perf(cuda): reduce MMQ stream-k overhead

.github/workflows/docker.yml

@@ -28,6 +28,7 @@ on:
       - .dockerignore
       - .github/workflows/docker.yml
       - server/CMakeLists.txt
+      - server/cmake/**
       - server/src/**
       - server/test/**
       - server/include/**
@@ -51,6 +52,7 @@ on:
       - docker-bake.hcl
       - .dockerignore
       - .github/workflows/docker.yml
+      - server/cmake/**
   # Manual trigger for one-off rebuilds or pre-release smoke tests. The
   # `push` input controls whether the resulting images land in GHCR or only
   # populate the buildx cache.

Dockerfile

@@ -55,7 +55,7 @@ WORKDIR /src
 
 # COPY ordering is structured to keep the CUDA build cached across
 # Python-only edits. The cmake build only depends on dflash/{CMakeLists,
-# include, src, test, hip_compat, deps}. Everything else (Python scripts,
+# cmake, include, src, test, hip_compat, deps}. Everything else (Python scripts,
 # workspace pyproject manifests, lockfile, READMEs) is copied later so
 # editing server.py / bench_*.py / lucebox sources doesn't invalidate the
 # ~25-minute CUDA template-instantiation layer below.
@@ -65,6 +65,7 @@ WORKDIR /src
 # uses server/; submodule binding names still write `dflash/deps/...`
 # inside .gitmodules (arbitrary identifiers; only paths matter).
 COPY server/CMakeLists.txt /src/server/CMakeLists.txt
+COPY server/cmake /src/server/cmake
 COPY server/include /src/server/include
 COPY server/src /src/server/src
 COPY server/test /src/server/test

Dockerfile.rocm

@@ -74,6 +74,7 @@ WORKDIR /src
 # COPY ordering mirrors the CUDA Dockerfile: C++ build inputs first so the
 # (slow) HIP build layer stays cached across Python-only edits.
 COPY server/CMakeLists.txt /src/server/CMakeLists.txt
+COPY server/cmake /src/server/cmake
 COPY server/include /src/server/include
 COPY server/src /src/server/src
 COPY server/test /src/server/test

server/CMakeLists.txt

@@ -806,6 +806,18 @@ if(DFLASH27B_TESTS)
         target_link_libraries(test_draft_topk_cuda PRIVATE dflash_common ${DFLASH27B_GGML_BACKEND_TARGET})
         list(APPEND _raw_unit_test_targets test_draft_topk_cuda)
     endif()
+    # Focused CUDA-vs-CPU oracle for the IQ4_XS MMQ stream-k scheduler. These
+    # shapes cover the no-fixup, partial-row, fixup, two-wave, and deeper-K
+    # transitions validated for the llama.cpp #22298 backport on sm_86.
+    if(DFLASH27B_GPU_BACKEND STREQUAL "cuda" AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_mmq_streamk_iq4_xs.cpp")
+        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)
+        list(APPEND _raw_unit_test_targets test_mmq_streamk_iq4_xs)
+    endif()
     if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_drafter_tail_capture_guard.cpp")
         # RED phase binary: same source WITHOUT the fix flag — documents the bug.
         add_executable(test_drafter_tail_capture_guard_red

server/deps/llama.cpp/VENDOR.md

@@ -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.