ggml : fix bug introduced in #1237

* Generalize quantize_fns for simpler FP16 handling

* Remove call to ggml_cuda_mul_mat_get_wsize

* ci : disable FMA for mac os actions

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.github/workflows/build.yml

@@ -111,6 +111,7 @@ jobs:
 
       - name: Dependencies
         id: depends
+        continue-on-error: true
         run: |
           brew update
 
@@ -129,15 +130,17 @@ jobs:
 
       - name: Dependencies
         id: depends
+        continue-on-error: true
         run: |
           brew update
 
       - name: Build
         id: cmake_build
         run: |
+          sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_AVX2=OFF ..
+          cmake -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF ..
           cmake --build . --config Release
 
       - name: Test

CMakeLists.txt

@@ -68,8 +68,9 @@ option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
+option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
-set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
+set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_DMMV_F16                   "llama: use 16 bit floats for dmmv CUDA kernels"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
@@ -79,7 +80,7 @@ option(LLAMA_QKK_64                          "llama: use super-block size of 64
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER               "llama: build server example"                           OFF)
+option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)
 
 #
 # Build info header
@@ -246,8 +247,14 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+        if (LLAMA_CUDA_FORCE_DMMV)
+            add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+        endif()
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-        add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
+        add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        if (DEFINED LLAMA_CUDA_DMMV_Y)
+            add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_DMMV_Y}) # for backwards compatibility
+        endif()
         if (LLAMA_CUDA_DMMV_F16)
             add_compile_definitions(GGML_CUDA_DMMV_F16)
         endif()
@@ -263,7 +270,7 @@ if (LLAMA_CUBLAS)
         if (LLAMA_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "61") # needed for f16 CUDA intrinsics
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "52") # lowest CUDA 12 standard
+            set(CMAKE_CUDA_ARCHITECTURES "52;61") # lowest CUDA 12 standard + lowest for integer intrinsics
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -386,11 +393,6 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES
     if (MSVC)
         # TODO: arm msvc?
     else()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
-            # Apple M1, M2, etc.
-            # Raspberry Pi 3, 4, Zero 2 (64-bit)
-            add_compile_options(-mcpu=native)
-        endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
             # Raspberry Pi 1, Zero
             add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access)

Makefile

@@ -1,11 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple libembdinput.so embd-input-test
-
-ifdef LLAMA_BUILD_SERVER
-	BUILD_TARGETS += server
-	LLAMA_SERVER_VERBOSE ?= 1
-server: private CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
-endif
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server libembdinput.so embd-input-test
 
 default: $(BUILD_TARGETS)
 
@@ -61,6 +55,10 @@ else
 	CXXFLAGS += -DNDEBUG
 endif
 
+ifdef LLAMA_SERVER_VERBOSE
+	CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
+endif
+
 # warnings
 CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith
 CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
@@ -166,16 +164,21 @@ ifdef LLAMA_CUBLAS
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
+ifdef LLAMA_CUDA_FORCE_DMMV
+	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+endif # LLAMA_CUDA_FORCE_DMMV
 ifdef LLAMA_CUDA_DMMV_X
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
-ifdef LLAMA_CUDA_DMMV_Y
-	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=$(LLAMA_CUDA_DMMV_Y)
+ifdef LLAMA_CUDA_MMV_Y
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+else ifdef LLAMA_CUDA_DMMV_Y
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
-endif # LLAMA_CUDA_DMMV_Y
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_DMMV_F16
 	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
 endif # LLAMA_CUDA_DMMV_F16
@@ -272,7 +275,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot train-text-from-scratch embd-input-test build-info.h
+	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h
 
 #
 # Examples

README.md

@@ -11,6 +11,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
+- Simple web chat example: https://github.com/ggerganov/llama.cpp/pull/1998
 - k-quants now support super-block size of 64: https://github.com/ggerganov/llama.cpp/pull/2001
 - New roadmap: https://github.com/users/ggerganov/projects/7
 - Azure CI brainstorming: https://github.com/ggerganov/llama.cpp/discussions/1985
@@ -85,6 +86,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
 - [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
+- [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B)
 
 **Bindings:**
 
@@ -343,8 +345,9 @@ Building the program with BLAS support may lead to some performance improvements
 
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 7.0/Turing/RTX 2000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA dequantization + mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |
   | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 

convert.py

@@ -136,7 +136,7 @@ def find_n_mult(n_ff: int, n_embd: int) -> int:
         calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
         if calc_ff == n_ff:
             return n_mult
-    return 1
+    raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).")
 
 @dataclass
 class Params:
@@ -321,6 +321,10 @@ class Tensor(metaclass=ABCMeta):
     @abstractmethod
     def permute(self, n_head: int) -> 'Tensor': ...
     @abstractmethod
+    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
+    @abstractmethod
+    def part(self, n_part: int) -> 'UnquantizedTensor': ...
+    @abstractmethod
     def to_ggml(self) -> 'GGMLCompatibleTensor': ...
 
 
@@ -345,6 +349,14 @@ class UnquantizedTensor(Tensor):
     def to_ggml(self) -> 'UnquantizedTensor':
         return self
 
+    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
+
+    def part(self, n_part: int) -> 'UnquantizedTensor':
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
+
     def permute(self, n_head: int) -> 'UnquantizedTensor':
         return UnquantizedTensor(permute(self.ndarray, n_head))
 
@@ -642,6 +654,19 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor:
         return lazy_tensor.load().permute(n_head)
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
 
