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

server/CMakeLists.txt

@@ -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 ${DFLASH27B_SRC_INCLUDE_DIRS})
@@ -1572,6 +1583,7 @@ if(DFLASH27B_TESTS)
             test/test_server_unit.cpp
             test/test_anchor_params.cpp
             test/test_derived_scalars.cpp
+            test/test_adaptive_spec_width.cpp
             test/test_adaptive_keep_ratio.cpp
             test/test_skip_park_guard.cpp
             test/test_bandit_integration.cpp
@@ -1955,6 +1967,16 @@ if(DFLASH27B_TESTS)
             add_dependencies(check test_concat_transpose)
         endif()
     endif()
+    if((DFLASH27B_GPU_BACKEND STREQUAL "cuda" OR
+        DFLASH27B_GPU_BACKEND STREQUAL "hip")
+       AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_moe_combine.cpp")
+        dflash_add_ggml_gpu_executable(
+            test_moe_combine test/test_moe_combine.cpp)
+        add_test(NAME moe_combine COMMAND test_moe_combine)
+        if(TARGET check)
+            add_dependencies(check test_moe_combine)
+        endif()
+    endif()
     if((DFLASH27B_GPU_BACKEND STREQUAL "cuda" OR
         DFLASH27B_GPU_BACKEND STREQUAL "hip")
        AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/bench_paged_attention.cpp")

server/deps/llama.cpp/ggml/include/ggml.h

@@ -2560,6 +2560,17 @@ extern "C" {
             int64_t               ff_dim,
             int64_t               n_expert_used);
 
+    // Apply one routing weight to every expert output column and reduce the
+    // expert axis. `experts` is [n_embd, n_used, n_tokens] and
+    // `expert_weights` is [n_used, n_tokens]. Backends may fuse the weighting
+    // and reduction into one kernel.
+    GGML_API struct ggml_tensor * ggml_moe_combine(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * experts,
+            struct ggml_tensor  * expert_weights);
+
+    // Compatibility alias for out-of-tree Laguna callers. New model code
+    // should use the model-neutral ggml_moe_combine API.
     GGML_API struct ggml_tensor * ggml_laguna_moe_combine(
             struct ggml_context * ctx,
             struct ggml_tensor  * experts,