vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors (#26371 )

* vulkan: add debug tooling to get more information about a DeviceLost error * fix submission threshold applied too late * use logging macros, throw instead of aborting * clean up circular dependency
mtmd/ggml: add ggml_build_forward_order (#26649 )
2026-08-07 04:08:04 +02:00 · 2026-08-06 10:24:13 +03:00 · 2026-08-06 00:47:59 +02:00 · 2026-08-05 21:31:54 +02:00 · 2026-08-05 17:58:34 +02:00 · 2026-08-05 15:30:14 +02:00
11 changed files with 502 additions and 136 deletions
@@ -17,8 +17,11 @@ from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logg
 from .qwen import QwenModel


-@ModelBase.register("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
+@ModelBase.register("DeepseekOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
+    # HF dynamic_preprocess() max_num, which differs per model
+    preproc_max_tiles = 9
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
@@ -43,6 +46,9 @@ class DeepseekOCRVisionModel(MmprojModel):
        # @bluebread: there's no window_size in config but just add it here anyway
        self.gguf_writer.add_vision_window_size(self.hparams.get("window_size", 14))

+        self.gguf_writer.add_vision_preproc_min_tiles(2)
+        self.gguf_writer.add_vision_preproc_max_tiles(self.preproc_max_tiles)
+
        # SAM configuration
        sam_hparams = hparams['sam']
        self.gguf_writer.add_vision_sam_layers_count(sam_hparams['layers'])
@@ -93,8 +99,15 @@ class DeepseekOCRVisionModel(MmprojModel):
        return super().filter_tensors((name, gen))


+@ModelBase.register("UnlimitedOCRForCausalLM")
+class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 32
+
+
@ModelBase.register("DeepseekOCR2ForCausalLM")
 class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 6
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2
@@ -2788,6 +2788,12 @@ extern "C" {
            struct ggml_cgraph * cgraph,
            struct ggml_tensor * tensor);

+    // add the tensor and its parents to the graph without marking them for compute
+    // the flag is set later, when the tensor is reached from a node that computes
+    GGML_API void ggml_build_forward_order(
+            struct ggml_cgraph * cgraph,
+            struct ggml_tensor * tensor);
+
    GGML_API void ggml_build_backward_expand(
        struct ggml_context *  ctx,        // context for gradient computation
        struct ggml_cgraph  *  cgraph,
@@ -186,13 +186,22 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }

 #define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256

-#define VK_CHECK(err, msg)                                          \
+#define VK_CHECK(err, msg, dev)                                     \
    do {                                                            \
-        vk::Result err_ = (err);                                    \
+        vk::Result err_;                                            \
+        try {                                                       \
+            err_ = (err);                                           \
+        } catch (vk::DeviceLostError &) {                           \
+            ggml_vk_print_device_lost_info(dev);                    \
+            GGML_LOG_ERROR("ggml_vulkan: %s at %s:%d\n",            \
+                #err, __FILE__, __LINE__);                          \
+            throw;                                                  \
+        }                                                           \
        if (err_ != vk::Result::eSuccess) {                         \
-            fprintf(stderr, "ggml_vulkan: %s error %s at %s:%d\n",  \
+            GGML_LOG_ERROR("ggml_vulkan: %s error %s at %s:%d\n",   \
                #err, to_string(err_).c_str(), __FILE__, __LINE__); \
-            exit(1);                                                \
+            throw vk::SystemError(vk::make_error_code(err_),        \
+                "ggml_vulkan: " msg);                               \
        }                                                           \
    } while (0)

@@ -302,9 +311,13 @@ struct vk_command_pool {
    }
 };

+static void ggml_vk_print_device_fault_info(const vk_device& device);
+static void ggml_vk_print_device_lost_info(const vk_device& device);
+
 // Prevent simultaneous submissions to the same queue.
 struct vk_queue_handle {
    vk::Queue queue;
+    vk_device_ref device;
    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() {}
@@ -315,7 +328,14 @@ struct vk_queue_handle_synchronized : vk_queue_handle {
    std::mutex mutex;
    void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
        std::lock_guard<std::mutex> guard(mutex);
-        queue.submit(submits, fence);
+        try {
+            queue.submit(submits, fence);
+        } catch (vk::DeviceLostError &) {
+            if (auto dev = device.lock()) {
+                ggml_vk_print_device_lost_info(dev);
+            }
+            throw;
+        }
    }
    void lock()   override { mutex.lock(); }
    void unlock() override { mutex.unlock(); }
@@ -324,7 +344,14 @@ struct vk_queue_handle_synchronized : vk_queue_handle {
 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
-        queue.submit(submits, fence);
+        try {
+            queue.submit(submits, fence);
+        } catch (vk::DeviceLostError &) {
+            if (auto dev = device.lock()) {
+                ggml_vk_print_device_lost_info(dev);
+            }
+            throw;
+        }
    }
    // lock()/unlock() inherited no-ops
 };
@@ -835,6 +862,15 @@ struct vk_device_struct {

    bool pipeline_executable_properties_support {};

+    bool device_fault {};
+    PFN_vkGetDeviceFaultInfoEXT pfn_vkGetDeviceFaultInfoEXT {};
+
+    bool serialize_submissions {};
+
+    const ggml_cgraph * diag_cgraph {};
+    int diag_prev_start = -1;
+    int diag_prev_end = -1;
+
    size_t idx;

    bool mul_mat_l[GGML_TYPE_COUNT];
@@ -1118,6 +1154,57 @@ void vk_command_pool::destroy(vk::Device& device) {
    cmd_buffers.clear();
 }

