[Kernel][Attention] Separate Attention.kv_scale into k_scale and v_scale
#6081
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
comaniac
reviewed
Jul 3, 2024
mgoin
changed the title
Attention.kv_scale into key_scale and value_scaleAttention.kv_scale into k_scale and v_scale
comaniac
reviewed
Jul 16, 2024
dtrifiro
pushed a commit
to opendatahub-io/vllm
that referenced
this pull request
Jul 17, 2024
fialhocoelho
pushed a commit
to opendatahub-io/vllm
that referenced
this pull request
Jul 19, 2024
xjpang
pushed a commit
to xjpang/vllm
that referenced
this pull request
Jul 24, 2024
Alvant
pushed a commit
to compressa-ai/vllm
that referenced
this pull request
Oct 26, 2024
…`v_scale` (vllm-project#6081) Signed-off-by: Alvant <[email protected]>
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Outdated suggestions cannot be applied.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.
Since we already quantize
key_cacheandvalue_cacheseparately in PagedAttention, there is "free accuracy on the table" for FP8 KV Cache quantization as we could use separate per-tensor scales for each.The FlashInfer FP8 attention kernel also uses separate
k_scaleandv_scalevalues, so this PR is in preparation to enable that usage. Source: https://github.com/flashinfer-ai/flashinfer/blob/dc2c76f8577d8695112b61d1fd43ef88569272ef/python/flashinfer/decode.py#L98-L101This PR will maintain backwards compatibility with FP8 model checkpoints that currently use
kv_scaleand duplicates that scale for both key+value if that is all that is available. However if a checkpoint hask_scaleandv_scalepresent on the attention module, we will prefer those values.