bf16+pipeline parallelism #1801
bf16 updates
fb0dc00f
Got bf16 working
6eb4f1fa
fp32 reduction; flattened tensors
a3d3576e
bf16+zero_stage_1 first cut
6f5ebc37
finish zero_stage 1 sharding
819abe2a
Matching fp16 with debugging codes
e48035b7
Matching loss with fp16
82450539
Fix gradient clipping
15293139
bf16 gradient clipping fix
27e5b956
tjruwase
force pushed
from
ed26ef43
to
27e5b956
3 years ago
Unscale grad norm
f4977024
Fix grad norm scaling
0ad7c7d3
Enable loading fp16_zero_1 into bf16_zero_1 engine and vice versa
b81d862f
Fix clip_grad key error
35ea3808
Reduce tied weight gradients
37011a92
Rebase with master
8fbd4bfd
Fix grad norm for moe
61d51fd6
Merge branch 'master' into olruwase/bf16-updates
3ee61cdb
Merge branch 'master' into olruwase/bf16-updates
46cc2ce3
Reduce specified gradients
de3616ca
Merge branch 'olruwase/reduce_specified_gradients' of github.com:micr…
89e054d8
Use O(n) instead of O(n^2)
ab61edb0
Remove optimizer restriction for bf16
b7d64fd7
Link bf16 & fp32 params
19198688
Clip gradients of last stage tied weights
77b649d1
Merge branch 'master' into olruwase/bf16-updates
4a505ecd
Merge branch 'master' into olruwase/bf16-updates
ff99cb25
jeffra
commented
on 2022-03-15
Merge branch 'master' into olruwase/bf16-updates
20fdba35
Merge branch 'master' into olruwase/bf16-updates
86fa437d
Merge branch 'master' into olruwase/bf16-updates
71499a8a
Merge branch 'master' into olruwase/bf16-updates
7e7fa60b
Simplify tied weights reduction logic
2aa612a6
Merge branch 'master' into olruwase/bf16-updates
2cd21f15
Merge branch 'olruwase/bf16-updates' of github.com:microsoft/DeepSpee…
a4cbf0c1
Merge branch 'master' into olruwase/bf16-updates
67ea260f
Merge branch 'master' into olruwase/bf16-updates
6a4d6e67
Merge branch 'master' into olruwase/bf16-updates
4e1dcfd1
Also clip all tp rank parameters
e24814a1
Merge branch 'olruwase/bf16-updates' of github.com:microsoft/DeepSpee…
88cdf61c
lp to hp mapping
20697bc4
Link lp/hp/optim state; Refresh links after checkpoint load
4e8f7fff
Merge branch 'master' into olruwase/bf16-updates
52a2f109
Merge branch 'olruwase/bf16-updates' of github.com:microsoft/DeepSpee…
3ed57035
Remove debug print
5481b864
Remove debug print
d911e672
Simplify zero_grad logic
144f6527
fp32 accessors
bb70816f
Merge branch 'master' into olruwase/bf16-updates
89b4b3f1
Merge branch 'master' into olruwase/bf16-updates
a9bfaee9
Fix update bug
fa4ff11d
Merge branch 'olruwase/bf16-updates' of github.com:microsoft/DeepSpee…
cfd56385
Merge branch 'master' into olruwase/bf16-updates
5ea1c60f
Merge branch 'master' into olruwase/bf16-updates
0e2a1c50
tjruwase
merged
56c52238
into master 3 years ago
mrwyattii
deleted the olruwase/bf16-updates branch 2 years ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub