github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-17 23:18:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

[npu] change device to accelerator api (#5239)

Browse Source 
* update accelerator

* fix timer

* fix amp

* update

* fix

* update bug

* add error raise

* fix autocast

* fix set device

* remove doc accelerator

* update doc

* update doc

* update doc

* use nullcontext

* update cpu

* update null context

* change time limit for example

* udpate

* update

* update

* update

* [npu] polish accelerator code

---------

Co-authored-by: Xuanlei Zhao <xuanlei.zhao@gmail.com>
Co-authored-by: zxl <43881818+oahzxl@users.noreply.github.com>
This commit is contained in:
Hongxin Liu
2024-01-09 10:20:05 +08:00
committed by GitHub
parent dd2c28a323
commit d202cc28c0
128 changed files with 1773 additions and 868 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
colossalai/legacy/utils/activation_checkpoint.py
View File

@@ -6,8 +6,8 @@ import weakref
import torch
from torch.utils.checkpoint import check_backward_validity, detach_variable
from colossalai.accelerator import get_accelerator
from colossalai.legacy.context.random import get_current_mode, get_states, set_mode, set_seed_states, sync_states
from colossalai.utils.device import autocast, get_current_device
def copy_to_device(obj, device):
@@ -33,7 +33,7 @@ class CheckpointFunction(torch.autograd.Function):
check_backward_validity(args)
ctx.run_function = run_function
ctx.activation_offload = activation_offload
ctx.device = get_current_device()
ctx.device = get_accelerator().get_current_device()
# preserve rng states
ctx.fwd_cpu_rng_state = torch.get_rng_state()
@@ -110,7 +110,7 @@ class CheckpointFunction(torch.autograd.Function):
inputs[idx] = tensors[i]
detached_inputs = detach_variable(tuple(inputs))
if ctx.had_autocast_in_fwd:
with torch.enable_grad(), autocast():
with torch.enable_grad(), get_accelerator().autocast()():
outputs = ctx.run_function(*detached_inputs)
else:
with torch.enable_grad():
@@ -226,7 +226,7 @@ def _checkpoint_without_reentrant(function, activation_offload=False, *args):
# rerun forward, the inner_pack will store all the activations in storage
if has_autocast_in_fwd:
with torch.enable_grad(), autocast(), torch.autograd.graph.saved_tensors_hooks(
with torch.enable_grad(), get_accelerator().autocast()(), torch.autograd.graph.saved_tensors_hooks(
inner_pack, inner_unpack
):
_unused = function(*args)
@@ -245,7 +245,7 @@ def _checkpoint_without_reentrant(function, activation_offload=False, *args):
# get device if we need to offload the activation
if activation_offload:
device = get_current_device()
device = get_accelerator().get_current_device()
# run function with pack and unpack as saved_tensors_hooks
with torch.autograd.graph.saved_tensors_hooks(pack, unpack):

9
colossalai/legacy/utils/memory.py
View File

@@ -6,9 +6,9 @@ import torch
import torch.distributed as dist
from packaging import version
from colossalai.accelerator import get_accelerator
from colossalai.legacy.core import global_context as gpc
from colossalai.logging import get_dist_logger
from colossalai.utils import get_current_device
_GLOBAL_CUDA_MEM_FRACTION = 1.0
_GLOBAL_CPU_MEM_CAPACITY = -1
@@ -112,7 +112,10 @@ def colo_device_memory_capacity(device: torch.device) -> int:
# In the context of 1-CPU-N-GPU, the memory capacity of the current process is 1/N overall CPU memory.
return colo_get_cpu_memory_capacity() / gpc.num_processes_on_current_node
if device.type == "cuda":
return torch.cuda.get_device_properties(get_current_device()).total_memory * _GLOBAL_CUDA_MEM_FRACTION
return (
torch.cuda.get_device_properties(get_accelerator().get_current_device()).total_memory
* _GLOBAL_CUDA_MEM_FRACTION
)
def colo_device_memory_used(device: torch.device) -> int:
@@ -153,7 +156,7 @@ def colo_set_process_memory_fraction(ratio: float) -> None:
return
global _GLOBAL_CUDA_MEM_FRACTION
_GLOBAL_CUDA_MEM_FRACTION = ratio
torch.cuda.set_per_process_memory_fraction(_GLOBAL_CUDA_MEM_FRACTION, get_current_device())
torch.cuda.set_per_process_memory_fraction(_GLOBAL_CUDA_MEM_FRACTION, get_accelerator().get_current_device())
def colo_set_cpu_memory_capacity(size: int) -> None:

4
colossalai/legacy/utils/profiler/legacy/comm_profiler.py
View File

@@ -8,7 +8,7 @@ import torch.distributed as dist
from torch.autograd.profiler import profile
from torch.distributed import ReduceOp
from colossalai.utils import get_current_device
from colossalai.accelerator import get_accelerator
from .prof_utils import BaseProfiler, _format_bandwidth, _format_memory, _format_time
@@ -177,7 +177,7 @@ class CommProfiler(BaseProfiler):
assert current_comm_event is not None, "dist op has not been found"
buffer = torch.tensor([current_comm_event.self_cuda_time], device=get_current_device())
buffer = torch.tensor([current_comm_event.self_cuda_time], device=get_accelerator().get_current_device())
torch_all_reduce(buffer, op=ReduceOp.MIN, group=group)
current_comm_event.self_cuda_time = buffer.item()