Megatron-DeepSpeed
[WIP] Hack my way to get OPT running
#301
Open
Overview
Commits
25
Changes
View On GitHub

[WIP] Hack my way to get OPT running #301

thomasw21 wants to merge 25 commits into main from thomas/opt
thomasw21
thomasw21 That's what I'm hoping is all I need
ce8d5e11
thomasw21 relu to work
b21d773e
thomasw21 Run inference for opt
43a9ab44
thomasw21 Test
8e095da2
thomasw21 I think OPT doesn't have shared vocabulary embedding
e742a3fe
thomasw21 Revert "I think OPT doesn't have shared vocabulary embedding"
55d1a4bf
thomasw21 Maybe this is what's wrong
b8090b76
thomasw21 Update model loading
dff9766f
thomasw21 WTF
01e269d8
thomasw21 Woops
17a075d2
thomasw21 Position offset
d45af145
thomasw21 Add bs eval harness code
73a79476
thomasw21 Missing import
f34841cd
thomasw21 Update position ids
1423bf9f
thomasw21 That pathing breaks everything
fae49cab
thomasw21 add offloading possitility
6a326afa
thomasw21 Black list all tasks that require greedy_until
b2caf4dd
thomasw21 Woops
48892b93
thomasw21 Disable CarbonEmissionTracker
1476691b
thomasw21 Enable CarbonEmissionTracker
5868d295
thomasw21 Woops
94ef27ad
thomasw21 Be very careful of random states
f713ff01
thomasw21 Woops
fab9fb96
thomasw21 Woops
23d3f4ca
thomasw21 Remove codecarbon
c6f9c021

Login to write a write a comment.

Login via GitHub

Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Terms Of Use Privacy Policy Refund Policy Impressum