Ma, Guokai
|
c08e69f212
Make op builder detection adapt to accelerator change (#5206)
|
7 月之前 |
ByronHsu
|
4578c2490b
[zero++] Synchronize at the end of secondary partitioning and simplify the logic (#5216)
|
7 月之前 |
Masahiro Tanaka
|
f295aea09e
Stop tracking backward chain of broadcast (ZeRO3) (#5113)
|
8 月之前 |
Masahiro Tanaka
|
c3cfe96bb3
Enable torch.compile with ZeRO (Experimental) (#4878)
|
8 月之前 |
ByronHsu
|
e81369318e
[minor] improve code quality and readablilty (#5011)
|
9 月之前 |
Heyang Qin
|
75ed63c94f
Enable hpz based on secondary tensor presence (#4906)
|
9 月之前 |
inkcherry
|
3110c38852
params partition for skip_init (#4722)
|
9 月之前 |
Max Kovalenko
|
81cc32075c
Partition parameters: Minor refactoring of use_secondary_tensor condition (#4868)
|
9 月之前 |
taozhiwei
|
fd0a52c1ac
use all_gather_into_tensor instead of all_gather (#4705)
|
10 月之前 |
Masahiro Tanaka
|
b8e1664232
Enable ZeRO3 allgather for multiple dtypes (#4647)
|
11 月之前 |
Abhishek Jindal
|
e339364127
Add torch no grad condition (#4391)
|
1 年之前 |
Ziyang
|
60bf78454c
Fix incorrect assignment of self.quantized_nontrainable_weights (#4399)
|
1 年之前 |
Olatunji Ruwase
|
aa4a7401f8
ZeRO-Inference refresh (#4197)
|
1 年之前 |
Heyang Qin
|
1f0a44d934
Keep hpz secondary tensor in forward pass (#4288)
|
1 年之前 |
Joe Mayer
|
57d629a17e
Empty tensor size check (#4186)
|
1 年之前 |
Sam Ade Jacobs
|
a855405e0b
DeepSpeed Ulysses release (#4198)
|
1 年之前 |
Xuehai Pan
|
426810a254
Fix ZeRO parameter initialization for tensors with `requires_grad=True` (#4138)
|
1 年之前 |
Heyang Qin
|
7711bdbbd2
MP ZeRO++ (#3954)
|
1 年之前 |
Olatunji Ruwase
|
7f90ef4bdd
Multiple zero stage 3 related fixes (#3886)
|
1 年之前 |
Ma, Guokai
|
0f5406323c
[CPU] FusedAdam and CPU training support (#3991)
|
1 年之前 |
digger yu
|
fc8de76f1d
Simplify chain comparisons, remove redundant parentheses (#3912)
|
1 年之前 |
hipudding
|
7528035c1e
Use device_name instead of device index to support other device (#3933)
|
1 年之前 |
Heyang Qin
|
e59f69a8ff
remove the call to param.ds_tensor from print (#3928)
|
1 年之前 |
hipudding
|
e292343d7b
Del comment deepspeed.zero.Init() can be used as a decorator (#3894)
|
1 年之前 |
Heyang Qin
|
f8551b439e
Fix racing condition in GatheredParameters (#3819)
|
1 年之前 |
Masahiro Tanaka
|
203ac9d7ac
support model declaration in zero.Init context (#3592)
|
1 年之前 |
Heyang Qin
|
d18aa2c79c
ZeRO++ (#3784)
|
1 年之前 |
Olatunji Ruwase
|
046afcedb4
Increase tensor creator coverage (#3684)
|
1 年之前 |
hablb
|
0977106ac9
zero3 performance optimizations (#3622)
|
1 年之前 |
digger yu
|
5d14afd26c
fix typo deepspeed/runtime (#3663)
|
1 年之前 |