[KVCache] Per Layer Sliding Window #17928
Merged
Conversation
This file contains hidden or 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
Update main
joshua-j-hong
changed the title
joshua-j-hong
force-pushed
the
jjhong_KV_alt_sliding_window
branch
from
3fb27bb to
936d500
Compare
|
With some further testing and investigation, there is an additional MLC-LLM/TVM bug related to excessive prefilling (even without the per-layer sliding window changes outlined here) that may be causing inference slowdown. |
MasterJH5574
approved these changes
Jun 11, 2025
|
Just see some conflicts with upstream. Likely we need to do a rebase. Related changes are the recent FFI refactor and a namespace rename from |
|
Conflicts and tests are fixed. Current plans are for a future change that will add optional parameters to the KVCache, some of which will be for layer sliding window. This will ensure that no hardcoding of values will be needed and the KVCache will be backwards compatible. |
MasterJH5574
approved these changes
Jun 23, 2025
ShiboXing
pushed a commit
to ShiboXing/tvm
that referenced
this pull request
Aug 10, 2025
Adds per layer sliding window functionality to the KV Cache. Correctness is mostly achieved, but there are some cases where single tokens are strange. The corresponding MLC-LLM PR is mlc-ai/mlc-llm#3248 A full list of changes and additions are below - Add a new attention type for per-layer sliding window called `MHA_SLIDING` - Add corresponding vectors for per-layer sliding window offset calculations - For sliding window attention enabled KV-cache, regular sliding window is disabled to prevent page eviction - Gemma3 has different rope parameters for local sliding window layers. This should be passed as a parameter for the KVCache, but currently these values are hardcoded
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.
Adds per layer sliding window functionality to the KV Cache. Correctness is mostly achieved, but there are some cases where single tokens are strange. The corresponding MLC-LLM PR is mlc-ai/mlc-llm#3248
A full list of changes and additions are below
MHA_SLIDING