+static void ggml_vk_print_device_fault_info(const vk_device& device) {
+    if (!device->device_fault || !device->pfn_vkGetDeviceFaultInfoEXT) {
+        return;
+    }
+
+    VkDeviceFaultCountsEXT fault_counts {};
+    fault_counts.sType = VK_STRUCTURE_TYPE_DEVICE_FAULT_COUNTS_EXT;
+    VkResult res = device->pfn_vkGetDeviceFaultInfoEXT(device->device, &fault_counts, nullptr);
+    if (res != VK_SUCCESS) {
+        GGML_LOG_ERROR("ggml_vulkan: vkGetDeviceFaultInfoEXT (counts) failed: %d\n", res);
+        return;
+    }
+
+    std::vector<VkDeviceFaultAddressInfoEXT> address_infos(fault_counts.addressInfoCount);
+    std::vector<VkDeviceFaultVendorInfoEXT> vendor_infos(fault_counts.vendorInfoCount);
+
+    VkDeviceFaultInfoEXT fault_info {};
+    fault_info.sType = VK_STRUCTURE_TYPE_DEVICE_FAULT_INFO_EXT;
+    fault_info.pAddressInfos = address_infos.data();
+    fault_info.pVendorInfos = vendor_infos.data();
+
+    res = device->pfn_vkGetDeviceFaultInfoEXT(device->device, &fault_counts, &fault_info);
+    if (res != VK_SUCCESS) {
+        GGML_LOG_ERROR("ggml_vulkan: vkGetDeviceFaultInfoEXT (info) failed: %d\n", res);
+        return;
+    }
+
+    if (fault_counts.addressInfoCount == 0 && fault_counts.vendorInfoCount == 0 && fault_info.description[0] == '\0') {
+        return;
+    }
+
+    if (fault_info.description[0] != '\0') {
+        GGML_LOG_ERROR("ggml_vulkan: device fault on %s: %s\n", device->name.c_str(), fault_info.description);
+    }
+
+    for (uint32_t i = 0; i < fault_counts.addressInfoCount; i++) {
+        const auto& info = address_infos[i];
+        GGML_LOG_CONT("  address fault %u: type=%d address=0x%llx precision=0x%llx\n",
+                i, (int)info.addressType,
+                (unsigned long long)info.reportedAddress,
+                (unsigned long long)info.addressPrecision);
+    }
+    for (uint32_t i = 0; i < fault_counts.vendorInfoCount; i++) {
+        const auto& info = vendor_infos[i];
+        GGML_LOG_CONT("  vendor fault %u: %s (code=0x%llx data=0x%llx)\n",
+                i, info.description,
+                (unsigned long long)info.vendorFaultCode,
+                (unsigned long long)info.vendorFaultData);
+    }
+}
+
 struct vk_buffer_struct {
    vk::Buffer buffer = VK_NULL_HANDLE;
    vk::DeviceMemory device_memory = VK_NULL_HANDLE;
@@ -2059,6 +2146,36 @@ static uint64_t ggml_vk_get_node_flops(const ggml_tensor * node) {
    return 0;
 }

+static void ggml_vk_print_node_list(const ggml_cgraph * cgraph, int start, int end) {
+    uint64_t total_flops = 0;
+    int n_ops = 0;
+    for (int j = start; j <= end && j < cgraph->n_nodes; j++) {
+        uint64_t flops = ggml_vk_get_node_flops(cgraph->nodes[j]);
+        total_flops += flops;
+        n_ops++;
+        if (flops > 0) {
+            GGML_LOG_CONT("  node %d: %s (%s) [%.2f GFLOP]\n",
+                    j, cgraph->nodes[j]->name, ggml_op_name(cgraph->nodes[j]->op),
+                    flops / 1e9);
+        } else {
+            GGML_LOG_CONT("  node %d: %s (%s)\n",
+                    j, cgraph->nodes[j]->name, ggml_op_name(cgraph->nodes[j]->op));
+        }
+    }
+    GGML_LOG_CONT("  total: %d ops, %.2f GFLOP\n", n_ops, total_flops / 1e9);
+}
+
+static void ggml_vk_print_device_lost_info(const vk_device& device) {
+    ggml_vk_print_device_fault_info(device);
+    if (device->serialize_submissions && device->diag_cgraph != nullptr && device->diag_prev_start >= 0) {
+        GGML_LOG_ERROR("ggml_vulkan: device lost on %s, likely caused by previous submission (nodes %d to %d):\n",
+                device->name.c_str(), device->diag_prev_start, device->diag_prev_end);
+        ggml_vk_print_node_list(device->diag_cgraph, device->diag_prev_start, device->diag_prev_end);
+    } else {
+        GGML_LOG_ERROR("ggml_vulkan: device lost on %s\n", device->name.c_str());
+    }
+}
+
 class vk_perf_logger {
  public:
    void print_timings(bool force = false) {
@@ -2471,17 +2588,27 @@ static void ggml_vk_wait_for_fence(ggml_backend_vk_context * ctx) {
    // Use waitForFences while most of the graph executes. Hopefully the CPU can sleep
    // during this wait.
    if (ctx->almost_ready_fence_pending) {
-        VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence");
+        VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence", ctx->device);
        ctx->device->device.resetFences({ ctx->almost_ready_fence });
        ctx->almost_ready_fence_pending = false;
    }

    // Spin (w/pause) waiting for the graph to finish executing.
    vk::Result result;
-    while ((result = ctx->device->device.getFenceStatus(ctx->fence)) != vk::Result::eSuccess) {
+    for (;;) {
+        try {
+            result = ctx->device->device.getFenceStatus(ctx->fence);
+        } catch (vk::DeviceLostError &) {
+            ggml_vk_print_device_lost_info(ctx->device);
+            GGML_LOG_ERROR("ggml_vulkan: getFenceStatus at %s:%d\n", __FILE__, __LINE__);
+            throw;
+        }
+        if (result == vk::Result::eSuccess) {
+            break;
+        }
        if (result != vk::Result::eNotReady) {
-            fprintf(stderr, "ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__);
-            exit(1);
+            GGML_LOG_ERROR("ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__);
+            throw vk::SystemError(vk::make_error_code(result), "ggml_vulkan: getFenceStatus");
        }
        for (uint32_t i = 0; i < 100; ++i) {
            YIELD();
@@ -3172,6 +3299,7 @@ static std::unique_ptr<vk_queue> ggml_vk_create_queue(vk_device& device, uint32_
    }

    h->queue = device->device.getQueue2(queue_info2);
+    h->device = device;
    q->handle = h;

    q->cmd_pool.init(device, q.get());
@@ -6117,6 +6245,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
 #endif
            } else if (strcmp(VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME, properties.extensionName) == 0) {
                internally_sync_support = true;
+            } else if (strcmp("VK_EXT_device_fault", properties.extensionName) == 0) {
+                device->device_fault = true;
            }
        }

@@ -6471,8 +6601,18 @@ static vk_device ggml_vk_get_device(size_t idx) {
        }
 #endif

+        VkPhysicalDeviceFaultFeaturesEXT fault_features {};
+        fault_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FAULT_FEATURES_EXT;
+        if (device->device_fault) {
+            last_struct->pNext = (VkBaseOutStructure *)&fault_features;
+            last_struct = (VkBaseOutStructure *)&fault_features;
+            device_extensions.push_back("VK_EXT_device_fault");
+        }
+
        vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);

+        device->device_fault = device->device_fault && fault_features.deviceFault;
+
        device->has_internally_synchronized_queues = internally_synchronized_queues_features.internallySynchronizedQueues;

        // Build queue create infos only after querying whether internally synchronized queues are enabled.
@@ -6771,6 +6911,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
        device_create_info.setPNext(&device_features2);
        device->device = device->physical_device.createDevice(device_create_info);