+def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
+    def load() -> Tensor:
+        return lazy_tensor.load().permute_part(n_part, n_head)
+    s = lazy_tensor.shape.copy()
+    s[0] = s[0] // 3
+    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+
+def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
+    def load() -> Tensor:
+        return lazy_tensor.load().part(n_part)
+    s = lazy_tensor.shape.copy()
+    s[0] = s[0] // 3
+    return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description)
 
 def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
     out: LazyModel = {}
@@ -650,11 +675,17 @@ def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
     out["output.weight"] = model["lm_head.weight"]
 
     for i in itertools.count():
-        if f"model.layers.{i}.self_attn.q_proj.weight" not in model:
+        if f"model.layers.{i}.self_attn.q_proj.weight" in model:
+            out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
+            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head)
+            out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
+        elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
+            out[f"layers.{i}.attention.wq.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)
+            out[f"layers.{i}.attention.wk.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head)
+            out[f"layers.{i}.attention.wv.weight"] = part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
+        else:
             break
-        out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
-        out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head)
-        out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
+
         out[f"layers.{i}.attention.wo.weight"] = model[f"model.layers.{i}.self_attn.o_proj.weight"]
 
         out[f"layers.{i}.feed_forward.w1.weight"] = model[f"model.layers.{i}.mlp.gate_proj.weight"]

examples/common.cpp

@@ -110,7 +110,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.seed = std::stoi(argv[i]);
+            params.seed = std::stoul(argv[i]);
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
                 invalid_param = true;

examples/common.h

@@ -22,7 +22,7 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed                            = -1;  // RNG seed
+    uint32_t seed                           = -1;  // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
     int32_t n_predict                       = -1;  // new tokens to predict
     int32_t n_ctx                           = 512; // context size
@@ -31,7 +31,7 @@ struct gpt_params {
     int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
     int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
-    bool    low_vram                        = 0;   // if true, reduce VRAM usage at the cost of performance
+    int32_t n_probs                         = 0;   // if greater than 0, output the probabilities of top n_probs tokens.
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
@@ -59,6 +59,7 @@ struct gpt_params {
     std::string lora_adapter = "";  // lora adapter path
     std::string lora_base    = "";  // base model path for the lora adapter
 
+    bool low_vram          = false;   // if true, reduce VRAM usage at the cost of performance
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs

examples/embd-input/embd-input-lib.cpp

@@ -29,7 +29,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
-    if (params.seed < 0) {
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
     fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
@@ -210,9 +210,12 @@ llama_token sampling_id(struct MyModel* mymodel) {
 const char * sampling(struct MyModel * mymodel) {
     llama_context * ctx = mymodel->ctx;
     int id = sampling_id(mymodel);
-    std::string ret;
-    if (id == llama_token_eos()) ret = "</s>";
-    else ret = llama_token_to_str(ctx, id);
+    static std::string ret;
+    if (id == llama_token_eos()) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_str(ctx, id);
+    }
     eval_id(mymodel, id);
     return ret.c_str();
 }

examples/embd-input/embd-input.h

@@ -5,7 +5,6 @@
 #include "llama.h"
 #include "build-info.h"
 
-
 extern "C" {
 
 typedef struct MyModel {
@@ -14,14 +13,13 @@ typedef struct MyModel {
     int n_past = 0;
 } MyModel;
 
-
 struct MyModel* create_mymodel(int argc, char ** argv);
 
 bool eval_float(void* model, float* input, int N);
 bool eval_tokens(void* model, std::vector<llama_token> tokens);
 bool eval_id(struct MyModel* mymodel, int id);
 bool eval_string(struct MyModel* mymodel, const char* str);
-const char* sampling(struct MyModel* mymodel);
+const char * sampling(struct MyModel* mymodel);
 llama_token sampling_id(struct MyModel* mymodel);
 void free_mymodel(struct MyModel* mymodel);
 

examples/embedding/embedding.cpp

@@ -24,11 +24,11 @@ int main(int argc, char ** argv) {
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
-    if (params.seed < 0) {
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {

examples/main/README.md

@@ -242,7 +242,7 @@ Example usage: `--logit-bias 29905-inf`
 
 ### RNG Seed
 
--   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
+-   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
 
 The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
 

examples/main/main.cpp

@@ -94,11 +94,11 @@ int main(int argc, char ** argv) {
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
-    if (params.seed < 0) {
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {

examples/perplexity/perplexity.cpp

@@ -136,11 +136,11 @@ int main(int argc, char ** argv) {
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
-    if (params.seed < 0) {
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {

examples/quantize-stats/quantize-stats.cpp

@@ -147,7 +147,7 @@ void test_roundtrip_on_chunk(
         const ggml_tensor * layer,
         int64_t offset,
         int64_t chunk_size,
-        const quantize_fns_t & qfns,
+        const ggml_type_traits_t & qfns,
         bool use_reference,
         float * input_scratch,
         char * quantized_scratch,
@@ -163,11 +163,11 @@ void test_roundtrip_on_chunk(
     }
 
     if (use_reference) {
-        qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
+        qfns.from_float_reference(input_scratch, quantized_scratch, chunk_size);
     } else {
-        qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
+        qfns.from_float(input_scratch, quantized_scratch, chunk_size);
     }
-    qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
+    qfns.to_float(quantized_scratch, output_scratch, chunk_size);
 
     update_error_stats(chunk_size, input_scratch, output_scratch, stats);
 }
@@ -177,7 +177,7 @@ void test_roundtrip_on_chunk(
 void test_roundtrip_on_layer(
         std::string & name,
         bool print_layer_stats,
-        const quantize_fns_t & qfns,
+        const ggml_type_traits_t & qfns,
         bool use_reference,
         const ggml_tensor * layer,
         std::vector<float> & input_scratch,
@@ -388,8 +388,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
-        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
+        if (qfns.from_float && qfns.to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
             }

examples/server/README.md

@@ -1,6 +1,6 @@
 # llama.cpp/example/server
 
-This example demonstrates a simple HTTP API server to interact with llama.cpp.
+This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
 
 Command line options:
 
@@ -21,6 +21,7 @@ Command line options:
 -   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 -   `--port`: Set the port to listen. Default: `8080`.
+-   `--public`: path from which to serve static files (default examples/server/public)
 -   `--embedding`: Enable embedding extraction, Default: disabled.
 
