Making large AI models cheaper, faster and more accessible
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 

433 lines
16 KiB

#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from collections import defaultdict
from contextlib import contextmanager
from typing import Dict, List, Optional, Union
import torch
import torch.distributed as dist
from torch import inf
from torch.nn.parameter import Parameter
from colossalai.legacy.constants import IS_TENSOR_PARALLEL, NUM_PARTITIONS, TENSOR_PARALLEL_ATTRIBUTES
from colossalai.legacy.context.parallel_mode import ParallelMode
from colossalai.legacy.core import global_context as gpc
from colossalai.legacy.global_variables import tensor_parallel_env as env
from colossalai.legacy.tensor import ProcessGroup
from colossalai.tensor import ColoParameter
from colossalai.utils.multi_tensor_apply import multi_tensor_applier
try:
from colossalai._C import fused_optim
except:
fused_optim = None
def print_rank_0(msg: str, logger=None):
"""Print messages and save logs(optional). This is executed only if you are the rank-0 gpu.
Args:
msg (str): A string message to output.
logger (:class:`colossalai.logging.DistributedLogger`, optional):
The logger to record the message, defaults to None.
"""
if gpc.get_global_rank() == 0:
if logger is None:
print(msg, flush=True)
else:
logger.info(msg)
def sync_model_param(model, parallel_mode):
r"""Make sure data parameters are consistent during Data Parallel Mode.
Args:
model (:class:`torch.nn.Module`): A pyTorch model on whose parameters you check the consistency.
parallel_mode (:class:`colossalai.legacy.context.ParallelMode`): Parallel mode to be checked.
Note:
The parallel_mode should be concluded in ``ParallelMode``. More details about ``ParallelMode`` could be found
in `parallel_mode <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/context/parallel_mode.py>`_
"""
if gpc.is_initialized(parallel_mode) and gpc.get_world_size(parallel_mode) > 1:
for param in model.parameters():
ranks = gpc.get_ranks_in_group(parallel_mode)
dist.broadcast(param, src=ranks[0], group=gpc.get_group(parallel_mode))
def is_dp_rank_0():
return not gpc.is_initialized(ParallelMode.DATA) or gpc.is_first_rank(ParallelMode.DATA)
def is_tp_rank_0():
return not gpc.is_initialized(ParallelMode.TENSOR) or gpc.is_first_rank(ParallelMode.TENSOR)
def is_no_pp_or_last_stage():
return not gpc.is_initialized(ParallelMode.PIPELINE) or gpc.is_last_rank(ParallelMode.PIPELINE)
def is_using_ddp():
return gpc.is_initialized(ParallelMode.DATA) and gpc.get_world_size(ParallelMode.DATA) > 1
def is_using_pp():
return gpc.is_initialized(ParallelMode.PIPELINE) and gpc.get_world_size(ParallelMode.PIPELINE) > 1
def is_using_sequence():
return gpc.is_initialized(ParallelMode.SEQUENCE) and gpc.get_world_size(ParallelMode.SEQUENCE) > 1
class model_branch_context(object):
def __enter__(self):
self.env_status = env.save()
def __exit__(self, *exc_info):
env.load(**self.env_status)
def is_model_parallel_parameter(p):
return hasattr(p, IS_TENSOR_PARALLEL) and getattr(p, IS_TENSOR_PARALLEL)
def _calc_l2_norm(grads):
# we should not
global fused_optim
if fused_optim is None:
from colossalai.kernel.kernel_loader import FusedOptimizerLoader
fused_optim = FusedOptimizerLoader().load()
norm = 0.0
if len(grads) > 0:
dummy_overflow_buf = torch.cuda.IntTensor([0])
norm, _ = multi_tensor_applier(
fused_optim.multi_tensor_l2norm, dummy_overflow_buf, [grads], False # no per-parameter norm
)
return norm
def _calc_lp(grads, norm_type):
norm = 0.0
for grad in grads:
grad_norm = torch.norm(grad, norm_type)
norm += grad_norm**norm_type
return norm
def _move_norm_to_cuda(norm: Union[float, torch.Tensor]) -> Union[float, torch.Tensor]:
if torch.is_tensor(norm) and norm.device.type != "cuda":
norm = norm.to(torch.cuda.current_device())
return norm
def _get_tensor_norm(norm: Union[float, torch.Tensor], move_to_cuda) -> torch.Tensor:
if isinstance(norm, float):
norm = torch.Tensor([norm])