+        if (device->device_fault) {
+            device->pfn_vkGetDeviceFaultInfoEXT = (PFN_vkGetDeviceFaultInfoEXT)
+                vkGetDeviceProcAddr(device->device, "vkGetDeviceFaultInfoEXT");
+        }
+
        // Queues
        device->compute_queue = ggml_vk_create_queue(device, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);

@@ -6893,6 +7038,8 @@ static vk_device ggml_vk_get_device(size_t idx) {

        device->idx = idx;

+        device->serialize_submissions = getenv("GGML_VK_SERIALIZE_SUBMISSIONS") != nullptr;
+
        device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;

        device->add_rms_fusion = !device->disable_fusion &&
@@ -8319,7 +8466,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
        }

        ggml_vk_submit(subctx, dst->device->fence);
-        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
+        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences", dst->device);
        dst->device->device.resetFences({ dst->device->fence });
        ggml_vk_queue_command_pools_cleanup(dst->device);
    }
@@ -8431,7 +8578,7 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
        ggml_vk_ctx_end(subctx);
        ggml_vk_submit(subctx, src->device->fence);
        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX),
-                 "vk_buffer_read_2d uma waitForFences");
+                 "vk_buffer_read_2d uma waitForFences", src->device);
        src->device->device.resetFences({ src->device->fence });
        ggml_vk_queue_command_pools_cleanup(src->device);

@@ -8452,7 +8599,7 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
        ggml_vk_ctx_end(subctx);

        ggml_vk_submit(subctx, src->device->fence);
-        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read_2d waitForFences");
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read_2d waitForFences", src->device);
        src->device->device.resetFences({ src->device->fence });
        ggml_vk_queue_command_pools_cleanup(src->device);

@@ -8487,7 +8634,7 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
        ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
        ggml_vk_ctx_end(subctx);
        ggml_vk_submit(subctx, src->device->fence);
-        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences", src->device);
        src->device->device.resetFences({ src->device->fence });
        ggml_vk_queue_command_pools_cleanup(src->device);
    } else {
@@ -8531,7 +8678,7 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
    ggml_vk_ctx_end(subctx);

    ggml_vk_submit(subctx, dst->device->fence);
-    VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences");
+    VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences", dst->device);
    dst->device->device.resetFences({ dst->device->fence });
    ggml_vk_queue_command_pools_cleanup(dst->device);
 }
@@ -14266,7 +14413,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t

    auto begin = std::chrono::high_resolution_clock::now();
    ggml_vk_submit(subctx, ctx->fence);
-    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences");
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences", ctx->device);
    ctx->device->device.resetFences({ ctx->fence });
    ggml_vk_queue_command_pools_cleanup(ctx->device);

@@ -14468,7 +14615,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
    auto begin = std::chrono::high_resolution_clock::now();

    ggml_vk_submit(subctx, ctx->fence);
-    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences", ctx->device);
    ctx->device->device.resetFences({ ctx->fence });
    ggml_vk_queue_command_pools_cleanup(ctx->device);

@@ -14754,7 +14901,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
    auto begin = std::chrono::high_resolution_clock::now();

    ggml_vk_submit(subctx, ctx->fence);
-    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences", ctx->device);
    ctx->device->device.resetFences({ ctx->fence });
    ggml_vk_queue_command_pools_cleanup(ctx->device);

@@ -15553,7 +15700,9 @@ static void ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
            memset(mset.dst, mset.val, mset.n);
        }

-        if (almost_ready && !ctx->almost_ready_fence_pending) {
+        if (ctx->device->serialize_submissions) {
+            ggml_vk_submit(subctx, ctx->fence);
+        } else if (almost_ready && !ctx->almost_ready_fence_pending) {
            ggml_vk_submit(subctx, ctx->almost_ready_fence);
            ctx->almost_ready_fence_pending = true;
        } else {
@@ -16164,12 +16313,20 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
            memcpy(cpy.dst, cpy.src, cpy.n);
        }

-        ggml_vk_submit(compute_ctx, {});
+        if (ctx->device->serialize_submissions) {
+            ggml_vk_submit(compute_ctx, ctx->fence);
+            VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "synchronize waitForFences", ctx->device);
+            ctx->device->device.resetFences({ ctx->fence });
+        } else {
+            ggml_vk_submit(compute_ctx, {});
+        }
        ctx->submit_pending = true;
    }

    if (ctx->submit_pending) {
-        if (ctx->device->async_use_transfer_queue && ctx->transfer_semaphore_last_submitted < ctx->transfer_semaphore.value) {
+        if (ctx->device->serialize_submissions) {
+            ctx->submit_pending = false;
+        } else if (ctx->device->async_use_transfer_queue && ctx->transfer_semaphore_last_submitted < ctx->transfer_semaphore.value) {
            vk::TimelineSemaphoreSubmitInfo tl_info{
                1, &ctx->transfer_semaphore.value,
                0, nullptr,
@@ -16186,7 +16343,9 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
        } else {
            ctx->device->compute_queue->handle->submit({}, ctx->fence);
        }
-        ggml_vk_wait_for_fence(ctx);
+        if (!ctx->device->serialize_submissions) {
+            ggml_vk_wait_for_fence(ctx);
+        }
        ctx->submit_pending = false;
        if (cmd_buf) {
            cmd_buf->in_use = false;
@@ -16758,6 +16917,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;

+    ctx->device->diag_cgraph = nullptr;
+    ctx->device->diag_prev_start = -1;
+    ctx->device->diag_prev_end = -1;
+
    if (vk_instance.debug_utils_support) {
        vk::DebugUtilsLabelEXT dul = {};
        dul.pLabelName = "ggml_backend_vk_graph_compute";
@@ -16849,6 +17012,36 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    }
    uint64_t flops_per_submit = std::min(flops_cap, ctx->last_total_flops / 40u);

+    auto const submit_after = [&](int start, int end) {
+        if (ctx->device->serialize_submissions) {
+            try {
+                auto res = ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX);
+                if (res != vk::Result::eSuccess) {
+                    GGML_LOG_ERROR("ggml_vulkan: waitForFences error during serialized submission\n");
+                    throw vk::SystemError(vk::make_error_code(res), "ggml_vulkan: waitForFences during serialized submission");
+                }
+            } catch (vk::DeviceLostError &) {
+                ggml_vk_print_device_fault_info(ctx->device);
+                GGML_LOG_ERROR("ggml_vulkan: device lost on %s waiting for submission (nodes %d to %d):\n",
+                        ctx->device->name.c_str(), start, end);
+                ggml_vk_print_node_list(cgraph, start, end);
+                throw;
+            }
+            ctx->device->device.resetFences({ ctx->fence });
+            ctx->submit_pending = false;
+            ctx->device->diag_cgraph = cgraph;
+            ctx->device->diag_prev_start = start;
+            ctx->device->diag_prev_end = end;
+        }
+        first_node_in_batch = true;
+        submitted_nodes = 0;
+        batch_flops = 0;
+        if (submit_count < 3) {
+            flops_per_submit *= 2;
+        }
+        submit_count++;
+    };
+
    for (int i = 0; i < cgraph->n_nodes; i++) {
        if (first_node_in_batch) {
            submit_node_idx = i;
@@ -16856,8 +17049,20 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg

        {
            auto node_flops = ggml_vk_get_node_flops(cgraph->nodes[i]);
-            batch_flops += node_flops;
            total_flops += node_flops;
+
+            // Flush the current batch before recording a node that would push it over the flop threshold
+            if (flops_per_submit != 0 && submitted_nodes > 0 && batch_flops + node_flops >= flops_per_submit) {
+                vk_context flush_ctx = ggml_vk_get_compute_ctx(ctx);
+                ggml_vk_ctx_end(flush_ctx);
+                flush_ctx->exit_tensor_idx = -1;
+                ctx->compute_ctx.reset();
+                ggml_vk_compute_forward(ctx, cgraph, cgraph->nodes[submit_node_idx], submit_node_idx, false);
+                submit_after(submit_node_idx, i - 1);
+                submit_node_idx = i;
+            }
+
+            batch_flops += node_flops;
        }

        // op_srcs_fused_elementwise indicates whether an op's srcs all contribute to
@@ -17111,13 +17316,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
        }

        if (submit && enqueued) {
-            first_node_in_batch = true;
-            submitted_nodes = 0;
-            batch_flops = 0;
-            if (submit_count < 3) {
-                flops_per_submit *= 2;
-            }
-            submit_count++;
+            submit_after(submit_node_idx, i + (int)ctx->num_additional_fused_ops);
        }
        i += ctx->num_additional_fused_ops;
        ctx->num_additional_fused_ops = 0;
@@ -17133,13 +17332,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
        ggml_vk_ctx_end(compute_ctx);

        ggml_vk_submit(compute_ctx, ctx->device->fence);
-        VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences");
+        VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences", ctx->device);
        ctx->device->device.resetFences({ ctx->device->fence });
        ctx->compute_ctx.reset();

        // Get the results and pass them to the logger
        std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
-        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->query_pool, 0, ctx->query_idx, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results");
+        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->query_pool, 0, ctx->query_idx, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results", ctx->device);
        if (!vk_perf_logger_concurrent) {
            // Log each op separately
            for (int i = 1; i < ctx->query_idx; i++) {
@@ -18366,7 +18565,7 @@ static void ggml_backend_vk_device_event_synchronize(ggml_backend_dev_t dev, ggm
        vk::Semaphore sem = vkev->tl_semaphore.s;
        uint64_t val = vkev->tl_semaphore.value;
        vk::SemaphoreWaitInfo swi{vk::SemaphoreWaitFlags{}, sem, val};
-        VK_CHECK(device->device.waitSemaphores(swi, UINT64_MAX), "event_synchronize");
+        VK_CHECK(device->device.waitSemaphores(swi, UINT64_MAX), "event_synchronize", device);

        // Reset and move submitted events
        for (auto& event : vkev->events_submitted) {
@@ -7200,6 +7200,10 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
    ggml_build_forward_impl(cgraph, tensor, true, true);
 }

+void ggml_build_forward_order(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
+    ggml_build_forward_impl(cgraph, tensor, true, false);
+}
+
 void ggml_build_backward_expand(
        struct ggml_context *  ctx,
        struct ggml_cgraph  *  cgraph,
@@ -648,10 +648,12 @@ const char * llama_grammar_parser::parse_sequence(
            } else {
                throw std::runtime_error(std::string("expecting ',' at ") + pos);
            }
-            bool has_max = max_times != UINT64_MAX;
-            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+            if (min_times > MAX_REPETITION_THRESHOLD) {
                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
            }
+            if (max_times != UINT64_MAX && max_times > MAX_REPETITION_THRESHOLD) {
+                max_times = UINT64_MAX;
+            }
            handle_repetitions(min_times, max_times);
        } else {
            break;
@@ -153,6 +153,53 @@ int main()
        root ::= "a"{,10}"
    )""");

+    verify_failure(R"""(
+        root ::= "a"{5000}
+    )""");
+
+    verify_failure(R"""(
+        root ::= "a"{5000,}
+    )""");
+
+    verify_failure(R"""(
+        root ::= "a"{5000,6000}
+    )""");
+
+    verify_parsing(R"""(
+        root ::= "a"{0,5000}
+    )""", {
+        {"root", 0},
+        {"root_1", 1},
+    }, {
+        // root (index 0)
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_END, 0},
+        // root_1 (index 1)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    });
+
+    verify_parsing(R"""(
+        root ::= "a"{3,5000}
+    )""", {
+        {"root", 0},
+        {"root_1", 1},
+    }, {
+        // root (index 0)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_END, 0},
+        // root_1 (index 1)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    });
+
    verify_parsing(R"""(
        root  ::= "a"
    )""", {
@@ -432,7 +432,7 @@ static bool arch_supported(const llm_arch arch) {

    // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MINIMAX_M3) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
        return false;
    }
 #endif // GGML_USE_WEBGPU
@@ -708,9 +708,10 @@ ggml_tensor * clip_graph::build_attn(
        ggml_tensor * sinks) const {
    // these nodes are added to the graph together so that they are not reordered
    // by doing so, the number of splits in the graph is reduced
-    ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
-    ggml_build_forward_expand(gf, v_cur);
+    // the order is fixed without the compute flag, so an unselected branch stays out of the compute set
+    ggml_build_forward_order(gf, q_cur);
+    ggml_build_forward_order(gf, k_cur);
+    ggml_build_forward_order(gf, v_cur);

    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
    //cb(q, "q", il);
@@ -1761,6 +1762,10 @@ struct clip_model_loader {
                            // qwen2 encoder is GQA, requires KEY_N_HEAD_KV
                            get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
                        }
+                        // unlimited-ocr shares the v1 projector but tiles up to 32
+                        get_u32(KEY_PREPROC_MIN_TILES, hparams.preproc_min_tiles, false);
+                        get_u32(KEY_PREPROC_MAX_TILES, hparams.preproc_max_tiles, false);
+                        GGML_ASSERT(hparams.preproc_min_tiles <= hparams.preproc_max_tiles);
                     } break;
                case PROJECTOR_TYPE_HUNYUANVL:
                    {
@@ -1909,6 +1914,9 @@ struct clip_model_loader {
                if (hparams.image_max_pixels > 0) {
                    LOG_INF("%s: image_max_pixels:   %d%s\n", __func__, hparams.image_max_pixels, hparams.custom_image_max_tokens > 0 ? " (custom value)" : "");
                }
+                if (hparams.preproc_max_tiles > 0) {
+                    LOG_INF("%s: preproc_tiles:      %d - %d\n", __func__, hparams.preproc_min_tiles, hparams.preproc_max_tiles);
+                }
            } else if (is_audio) {
                LOG_INF("\n--- audio hparams ---\n");
                LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);
@@ -30,20 +30,14 @@ namespace fs = std::filesystem;
 // internal helpers
 //

