support KV-Compress paged KV cache #27
Open
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
IsaacRe
force-pushed
the
irehg/vllm-kvc
branch
from
3aa31f1
to
ef92d5f
Compare
bgamari
reviewed
Nov 28, 2024
csrc/flash_attn/flash_api.cpp
Outdated
| @@ -159,6 +168,7 @@ void run_mha_fwd(Flash_fwd_params ¶ms, cudaStream_t stream, bool force_split | |||
| HEADDIM_SWITCH(params.d, [&] { | |||
| BOOL_SWITCH(params.is_causal, Is_causal, [&] { | |||
| if (params.num_splits <= 1 && !force_split_kernel) { // If we don't set it num_splits == 0 | |||
| assert(false); | |||
There was a problem hiding this comment.
What is going on here? Perhaps you mean to also remove the call that follows?
There was a problem hiding this comment.
Good catch. That was just to make sure it was launching the right kernel when I was testing. It should be removed
|
@IsaacRe Amazing! Excited to see this work 🚀 BTW, are you in the vLLM slack workspace? |
Signed-off-by: Isaac Rehg <[email protected]>
IsaacRe
force-pushed
the
irehg/vllm-kvc
branch
from
ef92d5f
to
28f6637
Compare
Thanks! Yes I am. I'm wrapping up chunked-prefill compat and will update in the channel when benchmarks are done |
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Outdated suggestions cannot be applied.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.
This PR adds support for paged KV cache following the structure of KV-Compress, where cache blocks are paged out on a per-head basis.
In this case the cache shape becomes [num_blocks, block_size, d] (rather than [num_blocks, block_size, num_heads, d]) and the block table shape becomes [num_seqs, num_heads, max_blocks_per_seq] (rather than [num_seqs, max_blocks_per_seq]). Sequence lengths also need to be specified per-head, so that a (num_seqs * num_heads) or (num_seqs * num_heads + 1)-sized tensor is provided (depending on whether the sequence length or cumulative offsets are used).
I configured it to use the dimensionality of the block tables tensor to detect whether a KV-Compress cache is being used, assuming KV-Compress when dim > 2 and following the existing logic otherwise.