Resolved ALIBI bias regression due to porting flat PA #34
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
tannervoas742
changed the title
tannervoas742
force-pushed
the
restore_alibi_for_flat_pa_final
branch
3 times, most recently
from
873fc19 to
1a652d3
Compare
|
@madamczykhabana, @libinta please help review. |
tannervoas742
force-pushed
the
restore_alibi_for_flat_pa_final
branch
4 times, most recently
from
67b3396 to
c0fb257
Compare
tannervoas742
force-pushed
the
restore_alibi_for_flat_pa_final
branch
3 times, most recently
from
d97b251 to
288b7c7
Compare
Changes:
- Added back alibi biases to decode stage.
- Optimized ALiBI memory usage.
- Added environment variable "VLLM_PROMPT_ALIBI_MAX_SEQ_LEN" to allow
large models to run with restricted prompt lengths.
- Prompt biases instantiated once in __init__ rather than each
forward.
- Prompt and decode biases are shared across encoder/decoder layers.
- Added environment variable "VLLM_ALIBI_USE_FLOAT32_BIASES" to resolve
accuracy issue on long sequences.
- Updated jais, mpt, falcon, baichuan, and bloom to work with ALiBI.
- Due to bloom's 176B parameter size I was unable to test this model.
Its changes are the simplest though.
- Works in lazy and eager mode.
- ALiBI is restricted to "VLLM_PROMPT_USE_FUSEDSDPA=false", and
"VLLM_CONTIGUOUS_PA=true".
- Add position offsets to improve quality on BS > 1 with sequences of
varying length.
- BS > 1 may have accuracy issues if on FW < 1.19.0. This is due to
limitation in softmax. Resolved on FW >= 1.19.0.
- NTT patch for GQA
Co-authored-by: Tanner Voas <[email protected]>
Co-authored-by: Haihao Xiang <[email protected]>
Signed-off-by: Tanner Voas <[email protected]>
tannervoas742
force-pushed
the
restore_alibi_for_flat_pa_final
branch
from
288b7c7 to
de937c2
Compare
itaraban
approved these changes
Dec 12, 2024
michalkuligowski
added a commit
that referenced
this pull request
Dec 13, 2024
This reverts commit 0766759.
michalkuligowski
added a commit
that referenced
this pull request
Dec 13, 2024
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.
Required for associated changes on vllm-fork PR
See the associated ticket above for further details.