From d0132a680a3364580b730bca16cd42aa7314423f Mon Sep 17 00:00:00 2001 From: Radoslav Gerganov Date: Wed, 26 Aug 2026 17:34:46 +0300 Subject: [PATCH 1/9] rpc : implement event and async backend APIs (#18626) * rpc : implement event and async backend APIs * cache responses from RPC_CMD_GET_ALLOC_SIZE --- ggml/include/ggml-rpc.h | 4 +- ggml/src/ggml-rpc/ggml-rpc.cpp | 605 ++++++++++++++++++++++++--------- 2 files changed, 443 insertions(+), 166 deletions(-) diff --git a/ggml/include/ggml-rpc.h b/ggml/include/ggml-rpc.h index 059e449626..cbfe400139 100644 --- a/ggml/include/ggml-rpc.h +++ b/ggml/include/ggml-rpc.h @@ -6,8 +6,8 @@ extern "C" { #endif -#define RPC_PROTO_MAJOR_VERSION 5 -#define RPC_PROTO_MINOR_VERSION 1 +#define RPC_PROTO_MAJOR_VERSION 6 +#define RPC_PROTO_MINOR_VERSION 0 #define RPC_PROTO_PATCH_VERSION 0 #ifdef __cplusplus diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp index 69a8a08ae1..9aa5883d80 100644 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp @@ -9,6 +9,9 @@ #include #include #include +#include +#include +#include #include #include #include @@ -17,6 +20,8 @@ #include #include #include +#include +#include static const char * RPC_DEBUG = std::getenv("GGML_RPC_DEBUG"); @@ -72,6 +77,7 @@ enum rpc_cmd { RPC_CMD_DEVICE_COUNT, RPC_CMD_GRAPH_RECOMPUTE, RPC_CMD_MEMSET_TENSOR, + RPC_CMD_NONE, RPC_CMD_COUNT, }; @@ -223,24 +229,24 @@ struct ggml_backend_rpc_buffer_type_context { size_t max_size; }; +class rpc_dispatcher; struct ggml_backend_rpc_context { - std::string endpoint; - uint32_t device; - std::string name; + std::shared_ptr dispatcher; + uint32_t device; + std::string name; }; struct ggml_backend_rpc_buffer_context { - std::shared_ptr sock; - void * base_ptr; - uint64_t remote_ptr; + std::shared_ptr dispatcher; + void * base_ptr; + uint64_t remote_ptr; }; // RPC helper functions // Computes FNV-1a hash of the data -static uint64_t fnv_hash(const uint8_t * data, size_t len) { +static uint64_t fnv_hash(const uint8_t * data, size_t len, uint64_t hash = 0xcbf29ce484222325ULL) { const uint64_t fnv_prime = 0x100000001b3ULL; - uint64_t hash = 0xcbf29ce484222325ULL; for (size_t i = 0; i < len; ++i) { hash ^= data[i]; @@ -357,44 +363,248 @@ static bool negotiate_hello(const std::shared_ptr & sock) { return true; } -static std::shared_ptr get_socket(const std::string & endpoint) { - static std::mutex mutex; - std::lock_guard lock(mutex); - static std::unordered_map> sockets; +template +class message_queue { +public: + message_queue() {} - auto it = sockets.find(endpoint); - if (it != sockets.end()) { - if (auto sock = it->second.lock()) { - return sock; + bool push(const T &value) { + std::unique_lock lock(mutex); + if (interrupted) { + return false; } + queue.push(value); + cvar.notify_all(); + return true; } + + bool pop(T* out) { + std::unique_lock lock(mutex); + cvar.wait(lock, [this] { return !queue.empty() || interrupted; }); + if (interrupted) { + return false; + } + *out = queue.front(); + queue.pop(); + return true; + } + + void interrupt() { + std::unique_lock lock(mutex); + interrupted = true; + lock.unlock(); + cvar.notify_all(); + } + +private: + bool interrupted = false; + std::queue queue; + std::mutex mutex; + std::condition_variable cvar; +}; + +class rpc_dispatcher { +public: + rpc_dispatcher() { + } + + void send(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size); + void send(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size, void * output, size_t output_size); + void send_async(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size); + void send_async(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size, void * output, size_t output_size); + + ggml_backend_event_t event_new(ggml_backend_dev_t dev); + void event_free(ggml_backend_event_t event); + void event_synchronize(ggml_backend_event_t event); + void event_record(ggml_backend_event_t event); + void synchronize(); + + void start(const std::string & endpoint); + void work(); + + ~rpc_dispatcher(); + +private: + struct rpc_msg { + rpc_cmd cmd; + std::shared_ptr input; + size_t input_size; + void * output; + size_t output_size; + std::promise completion; + }; + using rpc_msg_ptr = std::shared_ptr; + using rpc_msg_queue = message_queue; + struct rpc_event { + rpc_msg_ptr msg; + std::shared_future sf; + }; + rpc_msg_queue queue; + socket_ptr sock; + std::atomic_bool running; + std::thread thread; +}; + +static void rpc_dispatcher_trampoline(rpc_dispatcher * dispatcher) +{ + dispatcher->work(); +} + +void rpc_dispatcher::send(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size) { + auto msg = std::make_shared(); + msg->cmd = cmd; + msg->input = input; + msg->input_size = input_size; + msg->output = nullptr; + msg->output_size = 0; + GGML_ASSERT(queue.push(msg)); + auto future = msg->completion.get_future(); + future.wait(); +} + +void rpc_dispatcher::send_async(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size) { + auto msg = std::make_shared(); + msg->cmd = cmd; + msg->input = input; + msg->input_size = input_size; + msg->output = nullptr; + msg->output_size = 0; + GGML_ASSERT(queue.push(msg)); +} + +void rpc_dispatcher::send(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size, void * output, size_t output_size) { + auto msg = std::make_shared(); + msg->cmd = cmd; + msg->input = input; + msg->input_size = input_size; + msg->output = output; + msg->output_size = output_size; + GGML_ASSERT(queue.push(msg)); + auto future = msg->completion.get_future(); + future.wait(); +} + +void rpc_dispatcher::send_async(enum rpc_cmd cmd, std::shared_ptr input, size_t input_size, void * output, size_t output_size) { + auto msg = std::make_shared(); + msg->cmd = cmd; + msg->input = input; + msg->input_size = input_size; + msg->output = output; + msg->output_size = output_size; + GGML_ASSERT(queue.push(msg)); +} + +ggml_backend_event_t rpc_dispatcher::event_new(ggml_backend_dev_t dev) { + rpc_event * ev = new rpc_event; + ev->msg = std::make_shared(); + ev->msg->cmd = RPC_CMD_NONE; + ev->sf = ev->msg->completion.get_future().share(); + GGML_ASSERT(queue.push(ev->msg)); + return new ggml_backend_event { + /* .device = */ dev, + /* .context = */ ev, + }; +} + +void rpc_dispatcher::event_free(ggml_backend_event_t event) { + rpc_event * ev = (rpc_event *)event->context; + delete ev; +} + +void rpc_dispatcher::event_synchronize(ggml_backend_event_t event) { + rpc_event * ev = (rpc_event *)event->context; + ev->sf.wait(); +} + +void rpc_dispatcher::event_record(ggml_backend_event_t event) { + rpc_event * ev = (rpc_event *)event->context; + ev->msg = std::make_shared(); + ev->msg->cmd = RPC_CMD_NONE; + ev->sf = ev->msg->completion.get_future().share(); + GGML_ASSERT(queue.push(ev->msg)); +} + +void rpc_dispatcher::synchronize() { + // to ensure all messages are processed, submit dummy message and wait for it to complete + auto msg = std::make_shared(); + msg->cmd = RPC_CMD_NONE; + GGML_ASSERT(queue.push(msg)); + msg->completion.get_future().wait(); +} + +void rpc_dispatcher::start(const std::string & endpoint) { std::string host; int port; if (!parse_endpoint(endpoint, host, port)) { - GGML_LOG_ERROR("Failed to parse endpoint: %s\n", endpoint.c_str()); - return nullptr; + GGML_ABORT("Failed to parse endpoint: %s\n", endpoint.c_str()); + } + if (!rpc_transport_init()) { + GGML_ABORT("RPC transport initialization failed\n"); } - if (!rpc_transport_init()) { - return nullptr; - } - auto sock = socket_t::connect(host.c_str(), port); + sock = socket_t::connect(host.c_str(), port); if (sock == nullptr) { - return nullptr; + GGML_ABORT("Failed to connect to %s\n", endpoint.c_str()); } if (!negotiate_hello(sock)) { - return nullptr; + GGML_ABORT("RPC handshake failed for %s\n", endpoint.c_str()); } LOG_DBG("[%s] connected to %s\n", __func__, endpoint.c_str()); - sockets[endpoint] = sock; - return sock; + running = true; + thread = std::thread(rpc_dispatcher_trampoline, this); +} + +void rpc_dispatcher::work() { + while (running) { + rpc_msg_ptr msg_ptr; + if (!queue.pop(&msg_ptr)) { + break; + } + if (msg_ptr->cmd != RPC_CMD_NONE) { + if (msg_ptr->output) { + bool status = send_rpc_cmd(sock, msg_ptr->cmd, msg_ptr->input.get(), msg_ptr->input_size, msg_ptr->output, msg_ptr->output_size); + RPC_STATUS_ASSERT(status); + } else { + bool status = send_rpc_cmd(sock, msg_ptr->cmd, msg_ptr->input.get(), msg_ptr->input_size); + RPC_STATUS_ASSERT(status); + } + } + msg_ptr->completion.set_value(); + } +} + +rpc_dispatcher::~rpc_dispatcher() { + running = false; + queue.interrupt(); + sock = nullptr; + if (thread.joinable()) { + thread.join(); + } +} + +static std::shared_ptr get_dispatcher(const std::string & endpoint) { + static std::mutex mutex; + std::lock_guard lock(mutex); + static std::unordered_map> dispatchers; + + auto it = dispatchers.find(endpoint); + if (it != dispatchers.end()) { + if (auto dispatcher = it->second.lock()) { + return dispatcher; + } + } + + auto dispatcher = std::make_shared(); + dispatcher->start(endpoint); + dispatchers[endpoint] = dispatcher; + return dispatcher; } static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; - rpc_msg_free_buffer_req request = {ctx->remote_ptr}; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0); - RPC_STATUS_ASSERT(status); + auto request = std::make_shared(); + request->remote_ptr = ctx->remote_ptr; + ctx->dispatcher->send(RPC_CMD_FREE_BUFFER, request, sizeof(*request)); delete ctx; } @@ -403,10 +613,10 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) { if (ctx->base_ptr != nullptr) { return ctx->base_ptr; } - rpc_msg_buffer_get_base_req request = {ctx->remote_ptr}; + auto request = std::make_shared(); + request->remote_ptr = ctx->remote_ptr; rpc_msg_buffer_get_base_rsp response; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + ctx->dispatcher->send(RPC_CMD_BUFFER_GET_BASE, request, sizeof(*request), &response, sizeof(response)); ctx->base_ptr = reinterpret_cast(response.base_ptr); return ctx->base_ptr; } @@ -463,12 +673,9 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_ // Due to bandwidth constraints, we only call the server init tensor functions if necessary. // In particular, only quantized tensors need padding if (ggml_is_quantized(tensor->type) && (tensor->ne[0] % 512 != 0) && (tensor->view_src == nullptr)) { - rpc_msg_init_tensor_req request; - - request.tensor = serialize_tensor(tensor); - - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_INIT_TENSOR, &request, sizeof(request), nullptr, 0); - RPC_STATUS_ASSERT(status); + auto request = std::make_shared(); + request->tensor = serialize_tensor(tensor); + ctx->dispatcher->send(RPC_CMD_INIT_TENSOR, request, sizeof(*request)); } return GGML_STATUS_SUCCESS; } @@ -476,27 +683,24 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_ static void ggml_backend_rpc_buffer_memset_tensor( ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; - rpc_msg_memset_tensor_req request = { - /* .tensor = */ serialize_tensor(tensor), - /* .offset = */ offset, - /* .size = */ size, - /* .value = */ value, - }; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_MEMSET_TENSOR, &request, sizeof(request), nullptr, 0); - RPC_STATUS_ASSERT(status); + auto request = std::make_shared(); + request->tensor = serialize_tensor(tensor); + request->offset = offset; + request->size = size; + request->value = value; + ctx->dispatcher->send(RPC_CMD_MEMSET_TENSOR, request, sizeof(*request)); } static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; rpc_tensor rpc_tensor = serialize_tensor(tensor); if (size > HASH_THRESHOLD) { - rpc_msg_set_tensor_hash_req request; - request.tensor = rpc_tensor; - request.offset = offset; - request.hash = fnv_hash((const uint8_t*)data, size); + auto request = std::make_shared(); + request->tensor = rpc_tensor; + request->offset = offset; + request->hash = fnv_hash((const uint8_t*)data, size); rpc_msg_set_tensor_hash_rsp response; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + ctx->dispatcher->send(RPC_CMD_SET_TENSOR_HASH, request, sizeof(*request), &response, sizeof(response)); if (response.result) { // the server has the same data, no need to send it return; @@ -504,22 +708,21 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm } // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size; - std::vector input(input_size, 0); - memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor)); - memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset)); - memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size); - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size()); - RPC_STATUS_ASSERT(status); + uint8_t * input = new uint8_t[input_size](); + memcpy(input, &rpc_tensor, sizeof(rpc_tensor)); + memcpy(input + sizeof(rpc_tensor), &offset, sizeof(offset)); + memcpy(input + sizeof(rpc_tensor) + sizeof(offset), data, size); + std::shared_ptr input_ptr(input, std::default_delete()); + ctx->dispatcher->send(RPC_CMD_SET_TENSOR, input_ptr, input_size); } static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; - rpc_msg_get_tensor_req request; - request.tensor = serialize_tensor(tensor); - request.offset = offset; - request.size = size; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size); - RPC_STATUS_ASSERT(status); + auto request = std::make_shared(); + request->tensor = serialize_tensor(tensor); + request->offset = offset; + request->size = size; + ctx->dispatcher->send(RPC_CMD_GET_TENSOR, request, sizeof(*request), data, size); } static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { @@ -529,16 +732,15 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context; ggml_backend_buffer_t dst_buffer = dst->buffer; ggml_backend_rpc_buffer_context * dst_ctx = (ggml_backend_rpc_buffer_context *)dst_buffer->context; - if (src_ctx->sock != dst_ctx->sock) { + if (src_ctx->dispatcher != dst_ctx->dispatcher) { return false; } ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; - rpc_msg_copy_tensor_req request; - request.src = serialize_tensor(src); - request.dst = serialize_tensor(dst); + auto request = std::make_shared(); + request->src = serialize_tensor(src); + request->dst = serialize_tensor(dst); rpc_msg_copy_tensor_rsp response; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + ctx->dispatcher->send(RPC_CMD_COPY_TENSOR, request, sizeof(*request), &response, sizeof(response)); return response.result; } return false; @@ -546,9 +748,10 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; - rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value}; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0); - RPC_STATUS_ASSERT(status); + auto request = std::make_shared(); + request->remote_ptr = ctx->remote_ptr; + request->value = value; + ctx->dispatcher->send(RPC_CMD_BUFFER_CLEAR, request, sizeof(*request)); } static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = { @@ -572,15 +775,17 @@ static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; - rpc_msg_alloc_buffer_req request = {buft_ctx->device, size}; + auto request = std::make_shared(); + request->device = buft_ctx->device; + request->size = size; rpc_msg_alloc_buffer_rsp response; - auto sock = get_socket(buft_ctx->endpoint); - bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + + auto dispatcher = get_dispatcher(buft_ctx->endpoint); + dispatcher->send(RPC_CMD_ALLOC_BUFFER, request, sizeof(*request), &response, sizeof(response)); if (response.remote_ptr != 0) { ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft, ggml_backend_rpc_buffer_interface, - new ggml_backend_rpc_buffer_context{sock, nullptr, response.remote_ptr}, + new ggml_backend_rpc_buffer_context{dispatcher, nullptr, response.remote_ptr}, response.remote_size); return buffer; } else { @@ -588,11 +793,11 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back } } -static size_t get_alignment(const std::shared_ptr & sock, uint32_t device) { - rpc_msg_get_alignment_req request = {device}; +static size_t get_alignment(const std::shared_ptr & dispatcher, uint32_t device) { + auto request = std::make_shared(); + request->device = device; rpc_msg_get_alignment_rsp response; - bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + dispatcher->send(RPC_CMD_GET_ALIGNMENT, request, sizeof(*request), &response, sizeof(response)); return response.alignment; } @@ -601,11 +806,11 @@ static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_typ return buft_ctx->alignment; } -static size_t get_max_size(const std::shared_ptr & sock, uint32_t device) { - rpc_msg_get_max_size_req request = {device}; +static size_t get_max_size(const std::shared_ptr & dispatcher, uint32_t device) { + auto request = std::make_shared(); + request->device = device; rpc_msg_get_max_size_rsp response; - bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + dispatcher->send(RPC_CMD_GET_MAX_SIZE, request, sizeof(*request), &response, sizeof(response)); return response.max_size; } @@ -628,23 +833,63 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_ty if (rpc_get) { ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; - auto sock = get_socket(buft_ctx->endpoint); - rpc_msg_get_alloc_size_req request = { - /*.device =*/ buft_ctx->device, - /*.tensor =*/ serialize_tensor(tensor), - /*.srcs =*/ {}, + // Cache key for calls to read the alloc_size. + // We deliberately exclude src tensor dimensions from the key because: + // 1. For CPU backends, alloc_size = ggml_nbytes(output) regardless of src shapes + // 2. For GPU backends, the reservation graph uses max dimensions, so the + // cached value from reservation is always >= any subsequent request + // 3. Including src dims causes cache misses per-ubatch (e.g. growing KV cache) + // which blocks the main thread behind in-flight GRAPH_COMPUTE commands + struct alloc_size_cache_key { + uint32_t device; + uint32_t type; + uint32_t op; + int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)]; + uint32_t ne[GGML_MAX_DIMS]; }; + alloc_size_cache_key key = {}; + key.device = buft_ctx->device; + key.type = tensor->type; + key.op = tensor->op; + memcpy(key.op_params, tensor->op_params, sizeof(key.op_params)); + for (int i = 0; i < GGML_MAX_DIMS; i++) { + key.ne[i] = (uint32_t)tensor->ne[i]; + } + + uint64_t cache_hash = fnv_hash((const uint8_t *)&key, sizeof(key)); + cache_hash = fnv_hash((const uint8_t *)buft_ctx->endpoint.data(), buft_ctx->endpoint.size(), cache_hash); + + // alloc sizes are immutable for a given tensor configuration + static std::mutex cache_mutex; + static std::unordered_map cache; + + { + std::lock_guard lock(cache_mutex); + auto it = cache.find(cache_hash); + if (it != cache.end()) { + return it->second; + } + } + + auto request = std::make_shared(); + request->device = buft_ctx->device; + request->tensor = serialize_tensor(tensor); + // .get_alloc_size could be a function of the tensor's srcs, so we must serialize them as well for (int i = 0; i < GGML_MAX_SRC; i++) { - request.srcs[i] = serialize_tensor(tensor->src[i]); + request->srcs[i] = serialize_tensor(tensor->src[i]); } - // TODO: cache the alloc responses to avoid extra RPC calls? rpc_msg_get_alloc_size_rsp response; - bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALLOC_SIZE, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + auto dispatcher = get_dispatcher(buft_ctx->endpoint); + dispatcher->send(RPC_CMD_GET_ALLOC_SIZE, request, sizeof(*request), &response, sizeof(response)); + + { + std::lock_guard lock(cache_mutex); + cache[cache_hash] = response.alloc_size; + } return response.alloc_size; } @@ -673,9 +918,44 @@ static void ggml_backend_rpc_free(ggml_backend_t backend) { delete backend; } +static void ggml_backend_rpc_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context; + rpc_tensor rpc_tensor = serialize_tensor(tensor); + if (size > HASH_THRESHOLD) { + auto request = std::make_shared(); + request->tensor = rpc_tensor; + request->offset = offset; + request->hash = fnv_hash((const uint8_t*)data, size); + rpc_msg_set_tensor_hash_rsp response; + // TODO: make this async + ctx->dispatcher->send(RPC_CMD_SET_TENSOR_HASH, request, sizeof(*request), &response, sizeof(response)); + if (response.result) { + // the server has the same data, no need to send it + return; + } + } + // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) + size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size; + uint8_t * input = new uint8_t[input_size](); + memcpy(input, &rpc_tensor, sizeof(rpc_tensor)); + memcpy(input + sizeof(rpc_tensor), &offset, sizeof(offset)); + memcpy(input + sizeof(rpc_tensor) + sizeof(offset), data, size); + std::shared_ptr input_ptr(input, std::default_delete()); + ctx->dispatcher->send_async(RPC_CMD_SET_TENSOR, input_ptr, input_size); +} + +static void ggml_backend_rpc_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { + ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context; + auto request = std::make_shared(); + request->tensor = serialize_tensor(tensor); + request->offset = offset; + request->size = size; + ctx->dispatcher->send_async(RPC_CMD_GET_TENSOR, request, sizeof(*request), data, size); +} + static void ggml_backend_rpc_synchronize(ggml_backend_t backend) { - GGML_UNUSED(backend); - // this is no-op because we don't have any async operations + ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; + rpc_ctx->dispatcher->synchronize(); } static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::vector & tensors, std::unordered_set & visited) { @@ -698,7 +978,7 @@ static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::ve tensors.push_back(result); } -static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::vector & output) { +static uint8_t * serialize_graph(uint32_t device, const ggml_cgraph * cgraph, size_t * output_size) { uint32_t n_nodes = cgraph->n_nodes; std::vector tensors; std::unordered_set visited; @@ -708,9 +988,9 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve // serialization format: // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) | uint32_t n_tensors = tensors.size(); - int output_size = 2*sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor); - output.resize(output_size, 0); - uint8_t * dest = output.data(); + *output_size = 2*sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor); + uint8_t * output = new uint8_t[*output_size](); + uint8_t * dest = output; memcpy(dest, &device, sizeof(device)); dest += sizeof(device); memcpy(dest, &n_nodes, sizeof(n_nodes)); @@ -723,6 +1003,7 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve dest += sizeof(n_tensors); rpc_tensor * out_tensors = (rpc_tensor *)dest; memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor)); + return output; } static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { @@ -733,27 +1014,35 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g GGML_ASSERT(cgraph->n_nodes > 0); bool reuse = cgraph->uid != 0 && rpc_dev_ctx->last_graph_uid == cgraph->uid; if (reuse) { - rpc_msg_graph_recompute_req request; - request.device = rpc_ctx->device; - auto sock = get_socket(rpc_ctx->endpoint); - bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_RECOMPUTE, &request, sizeof(request)); - RPC_STATUS_ASSERT(status); + auto request = std::make_shared(); + request->device = rpc_ctx->device; + rpc_ctx->dispatcher->send_async(RPC_CMD_GRAPH_RECOMPUTE, request, sizeof(*request)); } else { rpc_dev_ctx->last_graph_uid = cgraph->uid; - std::vector input; - serialize_graph(rpc_ctx->device, cgraph, input); - auto sock = get_socket(rpc_ctx->endpoint); - bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size()); - RPC_STATUS_ASSERT(status); + size_t input_size = 0; + uint8_t * input = serialize_graph(rpc_ctx->device, cgraph, &input_size); + std::shared_ptr input_ptr(input, std::default_delete()); + rpc_ctx->dispatcher->send_async(RPC_CMD_GRAPH_COMPUTE, input_ptr, input_size); } return GGML_STATUS_SUCCESS; } +static void ggml_backend_rpc_event_record(ggml_backend_t backend, ggml_backend_event_t event) { + ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; + rpc_ctx->dispatcher->event_record(event); +} + +static void ggml_backend_rpc_event_wait(ggml_backend_t backend, ggml_backend_event_t event) { + // this is noop for RPC as we have a single stream + GGML_UNUSED(backend); + GGML_UNUSED(event); +} + static ggml_backend_i ggml_backend_rpc_interface = { /* .get_name = */ ggml_backend_rpc_name, /* .free = */ ggml_backend_rpc_free, - /* .set_tensor_async = */ NULL, - /* .get_tensor_async = */ NULL, + /* .set_tensor_async = */ ggml_backend_rpc_set_tensor_async, + /* .get_tensor_async = */ ggml_backend_rpc_get_tensor_async, /* .set_tensor_2d_async = */ NULL, /* .get_tensor_2d_async = */ NULL, /* .cpy_tensor_async = */ NULL, @@ -763,8 +1052,8 @@ static ggml_backend_i ggml_backend_rpc_interface = { /* .graph_plan_update = */ NULL, /* .graph_plan_compute = */ NULL, /* .graph_compute = */ ggml_backend_rpc_graph_compute, - /* .event_record = */ NULL, - /* .event_wait = */ NULL, + /* .event_record = */ ggml_backend_rpc_event_record, + /* .event_wait = */ ggml_backend_rpc_event_wait, /* .graph_optimize = */ NULL, }; @@ -778,13 +1067,9 @@ ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, u if (it != buft_map.end()) { return it->second; } - auto sock = get_socket(endpoint); - if (sock == nullptr) { - GGML_LOG_ERROR("Failed to connect to %s\n", endpoint); - return nullptr; - } - size_t alignment = get_alignment(sock, device); - size_t max_size = get_max_size(sock, device); + auto dispatcher = get_dispatcher(endpoint); + size_t alignment = get_alignment(dispatcher, device); + size_t