 ## Build
@@ -59,7 +60,7 @@ server.exe -m models\7B\ggml-model.bin -c 2048
 ```
 
 The above command will start a server that by default listens on `127.0.0.1:8080`.
-You can consume the endpoints with Postman or NodeJS with axios library.
+You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url.
 
 ## Testing with CURL
 
@@ -152,7 +153,7 @@ node .
 
     `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
 
-    `seed`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
+    `seed`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
 
     `ignore_eos`: Ignore end of stream token and continue generating (default: false).
 
@@ -190,3 +191,19 @@ Run with bash:
 ```sh
 bash chat.sh
 ```
+
+### API like OAI
+
+API example using Python Flask: [api_like_OAI.py](api_like_OAI.py)
+This example must be used with server.cpp
+
+```sh
+python api_like_OAI.py
+```
+
+After running the API server, you can use it in Python by setting the API base URL.
+```python
+openai.api_base = "http://<Your api-server IP>:port"
+```
+
+Then you can utilize llama.cpp as an OpenAI's **chat.completion** or **text_completion** API

examples/server/api_like_OAI.py

@@ -0,0 +1,219 @@
+import argparse
+from flask import Flask, jsonify, request, Response
+import urllib.parse
+import requests
+import time
+import json
+
+
+app = Flask(__name__)
+
+parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
+parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')
+parser.add_argument("--user-name", type=str, help="USER name in chat completions(default: '\\nUSER: ')", default="\\nUSER: ")
+parser.add_argument("--ai-name", type=str, help="ASSISTANT name in chat completions(default: '\\nASSISTANT: ')", default="\\nASSISTANT: ")
+parser.add_argument("--system-name", type=str, help="SYSTEM name in chat completions(default: '\\nASSISTANT's RULE: ')", default="\\nASSISTANT's RULE: ")
+parser.add_argument("--stop", type=str, help="the end of response in chat completions(default: '</s>')", default="</s>")
+parser.add_argument("--llama-api", type=str, help="Set the address of server.cpp in llama.cpp(default: http://127.0.0.1:8080)", default='http://127.0.0.1:8080')
+parser.add_argument("--api-key", type=str, help="Set the api key to allow only few user(default: NULL)", default="")
+parser.add_argument("--host", type=str, help="Set the ip address to listen.(default: 127.0.0.1)", default='127.0.0.1')
+parser.add_argument("--port", type=int, help="Set the port to listen.(default: 8081)", default=8081)
+
+args = parser.parse_args()
+
+def is_present(json, key):
+    try:
+        buf = json[key]
+    except KeyError:
+        return False
+    return True
+
+
+
+#convert chat to prompt
+def convert_chat(messages):
+    prompt = "" + args.chat_prompt.replace("\\n", "\n")
+
+    system_n = args.system_name.replace("\\n", "\n")
+    user_n = args.user_name.replace("\\n", "\n")
+    ai_n = args.ai_name.replace("\\n", "\n")
+    stop = args.stop.replace("\\n", "\n")
+
+
+    for line in messages:
+        if (line["role"] == "system"):
+            prompt += f"{system_n}{line['content']}"
+        if (line["role"] == "user"):
+            prompt += f"{user_n}{line['content']}"
+        if (line["role"] == "assistant"):
+            prompt += f"{ai_n}{line['content']}{stop}"
+    prompt += ai_n.rstrip()
+
+    return prompt
+
+def make_postData(body, chat=False, stream=False):
+    postData = {}
+    if (chat):
+        postData["prompt"] = convert_chat(body["messages"])
+    else:
+        postData["prompt"] = body["prompt"]
+    if(is_present(body, "temperature")): postData["temperature"] = body["temperature"]
+    if(is_present(body, "top_k")): postData["top_k"] = body["top_k"]
+    if(is_present(body, "top_p")): postData["top_p"] = body["top_p"]
+    if(is_present(body, "max_tokens")): postData["n_predict"] = body["max_tokens"]
+    if(is_present(body, "presence_penalty")): postData["presence_penalty"] = body["presence_penalty"]
+    if(is_present(body, "frequency_penalty")): postData["frequency_penalty"] = body["frequency_penalty"]
+    if(is_present(body, "repeat_penalty")): postData["repeat_penalty"] = body["repeat_penalty"]
+    if(is_present(body, "mirostat")): postData["mirostat"] = body["mirostat"]
+    if(is_present(body, "mirostat_tau")): postData["mirostat_tau"] = body["mirostat_tau"]
+    if(is_present(body, "mirostat_eta")): postData["mirostat_eta"] = body["mirostat_eta"]
+    if(is_present(body, "seed")): postData["seed"] = body["seed"]
+    if(is_present(body, "logit_bias")): postData["logit_bias"] = [[int(token), body["logit_bias"][token]] for token in body["logit_bias"].keys()]
+    if (args.stop != ""):
+        postData["stop"] = [args.stop]
+    else:
+        postData["stop"] = []
+    if(is_present(body, "stop")): postData["stop"] += body["stop"]
+    postData["n_keep"] = -1
+    postData["stream"] = stream
+
+    return postData
+
+def make_resData(data, chat=False, promptToken=[]):
+    resData = {
+        "id": "chatcmpl" if (chat) else "cmpl",
+        "object": "chat.completion" if (chat) else "text_completion",
+        "created": int(time.time()),
+        "truncated": data["truncated"],
+        "model": "LLaMA_CPP",
+        "usage": {
+            "prompt_tokens": data["tokens_evaluated"],
+            "completion_tokens": data["tokens_predicted"],
+            "total_tokens": data["tokens_evaluated"] + data["tokens_predicted"]
+        }
+    }
+    if (len(promptToken) != 0):
+        resData["promptToken"] = promptToken
+    if (chat):
+        #only one choice is supported
+        resData["choices"] = [{
+            "index": 0,
+            "message": {
+                "role": "assistant",
+                "content": data["content"],
+            },
+            "finish_reason": "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+        }]
+    else:
+        #only one choice is supported
+        resData["choices"] = [{
+            "text": data["content"],
+            "index": 0,
+            "logprobs": None,
+            "finish_reason": "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+        }]
+    return resData
+
+def make_resData_stream(data, chat=False, time_now = 0, start=False):
+    resData = {
+        "id": "chatcmpl" if (chat) else "cmpl",
+        "object": "chat.completion.chunk" if (chat) else "text_completion.chunk",
+        "created": time_now,
+        "model": "LLaMA_CPP",
+        "choices": [
+            {
+                "finish_reason": None,
+                "index": 0
+            }
+        ]
+    }
+    if (chat):
+        if (start):
+            resData["choices"][0]["delta"] =  {
+                "role": "assistant"
+            }
+        else:
+            resData["choices"][0]["delta"] =  {
+                "content": data["content"]
+            }
+            if (data["stop"]):
+                resData["choices"][0]["finish_reason"] = "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+    else:
+        resData["choices"][0]["text"] = data["content"]
+        if (data["stop"]):
+            resData["choices"][0]["finish_reason"] = "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+
+    return resData
+
+
+@app.route('/chat/completions', methods=['POST'])
+@app.route('/v1/chat/completions', methods=['POST'])
+def chat_completions():
+    if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
+        return Response(status=403)
+    body = request.get_json()
+    stream = False
+    tokenize = False
+    if(is_present(body, "stream")): stream = body["stream"]
+    if(is_present(body, "tokenize")): tokenize = body["tokenize"]
+    postData = make_postData(body, chat=True, stream=stream)
+
+    promptToken = []
+    if (tokenize):
+        tokenData = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/tokenize"), data=json.dumps({"content": postData["prompt"]})).json()
+        promptToken = tokenData["tokens"]
+
