Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[hotfix] doubled bias in FusedMLP #317

Merged
merged 1 commit into from
Jun 1, 2022
Merged

[hotfix] doubled bias in FusedMLP #317

merged 1 commit into from
Jun 1, 2022

Conversation

blefaudeux
Copy link
Contributor

@blefaudeux blefaudeux commented May 31, 2022 •
edited
Loading

What does this PR do?

Fixes the FusedMLP block having twice the bias layers, found randomly when working on the weight inits (#312)

Before submitting

  • Did you have fun?
    • Make sure you had fun coding 🙃
  • Did you read the contributor guideline?
  • Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
    • N/A
  • Did you make sure to update the docs?
    • N/A
  • Did you write any new necessary tests?
    • N/A
  • Did you update the changelog? (if needed)
    • N/A

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label May 31, 2022
@blefaudeux
Copy link
Contributor Author

Perf numbers without the doubled bias:

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.0 drop - fw 0.19 1.26 1.27 5.97
fused - gelu - no bias - 0.0 drop - fw 0.19 1.26 1.27 5.52

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.0 drop - fw 0.36 2.56 2.54 10.13
fused - gelu - no bias - 0.0 drop - fw 0.36 2.54 2.54 10.14

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.1 drop - fw 0.24 1.44 1.45 5.98
fused - gelu - no bias - 0.1 drop - fw 0.21 1.36 1.36 5.53

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.1 drop - fw 0.44 2.86 2.85 10.63
fused - gelu - no bias - 0.1 drop - fw 0.39 2.66 2.64 10.60

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.0 drop - fw 0.23 1.43 1.42 5.88
fused - gelu - bias - 0.0 drop - fw 0.23 1.43 1.43 5.92

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.0 drop - fw 0.43 2.83 2.83 10.76
fused - gelu - bias - 0.0 drop - fw 0.43 2.83 2.83 10.97

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.1 drop - fw 0.27 1.59 1.59 5.95
fused - gelu - bias - 0.1 drop - fw 0.22 1.36 1.36 5.54

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.1 drop - fw 0.50 3.10 3.12 11.60
fused - gelu - bias - 0.1 drop - fw 0.40 2.65 2.64 10.37

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.0 drop - fw+bw 0.60 4.26 4.27 16.46
fused - gelu - no bias - 0.0 drop - fw+bw 0.60 4.71 4.26 16.28

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.0 drop - fw+bw 1.14 9.02 8.45 31.93
fused - gelu - no bias - 0.0 drop - fw+bw 1.15 8.88 8.37 34.53

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.1 drop - fw+bw 0.67 4.56 4.58 16.84
fused - gelu - no bias - 0.1 drop - fw+bw 0.66 4.37 4.69 19.42

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - no bias - 0.1 drop - fw+bw 1.29 9.11 9.12 32.54
fused - gelu - no bias - 0.1 drop - fw+bw 1.18 8.49 8.69 34.63

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.0 drop - fw+bw 0.67 4.48 4.71 17.17
fused - gelu - bias - 0.0 drop - fw+bw 0.79 4.62 4.62 16.83

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.0 drop - fw+bw 1.24 9.81 8.81 32.42
fused - gelu - bias - 0.0 drop - fw+bw 1.35 9.20 9.01 32.51

--- Type: torch.float16 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.1 drop - fw+bw 0.75 5.26 5.16 17.36
fused - gelu - bias - 0.1 drop - fw+bw 0.96 4.55 4.57 16.64

--- Type: torch.float32 ---

Units: runtime in ms, lower is better. BMK - mul: 8 x 256 x 512 - 4 8 x 512 x 1024 - 4 4 x 1024 x 1024 - 4 2 x 2048 x 2048 - 4
standard - gelu - bias - 0.1 drop - fw+bw 1.40 9.47 9.72 33.54
fused - gelu - bias - 0.1 drop - fw+bw 1.30 8.56 8.56 32.66

If/when I get the time to revisit the fused linear this could probably be improved

blefaudeux
blefaudeux commented Jun 1, 2022
View reviewed changes
xformers/components/feedforward/fused_mlp.py
@@ -46,16 +46,26 @@ def __init__(
dim_mlp = hidden_layer_multiplier * dim_model

self.mlp = nn.Sequential(
nn.Linear(in_features=dim_model, out_features=dim_mlp, bias=bias),
nn.Linear(
in_features=dim_model, out_features=dim_mlp, bias=False
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

the gist of it, this was a typo, the bias is handled in the next layer already

@blefaudeux
Copy link
Contributor Author

insta-land @danthe3rd / @fmassa / @dianaml0 , I hope that's ok, semi-obvious typo

blefaudeux
blefaudeux commented Jun 1, 2022
View reviewed changes
xformers/benchmarks/benchmark_mlp.py
@@ -88,7 +88,7 @@ def mlp_fused():
),
]:
time = triton.testing.do_bench(testcase.function)[0]
key = f"B={B}, M={M}, K={K}, HLM={hlm}"
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

minor presentation changes

blefaudeux
blefaudeux commented Jun 1, 2022
View reviewed changes
xformers/benchmarks/benchmark_mlp.py
@@ -19,8 +19,8 @@
(8, 512, 1024),
(4, 1024, 1024),
(2, 2048, 2048),
(1, 2048, 12288),
(2, 4096, 4096),
(1, 2048, 4096),
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

trying to make this fit on a smaller GPU

@blefaudeux blefaudeux changed the title [hotfix] dual bias in FusedMLP [hotfix] doubled bias in FusedMLP Jun 1, 2022
@blefaudeux blefaudeux requested a review from danthe3rd June 1, 2022 00:06
@blefaudeux blefaudeux merged commit cbf4526 into main Jun 1, 2022
@blefaudeux blefaudeux deleted the hotfix_fused_mlp branch June 3, 2022 03:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@dianaml0 dianaml0 Awaiting requested review from dianaml0

@fmassa fmassa Awaiting requested review from fmassa

@danthe3rd danthe3rd Awaiting requested review from danthe3rd

Assignees
No one assigned
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@blefaudeux @facebook-github-bot