max_size = get_max_size(dispatcher, device); ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context { /* .endpoint = */ endpoint, /* .device = */ device, @@ -804,10 +1089,11 @@ ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, u ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device) { std::string dev_name = "RPC" + std::to_string(device) + "[" + std::string(endpoint) + "]"; + auto dispatcher = get_dispatcher(endpoint); ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context { - /* .endpoint = */ endpoint, - /* .device = */ device, - /* .name = */ dev_name, + /* .dispatcher = */ dispatcher, + /* .device = */ device, + /* .name = */ dev_name, }; auto reg = ggml_backend_rpc_add_server(endpoint); ggml_backend_t backend = new ggml_backend { @@ -823,26 +1109,16 @@ bool ggml_backend_is_rpc(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid()); } -static void get_device_memory(const std::shared_ptr & sock, uint32_t device, size_t * free, size_t * total) { - rpc_msg_get_device_memory_req request; - request.device = device; +void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total) { + auto dispatcher = get_dispatcher(endpoint); + auto request = std::make_shared(); + request->device = device; rpc_msg_get_device_memory_rsp response; - bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, &request, sizeof(request), &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + dispatcher->send(RPC_CMD_GET_DEVICE_MEMORY, request, sizeof(*request), &response, sizeof(response)); *free = response.free_mem; *total = response.total_mem; } -void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total) { - auto sock = get_socket(endpoint); - if (sock == nullptr) { - *free = 0; - *total = 0; - return; - } - get_device_memory(sock, device, free, total); -} - // RPC server-side implementation class rpc_server { @@ -1647,9 +1923,6 @@ static void rpc_serve_client(const std::vector & backends, const if (!server.free_buffer(request)) { return; } - if (!send_msg(sock, nullptr, 0)) { - return; - } break; } case RPC_CMD_BUFFER_CLEAR: { @@ -1660,9 +1933,6 @@ static void rpc_serve_client(const std::vector & backends, const if (!server.buffer_clear(request)) { return; } - if (!send_msg(sock, nullptr, 0)) { - return; - } break; } case RPC_CMD_MEMSET_TENSOR: { @@ -1673,9 +1943,6 @@ static void rpc_serve_client(const std::vector & backends, const if (!server.memset_tensor(request)) { return; } - if (!send_msg(sock, nullptr, 0)) { - return; - } break; } case RPC_CMD_SET_TENSOR: { @@ -1710,9 +1977,6 @@ static void rpc_serve_client(const std::vector & backends, const if (!server.init_tensor(request)) { return; } - if (!send_msg(sock, nullptr, 0)) { - return; - } break; } case RPC_CMD_GET_TENSOR: { @@ -1889,10 +2153,10 @@ static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggm props->type = ggml_backend_rpc_device_get_type(dev); ggml_backend_rpc_device_get_memory(dev, &props->memory_free, &props->memory_total); props->caps = { - /* .async = */ false, + /* .async = */ true, /* .host_buffer = */ false, /* .buffer_from_host_ptr = */ false, - /* .events = */ false, + /* .events = */ true, /* .mmap_support = */ true, }; } @@ -1929,6 +2193,24 @@ static bool ggml_backend_rpc_device_supports_buft(ggml_backend_dev_t dev, ggml_b return buft_ctx->endpoint == dev_ctx->endpoint && buft_ctx->device == dev_ctx->device; } +static ggml_backend_event_t ggml_backend_rpc_device_event_new(ggml_backend_dev_t dev) { + ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context; + auto dispatcher = get_dispatcher(ctx->endpoint); + return dispatcher->event_new(dev); +} + +static void ggml_backend_rpc_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) { + ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context; + auto dispatcher = get_dispatcher(ctx->endpoint); + dispatcher->event_free(event); +} + +static void ggml_backend_rpc_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) { + ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context; + auto dispatcher = get_dispatcher(ctx->endpoint); + dispatcher->event_synchronize(event); +} + static const struct ggml_backend_device_i ggml_backend_rpc_device_i = { /* .get_name = */ ggml_backend_rpc_device_get_name, /* .get_description = */ ggml_backend_rpc_device_get_description, @@ -1942,9 +2224,9 @@ static const struct ggml_backend_device_i ggml_backend_rpc_device_i = { /* .supports_op = */ ggml_backend_rpc_device_supports_op, /* .supports_buft = */ ggml_backend_rpc_device_supports_buft, /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, - /* .event_synchronize = */ NULL, + /* .event_new = */ ggml_backend_rpc_device_event_new, + /* .event_free = */ ggml_backend_rpc_device_event_free, + /* .event_synchronize = */ ggml_backend_rpc_device_event_synchronize, }; // backend reg interface @@ -2004,14 +2286,9 @@ ggml_backend_reg_t ggml_backend_rpc_reg(void) { } static uint32_t ggml_backend_rpc_get_device_count(const char * endpoint) { - auto sock = get_socket(endpoint); - if (sock == nullptr) { - GGML_LOG_ERROR("Failed to connect to %s\n", endpoint); - return 0; - } + auto dispatcher = get_dispatcher(endpoint); rpc_msg_device_count_rsp response; - bool status = send_rpc_cmd(sock, RPC_CMD_DEVICE_COUNT, nullptr, 0, &response, sizeof(response)); - RPC_STATUS_ASSERT(status); + dispatcher->send(RPC_CMD_DEVICE_COUNT, nullptr, 0, &response, sizeof(response)); return response.device_count; } From bf942164697d2d62c2237a17b677dc2c017ea8e7 Mon Sep 17 00:00:00 2001 From: Pranav Uttarkar <122235768+PranavUttarkar@users.noreply.github.com> Date: Wed, 26 Aug 2026 09:49:32 -0500 Subject: [PATCH 2/9] Implemented vulkan cross_entropy_loss and cross_entropy_loss_back (#27216) --- docs/ops.md | 4 +- docs/ops/Vulkan.csv | 8 +- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 138 ++++++++++++++++++ .../vulkan-shaders/cross_entropy_loss.comp | 78 ++++++++++ .../cross_entropy_loss_back.comp | 75 ++++++++++ .../vulkan-shaders/vulkan-shaders-gen.cpp | 2 + 6 files changed, 299 insertions(+), 6 deletions(-) create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp diff --git a/docs/ops.md b/docs/ops.md index 2c179dd01f..54a3602833 100644 --- a/docs/ops.md +++ b/docs/ops.md @@ -35,8 +35,8 @@ Legend: | COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | | COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | -| CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | -| CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | +| CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | +| CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | | DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | | DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ | diff --git a/docs/ops/Vulkan.csv b/docs/ops/Vulkan.csv index 59e67e1b20..13e67eb797 100644 --- a/docs/ops/Vulkan.csv +++ b/docs/ops/Vulkan.csv @@ -19292,10 +19292,10 @@ "Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=4096,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan" "Vulkan0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[6,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan" "Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan" -"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan" -"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan" -"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan" -"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan" +"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan" +"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan" +"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan" +"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan" "Vulkan0","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan" "Vulkan0","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan" "Vulkan0","GATED_DELTA_NET","type=f32,head_count=32,head_size=128,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0,K=1","support","1","yes","Vulkan" diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index c1d86aaac5..31923a2ba0 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1042,6 +1042,8 @@ struct vk_device_struct { vk_pipeline pipeline_argsort_large_f32[num_argsort_pipelines]; vk_pipeline pipeline_topk_f32[num_topk_pipelines]; vk_pipeline pipeline_sum_rows_f32; + vk_pipeline pipeline_cross_entropy_loss_f32, pipeline_cross_entropy_loss_f32_wg512; + vk_pipeline pipeline_cross_entropy_loss_back_f32, pipeline_cross_entropy_loss_back_f32_wg512; vk_pipeline pipeline_fwht_f32[4]; vk_pipeline