github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-06 19:40:28 +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

39
colossalai/legacy/nn/layer/parallel_2p5d/layers.py
View File

@@ -8,6 +8,7 @@ import torch.nn.functional as F
from torch import Tensor
from torch.nn import Parameter
from colossalai.accelerator import get_accelerator
from colossalai.legacy.communication import broadcast
from colossalai.legacy.context import ParallelMode, seed
from colossalai.legacy.core import global_context as gpc
@@ -19,7 +20,6 @@ from colossalai.legacy.utils.checkpointing import (
partition_tensor_parallel_state_dict,
)
from colossalai.nn import init as init
from colossalai.utils.device import get_current_device
from ..base_layer import ParallelLayer
from ..utils import divide, set_tensor_parallel_attribute_by_partition, to_2tuple
@@ -84,7 +84,7 @@ class Linear2p5D(ParallelLayer):
self.hidden_size_per_partition = divide(out_features, self.tesseract_dim)
# create weight, shape: [k/q, h/q]
factory_kwargs = {"device": get_current_device(), "dtype": dtype}
factory_kwargs = {"device": get_accelerator().get_current_device(), "dtype": dtype}
self.weight = Parameter(
torch.empty(self.input_size_per_partition, self.hidden_size_per_partition, **factory_kwargs)
)
@@ -272,7 +272,7 @@ class LayerNorm2p5D(ParallelLayer):
self.partitioned_partition = divide(normalized_shape, self.tesseract_dim) # *
# create parameters
factory_kwargs = {"device": get_current_device(), "dtype": dtype}
factory_kwargs = {"device": get_accelerator().get_current_device(), "dtype": dtype}
self.weight = Parameter(torch.ones(self.partitioned_partition, **factory_kwargs))
if bias:
@@ -451,18 +451,24 @@ class PatchEmbedding2p5D(ParallelLayer):
self.weight = Parameter(
torch.empty(
(self.embed_size_per_partition, in_chans, *self.patch_size),
device=get_current_device(),
device=get_accelerator().get_current_device(),
dtype=dtype,
)
)
self.bias = Parameter(torch.empty(self.embed_size_per_partition, device=get_current_device(), dtype=dtype))
self.bias = Parameter(
torch.empty(self.embed_size_per_partition, device=get_accelerator().get_current_device(), dtype=dtype)
)
self.cls_token = Parameter(
torch.zeros((1, 1, self.embed_size_per_partition), device=get_current_device(), dtype=dtype)
torch.zeros(
(1, 1, self.embed_size_per_partition), device=get_accelerator().get_current_device(), dtype=dtype
)
)
self.pos_embed = Parameter(
torch.zeros(
(1, self.num_patches + 1, self.embed_size_per_partition), device=get_current_device(), dtype=dtype
(1, self.num_patches + 1, self.embed_size_per_partition),
device=get_accelerator().get_current_device(),
dtype=dtype,
)
)
@@ -632,7 +638,9 @@ class Embedding2p5D(ParallelLayer):
self.embed_kwargs = kwargs
self.weight = Parameter(
torch.empty((num_embeddings, embed_dim_per_partition), device=get_current_device(), dtype=dtype)
torch.empty(
(num_embeddings, embed_dim_per_partition), device=get_accelerator().get_current_device(), dtype=dtype
)
)
self.reset_parameters(weight_initializer)
@@ -772,7 +780,7 @@ class VocabParallelEmbedding2p5D(ParallelLayer):
self.weight = Parameter(
torch.empty(
(self.num_embeddings_per_partition, self.embed_dim_per_partition),
device=get_current_device(),
device=get_accelerator().get_current_device(),
dtype=dtype,
)
)
@@ -910,11 +918,18 @@ class Classifier2p5D(ParallelLayer):
self.has_weight = False
else:
self.weight = Parameter(
torch.empty(self.num_classes, self.input_size_per_partition, device=get_current_device(), dtype=dtype)
torch.empty(
self.num_classes,
self.input_size_per_partition,
device=get_accelerator().get_current_device(),
dtype=dtype,
)
)
self.has_weight = True
if bias:
self.bias = Parameter(torch.zeros(self.num_classes, device=get_current_device(), dtype=dtype))
self.bias = Parameter(
torch.zeros(self.num_classes, device=get_accelerator().get_current_device(), dtype=dtype)
)
else:
self.bias = None
@@ -1068,7 +1083,7 @@ class VocabParallelClassifier2p5D(ParallelLayer):
self.hidden_size_per_partition = divide(num_classes, self.tesseract_dim)
# create weight, shape: [k/q, h/q]
factory_kwargs = {"device": get_current_device(), "dtype": dtype}
factory_kwargs = {"device": get_accelerator().get_current_device(), "dtype": dtype}
if weight is not None:
self.weight = weight
self.has_weight = False