feat(ds4): add concurrent serving on Strix Halo and R9700 + Strix

.github/workflows/ci.yml

@@ -84,13 +84,14 @@ jobs:
             test_dflash test_generate test_flash_attn_sparse test_server_unit \
             test_deepseek4_unit test_feature_gate test_seq_slot_manager \
             test_seq_engine_contract test_seq_batch_plan test_client_send_buffer \
+            test_deepseek4_page_layout test_deepseek4_paged_cache \
             test_model_smoke test_batched_gdn test_concat_transpose -j$(nproc)
 
       - name: Run C++ server unit tests
         run: |
           cd server/build
           ctest --output-on-failure \
-            -R "server_unit|deepseek4_unit|feature_gate|seq_slot_manager|seq_engine_contract|seq_batch_plan|client_send_buffer|batched_gdn_cpu" \
+            -R "server_unit|deepseek4_unit|feature_gate|seq_slot_manager|seq_engine_contract|seq_batch_plan|client_send_buffer|deepseek4_page_layout|deepseek4_paged_cache|batched_gdn_cpu" \
             --no-tests=error
 
       - name: Populate venv with cu128 torch + setuptools

server/CMakeLists.txt

@@ -436,6 +436,8 @@ add_library(dflash_common STATIC
     src/deepseek4/deepseek4_loader.cpp
     src/deepseek4/deepseek4_graph.cpp
     src/deepseek4/deepseek4_roctx.cpp
+    src/deepseek4/deepseek4_paged_cache.cpp
+    src/deepseek4/deepseek4_seq_engine.cpp
     src/deepseek4/deepseek4_backend.cpp
     src/deepseek4/deepseek4_daemon.cpp
     src/deepseek4/deepseek4_layer_split_adapter.cpp
@@ -472,7 +474,7 @@ add_library(dflash_common STATIC
     src/common/draft_swa.cpp
     src/common/dflash_spec_decode.cpp
     src/common/concurrency/paged_kv_pool.cpp
-    src/qwen35/concurrency/qwen35_slot_manager.cpp
+    src/common/concurrency/seq_slot_manager.cpp
     src/common/layer_split_backend.cpp
     src/common/layer_split_runtime.cpp
     src/qwen35/graph_builders.cpp
@@ -1426,11 +1428,29 @@ if(DFLASH27B_TESTS)
             ${CMAKE_CURRENT_SOURCE_DIR}/src)
         list(APPEND _raw_unit_test_targets test_paged_kv_pool)
     endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_deepseek4_page_layout.cpp")
+        add_executable(test_deepseek4_page_layout
+            test/test_deepseek4_page_layout.cpp)
+        target_include_directories(test_deepseek4_page_layout PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}/src
+            ${CMAKE_CURRENT_SOURCE_DIR}/include)
+        list(APPEND _raw_unit_test_targets test_deepseek4_page_layout)
+    endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_deepseek4_paged_cache.cpp")
+        add_executable(test_deepseek4_paged_cache
+            test/test_deepseek4_paged_cache.cpp
+            src/deepseek4/deepseek4_paged_cache.cpp)
+        target_compile_definitions(test_deepseek4_paged_cache PRIVATE
+            DFLASH_DS4_PLAN_ONLY=1)
+        target_include_directories(test_deepseek4_paged_cache PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}/src)
+        list(APPEND _raw_unit_test_targets test_deepseek4_paged_cache)
+    endif()
     if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_seq_slot_manager.cpp")
         # Host-side slot bookkeeping test (concurrent serving): no GPU.
         add_executable(test_seq_slot_manager
             test/test_seq_slot_manager.cpp
-            src/qwen35/concurrency/qwen35_slot_manager.cpp
+            src/common/concurrency/seq_slot_manager.cpp
             src/common/concurrency/paged_kv_pool.cpp)
         target_include_directories(test_seq_slot_manager PRIVATE
             ${CMAKE_CURRENT_SOURCE_DIR}/src)

server/README.md

@@ -299,10 +299,10 @@ See the [current six-expert Strix Halo profile](https://www.lucebox.com/blog/dee
 | `--prefix-cache-slots <N>` | `32` | In-memory prefix-cache slots; `0` disables. |
 | `--agent-turn-cache` | off | Extend prefix caching through generated tool calls. |
 | `--prefill-cache-slots <N>` | `0` | Full-prompt cache slots. |
-| `--paged-attention` | off | Enable 16-token paged KV blocks for supported Qwen targets. |
-| `--max-concurrency <N>` | `1` | Maximum concurrent decode sequences; values above 1 enable paged attention. |
+| `--paged-attention` | off | Enable paged KV for supported Qwen targets (16-token blocks) or DeepSeek4 on Strix Halo and R9700 plus Strix Halo (128-token pages). DeepSeek4 paged serving is AR-only. |
+| `--max-concurrency <N>` | `1` | Maximum concurrent decode sequences. Qwen supports up to 64; DeepSeek4 supports up to 6. Values above 1 enable paged attention. |
 | `--admission-coalesce-ms <N>` | `20` | Idle-to-busy batching window from 0 through 1000 ms. |
-| `--kv-pool-tokens <N>` | auto | Shared physical K/V capacity for concurrent serving. |
+| `--kv-pool-tokens <N>` | auto | Shared physical K/V capacity for concurrent serving, rounded to the backend page size. |
 | `--kv-cache-dir <path>` | none | Enable persistent disk KV cache in this directory. |
 | `--kv-cache-budget <MB>` | `4096` | Disk KV-cache size cap. |
 | `--kv-cache-min-tokens <N>` | `512` | Minimum prefix length to persist. |

server/deps/llama.cpp/ggml/src/ggml-backend.cpp

@@ -2322,7 +2322,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
     std::vector<int32_t> ids;
     std::vector<ggml_bitset_t> used_ids;
     size_t batch_staging_cursors[GGML_SCHED_MAX_BACKENDS] = {};
-    const bool copy_destinations_ready = sched->backends_synchronized;
+    const bool staging_arena_reusable = sched->backends_synchronized;
     // From this point onward an early return must remain conservative: a
     // backend may have accepted work even if a later split fails.
     sched->backends_synchronized = false;
@@ -2344,7 +2344,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
 
         ggml_backend_sched_split_copy_state copy_state{
             sched, split_backend_id, split_backend,
-            copy_destinations_ready, copy_destinations_ready,
+            staging_arena_reusable,
+            // Earlier splits may have submitted work to this backend after
+            // the entry synchronization. Establish readiness once per split.
+            false,
             batch_staging_cursors};
 
         // copy the input tensors to the split backend