pipeline_cumsum_f32; vk_pipeline pipeline_cumsum_small_f32; @@ -5758,6 +5760,10 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_f32, "cross_entropy_loss_f32", cross_entropy_loss_f32_len, cross_entropy_loss_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_f32_wg512, "cross_entropy_loss_f32_wg512", cross_entropy_loss_f32_len, cross_entropy_loss_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { 512 }, 1); + ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_back_f32, "cross_entropy_loss_back_f32", cross_entropy_loss_back_f32_len, cross_entropy_loss_back_f32_data, "main", 4, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_back_f32_wg512, "cross_entropy_loss_back_f32_wg512", cross_entropy_loss_back_f32_len, cross_entropy_loss_back_f32_data, "main", 4, sizeof(vk_op_push_constants), {1, 1, 1}, { 512 }, 1); // Intel Windows driver in range [32.0.101.8509, 32.0.101.8860) will crash when using fwht kernels so we gate that here const bool can_use_fwht = device->driver_id != vk::DriverId::eIntelProprietaryWindows || !ggml_vk_intel_windows_driver_in_range(device->properties.driverVersion, 101, 8509, 101, 8860); @@ -11577,6 +11583,17 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_sum_rows_f32; } return nullptr; + case GGML_OP_CROSS_ENTROPY_LOSS: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return src0->ne[0] > 1024 ? ctx->device->pipeline_cross_entropy_loss_f32_wg512 : ctx->device->pipeline_cross_entropy_loss_f32; + } + return nullptr; + case GGML_OP_CROSS_ENTROPY_LOSS_BACK: + // src0 is the scalar grad; src1 is logits + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && src2 && src2->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return src1->ne[0] > 1024 ? ctx->device->pipeline_cross_entropy_loss_back_f32_wg512 : ctx->device->pipeline_cross_entropy_loss_back_f32; + } + return nullptr; case GGML_OP_CUMSUM: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { if (src0->ne[0] <= 512) { @@ -13942,6 +13959,103 @@ static void ggml_vk_cumsum(ggml_backend_vk_context * ctx, vk_context& subctx, co ctx->prealloc_split_k_need_sync = true; } +static std::array ggml_vk_nrows_elements(uint32_t nr) { + if (nr > 262144) { + return { 512, 512, CEIL_DIV(nr, 262144) }; + } + if (nr > 512) { + return { 512, CEIL_DIV(nr, 512), 1 }; + } + return { nr, 1, 1 }; +} + +static void ggml_vk_cross_entropy_loss(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + GGML_ASSERT(ggml_is_scalar(dst)); + + const uint32_t nclasses = (uint32_t)src0->ne[0]; + const uint32_t nrows = (uint32_t)ggml_nrows(src0); + + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, nullptr, dst, GGML_OP_CROSS_ENTROPY_LOSS); + GGML_ASSERT(pipeline != nullptr); + + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); + ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_sum_rows_f32, 1); + + vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0); + vk_subbuffer src1_buf = ggml_vk_tensor_subbuffer(ctx, src1); + vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, true); + + const vk_op_push_constants pc = { nclasses, nrows, 0.0f, 0.0f, 0.0f, 0.0f }; + + const size_t tmp_size = (size_t)nrows * sizeof(float); + if (ctx->prealloc_size_x < tmp_size) { + ctx->prealloc_size_x = tmp_size; + ggml_vk_preallocate_buffers(ctx, subctx); + } + if (ctx->prealloc_x_need_sync) { + ggml_vk_sync_buffers(ctx, subctx); + } + + vk_subbuffer tmp_buf = { ctx->prealloc_x, 0, tmp_size }; + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, tmp_buf }, pc, ggml_vk_nrows_elements(nrows)); + ggml_vk_sync_buffers(ctx, subctx); + + vk_op_sum_rows_push_constants sp = {}; + sp.n_cols = nrows; + sp.ne01 = 1; + sp.ne02 = 1; + sp.weight = 1.0f; + init_pushconst_fastdiv(sp); + sp.misalign_offsets = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type); + + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_sum_rows_f32, { tmp_buf, dst_buf }, sp, { 1, 1, 1 }); + ctx->prealloc_x_need_sync = true; +} + +static void ggml_vk_cross_entropy_loss_back(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) { + const ggml_tensor * grad = dst->src[0]; + const ggml_tensor * logits = dst->src[1]; + const ggml_tensor * labels = dst->src[2]; + + GGML_ASSERT(grad->type == GGML_TYPE_F32); + GGML_ASSERT(logits->type == GGML_TYPE_F32); + GGML_ASSERT(labels->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_scalar(grad)); + GGML_ASSERT(ggml_is_contiguous(grad)); + GGML_ASSERT(ggml_is_contiguous(logits)); + GGML_ASSERT(ggml_is_contiguous(labels)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(logits, labels)); + GGML_ASSERT(ggml_are_same_shape(logits, dst)); + + const uint32_t nclasses = (uint32_t)logits->ne[0]; + const uint32_t nrows = (uint32_t)ggml_nrows(logits); + + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, grad, logits, labels, dst, GGML_OP_CROSS_ENTROPY_LOSS_BACK); + GGML_ASSERT(pipeline != nullptr); + + ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1); + + vk_subbuffer grad_buf = ggml_vk_tensor_subbuffer(ctx, grad); + vk_subbuffer logits_buf = ggml_vk_tensor_subbuffer(ctx, logits); + vk_subbuffer labels_buf = ggml_vk_tensor_subbuffer(ctx, labels); + vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst); + + const vk_op_push_constants pc = { nclasses, nrows, 0.0f, 0.0f, 0.0f, 0.0f }; + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { grad_buf, logits_buf, labels_buf, dst_buf }, pc, ggml_vk_nrows_elements(nrows)); +} + static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f, 0.0f, 0.0f }); } @@ -15687,6 +15801,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr case GGML_OP_ARGMAX: ggml_vk_argmax(ctx, compute_ctx, src0, node); + break; + case GGML_OP_CROSS_ENTROPY_LOSS: + ggml_vk_cross_entropy_loss(ctx, compute_ctx, node); + + break; + case GGML_OP_CROSS_ENTROPY_LOSS_BACK: + ggml_vk_cross_entropy_loss_back(ctx, compute_ctx, node); + break; case GGML_OP_COUNT_EQUAL: ggml_vk_count_equal(ctx, compute_ctx, src0, src1, node); @@ -18511,6 +18633,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm } case GGML_OP_ARGMAX: return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; + case GGML_OP_CROSS_ENTROPY_LOSS: + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32 + && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 + && ggml_are_same_shape(op->src[0], op->src[1]) + && ggml_is_contiguous(op) && ggml_is_scalar(op) && op->type == GGML_TYPE_F32; + case GGML_OP_CROSS_ENTROPY_LOSS_BACK: + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32 && ggml_is_scalar(op->src[0]) + && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 + && ggml_is_contiguous(op->src[2]) && op->src[2]->type == GGML_TYPE_F32 + && ggml_are_same_shape(op->src[1], op->src[2]) + && ggml_are_same_shape(op->src[1], op) + && ggml_is_contiguous(op) && op->type == GGML_TYPE_F32; case GGML_OP_COUNT_EQUAL: return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_I32 && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_I32; @@ -19437,6 +19571,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * tensor_clone = ggml_mean(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_ARGMAX) { tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]); + } else if (tensor->op == GGML_OP_CROSS_ENTROPY_LOSS) { + tensor_clone = ggml_cross_entropy_loss(ggml_ctx, src_clone[0], src_clone[1]); + } else if (tensor->op == GGML_OP_CROSS_ENTROPY_LOSS_BACK) { + tensor_clone = ggml_cross_entropy_loss_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]); } else if (tensor->op == GGML_OP_COUNT_EQUAL) { tensor_clone = ggml_count_equal(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_SOLVE_TRI) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp new file mode 100644 index 0000000000..0c135c6fd2 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp @@ -0,0 +1,78 @@ +#version 450 + +#include "generic_head.glsl" +#include "types.glsl" + +#extension GL_EXT_control_flow_attributes : enable + +layout(constant_id = 0) const uint BLOCK_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; +layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE tmp[BLOCK_SIZE]; + +FLOAT_TYPE wg_reduce_max(FLOAT_TYPE v) { + const uint tid = gl_LocalInvocationID.x; + tmp[tid] = v; + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + tmp[tid] = max(tmp[tid], tmp[tid + s]); + } + barrier(); + } + v = tmp[0]; + barrier(); + return v; +} + +FLOAT_TYPE wg_reduce_sum(FLOAT_TYPE v) { + const uint tid = gl_LocalInvocationID.x; + tmp[tid] = v; + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(); + } + v = tmp[0]; + barrier(); + return v; +} + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + if (row >= p.KY) { + return; + } + + const uint off = row * p.KX; + + FLOAT_TYPE max_logit = FLOAT_TYPE(uintBitsToFloat(0xFF800000)); + for (uint i = tid; i < p.KX; i += BLOCK_SIZE) { + max_logit = max(max_logit, FLOAT_TYPE(data_a[off + i])); + } + max_logit = wg_reduce_max(max_logit); + + FLOAT_TYPE sum_exp = FLOAT_TYPE(0.0f); + for (uint i = tid; i < p.KX; i += BLOCK_SIZE) { + sum_exp += exp(FLOAT_TYPE(data_a[off + i]) - max_logit); + } + const FLOAT_TYPE log_sum = log(wg_reduce_sum(sum_exp)); + + FLOAT_TYPE loss = FLOAT_TYPE(0.0f); + for (uint i = tid; i < p.KX; i += BLOCK_SIZE) { + loss += (FLOAT_TYPE(data_a[off + i]) - max_logit - log_sum) * FLOAT_TYPE(data_b[off + i]); + } + loss = -wg_reduce_sum(loss) / FLOAT_TYPE(p.KY); + + if (tid == 0) { + data_d[row] = D_TYPE(loss); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp new file mode 100644 index 0000000000..3cdebe86e4 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp @@ -0,0 +1,75 @@ +#version 450 + +#include "generic_head.glsl" +#include "types.glsl" + +#extension GL_EXT_control_flow_attributes : enable + +layout(constant_id = 0) const uint BLOCK_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer G {A_TYPE data_g[];}; +layout (binding = 1) readonly buffer X {B_TYPE data_x[];}; +layout (binding = 2) readonly buffer Y {B_TYPE data_y[];}; +layout (binding = 3) writeonly buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE tmp[BLOCK_SIZE]; + +FLOAT_TYPE wg_reduce_max(FLOAT_TYPE v) { + const uint tid = gl_LocalInvocationID.x; + tmp[tid] = v; + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + tmp[tid] = max(tmp[tid], tmp[tid + s]); + } + barrier(); + } + v = tmp[0]; + barrier(); + return v; +} + +FLOAT_TYPE wg_reduce_sum(FLOAT_TYPE v) { + const uint tid = gl_LocalInvocationID.x; + tmp[tid] = v; + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(); + } + v = tmp[0]; + barrier(); + return v; +} + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + if (row >= p.KY) { + return; + } + + const uint off = row * p.KX; + const FLOAT_TYPE d_by_nrows = FLOAT_TYPE(data_g[0]) / FLOAT_TYPE(p.KY); + + FLOAT_TYPE max_logit = FLOAT_TYPE(uintBitsToFloat(0xFF800000)); + for (uint i = tid; i < p.KX; i += BLOCK_SIZE) { + max_logit = max(max_logit, FLOAT_TYPE(data_x[off + i])); + } + max_logit = wg_reduce_max(max_logit); + + FLOAT_TYPE sum_exp = FLOAT_TYPE(0.0f); + for (uint i = tid; i < p.KX; i += BLOCK_SIZE) { + sum_exp += exp(FLOAT_TYPE(data_x[off + i]) - max_logit); + } + const FLOAT_TYPE inv_sum = FLOAT_TYPE(1.0f) / wg_reduce_sum(sum_exp); + + for (uint i = tid; i < p.KX; i += BLOCK_SIZE) { + const FLOAT_TYPE sm = exp(FLOAT_TYPE(data_x[off + i]) - max_logit) * inv_sum; + data_d[off + i] = D_TYPE((sm - FLOAT_TYPE(data_y[off + i])) * d_by_nrows); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index 17d57d5a18..dbb99782cf 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -1029,6 +1029,8 @@ void process_shaders() { string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}})); string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("cross_entropy_loss_f32", "cross_entropy_loss.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("cross_entropy_loss_back_f32", "cross_entropy_loss_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("fwht_f32", "fwht.comp", {}); string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}}); string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}})); From 5e6a37cb115dc1074e274ac004373f5661909695 Mon Sep 17 00:00:00 2001 From: Ruben Ortlam Date: Wed, 26 Aug 2026 18:02:06 +0200 Subject: [PATCH 3/9] vulkan: warptiles currently assume warp sizes <= 64, clamp to work around larger warps (#27726) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 68 +++++++++++++++------------- 1 file changed, 37 insertions(+), 31 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 31923a2ba0..8108e94c16 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -4171,10 +4171,16 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u); const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u); + // clamp WARP for l_/m_ warptiles so WM <= BM (breaks on subgroupSize > 64) + const uint32_t mm_warp_8 = std::min(subgroup_size_8, 64u); + const uint32_t mm_warp_16 = std::min(subgroup_size_16, 64u); + const uint32_t mul_mat_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size; const uint32_t mul_mat_subgroup_size_8 = std::max(mul_mat_subgroup_size, 8u); const uint32_t mul_mat_subgroup_size_16 = std::max(mul_mat_subgroup_size, 16u); const uint32_t mul_mat_subgroup_size_32 = std::max(mul_mat_subgroup_size, 32u); + const uint32_t mul_mat_mm_warp_8 = std::min(mul_mat_subgroup_size_8, 64u); + const uint32_t mul_mat_mm_warp_16 = std::min(mul_mat_subgroup_size_16, 64u); const bool subgroup_min_size_16 = (!device->subgroup_size_control && device->subgroup_size >= 16) || (device->subgroup_size_control && device->subgroup_max_size >= 16); @@ -4255,39 +4261,39 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32; - l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 }; - m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; - s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; + l_warptile = { 128, 128, 128, 16, mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mm_warp_8 }; + m_warptile = { 128, 64, 64, 16, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 }; + s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; - l_warptile_mmq = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 }; - m_warptile_mmq = { 128, 64, 64, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; - s_warptile_mmq = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; + l_warptile_mmq = { 128, 128, 128, 32, mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mm_warp_8 }; + m_warptile_mmq = { 128, 64, 64, 32, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 }; + s_warptile_mmq = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; // Integer MMQ has a smaller shared memory profile, but heavier register use - l_warptile_mmq_int = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; - m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, 2, 2, 1, subgroup_size_8 }; - s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 1, subgroup_size_8 }; + l_warptile_mmq_int = { 128, 128, 128, 32, mm_warp_8 * 2, 64, 2, 4, 4, 1, mm_warp_8 }; + m_warptile_mmq_int = { 128, 64, 64, 32, mm_warp_8, 32, 2, 2, 2, 1, mm_warp_8 }; + s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 1, subgroup_size_8 }; // K-quants use even more registers, mitigate by setting WMITER to 1 - l_warptile_mmq_int_k = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 }; - m_warptile_mmq_int_k = { 128, 64, 64, 32, subgroup_size_8, 32, 1, 2, 2, 1, subgroup_size_8 }; - s_warptile_mmq_int_k = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, subgroup_size_8 }; + l_warptile_mmq_int_k = { 128, 128, 128, 32, mm_warp_8 * 2, 64, 1, 4, 4, 1, mm_warp_8 }; + m_warptile_mmq_int_k = { 128, 64, 64, 32, mm_warp_8, 32, 1, 2, 2, 1, mm_warp_8 }; + s_warptile_mmq_int_k = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, subgroup_size_8 }; - l_warptile_id = { 128, 128, 128, 16, mul_mat_subgroup_size_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_16 }; - m_warptile_id = { 128, 64, 64, 16, mul_mat_subgroup_size_16, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_16 }; - s_warptile_id = { mul_mat_subgroup_size_16, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 }; + l_warptile_id = { 128, 128, 128, 16, mul_mat_mm_warp_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_mm_warp_16 }; + m_warptile_id = { 128, 64, 64, 16, mul_mat_mm_warp_16, 32, 2, tm_m, tn_m, tk_m, mul_mat_mm_warp_16 }; + s_warptile_id = { mul_mat_subgroup_size_16, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 }; - l_warptile_mmqid = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_8 }; - m_warptile_mmqid = { 128, 64, 64, 32, mul_mat_subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_8 }; - s_warptile_mmqid = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 }; + l_warptile_mmqid = { 128, 128, 128, 32, mul_mat_mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_mm_warp_8 }; + m_warptile_mmqid = { 128, 64, 64, 32, mul_mat_mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mul_mat_mm_warp_8 }; + s_warptile_mmqid = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 }; - l_warptile_mmqid_int = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, 