CUDA: add small-k optimization for mul-mat-id bs 2-4#20885
Closed
am17an wants to merge 2 commits intoggml-org:masterfrom
Closed
CUDA: add small-k optimization for mul-mat-id bs 2-4#20885am17an wants to merge 2 commits intoggml-org:masterfrom
am17an wants to merge 2 commits intoggml-org:masterfrom
Conversation
github-actions
Bot
added
Nvidia GPU
Contributor
JohannesGaessler
left a comment
JohannesGaessler
left a comment
There was a problem hiding this comment.
I think you also need to edit the switch cases for 2-4 columns for some of the GPUs.
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Contributor
Author
MUL-MAT-ID takes the ncols_dst=1 path, which we added in #18958 |
Contributor
|
Sorry, I didn't realize you had created a PR for the same issue. I filed one just now: #20905 The implementation is very different, though. So it will be good to do a perf comparison. |
Contributor
|
Checked the perf data. PR #20905 shows better perf than #20885 across models and GPUs. Performance
|
gaugarg-nv
pushed a commit
to gaugarg-nv/llama.cpp
that referenced
this pull request
Mar 25, 2026
…optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8
Contributor
Author
|
Superseded by #20905 |
JohannesGaessler
pushed a commit
that referenced
this pull request
Mar 29, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR #20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
ggerganov
pushed a commit
to ggml-org/ggml
that referenced
this pull request
Apr 1, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR ggml-org/llama.cpp#20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
ggerganov
pushed a commit
to ggml-org/ggml
that referenced
this pull request
Apr 1, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR ggml-org/llama.cpp#20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
slartibardfast
pushed a commit
to slartibardfast/llama.cpp
that referenced
this pull request
Apr 12, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR ggml-org#20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
Seunghhon
pushed a commit
to Seunghhon/llama.cpp
that referenced
this pull request
Apr 26, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR ggml-org#20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
pull Bot
pushed a commit
to sh1970/whisper.cpp
that referenced
this pull request
May 1, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR ggml-org/llama.cpp#20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
rsenthilkumar6
pushed a commit
to rsenthilkumar6/llama.cpp
that referenced
this pull request
May 1, 2026
* Optimize MOE GEMV kernel for BS > 1. The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row. New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync). This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization. * Remove em-dashes * Cherry-pick changes from @am17an PR ggml-org#20885 to enable small_k optimization only for cases where it benefits Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8 * Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com>
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
3 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.
Following on #20635
I just did a basic test on a 5090, no compilation slow-down