+    if (not stream):
+        data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData))
+        print(data.json())
+        resData = make_resData(data.json(), chat=True, promptToken=promptToken)
+        return jsonify(resData)
+    else:
+        def generate():
+            data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
+            time_now = int(time.time())
+            resData = make_resData_stream({}, chat=True, time_now=time_now, start=True)
+            yield 'data: {}\n'.format(json.dumps(resData))
+            for line in data.iter_lines():
+                if line:
+                    decoded_line = line.decode('utf-8')
+                    resData = make_resData_stream(json.loads(decoded_line[6:]), chat=True, time_now=time_now)
+                    yield 'data: {}\n'.format(json.dumps(resData))
+        return Response(generate(), mimetype='text/event-stream')
+
+
+@app.route('/completions', methods=['POST'])
+@app.route('/v1/completions', methods=['POST'])
+def completion():
+    if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
+        return Response(status=403)
+    body = request.get_json()
+    stream = False
+    tokenize = False
+    if(is_present(body, "stream")): stream = body["stream"]
+    if(is_present(body, "tokenize")): tokenize = body["tokenize"]
+    postData = make_postData(body, chat=False, stream=stream)
+
+    promptToken = []
+    if (tokenize):
+        tokenData = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/tokenize"), data=json.dumps({"content": postData["prompt"]})).json()
+        promptToken = tokenData["tokens"]
+
+    if (not stream):
+        data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData))
+        print(data.json())
+        resData = make_resData(data.json(), chat=False, promptToken=promptToken)
+        return jsonify(resData)
+    else:
+        def generate():
+            data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
+            time_now = int(time.time())
+            for line in data.iter_lines():
+                if line:
+                    decoded_line = line.decode('utf-8')
+                    resData = make_resData_stream(json.loads(decoded_line[6:]), chat=False, time_now=time_now)
+                    yield 'data: {}\n'.format(json.dumps(resData))
+        return Response(generate(), mimetype='text/event-stream')
+
+if __name__ == '__main__':
+    app.run(args.host, port=args.port)

examples/server/completion.js.hpp

@@ -0,0 +1,193 @@
+unsigned char completion_js[] = {
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x44,
+  0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x3a, 0x20, 0x74, 0x72,
+  0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x3a, 0x20, 0x35, 0x30, 0x30, 0x2c, 0x0a, 0x20, 0x20,
+  0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x3a,
+  0x20, 0x30, 0x2e, 0x32, 0x2c, 0x0a, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70,
+  0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x5d, 0x0a, 0x7d,
+  0x3b, 0x0a, 0x0a, 0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x54, 0x68,
+  0x69, 0x73, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20,
+  0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x73, 0x20, 0x74, 0x68,
+  0x65, 0x20, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x20, 0x6c, 0x6c, 0x61,
+  0x6d, 0x61, 0x20, 0x64, 0x69, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x72,
+  0x79, 0x2e, 0x0a, 0x20, 0x2a, 0x20, 0x40, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x20, 0x7b, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x7d, 0x20, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x20, 0x2d, 0x20, 0x54, 0x68, 0x65, 0x20, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x66, 0x6f,
+  0x72, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x69, 0x6f, 0x6e, 0x20, 0x72, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74,
+  0x2e, 0x0a, 0x20, 0x2a, 0x20, 0x40, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x20,
+  0x7b, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x7d, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x2d, 0x20, 0x61, 0x6e,
+  0x20, 0x69, 0x6e, 0x73, 0x74, 0x61, 0x6e, 0x63, 0x65, 0x20, 0x6f, 0x66,
+  0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f,
+  0x6c, 0x6c, 0x65, 0x72, 0x20, 0x69, 0x66, 0x20, 0x79, 0x6f, 0x75, 0x20,
+  0x6e, 0x65, 0x65, 0x64, 0x20, 0x6f, 0x6e, 0x65, 0x2c, 0x20, 0x6f, 0x72,
+  0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x2e, 0x0a, 0x20, 0x2a, 0x20, 0x40, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x20, 0x7b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69,
+  0x6f, 0x6e, 0x7d, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b,
+  0x20, 0x2d, 0x20, 0x54, 0x68, 0x65, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
+  0x61, 0x63, 0x6b, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e,
+  0x20, 0x74, 0x6f, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x20, 0x77, 0x68, 0x65,
+  0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x6e, 0x65,
+  0x2e, 0x0a, 0x20, 0x2a, 0x20, 0x40, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
+  0x73, 0x20, 0x7b, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x7d, 0x20, 0x74,
+  0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64,
+  0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x73,
+  0x74, 0x72, 0x69, 0x6e, 0x67, 0x2e, 0x20, 0x49, 0x64, 0x65, 0x61, 0x6c,
+  0x6c, 0x79, 0x20, 0x69, 0x67, 0x6e, 0x6f, 0x72, 0x65, 0x64, 0x2c, 0x20,
+  0x61, 0x6e, 0x64, 0x20, 0x79, 0x6f, 0x75, 0x20, 0x67, 0x65, 0x74, 0x20,
+  0x61, 0x74, 0x20, 0x69, 0x74, 0x20, 0x76, 0x69, 0x61, 0x20, 0x74, 0x68,
+  0x65, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x2e, 0x0a,
+  0x20, 0x2a, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79,
+  0x6e, 0x63, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20,
+  0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x63, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f,
+  0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72,
+  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69,
+  0x6f, 0x6e, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b,
+  0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x44, 0x65, 0x66,
+  0x61, 0x75, 0x6c, 0x74, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x77, 0x65, 0x20, 0x75, 0x73, 0x65, 0x20, 0x66, 0x65, 0x74,
+  0x63, 0x68, 0x20, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x6c, 0x79, 0x20,
+  0x68, 0x65, 0x72, 0x65, 0x20, 0x62, 0x65, 0x63, 0x61, 0x73, 0x75, 0x65,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x62, 0x75, 0x69, 0x6c, 0x74, 0x20, 0x69,
+  0x6e, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74,
+  0x53, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20,
+  0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20,
+  0x50, 0x4f, 0x53, 0x54, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x72, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x20, 0x3d, 0x20,
+  0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28,
+  0x22, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e,
+  0x22, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x74,
+  0x68, 0x6f, 0x64, 0x3a, 0x20, 0x27, 0x50, 0x4f, 0x53, 0x54, 0x27, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x64, 0x79, 0x3a, 0x20, 0x4a,
+  0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66,
+  0x79, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e,
+  0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x73, 0x3a, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x43, 0x6f, 0x6e, 0x6e, 0x65,
+  0x63, 0x74, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x27, 0x6b, 0x65, 0x65,
+  0x70, 0x2d, 0x61, 0x6c, 0x69, 0x76, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x27, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x2d, 0x54, 0x79, 0x70, 0x65, 0x27, 0x3a, 0x20, 0x27, 0x61, 0x70, 0x70,
+  0x6c, 0x69, 0x63, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x2f, 0x6a, 0x73, 0x6f,
+  0x6e, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x41,
+  0x63, 0x63, 0x65, 0x70, 0x74, 0x27, 0x3a, 0x20, 0x27, 0x74, 0x65, 0x78,
+  0x74, 0x2f, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2d, 0x73, 0x74, 0x72, 0x65,
+  0x61, 0x6d, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x3a, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x73, 0x69,
+  0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x61,
+  0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x70, 0x6f, 0x6e,
+  0x73, 0x65, 0x2e, 0x62, 0x6f, 0x64, 0x79, 0x2e, 0x67, 0x65, 0x74, 0x52,
+  0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72,
+  0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x54, 0x65, 0x78, 0x74, 0x44,
+  0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72,
+  0x79, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65,
+  0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74,
+  0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64,
+  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f, 0x6e,
+  0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x73, 0x73, 0x65, 0x20, 0x61, 0x6e, 0x73, 0x77, 0x65, 0x72,
+  0x73, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x66, 0x6f, 0x72,
+  0x6d, 0x20, 0x6d, 0x75, 0x6c, 0x74, 0x69, 0x70, 0x6c, 0x65, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x73, 0x20, 0x6f, 0x66, 0x3a, 0x20, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x5c, 0x6e, 0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x64, 0x61,
+  0x74, 0x61, 0x20, 0x61, 0x6c, 0x77, 0x61, 0x79, 0x73, 0x20, 0x70, 0x72,
+  0x65, 0x73, 0x65, 0x6e, 0x74, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x6b,
+  0x65, 0x79, 0x2e, 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x75, 0x72, 0x20, 0x63,
+  0x61, 0x73, 0x65, 0x20, 0x77, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x6d, 0x61, 0x69, 0x6e, 0x6c, 0x79, 0x20, 0x63,
+  0x61, 0x72, 0x65, 0x20, 0x61, 0x62, 0x6f, 0x75, 0x74, 0x20, 0x74, 0x68,
+  0x65, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x6b, 0x65, 0x79, 0x20,
+  0x68, 0x65, 0x72, 0x65, 0x2c, 0x20, 0x77, 0x68, 0x69, 0x63, 0x68, 0x20,
+  0x77, 0x65, 0x20, 0x65, 0x78, 0x70, 0x65, 0x63, 0x74, 0x20, 0x61, 0x73,
+  0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d,
+  0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63,
+  0x6f, 0x64, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x20, 0x61,
+  0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
+  0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64, 0x20, 0x74, 0x68,
+  0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20, 0x2f, 0x5e, 0x28,
+  0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e, 0x2a, 0x29, 0x24,
+  0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66,
+  0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x20, 0x6f, 0x66, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e,
+  0x6d, 0x61, 0x74, 0x63, 0x68, 0x41, 0x6c, 0x6c, 0x28, 0x72, 0x65, 0x67,
+  0x65, 0x78, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x73, 0x69, 0x6e, 0x63, 0x65, 0x20, 0x77, 0x65, 0x20, 0x6b, 0x6e,
+  0x6f, 0x77, 0x20, 0x74, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x2c, 0x20, 0x6c, 0x65,
+  0x74, 0x27, 0x73, 0x20, 0x6a, 0x75, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63,
+  0x6f, 0x64, 0x65, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6a, 0x73, 0x6f, 0x6e,
+  0x20, 0x69, 0x6e, 0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61,
+  0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61,
+  0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
+  0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x61, 0x63,
+  0x6b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x20, 0x3d, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x28,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x29, 0x20, 0x21, 0x3d, 0x20, 0x66,
+  0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20,
+  0x73, 0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20,
+  0x77, 0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61,
+  0x6b, 0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e,
+  0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65,
+  0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61,
+  0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65, 0x72,
+  0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x65, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65,
+  0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61,
+  0x6c, 0x6c, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f,
+  0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x7d, 0x0a
+};
+unsigned int completion_js_len = 2275;

examples/server/deps.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+# Download and update deps for binary
+
+# get the directory of this script file
+DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
+PUBLIC=$DIR/public
+OUTPUT=$DIR/templats.hpp
+
+echo "// Generated file, do not edit" > $OUTPUT
+echo "" > $OUTPUT
+
+echo "download js bundle files"
+curl https://npm.reversehttp.com/@preact/signals-core,@preact/signals,htm/preact,preact,preact/hooks > $PUBLIC/index.js
+echo >> $PUBLIC/index.js # add newline
+
+FILES=$(ls $PUBLIC)
+
+for FILE in $FILES; do
+  func=$(echo $FILE | tr '.' '_')
+  echo "generate $FILE.hpp ($func)"
+  xxd -n $func -i $PUBLIC/$FILE > $DIR/$FILE.hpp
+done

examples/server/index.html.hpp

@@ -0,0 +1,846 @@
+unsigned char index_html[] = {
+  0x3c, 0x68, 0x74, 0x6d, 0x6c, 0x3e, 0x0a, 0x0a, 0x3c, 0x68, 0x65, 0x61,
+  0x64, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x6d, 0x65, 0x74, 0x61, 0x20, 0x63,
+  0x68, 0x61, 0x72, 0x73, 0x65, 0x74, 0x3d, 0x22, 0x55, 0x54, 0x46, 0x2d,
+  0x38, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x6d, 0x65, 0x74, 0x61, 0x20,
+  0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x76, 0x69, 0x65, 0x77, 0x70, 0x6f,
+  0x72, 0x74, 0x22, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3d,
+  0x22, 0x77, 0x69, 0x64, 0x74, 0x68, 0x3d, 0x64, 0x65, 0x76, 0x69, 0x63,
+  0x65, 0x2d, 0x77, 0x69, 0x64, 0x74, 0x68, 0x2c, 0x20, 0x69, 0x6e, 0x69,
+  0x74, 0x69, 0x61, 0x6c, 0x2d, 0x73, 0x63, 0x61, 0x6c, 0x65, 0x3d, 0x31,
+  0x2c, 0x20, 0x6d, 0x61, 0x78, 0x69, 0x6d, 0x75, 0x6d, 0x2d, 0x73, 0x63,
+  0x61, 0x6c, 0x65, 0x3d, 0x31, 0x22, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20,
+  0x3c, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x2e, 0x63, 0x70, 0x70, 0x20, 0x2d, 0x20, 0x63, 0x68, 0x61, 0x74, 0x3c,
+  0x2f, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x3c,
+  0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x62, 0x6f, 0x64, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d,
+  0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x66, 0x66, 0x66, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6c, 0x6f, 0x72,
+  0x3a, 0x20, 0x23, 0x30, 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c,
+  0x79, 0x3a, 0x20, 0x73, 0x79, 0x73, 0x74, 0x65, 0x6d, 0x2d, 0x75, 0x69,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, 0x74,
+  0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x39, 0x30, 0x25, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x23,
+  0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e,
+  0x3a, 0x20, 0x30, 0x65, 0x6d, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61,
+  0x79, 0x3a, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d, 0x64, 0x69, 0x72, 0x65,
+  0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x63, 0x6f, 0x6c, 0x75, 0x6d,
+  0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6a, 0x75, 0x73,
+  0x74, 0x69, 0x66, 0x79, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x3a, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x2d, 0x62, 0x65, 0x74, 0x77,
+  0x65, 0x65, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68,
