0)]
|
|
|
|
assert grad.dtype == zero_grad.dtype
|
2022-03-09 02:39:02 +00:00
|
|
|
assert allclose(grad, zero_grad, loose=loose), f'diff: {grad - zero_grad}'
|
2022-03-02 10:28:29 +00:00
|
|
|
|
|
|
|
|
|
|
|
def check_params_padding(model, zero_model, loose=False):
|
|
|
|
rank = dist.get_rank()
|
|
|
|
for p, zero_p in zip(model.parameters(), zero_model.parameters()):
|
|
|
|
zero_p = zero_p.clone().to(p.device)
|
|
|
|
chunks = torch.flatten(p).chunk(dist.get_world_size())
|
|
|
|
if rank >= len(chunks):
|
|
|
|
continue
|
|
|
|
p = chunks[rank]
|
|
|
|
if zero_p.size(0) > p.size(0):
|
|
|
|
zero_p = zero_p[:p.size(0)]
|
|
|
|
assert p.dtype == zero_p.dtype
|
|
|
|
assert allclose(p, zero_p, loose=loose)
|
2022-03-04 05:40:48 +00:00
|
|
|
|
|
|
|
|
2022-03-23 06:59:59 +00:00
|
|
|
def check_sharded_model_params(model, zero_model, loose=False, reuse_fp16_shard=False):
|
2022-03-04 05:40:48 +00:00
|
|
|
rank = dist.get_rank()
|
|
|
|
for p, zero_p in zip(model.parameters(), zero_model.parameters()):
|
2022-03-23 06:59:59 +00:00
|
|
|
if reuse_fp16_shard:
|
|
|
|
zero_p = zero_p.data.to(p.device).float()
|
|
|
|
else:
|
|
|
|
zero_p = zero_p.col_attr.sharded_data_tensor.payload.to(p.device).float()
|
2022-03-04 05:40:48 +00:00
|
|
|
chunks = torch.flatten(p).chunk(dist.get_world_size())
|
|
|
|
if rank >= len(chunks):
|
|
|
|
continue
|
2022-03-09 08:09:36 +00:00
|
|
|
p = chunks[rank].float()
|
2022-03-04 05:40:48 +00:00
|
|
|
if zero_p.size(0) > p.size(0):
|
|
|
|
zero_p = zero_p[:p.size(0)]
|
|
|
|
assert p.dtype == zero_p.dtype
|
2022-03-09 08:09:36 +00:00
|
|
|
assert allclose(p, zero_p, loose=loose), f'{p} vs {zero_p}'
|