if move_to_cuda:
norm = norm.to(torch.cuda.current_device())
return norm
# ======== Gradient Clipping =========
def _compute_local_lp(params: List[ColoParameter], norm_type: float) -> float:
if len(params) == 0:
return 0.0
grads = [p.grad for p in params]
use_cuda_kernel = grads[0].device.type == "cuda"
if norm_type == inf:
local_lp = max([g.abs().max() for g in grads])
elif norm_type == 2.0 and use_cuda_kernel:
local_lp = _calc_l2_norm(grads) ** norm_type
else:
local_lp = _calc_lp(grads, norm_type)
if isinstance(local_lp, torch.Tensor):
return local_lp.item()
return local_lp
def _compute_buckets_lp(params: List[ColoParameter], norm_type: float) -> float:
if len(params) == 0:
return 0.0
buckets: Dict[Optional[ProcessGroup], List[ColoParameter]] = defaultdict(list)
for p in params:
if p.is_replicate():
buckets[None].append(p)
else:
buckets[p.get_process_group().tp_process_group()].append(p)
total_lp = 0.0
for group, bucket in buckets.items():
local_lp = _compute_local_lp(bucket, norm_type)
if group is not None:
local_lp_tensor = torch.tensor([local_lp], device=torch.cuda.current_device())
if norm_type == inf:
dist.all_reduce(local_lp_tensor, op=dist.ReduceOp.MAX, group=group)
else:
dist.all_reduce(local_lp_tensor, group=group)
local_lp = local_lp_tensor.item()
if norm_type == inf:
total_lp = max(total_lp, local_lp)
else:
total_lp += local_lp
return total_lp
def _compute_pp_grad_lp(total_lp: float, norm_type: float) -> float:
if gpc.is_initialized(ParallelMode.PIPELINE) and gpc.get_world_size(ParallelMode.PIPELINE) > 1:
total_lp_tensor = torch.tensor([total_lp], device=torch.cuda.current_device())
if norm_type == inf:
dist.all_reduce(total_lp_tensor, op=dist.ReduceOp.MAX, group=gpc.get_group(ParallelMode.PIPELINE))
else:
dist.all_reduce(total_lp_tensor, group=gpc.get_group(ParallelMode.PIPELINE))
total_lp = total_lp_tensor.item()
return total_lp
def _compute_grad_lp(parameters, norm_type: float = 2.0) -> float:
if isinstance(parameters, torch.Tensor):
parameters = [parameters]
grad_dtype = None
cpu_grad_params: List[ColoParameter] = []
cuda_grad_params: List[ColoParameter] = []
for p in parameters:
if p.grad is None:
continue
assert isinstance(p, ColoParameter)
if grad_dtype is None:
grad_dtype = p.grad.dtype
assert p.grad.dtype == grad_dtype, f"Expected all grads are {grad_dtype}, got {p.grad.dtype}"
if p.grad.device.type == "cuda":
cuda_grad_params.append(p)
else:
cpu_grad_params.append(p)
norm_type = float(norm_type)
cpu_lp = _compute_buckets_lp(cpu_grad_params, norm_type)
cuda_lp = _compute_buckets_lp(cuda_grad_params, norm_type)
if norm_type == inf:
total_lp = max(cpu_lp, cuda_lp)
else:
total_lp = cpu_lp + cuda_lp
return _compute_pp_grad_lp(total_lp, norm_type)
def compute_grad_norm(parameters, norm_type: float = 2.0) -> float:
norm_type = float(norm_type)
total_norm = _compute_grad_lp(parameters, norm_type)
if norm_type != inf:
total_norm = total_norm ** (1 / norm_type)
return total_norm
def _clip_grad_norm(parameters, max_norm: float, total_norm: float) -> None:
clip_coef = max_norm / (total_norm + 1e-6)
if clip_coef < 1.0:
cuda_grads: List[torch.Tensor] = []
cpu_grads: List[torch.Tensor] = []
if isinstance(parameters, torch.Tensor):
parameters = [parameters]
for p in parameters:
if p.grad is None:
continue
if p.grad.device.type == "cuda":
cuda_grads.append(p.grad.detach())
else:
cpu_grads.append(p.grad.detach())
if len(cuda_grads) > 0:
dummy_overflow_buf = torch.cuda.IntTensor([0])
multi_tensor_applier(
fused_optim.multi_tensor_scale, dummy_overflow_buf, [cuda_grads, cuda_grads], clip_coef
)
for g in cpu_grads:
g.mul_(clip_coef)
def clip_grad_norm(parameters, max_norm: float, norm_type: float = 2.0) -> float:
total_norm = compute_grad_norm(parameters, norm_type)
_clip_grad_norm(parameters, max_norm, total_norm)
return total_norm
def clip_grad_norm_fp32(parameters, max_norm, norm_type=2):
"""Clips gradient norm of an iterable of parameters whose gradients are in fp32.