-#if defined(_WIN32)
-// A chunk can end in the middle of a multi-byte sequence, so the incomplete
-// tail is dropped before validating what precedes it.
-static bool is_utf8_text(const std::string & text) {
-    return is_valid_utf8(text.substr(0, validate_utf8(text)));
-}
-
-// A child process writes its output in the OEM code page, which is not UTF-8
-// on a western Windows install, so accented text reaches the JSON layer as
-// invalid bytes and is replaced there. Text that already decodes as UTF-8 is
-// returned untouched, so a child that emits UTF-8 is never decoded twice.
-// run() spawns without a console, so the console code page does not apply.
+// a child process writes in the OEM code page, so accented output would reach
+// the JSON layer as invalid bytes. run() spawns without a console, so the
+// console code page never applies
 static std::string console_output_to_utf8(const std::string & text) {
-    if (text.empty() || is_utf8_text(text)) {
+#if defined(_WIN32)
+    // a chunk can end mid sequence, so the incomplete tail is dropped first
+    if (text.empty() || is_valid_utf8(text.substr(0, validate_utf8(text)))) {
+        // never decode twice a child that already emits UTF-8
        return text;
    }

@@ -64,12 +58,10 @@ static std::string console_output_to_utf8(const std::string & text) {
    std::string utf8(utf8_len, '\0');
    WideCharToMultiByte(CP_UTF8, 0, wide.data(), wide_len, utf8.data(), utf8_len, nullptr, nullptr);
    return utf8;
-}
 #else
-static std::string console_output_to_utf8(const std::string & text) {
    return text;
-}
 #endif
+}

 json server_tool::to_json() const {
    return {
@@ -94,14 +86,30 @@ enum class list_kind {
    all,   // both
 };

+// a narrow path uses the active code page on Windows, so every crossing between
+// a std::string (always UTF-8 here) and fs::path is converted explicitly
+static fs::path path_from_utf8(const std::string & s) {
+    return fs::u8path(s);
+}
+
+// '/' separators on every platform: Windows accepts them, the web UI needs them
+static std::string path_to_utf8(const fs::path & p) {
+    const auto s = p.generic_u8string();
+    return std::string(s.begin(), s.end());
+}
+
 // home directory, read once at first use (getenv is not thread safe against setenv)
 static const std::string & home_dir() {
    static const std::string home = [] {
-        const char * h = getenv("HOME");
 #ifdef _WIN32
-        if (h == nullptr) h = getenv("USERPROFILE");
-#endif
+        // the narrow getenv would return the profile path in the active code page
+        const wchar_t * w = _wgetenv(L"HOME");
+        if (w == nullptr) w = _wgetenv(L"USERPROFILE");
+        return w ? path_to_utf8(fs::path(w)) : std::string();
+#else
+        const char * h = getenv("HOME");
        return h ? std::string(h) : std::string();
+#endif
    }();
    return home;
 }
@@ -140,11 +148,14 @@ public:
        std::string rel; // '/'-separated, relative to `base`
        bool is_dir = false;
    };
-    // entries relative to `base`; sets `err` if `base` isn't a directory
+    struct list_result {
+        std::vector<list_entry> entries;
+        std::string err;        // set when `base` is not a directory
+        bool truncated = false; // set when the walk could not see everything
+    };
+    // entries relative to `base`, which must already be resolved (absolute)
    // max_depth == 0 means unlimited, 1 means direct children of `base` only
-    // `base` must already be resolved (absolute); `caller_path` is the path the
-    // caller passed, used only for error messages
-    virtual std::vector<list_entry> list_entries(const std::string & base, const std::string & caller_path, int max_depth, list_kind kind, std::string & err, bool & truncated) const = 0;
+    virtual list_result list_entries(const std::string & base, int max_depth, list_kind kind) const = 0;
    // on_chunk, if set, is called with each chunk of output as it is read (before truncation cuts in);
    // returning false terminates the process early (e.g. the client disconnected)
    virtual exec_result run(
@@ -162,37 +173,47 @@ public:
    // expands a leading `~`, then resolves `path` against `cwd` (or the server
    // working directory when `cwd` is unset); the result is always absolute
    std::string resolve(const std::string & path) const override {
-        std::string p = expand_home(path);
-        if (fs::path(p).is_absolute()) {
-            return p;
+        const std::string p = expand_home(path);
+
+        fs::path full = path_from_utf8(p);
+        if (!full.is_absolute()) {
+            if (cwd.empty()) {
+                std::error_code ec;
+                const fs::path cur = fs::current_path(ec);
+                if (ec) return p;
+                full = cur / full;
+            } else {
+                full = path_from_utf8(cwd) / full;
+            }
        }
-        if (cwd.empty()) {
-            std::error_code ec;
-            fs::path cur = fs::current_path(ec);
-            if (ec) return p;
-            return (cur / p).string();
+
+        // drop "." and ".." so they never reach git or the client
+        full = full.lexically_normal();
+        // a trailing ".." normalizes to a path that ends with a separator
+        if (!full.has_filename() && full != full.root_path()) {
+            full = full.parent_path();
        }
-        return (fs::path(cwd) / p).string();
+        return path_to_utf8(full);
    }

    bool is_directory(const std::string & path) const override {
        std::error_code ec;
-        return fs::is_directory(resolve(path), ec) && !ec;
+        return fs::is_directory(path_from_utf8(resolve(path)), ec) && !ec;
    }

    bool is_regular_file(const std::string & path) const override {
        std::error_code ec;
-        return fs::is_regular_file(resolve(path), ec) && !ec;
+        return fs::is_regular_file(path_from_utf8(resolve(path)), ec) && !ec;
    }

    bool file_size(const std::string & path, uintmax_t & out_size) const override {
        std::error_code ec;
-        out_size = fs::file_size(resolve(path), ec);
+        out_size = fs::file_size(path_from_utf8(resolve(path)), ec);
        return !ec;
    }

    bool read_file(const std::string & path, std::string & out) const override {
-        std::ifstream f(resolve(path), std::ios::binary);
+        std::ifstream f(path_from_utf8(resolve(path)), std::ios::binary);
        if (!f) return false;
        std::ostringstream ss;
        ss << f.rdbuf();
@@ -202,7 +223,7 @@ public:

    bool write_file(const std::string & path, const std::string & content) const override {
        std::error_code ec;
-        fs::path fpath(resolve(path));
+        fs::path fpath = path_from_utf8(resolve(path));
        if (fpath.has_parent_path()) {
            fs::create_directories(fpath.parent_path(), ec);
            if (ec) return false;
@@ -213,13 +234,13 @@ public:
        return (bool) f;
    }

