tests : reduce FA test sizes (#28842)

- Clamp the -j parallelism to min(nproc, 2) so a single-core runner
  uses -j 1 and multi-core runners use at most -j 2, instead of
  unconditionally using $(nproc).
- Add a 3600s timeout to both test-backend-ops runs (the high-perf CPU
  path and the default path) so a hung test cannot stall CI indefinitely.
- Note a TODO to reduce the timeout to 1800s in the future.

Assisted-by: pi:llama.cpp/Qwen3.8-27B

.github/workflows/server-sanitize.yml

@@ -116,7 +116,6 @@ jobs:
         run: |
           source .venv/bin/activate
           cd tools/server/tests
-          export ${{ matrix.extra_args }}
           PYTEST_WORKERS=1 ./tests.sh
 
       - name: Slow tests
@@ -125,5 +124,4 @@ jobs:
         run: |
           source .venv/bin/activate
           cd tools/server/tests
-          export ${{ matrix.extra_args }}
           PYTEST_WORKERS=1 SLOW_TESTS=1 ./tests.sh

ci/run.sh

@@ -775,7 +775,11 @@ function gg_run_test_backend_ops {
 
     set -e
 
-    local args_extra="-j $(nproc)"
+    local n_jobs=$(nproc)
+    if [ "${n_jobs}" -gt 2 ]; then
+        n_jobs=2
+    fi
+    local args_extra="-j ${n_jobs}"
 
     # TODO: fix multi-threaded for ROCm
     #       https://github.com/ggml-org/llama.cpp/actions/runs/34576278519/job/103297889044?pr=28740#step:3:4865
@@ -789,10 +793,11 @@ function gg_run_test_backend_ops {
         args_extra=""
     fi
 
+    # TODO: reduce the test-backend-ops timeout to 1800s
     if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
-        (time ./bin/test-backend-ops ${args_extra} -b CPU) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
+        (time timeout 3600 ./bin/test-backend-ops ${args_extra} -b CPU) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
     else
-        (time ./bin/test-backend-ops ${args_extra}       ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
+        (time timeout 3600 ./bin/test-backend-ops ${args_extra}       ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops.log
     fi
 
     set +e

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -333,6 +333,7 @@ static void ggml_vk_print_device_lost_info(const vk_device& device);
 struct vk_queue_handle {
     vk::Queue queue;
     vk_device_ref device;
+    std::mutex * device_submit_mutex = nullptr;
     virtual void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) = 0;
     virtual void lock()   {}   // no-op by default (internally synchronized case)
     virtual void unlock() {}
@@ -342,6 +343,11 @@ struct vk_queue_handle {
 struct vk_queue_handle_synchronized : vk_queue_handle {
     std::mutex mutex;
     void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
+        // Workaround for NVIDIA driver bug
+        std::unique_lock<std::mutex> device_guard;
+        if (device_submit_mutex) {
+            device_guard = std::unique_lock<std::mutex>(*device_submit_mutex);
+        }
         std::lock_guard<std::mutex> guard(mutex);
         try {
             queue.submit(submits, fence);
@@ -356,9 +362,14 @@ struct vk_queue_handle_synchronized : vk_queue_handle {
     void unlock() override { mutex.unlock(); }
 };
 
+// Driver guarantees internal synchronization via VK_KHR_internally_synchronized_queues
 struct vk_queue_handle_unsynchronized : vk_queue_handle {
     void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
-        // Driver guarantees internal synchronization via VK_KHR_internally_synchronized_queues
+        // Workaround for NVIDIA driver bug
+        std::unique_lock<std::mutex> device_guard;
+        if (device_submit_mutex) {
+            device_guard = std::unique_lock<std::mutex>(*device_submit_mutex);
+        }
         try {
             queue.submit(submits, fence);
         } catch (vk::DeviceLostError &) {
@@ -835,6 +846,7 @@ static bool ggml_vk_lightning_indexer_k_type_supported(ggml_type type) {
 
 struct vk_device_struct {
     std::recursive_mutex mutex;
+    std::mutex queue_submit_mutex;
     mutable std::shared_mutex pinned_memory_mutex;
 
     // Guards compile_pending, all_pipelines, and the dynamic pipeline maps
@@ -3520,6 +3532,10 @@ static std::unique_ptr<vk_queue> ggml_vk_create_queue(vk_device& device, uint32_
 
     h->queue = device->device.getQueue2(queue_info2);
     h->device = device;
+    // Avoid concurrent submissions on NVIDIA due to driver bug.
+    if (device->vendor_id == VK_VENDOR_ID_NVIDIA) {
+        h->device_submit_mutex = &device->queue_submit_mutex;
+    }
     q->handle = h;
 
     q->cmd_pool.init(device, q.get());

tests/test-backend-ops.cpp

@@ -10678,9 +10678,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_flash_attn_ext(256, 256, 2, {16, 1}, 16384,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0));
 
     // MLA shape: the V cache is a sub-view of the K cache, with quantized KV
-    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1},  113,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
-    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
-    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {20, 1}, 1024,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {8, 1},  113,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {8, 1}, 1024,   1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
+    test_cases.emplace_back(new test_flash_attn_ext(576, 512, 1, {8, 1}, 1024,  64, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0, {0, 1, 2, 3}, true, true));
 
     // Sparse mask hint: supported decode/prefill layouts and dense fallbacks.
     test_cases.emplace_back(new test_flash_attn_ext(512, 512, 1, { 8, 1}, 4096, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16, GGML_TYPE_F16, {0, 1, 2, 3}, true, false,  512));