+  0x65, 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x66, 0x6f, 0x6f, 0x74,
+  0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x2d, 0x61, 0x6c, 0x69, 0x67, 0x6e, 0x3a, 0x20, 0x63,
+  0x65, 0x6e, 0x74, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x69, 0x6e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69,
+  0x6e, 0x3a, 0x20, 0x33, 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x66,
+  0x6c, 0x65, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66,
+  0x6c, 0x65, 0x78, 0x2d, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x69, 0x6f,
+  0x6e, 0x3a, 0x20, 0x63, 0x6f, 0x6c, 0x75, 0x6d, 0x6e, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x6a, 0x75, 0x73, 0x74, 0x69, 0x66, 0x79,
+  0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3a, 0x20, 0x73, 0x70,
+  0x61, 0x63, 0x65, 0x2d, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x61, 0x70, 0x3a, 0x20,
+  0x31, 0x65, 0x6d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x66, 0x6c, 0x65, 0x78, 0x2d, 0x67, 0x72, 0x6f, 0x77, 0x3a, 0x20, 0x31,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x76, 0x65, 0x72,
+  0x66, 0x6c, 0x6f, 0x77, 0x2d, 0x79, 0x3a, 0x20, 0x61, 0x75, 0x74, 0x6f,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72,
+  0x64, 0x65, 0x72, 0x3a, 0x20, 0x31, 0x70, 0x78, 0x20, 0x73, 0x6f, 0x6c,
+  0x69, 0x64, 0x20, 0x23, 0x63, 0x63, 0x63, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x2d, 0x72, 0x61,
+  0x64, 0x69, 0x75, 0x73, 0x3a, 0x20, 0x35, 0x70, 0x78, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67,
+  0x3a, 0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x64, 0x79,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x78,
+  0x2d, 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x36, 0x30, 0x30, 0x70,
+  0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x6e,
+  0x2d, 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x33, 0x30, 0x30, 0x70,
+  0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x2d, 0x68, 0x65, 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, 0x31, 0x2e,
+  0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72,
+  0x67, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69,
+  0x6e, 0x67, 0x3a, 0x20, 0x30, 0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x70, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x76,
+  0x65, 0x72, 0x66, 0x6c, 0x6f, 0x77, 0x2d, 0x77, 0x72, 0x61, 0x70, 0x3a,
+  0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2d, 0x77, 0x6f, 0x72, 0x64, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x6f, 0x72, 0x64, 0x2d,
+  0x77, 0x72, 0x61, 0x70, 0x3a, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2d,
+  0x77, 0x6f, 0x72, 0x64, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x68, 0x79, 0x70, 0x68, 0x65, 0x6e, 0x73, 0x3a, 0x20, 0x61, 0x75, 0x74,
+  0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72,
+  0x67, 0x69, 0x6e, 0x2d, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x35,
+  0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61,
+  0x72, 0x67, 0x69, 0x6e, 0x2d, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x3a,
+  0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x23, 0x77, 0x72, 0x69, 0x74,
+  0x65, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, 0x31,
+  0x65, 0x6d, 0x20, 0x30, 0x20, 0x30, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a,
+  0x20, 0x66, 0x6c, 0x65, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74,
+  0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x63, 0x6f, 0x6c, 0x75, 0x6d, 0x6e, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x61, 0x70, 0x3a, 0x20,
+  0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x61, 0x6c, 0x69, 0x67, 0x6e, 0x2d, 0x69, 0x74, 0x65, 0x6d, 0x73,
+  0x3a, 0x20, 0x73, 0x74, 0x72, 0x65, 0x74, 0x63, 0x68, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72,
+  0x69, 0x67, 0x68, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x66, 0x6c,
+  0x65, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c,
+  0x65, 0x78, 0x2d, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e,
+  0x3a, 0x20, 0x72, 0x6f, 0x77, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x67, 0x61, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6a, 0x75, 0x73, 0x74, 0x69,
+  0x66, 0x79, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3a, 0x20,
+  0x66, 0x6c, 0x65, 0x78, 0x2d, 0x65, 0x6e, 0x64, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x65,
+  0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20, 0x6e, 0x6f,
+  0x6e, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61,
+  0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20,
+  0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69,
+  0x6e, 0x67, 0x3a, 0x20, 0x35, 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d, 0x67, 0x72, 0x6f, 0x77,
+  0x3a, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77,
+  0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x70,
+  0x72, 0x65, 0x20, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a,
+  0x20, 0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64,
+  0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6c, 0x6f,
+  0x72, 0x3a, 0x20, 0x23, 0x64, 0x64, 0x64, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x64, 0x65, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, 0x74,
+  0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a, 0x20, 0x6d, 0x6f, 0x6e,
+  0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30,
+  0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, 0x65, 0x72,
+  0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a, 0x20, 0x33, 0x70, 0x78,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x6c, 0x61,
+  0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x35, 0x65,
+  0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
+  0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, 0x6f, 0x63,
+  0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x3c,
+  0x2f, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x3c,
+  0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d,
+  0x22, 0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20, 0x68,
+  0x2c, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x65, 0x66,
+  0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74,
+  0x65, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c, 0x20,
+  0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x75,
+  0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75, 0x73,
+  0x65, 0x52, 0x65, 0x66, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e,
+  0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x20, 0x7d, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61,
+  0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72,
+  0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, 0x20,
+  0x69, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73,
+  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65,
+  0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, 0x65,
+  0x6e, 0x64, 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, 0x74,
+  0x2e, 0x20, 0x72, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x64, 0x20, 0x69, 0x6e,
+  0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x2e, 0x22, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, 0x6e, 0x7b, 0x7b, 0x68, 0x69,
+  0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, 0x5c, 0x6e, 0x7b, 0x7b, 0x63,
+  0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, 0x6e,
+  0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, 0x7b, 0x7b, 0x6d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70,
+  0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74,
+  0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61,
+  0x72, 0x3a, 0x20, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x22, 0x2c, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x3a, 0x20,
+  0x22, 0x55, 0x73, 0x65, 0x72, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74,