-    std::vector<list_entry> list_entries(const std::string & base, const std::string & caller_path, int max_depth, list_kind kind, std::string & err, bool & truncated) const override {
-        err.clear();
-        truncated = false;
+    list_result list_entries(const std::string & base, int max_depth, list_kind kind) const override {
+        list_result out;
+
        std::error_code ec;
        if (!fs::is_directory(base, ec) || ec) {
-            err = "path does not exist or is not a directory: " + caller_path;
-            return {};
+            out.err = "path does not exist or is not a directory";
+            return out;
        }

        const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(SERVER_TOOL_LIST_ENTRIES_TIMEOUT);
@@ -231,7 +252,6 @@ public:
                SERVER_TOOL_GIT_LS_FILES_MAX_OUTPUT, SERVER_TOOL_LIST_ENTRIES_TIMEOUT);

            if (res.exit_code == 0 && !res.timed_out) {
-                std::vector<list_entry> result;
                std::istringstream iss(res.output);
                std::string line;
                while (std::getline(iss, line)) {
@@ -239,15 +259,16 @@ public:
                    if (line.empty()) continue;
                    std::replace(line.begin(), line.end(), '\\', '/');
                    if (max_depth > 0 && entry_depth(line) > max_depth) continue;
-                    if (is_regular_file((fs::path(base) / line).string())) {
-                        result.push_back({line, false});
+                    if (is_regular_file(path_to_utf8(path_from_utf8(base) / path_from_utf8(line)))) {
+                        out.entries.push_back({line, false});
                    }
                }
-                return result;
+                return out;
            }
        }

-        return list_entries_fallback(base, max_depth, kind, deadline, truncated);
+        out.entries = list_entries_fallback(base, max_depth, kind, deadline, out.truncated);
+        return out;
    }

    exec_result run(
@@ -326,6 +347,42 @@ public:
 private:
    std::string cwd;

+    // a link can point back to an ancestor and loop forever, so it is never walked
+    static bool is_link(const fs::directory_entry & entry) {
+        std::error_code ec;
+        if (entry.is_symlink(ec) || ec) {
+            return true;
+        }
+#if defined(_WIN32)
+        // a junction looks like a plain directory to std::filesystem, so read the reparse tag
+        WIN32_FIND_DATAW data;
+        const HANDLE h = FindFirstFileW(entry.path().c_str(), &data);
+        if (h == INVALID_HANDLE_VALUE) {
+            return false;
+        }
+        FindClose(h);
+        if ((data.dwFileAttributes & FILE_ATTRIBUTE_REPARSE_POINT) == 0) {
+            return false;
+        }
+        // other reparse points (cloud placeholder, dedup stub) are real directories
+        return data.dwReserved0 == IO_REPARSE_TAG_SYMLINK || data.dwReserved0 == IO_REPARSE_TAG_MOUNT_POINT;
+#else
+        return false;
+#endif
+    }
+
+    // NTFS is case insensitive, so Build and build are the same directory
+    static std::string get_effective_name(const std::string & fname) {
+#if defined(_WIN32)
+        std::string lowered = fname;
+        std::transform(lowered.begin(), lowered.end(), lowered.begin(),
+                       [](unsigned char c) { return (char) std::tolower(c); });
+        return lowered;
+#else
+        return fname;
+#endif
+    }
+
    static const std::unordered_set<std::string> & junk_dir_names() {
        static const std::unordered_set<std::string> names = {
            ".git", ".svn", ".hg", "node_modules", "__pycache__",
@@ -337,46 +394,53 @@ private:
    std::vector<list_entry> list_entries_fallback(const std::string & base, int max_depth, list_kind kind,
                                                  std::chrono::steady_clock::time_point deadline, bool & truncated) const {
        std::vector<list_entry> result;
-        std::error_code ec;

        std::vector<std::tuple<fs::path, fs::path, int>> stack;
-        stack.emplace_back(fs::path(base), fs::path(), 0);
+        stack.emplace_back(path_from_utf8(base), fs::path(), 0);

        while (!stack.empty()) {
-            auto [dir, rel_dir, depth] = stack.back();
+            if (std::chrono::steady_clock::now() >= deadline) {
+                truncated = true;
+                return result;
+            }
+
+            auto [dir, rel_dir, depth] = std::move(stack.back());
            stack.pop_back();

-            // the throwing increment would escape the tool on a directory that
-            // goes away mid walk, so step the iterator explicitly
+            std::error_code ec;
+            // step the iterator by hand: the throwing increment escapes on a directory that goes away
            fs::directory_iterator it(dir, fs::directory_options::skip_permission_denied, ec);
+            // permission errors are skipped above, so this is a subtree the caller never sees
+            if (ec) {
+                truncated = true;
+                continue;
+            }
            for (const fs::directory_iterator end; it != end; it.increment(ec)) {
-                if (ec) break;
+                if (ec) {
+                    truncated = true;
+                    break;
+                }
                if (std::chrono::steady_clock::now() >= deadline) {
                    truncated = true;
                    return result;
                }
                const fs::directory_entry & entry = *it;
-                std::string fname = entry.path().filename().string();
+                const fs::path fname = entry.path().filename();
                std::error_code tec;
-                if (entry.is_directory(tec)) {
-                    std::string rel = (rel_dir / fname).string();
-                    std::replace(rel.begin(), rel.end(), '\\', '/');
+                const bool is_dir = entry.is_directory(tec);
+                if (tec) continue;
+                if (is_dir) {
                    if (kind == list_kind::dirs || kind == list_kind::all) {
-                        result.push_back({rel, true});
+                        result.push_back({path_to_utf8(rel_dir / fname), true});
                    }
-                    // junk directories stay selectable but are never walked: they
-                    // hold nothing worth searching and can be enormous
-                    if (junk_dir_names().count(fname) > 0) continue;
-                    // do not descend into symlinks: a link can point back to an
-                    // ancestor and loop forever
-                    if (!entry.is_symlink(tec) && (max_depth == 0 || depth + 1 < max_depth)) {
+                    // junk directories stay selectable but are never walked: they can be enormous
+                    if (junk_dir_names().count(get_effective_name(path_to_utf8(fname))) > 0) continue;
+                    if (!is_link(entry) && (max_depth == 0 || depth + 1 < max_depth)) {
                        stack.emplace_back(entry.path(), rel_dir / fname, depth + 1);
                    }
                } else if (entry.is_regular_file(tec)) {
-                    std::string rel = (rel_dir / fname).string();
-                    std::replace(rel.begin(), rel.end(), '\\', '/');
                    if (kind == list_kind::files || kind == list_kind::all) {
-                        result.push_back({rel, false});
+                        result.push_back({path_to_utf8(rel_dir / fname), false});
                    }
                }
            }
@@ -394,7 +458,7 @@ static std::unique_ptr<tools_io> make_tools_io(const json & params) {
 // no '/' in pattern -> match basename at any depth; else match full relative path
 static bool path_glob_match(const std::string & pattern, const std::string & rel_path) {
    if (pattern.find('/') == std::string::npos) {
-        return glob_match(pattern, fs::path(rel_path).filename().string());
+        return glob_match(pattern, path_to_utf8(path_from_utf8(rel_path).filename()));
    }
    if (pattern == "**" || pattern.rfind("**/", 0) == 0 || pattern.rfind('/', 0) == 0) {
        return glob_match(pattern, rel_path);
@@ -491,7 +555,7 @@ struct server_tool_read_file : server_tool {
 // file_glob_search: find files matching a glob pattern under a base directory
 //

