Skip to content

perf(ds4): accelerate ROCm decode and long-context selection #667

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
davide221 wants to merge 2 commits into
base: codex/ds4-pflash-long-context
Choose a base branch
from
Open

perf(ds4): accelerate ROCm decode and long-context selection #667

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
179a90f
feat(spec): add opt-in adaptive verification width
davide221 Aug 29, 2026
5eb4fbe
perf(ds4): accelerate gfx1151 verification
davide221 Aug 29, 2026
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 @@ -946,6 +946,17 @@ if(DFLASH27B_TESTS)
${CMAKE_CURRENT_SOURCE_DIR}/deps/llama.cpp/ggml/include)
target_link_libraries(bench_rocmfp_mix_gateup_glu PRIVATE
ggml ggml-base ${DFLASH27B_GGML_BACKEND_TARGET} hip::host)

# Benchmark, not a test: compares the generic segmented full sort with
# the exact two-stage block-radix TOP_K used by the DS4 long-context
# indexer. It also checks selected-index set parity at every tile count.
add_executable(bench_ds4_topk test/bench_ds4_topk.cpp)
set_source_files_properties(test/bench_ds4_topk.cpp PROPERTIES LANGUAGE HIP)
set_target_properties(bench_ds4_topk PROPERTIES HIP_ARCHITECTURES "${_dflash_archs}")
target_include_directories(bench_ds4_topk PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/deps/llama.cpp/ggml/include)
target_link_libraries(bench_ds4_topk PRIVATE
ggml ggml-base ${DFLASH27B_GGML_BACKEND_TARGET} hip::host)
endif()
add_executable(test_qwen35_tensor_parallel test/test_qwen35_tensor_parallel.cpp)
target_include_directories(test_qwen35_tensor_parallel PRIVATE ${DFLASH27B_SRC_INCLUDE_DIRS})
Expand Down Expand Up @@ -1572,6 +1583,7 @@ if(DFLASH27B_TESTS)
test/test_server_unit.cpp
test/test_anchor_params.cpp
test/test_derived_scalars.cpp
test/test_adaptive_spec_width.cpp
test/test_adaptive_keep_ratio.cpp
test/test_skip_park_guard.cpp
test/test_bandit_integration.cpp
Expand Down
Loading