Support PP->TP resharding for token-major KV cache layout - #1139
Merged
Merged
Conversation
copybara-service
Bot
force-pushed
the
test_991894537
branch
from
October 1, 2026 22:54
d9837e0 to
83ea841
Compare
copybara-service
Bot
changed the title
copybara-service
Bot
force-pushed
the
test_991894537
branch
from
October 2, 2026 04:51
83ea841 to
0521920
Compare
Currently, the pool reshard planner takes only plain contiguous byte ranges. However, with a token-major KV page (tokens, then heads), a tensor-parallel decode's share of a pipeline-parallel prefill's page is one head slice per token, which is a strided range, so PP->TP transfers are rejected at planning. This PR fixes that by: accepting a strided span that lies inside one destination page; sending it as one strided entry per destination rank instead of one entry per token, when both pools store a block as one contiguous region; raising the receiver's repeat bound from 2^20 to 2^24, enough for a 32K-token request. Used by sgl-project/sglang-torchtpu#748 (PP8 prefill to TP8 decode). PiperOrigin-RevId: 992103162
copybara-service
Bot
force-pushed
the
test_991894537
branch
from
October 2, 2026 05:03
0521920 to
120bab9
Compare
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Support PP->TP resharding for token-major KV cache layout
Currently, the pool reshard planner takes only plain contiguous byte ranges. However, with a token-major KV page (tokens, then heads), a tensor-parallel decode's share of a pipeline-parallel prefill's page is one head slice per token, which is a strided range, so PP->TP transfers are rejected at planning.
This PR fixes that by:
accepting a strided span that lies inside one destination page;
sending it as one strided entry per destination rank instead of one entry per token, when both pools store a block as one contiguous region;
raising the receiver's repeat bound from 2^20 to 2^24, enough for a 32K-token request.
Used by sgl-project/sglang-torchtpu#748 (PP8 prefill to TP8 decode).