-static constexpr size_t SERVER_TOOL_FILE_SEARCH_MAX_RESULTS = 100;
+static constexpr int SERVER_TOOL_FILE_SEARCH_MAX_RESULTS = 100;
 static constexpr const char * SERVER_TOOL_FILE_SEARCH_TYPE_FILE = "file";
 static constexpr const char * SERVER_TOOL_FILE_SEARCH_TYPE_DIR  = "dir";
 static constexpr const char * SERVER_TOOL_FILE_SEARCH_TYPE_ALL  = "all";
@@ -525,7 +589,7 @@ struct server_tool_file_glob_search : server_tool {
                        {"exclude",   {{"type", "string"},  {"description", "Glob pattern for files to exclude"}}},
                        {"type",      {{"type", "string"},  {"description", "Entry type to return: \"file\" (default), \"dir\" or \"all\""}}},
                        {"max_depth", {{"type", "integer"}, {"description", "Maximum depth to descend into subdirectories (default: 0 = unlimited; 1 = direct children only)"}}},
-                        {"limit",     {{"type", "integer"}, {"description", string_format("Maximum number of results to return (default %zu; values below 1 fall back to the default)", SERVER_TOOL_FILE_SEARCH_MAX_RESULTS)}}},
+                        {"limit",     {{"type", "integer"}, {"description", string_format("Maximum number of results to return, capped at %d (default %d)", SERVER_TOOL_FILE_SEARCH_MAX_RESULTS, SERVER_TOOL_FILE_SEARCH_MAX_RESULTS)}}},
                    }},
                    {"required", json::array({"path"})},
                }},
@@ -536,17 +600,18 @@ struct server_tool_file_glob_search : server_tool {
    json invoke(json params, server_tool::stream *) const override {
        auto io = make_tools_io(params);

-        std::string base      = io->resolve(params.at("path").get<std::string>());
-        // normalize to forward slashes so the web UI (which assumes '/') can
-        // join the relative entries into absolute paths on Windows too
-        std::replace(base.begin(), base.end(), '\\', '/');
+        const std::string path = params.at("path").get<std::string>();
+
+        std::string base      = io->resolve(path);
        std::string include   = json_value(params, "include", std::string("**"));
        std::string exclude   = json_value(params, "exclude", std::string(""));
        std::string type      = json_value(params, "type",    std::string("file"));
        int         max_depth = std::max(0, json_value(params, "max_depth", 0));
-        int         limit     = json_value(params, "limit", (int) SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);
-        if (limit < 1) limit = SERVER_TOOL_FILE_SEARCH_MAX_RESULTS;
-        limit = std::min(limit, (int) SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);
+        const int   limit_req = json_value(params, "limit", SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);
+        if (limit_req < 1) {
+            return {{"error", "invalid limit: " + std::to_string(limit_req) + " (expected 1 or more)"}};
+        }
+        const int   limit     = std::min(limit_req, SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);

        list_kind kind;
        if (type == SERVER_TOOL_FILE_SEARCH_TYPE_FILE) {
@@ -559,15 +624,13 @@ struct server_tool_file_glob_search : server_tool {
            return {{"error", "invalid type: " + type + " (expected \"file\", \"dir\" or \"all\")"}};
        }

-        std::string err;
-        bool truncated = false;
-        auto entries = io->list_entries(base, params.at("path").get<std::string>(), max_depth, kind, err, truncated);
-        if (!err.empty()) {
-            return {{"error", err}};
+        const auto listing = io->list_entries(base, max_depth, kind);
+        if (!listing.err.empty()) {
+            return {{"error", listing.err + ": " + path}};
        }

        std::vector<tools_io::list_entry> matches;
-        for (const auto & entry : entries) {
+        for (const auto & entry : listing.entries) {
            if (!path_glob_match(include, entry.rel)) continue;
            if (!exclude.empty() && path_glob_match(exclude, entry.rel)) continue;
            matches.push_back(entry);
@@ -592,8 +655,8 @@ struct server_tool_file_glob_search : server_tool {
                "[%zu results limit reached (%zu total matches). Refine the glob pattern to narrow the search.]\n",
                shown, total);
        }
-        if (truncated) {
-            output_text << "[search timed out, results truncated]\n";
+        if (listing.truncated) {
+            output_text << "[results truncated: time budget or unreadable directory]\n";
        }

        // `base` is always absolute (resolve falls back to the server cwd), so
@@ -688,16 +751,14 @@ struct server_tool_grep_search : server_tool {
        if (io->is_regular_file(abs_path)) {
            files.emplace_back(abs_path, path);
        } else if (io->is_directory(abs_path)) {
-            std::string err;
-            bool truncated = false;
-            auto candidates = io->list_entries(abs_path, path, 0, list_kind::files, err, truncated);
-            if (!err.empty()) {
-                return {{"error", err}};
+            const auto listing = io->list_entries(abs_path, 0, list_kind::files);
+            if (!listing.err.empty()) {
+                return {{"error", listing.err + ": " + path}};
            }
-            for (const auto & entry : candidates) {
+            for (const auto & entry : listing.entries) {
                if (!path_glob_match(include, entry.rel)) continue;
                if (!exclude.empty() && path_glob_match(exclude, entry.rel)) continue;
-                files.emplace_back((fs::path(abs_path) / entry.rel).string(), entry.rel);
+                files.emplace_back(path_to_utf8(path_from_utf8(abs_path) / path_from_utf8(entry.rel)), entry.rel);
            }
        } else {
            return {{"error", "path does not exist: " + path}};
@@ -1306,7 +1367,7 @@ struct server_tool_get_info : server_tool {
        std::string cwd = json_value(params, "cwd", std::string());
        if (cwd.empty()) {
            std::error_code ec;
-            cwd = fs::current_path(ec).string();
+            cwd = path_to_utf8(fs::current_path(ec));
        }

        return {
@@ -214,6 +214,27 @@ def test_tools_builtin_file_glob_search_max_depth_and_limit(tmp_path):
    assert "Total matches: 3" in res["plain_text_response"]