4, 4, 1, mul_mat_subgroup_size_8 }; - m_warptile_mmqid_int = { 128, 64, 64, 32, mul_mat_subgroup_size_8, 32, 2, 2, 2, 1, mul_mat_subgroup_size_8 }; - s_warptile_mmqid_int = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 1, mul_mat_subgroup_size_8 }; + l_warptile_mmqid_int = { 128, 128, 128, 32, mul_mat_mm_warp_8 * 2, 64, 2, 4, 4, 1, mul_mat_mm_warp_8 }; + m_warptile_mmqid_int = { 128, 64, 64, 32, mul_mat_mm_warp_8, 32, 2, 2, 2, 1, mul_mat_mm_warp_8 }; + s_warptile_mmqid_int = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 1, mul_mat_subgroup_size_8 }; - l_warptile_mmqid_int_k = { 128, 128, 128, 32, mul_mat_subgroup_size_16 * 2, 64, 1, 4, 4, 1, mul_mat_subgroup_size_16 }; - m_warptile_mmqid_int_k = { 128, 64, 64, 32, mul_mat_subgroup_size_16, 32, 1, 2, 2, 1, mul_mat_subgroup_size_16 }; - s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, mul_mat_subgroup_size_16 }; + l_warptile_mmqid_int_k = { 128, 128, 128, 32, mul_mat_mm_warp_16 * 2, 64, 1, 4, 4, 1, mul_mat_mm_warp_16 }; + m_warptile_mmqid_int_k = { 128, 64, 64, 32, mul_mat_mm_warp_16, 32, 1, 2, 2, 1, mul_mat_mm_warp_16 }; + s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, mul_mat_subgroup_size_16 }; // chip specific tuning if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) { @@ -4295,13 +4301,13 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 }; } else if (device->vendor_id == VK_VENDOR_ID_AMD && device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary) { // This is intentionally using tx_m values, slight performance increase - l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; - l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; - l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 }; + l_warptile = { 256, 128, 128, 16, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 }; + l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 }; + l_warptile_mmq_int_k = { 256, 128, 128, 32, mm_warp_16, 64, 1, 4, 2, 1, mm_warp_16 }; } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support) { // Xe2/Xe3 with coopmat enabled - warptile performance tuning - l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; - l_warptile_mmq = { 512, 128, 128, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; + l_warptile = { 512, 128, 128, 16, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 }; + l_warptile_mmq = { 512, 128, 128, 32, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 }; } l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 }; @@ -5174,8 +5180,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32; // use scalar tile sizes - l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; - m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 }; + l_warptile = { 128, 128, 128, 16, mm_warp_8 * 2, 64, 2, 4, 4, 1, mm_warp_8 }; + m_warptile = { 128, 64, 64, 16, mm_warp_8, 32, 2, 4, 2, 1, mm_warp_8 }; s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, 2, 2, 1, subgroup_size_8 }; l_wg_denoms = {128, 128, 1 }; From 0379a19f09495da510e11785261b194ecb99097f Mon Sep 17 00:00:00 2001 From: Aleksander Grygier Date: Wed, 26 Aug 2026 20:19:19 +0200 Subject: [PATCH 4/9] ui: Update Dialog component styling (#27743) * feat(ui): make base dialog responsive and support sticky headers * ui: move dialog close button to the sticky header Assisted-by: pi * chore: Formatting & linting --- .../app/dialogs/DialogMermaidPreview.svelte | 1 + .../components/ui/dialog/dialog-close.svelte | 6 +++++ .../ui/dialog/dialog-content.svelte | 5 ++-- .../components/ui/dialog/dialog-header.svelte | 27 +++++++++++++++++-- 4 files changed, 35 insertions(+), 4 deletions(-) diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMermaidPreview.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMermaidPreview.svelte index 09e53442ac..e741373497 100644 --- a/tools/ui/src/lib/components/app/dialogs/DialogMermaidPreview.svelte +++ b/tools/ui/src/lib/components/app/dialogs/DialogMermaidPreview.svelte @@ -14,6 +14,7 @@ diff --git a/tools/ui/src/lib/components/ui/dialog/dialog-close.svelte b/tools/ui/src/lib/components/ui/dialog/dialog-close.svelte index e8a96a7731..cd957a9bf8 100644 --- a/tools/ui/src/lib/components/ui/dialog/dialog-close.svelte +++ b/tools/ui/src/lib/components/ui/dialog/dialog-close.svelte @@ -5,3 +5,9 @@ + + diff --git a/tools/ui/src/lib/components/ui/dialog/dialog-content.svelte b/tools/ui/src/lib/components/ui/dialog/dialog-content.svelte index e6b2a641d5..8a1a5909d5 100644 --- a/tools/ui/src/lib/components/ui/dialog/dialog-content.svelte +++ b/tools/ui/src/lib/components/ui/dialog/dialog-content.svelte @@ -10,7 +10,7 @@ class: className, portalProps, ref = $bindable(null), - showCloseButton = true, + showCloseButton = false, ...restProps }: WithoutChildrenOrChild & { portalProps?: DialogPrimitive.PortalProps; @@ -25,7 +25,8 @@ + import XIcon from '@lucide/svelte/icons/x'; import { cn, type WithElementRef } from '$lib/components/ui/utils'; + import { Dialog as DialogPrimitive } from 'bits-ui'; import type { HTMLAttributes } from 'svelte/elements'; let { children, class: className, ref = $bindable(null), + showCloseButton = true, ...restProps - }: WithElementRef> = $props(); + }: WithElementRef> & { + showCloseButton?: boolean; + } = $props(); +
{@render children?.()} + + {#if showCloseButton} + + + + Close + + {/if}
From 539f24529bdf99e0baefd54fefff1660034bfe7b Mon Sep 17 00:00:00 2001 From: Aleksander Grygier Date: Wed, 26 Aug 2026 21:07:24 +0200 Subject: [PATCH 5/9] ui: Move Settings and MCP Servers routes to dialog-based views (#27744) * ui : open MCP servers in a dialog from the chat form Replace the MCP servers submenu with a single "MCP Servers" item that opens a new DialogMcpServers dialog instead of navigating to the /mcp-servers route. Assisted-by: pi * ui : browse MCP resources from the server card Make the Resources capability badge clickable so it opens the MCP resources browser dialog, and drop the page-only chrome from SettingsMcpServers. Assisted-by: pi * ui : remove mcp-servers route and sidebar entry MCP servers are now managed in a dialog, so drop the dedicated route and the sidebar icon that navigated to it. Assisted-by: pi * ui : remove unused MCP servers submenu component The submenu was replaced by the MCP servers dialog, so delete the component and its export. Assisted-by: pi * feat(ui): add DialogSettingsChat dialog * refactor(ui): switch SettingsChat to in-app section navigation * feat(ui): open settings as dialog from sidebar * refactor(ui): remove settings route and URL-based settings navigation * fix(ui): adjust MCP dialogs for new base sizing * chore: Formatting & linting --- .../app/chat/ChatForm/ChatForm.svelte | 9 +- .../ChatFormActionAddDropdown.svelte | 48 ++---- .../ChatFormActionAddMcpServersSubmenu.svelte | 152 ------------------ .../ChatFormActionAddMcpSubmenu.svelte | 51 ++++++ .../ChatFormActions/ChatFormActions.svelte | 7 +- tools/ui/src/lib/components/app/chat/index.ts | 18 +-- .../dialogs/DialogMcpResourcesBrowser.svelte | 2 +- .../app/dialogs/DialogMcpServerAddNew.svelte | 2 +- .../app/dialogs/DialogMcpServers.svelte | 33 ++++ .../app/dialogs/DialogSettingsChat.svelte | 34 ++++ .../src/lib/components/app/dialogs/index.ts | 17 ++ .../app/mcp/McpCapabilitiesBadges.svelte | 29 +++- .../mcp/McpServerCard/McpServerCard.svelte | 18 ++- .../McpServerCard/McpServerCardHeader.svelte | 4 +- .../SidebarNavigation.svelte | 7 +- .../SidebarNavigationActions.svelte | 38 +++-- .../settings/SettingsChat/SettingsChat.svelte | 48 +++--- .../SettingsChatDesktopSidebar.svelte | 49 ++---- .../settings/SettingsChatMobileHeader.svelte | 63 +++----- .../app/settings/SettingsFooter.svelte | 2 +- .../app/settings/SettingsMcpServers.svelte | 57 ++----- .../src/lib/components/app/settings/index.ts | 6 +- .../ui/src/lib/constants/routes.constants.ts | 4 - tools/ui/src/lib/constants/ui.constants.ts | 11 +- tools/ui/src/lib/enums/ui.enums.ts | 3 +- .../hooks/use-settings-navigation.svelte.ts | 45 ------ tools/ui/src/lib/services/index.ts | 1 - tools/ui/src/lib/services/router.service.ts | 7 +- tools/ui/src/lib/stores/index.ts | 2 - .../lib/stores/settings/referrer.svelte.ts | 19 --- tools/ui/src/routes/mcp-servers/+page.svelte | 5 - tools/ui/src/routes/settings/+layout.svelte | 38 ----- .../routes/settings/[[section]]/+page.svelte | 15 -- 33 files changed, 305 insertions(+), 539 deletions(-) delete mode 100644 tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpServersSubmenu.svelte create mode 100644 tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpSubmenu.svelte create mode 100644 tools/ui/src/lib/components/app/dialogs/DialogMcpServers.svelte create mode 100644 tools/ui/src/lib/components/app/dialogs/DialogSettingsChat.svelte delete mode 100644 tools/ui/src/lib/hooks/use-settings-navigation.svelte.ts delete mode 100644 tools/ui/src/lib/stores/settings/referrer.svelte.ts delete mode 100644 tools/ui/src/routes/mcp-servers/+page.svelte delete mode 100644 tools/ui/src/routes/settings/+layout.svelte delete mode 100644 tools/ui/src/routes/settings/[[section]]/+page.svelte diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte index 9152935f07..40819d6f1a 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte @@ -8,7 +8,8 @@ ChatFormInputFileInputInvisible, ChatFormMcpResourcesList, ChatFormPickers, - DialogMcpResourcesBrowser + DialogMcpResourcesBrowser, + DialogMcpServers } from '$lib/components/app'; import { CLIPBOARD_CONTENT_QUOTE_PREFIX, @@ -183,6 +184,9 @@ let isResourceDialogOpen = $state(false); let preSelectedResourceUri = $state(undefined); + // MCP Servers Dialog State + let isMcpServersDialogOpen = $state(false); + let currentConfig = $derived(settingsStore.config); let pasteLongTextToFileLength = $derived.by(() => { @@ -618,6 +622,7 @@ onFileUpload={handleFileUpload} onMcpPromptClick={showMcpPromptButton ? () => pickers.openPromptPicker() : undefined} onMcpResourcesClick={() => (isResourceDialogOpen = true)} + onMcpSettingsClick={() => (isMcpServersDialogOpen = true)} onMicClick={handleMicClick} {onStop} onSystemPromptClick={() => onSystemPromptClick?.({ files: uploadedFiles, message: value })} @@ -656,3 +661,5 @@ }} preSelectedUri={preSelectedResourceUri} /> + + diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddDropdown.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddDropdown.svelte index 02bfadb7e4..5ff3c6a908 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddDropdown.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddDropdown.svelte @@ -1,10 +1,6 @@ - - - - - - - MCP Servers - - - - {#if hasMcpServers} - -
- {#each filteredMcpServers as server (server.id)} - {@const healthState = mcpStore.getHealthCheckState(server.id)} - {@const hasError = healthState.status === HealthCheckStatus.ERROR} - {@const isEnabledForChat = isServerEnabledForChat(server.id)} - {@const displayName = getServerLabel(server)} - {@const faviconUrl = mcpStore.getServerFavicon(server.id)} - - - {/each} -
- - {#snippet footer()} - - - - Manage MCP Servers - - {/snippet} - - {:else} -
- No MCP servers configured -
- - - - - - - Add MCP Servers - - {/if} - - - diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpSubmenu.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpSubmenu.svelte new file mode 100644 index 0000000000..07439afd61 --- /dev/null +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddMcpSubmenu.svelte @@ -0,0 +1,51 @@ + + + + + + + MCP + + + + + + + Servers + + + {#if chatFormActions.hasMcpPromptsSupport} + + + + Prompts + + {/if} + + {#if chatFormActions.hasMcpResourcesSupport} + + + + Resources + + {/if} + + diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte index 97351b3a6c..f1aa743693 100644 --- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte +++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte @@ -1,6 +1,5 @@ - + diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte index ae8b24cb2d..fab45aa970 100644 --- a/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte +++ b/tools/ui/src/lib/components/app/dialogs/DialogMcpServerAddNew.svelte @@ -246,7 +246,7 @@ - + Add New MCP Server diff --git a/tools/ui/src/lib/components/app/dialogs/DialogMcpServers.svelte b/tools/ui/src/lib/components/app/dialogs/DialogMcpServers.svelte new file mode 100644 index 0000000000..2eaaa0a657 --- /dev/null +++ b/tools/ui/src/lib/components/app/dialogs/DialogMcpServers.svelte @@ -0,0 +1,33 @@ + + + + + + + + + MCP Servers + + + + + + diff --git a/tools/ui/src/lib/components/app/dialogs/DialogSettingsChat.svelte b/tools/ui/src/lib/components/app/dialogs/DialogSettingsChat.svelte new file mode 100644 index 0000000000..5e0046aae5 --- /dev/null +++ b/tools/ui/src/lib/components/app/dialogs/DialogSettingsChat.svelte @@ -0,0 +1,34 @@ + + + + + + + + + Settings + + + + (open = false)} onSectionChange={() => {}} /> + + diff --git a/tools/ui/src/lib/components/app/dialogs/index.ts b/tools/ui/src/lib/components/app/dialogs/index.ts index 38aa099354..1ffa7a2b9e 100644 --- a/tools/ui/src/lib/components/app/dialogs/index.ts +++ b/tools/ui/src/lib/components/app/dialogs/index.ts @@ -18,6 +18,23 @@ */ export { default as DialogMcpServerAddNew } from './DialogMcpServerAddNew.svelte'; +/** + * **DialogMcpServers** - MCP servers dialog shown from the chat form + * + * Shows the same MCP server list as the `/mcp-servers` route inside a modal + * dialog. + */ +export { default as DialogMcpServers } from './DialogMcpServers.svelte'; + +/** + * **DialogSettingsChat** - Chat settings shown in a modal dialog + * + * Wraps the full SettingsChat layout (sidebar, mobile header, fields, footer) + * inside a ShadCN Dialog instead of a dedicated route. Section switching is + * handled in-app via `onSectionChange` rather than URL navigation. + */ +export { default as DialogSettingsChat } from './DialogSettingsChat.svelte'; + /** * **DialogExportSettings** - Settings export dialog with sensitive data warning * diff --git a/tools/ui/src/lib/components/app/mcp/McpCapabilitiesBadges.svelte b/tools/ui/src/lib/components/app/mcp/McpCapabilitiesBadges.svelte index 632cd3e8a7..5969547036 100644 --- a/tools/ui/src/lib/components/app/mcp/McpCapabilitiesBadges.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpCapabilitiesBadges.svelte @@ -1,13 +1,22 @@ {#if capabilities} @@ -20,10 +29,24 @@ {/if} {#if capabilities.server.resources} - + { + if (e.key === 'Enter' || e.key === ' ') { + e.preventDefault(); + onBrowseResources?.(); + } + }} + role="button" + tabindex={0} + variant="outline" + > Resources + + {/if} diff --git a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte index 84383d59ea..6e9a995385 100644 --- a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCard.svelte @@ -22,9 +22,10 @@ onToggle: (enabled: boolean) => void; onUpdate: (updates: Partial) => void; onDelete: () => void; + onBrowseResources?: () => void; } - let { enabled, onDelete, onToggle, onUpdate, server }: Props = $props(); + let { enabled, onBrowseResources, onDelete, onToggle, onUpdate, server }: Props = $props(); let healthState = $derived(mcpStore.getHealthCheckState(server.id)); let displayName = $derived(mcpStore.getServerLabel(server)); @@ -125,6 +126,7 @@ {displayName} enabled={enabled ?? server.enabled} {faviconUrl} + {onBrowseResources} {onToggle} {serverInfo} {transportType} @@ -191,12 +193,14 @@ {/if} - +
+ +
{/if} diff --git a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardHeader.svelte b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardHeader.svelte index 30369d6509..9baded678e 100644 --- a/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardHeader.svelte +++ b/tools/ui/src/lib/components/app/mcp/McpServerCard/McpServerCardHeader.svelte @@ -12,6 +12,7 @@ enabled: boolean; disabled?: boolean; onToggle: (enabled: boolean) => void; + onBrowseResources?: () => void; serverInfo?: MCPServerInfo; capabilities?: MCPCapabilitiesInfo; transportType?: MCPTransportType; @@ -23,6 +24,7 @@ displayName, enabled, faviconUrl, + onBrowseResources, onToggle, serverInfo, transportType @@ -57,7 +59,7 @@ {/if} {#if capabilities} - + {/if} {/if} diff --git a/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte b/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte index dfe55b5e29..8bf12279e3 100644 --- a/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte +++ b/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte @@ -5,6 +5,7 @@ import { ActionIcon, DialogConversationRename, + DialogSettingsChat, Logo, SidebarNavigationActions, SidebarNavigationConversationList @@ -91,6 +92,7 @@ let selectedIds = new SvelteSet(); let renameDialogOpen = $state(false); + let settingsDialogOpen = $state(false); let renameTargetConversationId = $state(null); let renameDraft = $state(''); let renameOriginalTitle = $state(''); @@ -308,7 +310,7 @@ -{#if innerWidth > 768 || (!page.url.hash.includes(ROUTES.SETTINGS) && !page.url.hash.includes(ROUTES.MCP_SERVERS) && !page.url.hash.includes(ROUTES.SEARCH))} +{#if innerWidth > 768 || !page.url.hash.includes(ROUTES.SEARCH)}