Fix RLHF slowdown in attention multi steps extend_step. #849
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jax.lax.dynamic_update_slice_in_dim is generally faster than advanced indexing, but an unusual slowdown was observed, with RLHF sampling taking up to 3 hours per run. Investigate and fix it. https://a1350286.slack.com/archives/C03HJAYC7JA/p1731998432387409?thread_ts=1731968765.840839&cid=C03HJAYC7JA For your information, in https://github.pie.apple.com/foundation-models/axlearn/pull/894, I experimented with both dynamic_update_slice and advanced indexing on TPUv4 and chose the faster option. It's also known that dynamic_update_slice performs better when copying contiguous memory. This is a very surprising case. Advanced Indexing ---------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------- QkvLinearExtendStepBenchmark/2048/16/1024/1 7.16 ms 0.623 ms 492 QkvLinearExtendStepBenchmark/2048/16/4096/1 8.52 ms 0.624 ms 561 QkvLinearExtendStepBenchmark/2048/16/32768/1 34.6 ms 1.64 ms 78 QkvLinearExtendStepBenchmark/2048/16/4096/8 63.6 ms 1.74 ms 81 QkvLinearExtendStepBenchmark/2048/16/4096/64 276 ms 2.40 ms 81 QkvLinearExtendStepBenchmark/2048/16/4096/512 2541 ms 81.6 ms 1 dynamic_update_slice ---------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------- QkvLinearExtendStepBenchmark/2048/16/1024/1 1.70 ms 0.513 ms 1125 QkvLinearExtendStepBenchmark/2048/16/4096/1 3.40 ms 0.519 ms 1174 QkvLinearExtendStepBenchmark/2048/16/32768/1 20.1 ms 0.930 ms 404 QkvLinearExtendStepBenchmark/2048/16/4096/8 3.68 ms 0.524 ms 1139 QkvLinearExtendStepBenchmark/2048/16/4096/64 3.74 ms 0.532 ms 1125 QkvLinearExtendStepBenchmark/2048/16/4096/512 2530 ms 80.4 ms 1
ruomingp
approved these changes
Nov 19, 2024
Yes :) More details in 918 |
ds-hwang
added a commit
to ds-hwang/axlearn
that referenced
this pull request
Dec 2, 2024
`k_proj` is not properly set sharding hints, so QKVLinear.extend_step cannot
create next `cached_key` with proper hints.
This causes OOM for diffusion model, because the code cannot know the local
batch size.
Shape: f32[1024,2048,8,128]{3,2,1,0:T(8,128)}
Unpadded size: 8.00G
To fix it, copy `cached_key.sharding` to `k_proj.sharding`, as `cached_key`
sharding is properly set up.
In addition, this is the reason of RLHF slowdown, so revert the workaround change.
apple#849
qdavid1
pushed a commit
to qdavid1/axlearn
that referenced
this pull request
Dec 11, 2024
jax.lax.dynamic_update_slice_in_dim is generally faster than advanced indexing, but an unusual slowdown was observed, with RLHF sampling taking up to 3 hours per run. Investigate and fix it. https://a1350286.slack.com/archives/C03HJAYC7JA/p1731998432387409?thread_ts=1731968765.840839&cid=C03HJAYC7JA For your information, in https://github.pie.apple.com/foundation-models/axlearn/pull/894, I experimented with both dynamic_update_slice and advanced indexing on TPUv4 and chose the faster option. It's also known that dynamic_update_slice performs better when copying contiguous memory. This is a very surprising case. Advanced Indexing ---------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------- QkvLinearExtendStepBenchmark/2048/16/1024/1 7.16 ms 0.623 ms 492 QkvLinearExtendStepBenchmark/2048/16/4096/1 8.52 ms 0.624 ms 561 QkvLinearExtendStepBenchmark/2048/16/32768/1 34.6 ms 1.64 ms 78 QkvLinearExtendStepBenchmark/2048/16/4096/8 63.6 ms 1.74 ms 81 QkvLinearExtendStepBenchmark/2048/16/4096/64 276 ms 2.40 ms 81 QkvLinearExtendStepBenchmark/2048/16/4096/512 2541 ms 81.6 ms 1 dynamic_update_slice ---------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------- QkvLinearExtendStepBenchmark/2048/16/1024/1 1.70 ms 0.513 ms 1125 QkvLinearExtendStepBenchmark/2048/16/4096/1 3.40 ms 0.519 ms 1174 QkvLinearExtendStepBenchmark/2048/16/32768/1 20.1 ms 0.930 ms 404 QkvLinearExtendStepBenchmark/2048/16/4096/8 3.68 ms 0.524 ms 1139 QkvLinearExtendStepBenchmark/2048/16/4096/64 3.74 ms 0.532 ms 1125 QkvLinearExtendStepBenchmark/2048/16/4096/512 2530 ms 80.4 ms 1
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fix RLHF slowdown in attention multi steps extend_step.
jax.lax.dynamic_update_slice_in_dim is generally faster than advanced indexing,
but an unusual slowdown was observed, with RLHF sampling taking up to 3 hours
per run. TODO: Investigate and fix it.
For your information, in #831, I experimented
with both dynamic_update_slice and advanced indexing on TPUv4 and chose the
faster option. It's also known that dynamic_update_slice performs better when
copying contiguous memory. This is a very surprising case.
Advanced Indexing
dynamic_update_slice