This is adapted from :func:`torch.nn.utils.clip_grad.clip_grad_norm_` and
added functionality to handle model parallel parameters.
Note:
the gradients are modified in place.
Args:
parameters (Iterable[:class:`torch.tensor`] or :class:`torch.tensor`):
An iterable of Tensors or a single Tensor that will have gradients normalized.
max_norm (Union[float, int]): Max norm of the gradients.
norm_type (Union[float, int, 'inf']): Type of the used p-norm. Can be ``'inf'`` for infinity norm.
Returns:
float: Total norm of the parameters.
"""
if isinstance(parameters, torch.Tensor):
parameters = [parameters]
# Filter parameters based on:
# - grad should not be none
# - parameter should not be shared
# - should not be a replica due to tensor model parallelism
params: List[Parameter] = []
has_zero_shared_param: bool = False
for param in parameters:
if param.grad is not None:
# Make sure the grads are in fp32
assert (
param.grad.dtype == torch.float
), f"expected gradient to be dtype torch.float, but got {param.grad.type()}"
if hasattr(param, "colo_attr") and param.colo_attr.sharded_data_tensor.is_sharded:
has_zero_shared_param = True
params.append(param)
if len(params) == 0:
enable_cuda_kernels = False
else:
enable_cuda_kernels = params[0].grad.device.type == "cuda"
# Norm parameters.
max_norm = float(max_norm)
norm_type = float(norm_type)
# Parameters can be on CPU or CUDA
# If parameters are on CPU, disable CUDA kernels
# Calculate norm.
if norm_type == inf:
total_norm = max(p.grad.data.abs().max() for p in params)
total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)])
# Take max across all model-parallel GPUs.
if gpc.is_initialized(ParallelMode.MODEL) and gpc.get_world_size(ParallelMode.MODEL) > 1:
dist.all_reduce(
total_norm_cuda, op=dist.ReduceOp.MAX, group=gpc.get_group(ParallelMode.MODEL), async_op=False
)
if has_zero_shared_param:
dist.all_reduce(
total_norm_cuda, op=dist.ReduceOp.MAX, group=gpc.get_group(ParallelMode.DATA), async_op=False
)
total_norm = total_norm_cuda[0].item()
else:
tensor_parallel_grads = []
no_tensor_parallel_grads = []
zero_sharded_grads = []
for p in params:
if is_model_parallel_parameter(p):
reductor = (gpc.get_world_size(ParallelMode.TENSOR) / getattr(p, NUM_PARTITIONS)) ** (1 / norm_type)
tensor_parallel_grads.append(p.grad.data / reductor)
elif hasattr(p, "colo_attr") and p.colo_attr.sharded_data_tensor.is_sharded:
zero_sharded_grads.append(p.grad.data)
else:
no_tensor_parallel_grads.append(p.grad.data)
if norm_type == 2.0 and enable_cuda_kernels:
tensor_parallel_norm = _calc_l2_norm(tensor_parallel_grads) ** norm_type
no_tensor_parallel_norm = _calc_l2_norm(no_tensor_parallel_grads) ** norm_type
zero_sharded_norm = _calc_l2_norm(zero_sharded_grads) ** norm_type
else:
tensor_parallel_norm = _calc_lp(tensor_parallel_grads, norm_type)
no_tensor_parallel_norm = _calc_lp(no_tensor_parallel_grads, norm_type)
zero_sharded_norm = _calc_lp(zero_sharded_grads, norm_type)
