Support RoPE position info in batch prefill/decode kernels #69
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
Collaborator
|
I'll merge this into the mainline after #75 gets merged. |
MasterJH5574
force-pushed
the
qk-rope-info
branch
2 times, most recently
from
5b189f5 to
47686ef
Compare
Collaborator
|
Sorry about the new conflicts, I'll take care of them tmr. |
This PR adds q/k position information to batch prefill/decode kernels. More specifically, the kernel now accepts two additional arrays: * `q_rope_position` with shape `(total_q_len,)`, denoting the in-sequence position of each position in the input q. * `k_rope_pos_offset` with shape `(num_sequence,)`, denoting the start position of each sequence in k. These two arrays helps on-the-fly calculate RoPE in multi-level cases. Tests `test_batch_prefill` and `test_batch_decode` can pass. Performance is not validated yet. Per discussion with Zihao, this change is not very likely to incur significant perf regression.
MasterJH5574
force-pushed
the
qk-rope-info
branch
from
47686ef to
d89fe51
Compare
yzh119
approved these changes
Feb 1, 2024
Collaborator
yzh119
left a comment
yzh119
left a comment
There was a problem hiding this comment.
LGTM, thank you @MasterJH5574 !
yzh119
added a commit
that referenced
this pull request
Feb 16, 2024
diptorupd
referenced
this pull request
in ROCm/flashinfer
Sep 29, 2025
Adds a common wrapper function to mma_ops.hpp for hgemm kernels that works for both CUDA and HIP. Replaces `mma_sync_m16n16k16_row_col_f16f16f32`
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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 q/k position information to batch prefill/decode kernels. More specifically, the kernel now accepts two additional arrays:
q_rope_positionwith shape(total_q_len,), denoting the in-sequence position of each position in the input q.k_rope_pos_offsetwith shape(num_sequence,), denoting the start position of each sequence in k.These two arrays helps on-the-fly calculate RoPE in multi-level cases.
Tests
test_batch_prefillandtest_batch_decodecan pass. Performance is not validated yet. Per discussion with Zihao, this change is not very likely to incur significant perf regression.