+  0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x74, 0x72,
+  0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
+  0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64,
+  0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28,
+  0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f,
+  0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e,
+  0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74,
+  0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74,
+  0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c,
+  0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, 0x2c, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74,
+  0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e,
+  0x31, 0x38, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f,
+  0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x3a, 0x20, 0x30, 0x2e,
+  0x35, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69,
+  0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e,
+  0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f,
+  0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x6e, 0x75, 0x6c,
+  0x6c, 0x20, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x74,
+  0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x74,
+  0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73,
+  0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, 0x74,
+  0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65,
+  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e,
+  0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, 0x20,
+  0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28,
+  0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65,
+  0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, 0x2a,
+  0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, 0x5f,
+  0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
+  0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
+  0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
+  0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e,
+  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67,
+  0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75,
+  0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f,
+  0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d,
+  0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73,
+  0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28,
+  0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63,
+  0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, 0x73,
+  0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x5d,
+  0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x70, 0x61, 0x79, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x3d,
+  0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65,
+  0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72,
+  0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72,
+  0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28,
+  0x28, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x6d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f,
+  0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74,
+  0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c,
+  0x20, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x6d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x7d, 0x29, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28,
+  0x22, 0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6c, 0x65, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d,
+  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73,
+  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x70, 0x61,
+  0x79, 0x6c, 0x6f, 0x61, 0x64, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x22, 0x3c,
+  0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d,
+  0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x3a,
+  0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69,
+  0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c,
+  0x65, 0x74, 0x65, 0x28, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
+  0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x28,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x6d,
+  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72,
+  0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20,
+  0x2b, 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x20, 0x6c,
+  0x65, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20, 0x77, 0x68, 0x69, 0x74, 0x65,
+  0x73, 0x70, 0x61, 0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65,
+  0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x72, 0x65,
+  0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2b, 0x2f,
+  0x2c, 0x20, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70,
+  0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e,
+  0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x22, 0x7b,
+  0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x63, 0x75,
+  0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x5d, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74,
+  0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e,
+  0x6c, 0x6f, 0x67, 0x28, 0x22, 0x2d, 0x2d, 0x3e, 0x22, 0x2c, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x2c, 0x20, 0x27, 0x20, 0x72, 0x65, 0x73, 0x70, 0x6f,
+  0x6e, 0x73, 0x65, 0x20, 0x77, 0x61, 0x73, 0x3a, 0x27, 0x2c, 0x20, 0x63,
+  0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67,
+  0x65, 0x2c, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69,
+  0x70, 0x74, 0x20, 0x73, 0x74, 0x61, 0x74, 0x65, 0x3a, 0x27, 0x2c, 0x20,
+  0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
+  0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
+  0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f,
+  0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70,
+  0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
+  0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e,
+  0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70,
+  0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72,
+  0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74,
+  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c,
+  0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65, 0x74, 0x20,
+  0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28,
+  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70,
+  0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69,
+  0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f,
+  0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73, 0x73, 0x61,
+  0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
+  0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x22,
+  0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73,
+  0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x68,
+  0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33, 0x20, 0x26,
+  0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x68, 0x69,
+  0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69,
+  0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72,
+  0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x3d, 0x24,
+  0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79,
+  0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x72, 0x6f,
+  0x77, 0x73, 0x3d, 0x32, 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72,
+  0x65, 0x73, 0x73, 0x3d, 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53,
+  0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, 0x7d, 0x20, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24,
+  0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
+  0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x7d, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c,
+  0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20, 0x73, 0x6f, 0x6d,
+  0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x22, 0x2f, 0x3e,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c,