# If norm is type of float, then we convert them into torch.Tensor.
tensor_parallel_norm = _get_tensor_norm(tensor_parallel_norm, enable_cuda_kernels)
no_tensor_parallel_norm = _get_tensor_norm(no_tensor_parallel_norm, enable_cuda_kernels)
zero_sharded_norm = _get_tensor_norm(zero_sharded_norm, enable_cuda_kernels)
# If grads are on CPU, the norms is also on CPU. Cast them to CUDA tensors
if not enable_cuda_kernels:
tensor_parallel_norm = _move_norm_to_cuda(tensor_parallel_norm)
no_tensor_parallel_norm = _move_norm_to_cuda(no_tensor_parallel_norm)
zero_sharded_norm = _move_norm_to_cuda(zero_sharded_norm)
# Sum across all model-parallel GPUs.
if gpc.is_initialized(ParallelMode.TENSOR) and len(tensor_parallel_grads) > 0:
dist.all_reduce(tensor_parallel_norm, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.TENSOR))
# Sum across all zero sharded GPUs
if len(zero_sharded_grads) > 0:
dist.all_reduce(zero_sharded_norm, group=gpc.get_group(ParallelMode.DATA))
no_tensor_parallel_norm += zero_sharded_norm
total_norm = tensor_parallel_norm + no_tensor_parallel_norm
if gpc.is_initialized(ParallelMode.PIPELINE) and gpc.get_world_size(ParallelMode.PIPELINE) > 1:
dist.all_reduce(total_norm, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.PIPELINE))
total_norm = total_norm ** (1.0 / norm_type)
if torch.is_tensor(total_norm):
total_norm = total_norm.item()
# Scale.
clip_coeff = max_norm / (total_norm + 1.0e-6)
if clip_coeff < 1.0:
if enable_cuda_kernels:
grads = [p.grad.detach() for p in params]
dummy_overflow_buf = torch.cuda.IntTensor([0])
multi_tensor_applier(fused_optim.multi_tensor_scale, dummy_overflow_buf, [grads, grads], clip_coeff)
else:
for p in params:
p.grad.detach().mul_(clip_coeff)
return total_norm
def count_zeros_fp32(parameters):
if isinstance(parameters, torch.Tensor):
parameters = [parameters]
# Filter parameters based on:
# - grad should not be none
# - parameter should not be shared
# - should not be a replica due to tensor model parallelism
total_num_zeros = 0.0
for param in parameters:
grad_not_none = param.grad is not None
is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
if grad_not_none and is_not_tp_duplicate:
grad = param.grad.detach()
num_zeros = grad.numel() - torch.count_nonzero(grad)
total_num_zeros = num_zeros + total_num_zeros
total_num_zeros = torch.IntTensor([int(total_num_zeros)]).cuda()
# Sum across all model-parallel GPUs.
ops = []
ops.append(
dist.all_reduce(total_num_zeros, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.TENSOR), async_op=True)
)
if gpc.is_initialized(ParallelMode.PIPELINE):
ops.append(
dist.all_reduce(
total_num_zeros, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.PIPELINE), async_op=True
)
)
for req in ops:
req.wait()
total_num_zeros = total_num_zeros.item()
return total_num_zeros
def copy_tensor_parallel_attributes(src_tensor, dst_tensor):
for attr in TENSOR_PARALLEL_ATTRIBUTES:
if hasattr(src_tensor, attr):
val = getattr(src_tensor, attr)
setattr(dst_tensor, attr, val)
def param_is_not_tensor_parallel_duplicate(param):
return (hasattr(param, IS_TENSOR_PARALLEL) and getattr(param, IS_TENSOR_PARALLEL)) or (
gpc.get_local_rank(ParallelMode.TENSOR) == 0
)
@contextmanager
def switch_virtual_pipeline_parallel_rank(rank):
prev_rank = gpc.virtual_pipeline_parallel_rank
try:
gpc.set_virtual_pipeline_parallel_rank(rank)
yield
finally:
gpc.set_virtual_pipeline_parallel_rank(prev_rank)