Skip to content

perf(cuda): reduce MMQ stream-k overhead #518

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
dusterbloom wants to merge 1 commit into
base: main
Choose a base branch
from
+407 −140
Open

perf(cuda): reduce MMQ stream-k overhead #518

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions server/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -712,6 +712,18 @@ if(DFLASH27B_TESTS)
target_link_libraries(test_draft_topk_cuda PRIVATE dflash_common ${DFLASH27B_GGML_BACKEND_TARGET})
add_test(NAME draft_topk_cuda COMMAND test_draft_topk_cuda)
endif()
# Focused CUDA-vs-CPU oracle for the IQ4_XS MMQ stream-k scheduler. These
# shapes cover the no-fixup, partial-row, fixup, two-wave, and deeper-K
# transitions validated for the llama.cpp #22298 backport on sm_86.
if(DFLASH27B_GPU_BACKEND STREQUAL "cuda")
add_executable(test_mmq_streamk_iq4_xs test/test_mmq_streamk_iq4_xs.cpp)
target_link_libraries(test_mmq_streamk_iq4_xs PRIVATE
ggml
ggml-cpu
ggml-cuda
ggml-base)
add_test(NAME mmq_streamk_iq4_xs COMMAND test_mmq_streamk_iq4_xs)
endif()
# GPU port of the sample_logits chain vs the CPU reference. CUDA only:
# geometric_sampler_cuda.cu is compiled into dflash_common solely on the cuda backend.
if(DFLASH27B_GPU_BACKEND STREQUAL "cuda" AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_gpu_sampler_cuda.cpp")
Expand Down
3 changes: 2 additions & 1 deletion server/deps/llama.cpp/VENDOR.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,8 @@ This directory contains the ggml-only subset used by Lucebox Hub.
- Source base commit: `6fbe72d67069136bbd370be703e1d4f441b5e942`
- Included merged PR: `#35` (`0fe65d9354b7c5da52a7741d2e37ba85f0d0c925`)
- Included test PR: `#37` (`0699be81480428f01b9b7ac49a09a2d51c77f8df`)
- Reconstruction: `luce-dflash@6fbe72d67069136bbd370be703e1d4f441b5e942` plus cherry-picked PRs `#35` and `#37`
- Included upstream backport: `llama.cpp #22298` (`9725a313be0528214c4a02fed906ddaf7b3f712e`)
- Reconstruction: `luce-dflash@6fbe72d67069136bbd370be703e1d4f441b5e942` plus cherry-picked PRs `#35`, `#37`, and upstream `llama.cpp #22298`
- Vendored paths: `LICENSE`, `common/jinja`, `common/log.h`, `common/unicode.*`, `ggml`, `gguf-py`

Open ggml feature PRs are intentionally not included until they are merged, except for explicitly listed hub test PRs.
Loading
Loading