+def test_tools_builtin_file_glob_search_junk_dirs(tmp_path):
+    global server
+    server.start()
+
+    (tmp_path / "build" / "nested").mkdir(parents=True)
+    (tmp_path / "build" / "artifact.txt").write_text("built")
+    (tmp_path / "src").mkdir()
+    (tmp_path / "src" / "main.cpp").write_text("int main() {}")
+
+    # a junk directory stays selectable as a working directory
+    res = call_tool("file_glob_search", {"path": str(tmp_path), "type": "dir", "max_depth": 1})
+    assert "build" in [e["path"] for e in res["entries"]]
+
+    # but it is never walked, so nothing inside it shows up
+    res = call_tool("file_glob_search", {"path": str(tmp_path), "type": "all"})
+    paths = [e["path"] for e in res["entries"]]
+    assert "src/main.cpp" in paths
+    assert "build/artifact.txt" not in paths
+    assert "build/nested" not in paths
+
+
 def test_tools_builtin_file_glob_search_rejects_invalid_type(tmp_path):
    global server
    server.start()
@@ -87,8 +87,9 @@
 	let searchSeq = 0;

 	// Cache of the last file_glob_search result per (parent, include, max_depth),
-	// so repeated queries in the same directory don't re-walk the tree. Entries
-	// expire after a short TTL.
+	// so repeated queries in the same directory don't re-walk the tree. Entering
+	// a directory hits it every time: the children listed for an exactly typed
+	// segment are what the next keystroke, the trailing slash, asks for again.
 	const SEARCH_CACHE_TTL_MS = 2000;
 	const searchCache = new SvelteMap<string, { results: GlobEntry[]; base: string; at: number }>();

@@ -161,7 +162,11 @@
 		if (typeof res.error === 'string') return { base: '', entries: [], error: res.error };
 		const base = typeof res.base === 'string' ? res.base : '';
 		const entries = Array.isArray(res.entries) ? (res.entries as GlobEntry[]) : [];
-		searchCache.set(key, { results: entries, base, at: Date.now() });
+		const now = Date.now();
+		for (const [k, v] of searchCache) {
+			if (now - v.at >= SEARCH_CACHE_TTL_MS) searchCache.delete(k);
+		}
+		searchCache.set(key, { results: entries, base, at: now });
 		return { base, entries };
 	}
Author	SHA1	Message	Date
Ruben OrtlamandGitHub	803b7fcae8	vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors (#26371 ) * vulkan: add debug tooling to get more information about a DeviceLost error * fix submission threshold applied too late * use logging macros, throw instead of aborting * clean up circular dependency	2026-08-06 10:24:13 +03:00
PascalandGitHub	c8e03ce812	mtmd/ggml: add ggml_build_forward_order (#26649 ) * ggml: add ggml_build_forward_order ggml_build_forward_expand marks the tensor and all its ancestors for compute, so using it as a pure ordering hint (keeping q, k and v together) defeats ggml_build_forward_select: the unselected branch is forced to run with inputs that were never uploaded. In the mtmd audio graph this makes GEN_WAV calls execute the GEN_CODE branch with a stale inp_code0, hitting the get_rows bound assert on CPU. Add ggml_build_forward_order, which inserts nodes without the compute flag; the flag is restored when the branch is actually selected. Switch the q/k/v hints in clip_graph::build_attn to it. * nit: reduce comments (AGENTS.md)	2026-08-06 00:47:59 +02:00
PascalandGitHub	f9e832c10e	server: harden the file_glob_search directory walk (#26626 ) * server: don't walk Windows junctions in file_glob_search std::filesystem reports a junction as a plain directory, so the symlink guard misses it and a junction pointing back at an ancestor is walked until the path length gives out read the reparse tag and treat a symlink and a mount point as links, leaving any other reparse point walkable so cloud placeholders and dedup stubs still get searched look junk directory names up case insensitively on Windows, where NTFS makes Build the same directory as build test that a junk directory stays selectable while its contents stay out of search results * server: report a directory the walk could not read a directory that fails to open or to iterate was skipped in silence, so a caller got a listing that looked complete while a whole subtree was missing: a path over the platform limit, a volume going away, a name the filesystem rejects skip_permission_denied never reaches this path, so an error here is an incomplete answer rather than a deliberate omission, and it now sets the truncated flag * server: simplify the file_glob_search listing plumbing return a small result struct instead of two out params and a caller path that only fed an error string, taking list_entries from six parameters down to three scope the error code to the directory being read, act on the status code the entry lookups already returned, and treat an unreadable link state as a link so the walk never descends on a guess check the deadline when a directory is popped, not only per entry, so a tree of empty directories cannot outlive the budget read the path parameter once, and reject an invalid limit the way an invalid type is already rejected, instead of silently falling back normalize the resolved path, so a "." or ".." a caller typed reaches neither git nor the client, and return the generic path form with '/' separators on every platform, so the base sent to clients no longer needs a local fixup * ui: expire cached picker searches the cache grew for the lifetime of the component: entries went stale after the TTL but were never removed, so every distinct query typed in a session stayed in memory drop expired entries when a new result is stored * server: address review from @ngxson trim comments to one line each, and drop two that restate the code rename junk_lookup_name to get_effective_name, and move it and the link check to private static members next to junk_dir_names merge the Windows and Linux link checks into one is_link, so symlinks are checked everywhere and junctions only add to it on Windows * server: convert tool paths as UTF-8 on Windows a narrow path uses the active code page there, so a file name came back mangled and a path with an accent could not be opened at all convert explicitly at every crossing between a std::string, which always carries UTF-8 here, and fs::path read the home directory through the wide environment, since the narrow one returns the profile path in the active code page too the walker no longer normalizes separators by hand, since paths now come back in generic form * server: fold the platform branch inside console_output_to_utf8 match the shape of the other helpers, one definition with the #if inside, instead of two definitions wrapped in #if and #else inline the single caller helper and trim the comment	2026-08-05 21:31:54 +02:00
Niklas WenzelandGitHub	360e1349f0	tests: re-enable MiniMax M3 in `test-llama-archs` (#26633 )	2026-08-05 17:58:34 +02:00
Saba FallahandGitHub	b06aa774c0	mtmd: Unlimited-OCR fix max_tiles, setting in converter (#25614 )	2026-08-05 15:30:14 +02:00
Aldehir RojasandGitHub	cd0fa6051a	grammar : degrade max repetition >= 2000 to unbounded (#26613 )	2026-08-05 07:39:10 -05:00