+  0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74, 0x22, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d,
+  0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22, 0x20, 0x64, 0x69, 0x73,
+  0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x21, 0x67, 0x65, 0x6e,
+  0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x7d, 0x20, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c, 0x2f, 0x62, 0x75,
+  0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20,
+  0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, 0x74,
+  0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64,
+  0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e,
+  0x67, 0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70, 0x3c, 0x2f, 0x62, 0x75, 0x74,
+  0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f,
+  0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x72, 0x65, 0x73,
+  0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x3c, 0x2f, 0x62,
+  0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f,
+  0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67,
+  0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73,
+  0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72,
+  0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28,
+  0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28,
+  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c,
+  0x20, 0x74, 0x6f, 0x20, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x28,
+  0x69, 0x66, 0x20, 0x6e, 0x65, 0x65, 0x64, 0x65, 0x64, 0x29, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63,
+  0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72,
+  0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x61, 0x69, 0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e,
+  0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67,
+  0x68, 0x74, 0x20, 0x3c, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69,
+  0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e,
+  0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, 0x70, 0x20, 0x2b, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75,
+  0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x6f, 0x66, 0x66, 0x73, 0x65, 0x74,
+  0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x2b, 0x20, 0x33, 0x30, 0x30,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x2e,
+  0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f,
+  0x6c, 0x6c, 0x54, 0x6f, 0x28, 0x30, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x61, 0x69, 0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e,
+  0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67,
+  0x68, 0x74, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x20, 0x3d, 0x20, 0x28,
+  0x5b, 0x75, 0x73, 0x65, 0x72, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x29,
+  0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d,
+  0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x6d,
+  0x73, 0x67, 0x7d, 0x3e, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e,
+  0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75,
+  0x73, 0x65, 0x72, 0x29, 0x7d, 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f,
+  0x6e, 0x67, 0x3e, 0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x61, 0x72, 0x6b, 0x64,
+  0x6f, 0x77, 0x6e, 0x7d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x6d, 0x73, 0x67,
+  0x29, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74,
+  0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d,
+  0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24,
+  0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24,
+  0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c,
+  0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69,
+  0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x60,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66,
+  0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72,
+  0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70,
+  0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20,
+  0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65,
+  0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
+  0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69,
+  0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65,
+  0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d,
+  0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65,
+  0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70,
+  0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
+  0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
+  0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a,
+  0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74,
+  0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74,
+  0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
+  0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c,
+  0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65,
+  0x5d, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61,
+  0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68,
+  0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64,
+  0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d,
+  0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f,
+  0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20,
+  0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20,
+  0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
+  0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73,
+  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, 0x72, 0x6f,
+  0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
+  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69,
+  0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
+  0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, 0x6e,
+  0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70,
+  0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d,
+  0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f,
+  0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, 0x72,
+  0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24,
+  0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69,
+  0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76,
+  0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x62,
+  0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
+  0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22,
+  0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d,
+  0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, 0x20,
+  0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70,
+  0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d,
+  0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70,
+  0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74,
+  0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78,
+  0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f,
+  0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
+  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69,
+  0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68,
+  0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61,
+  0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70,
+  0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22,
+  0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22,
+  0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f,
+  0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
+  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69,
+  0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
+  0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22,
+  0x3e, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65,
+  0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
+  0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74,
+  0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x20,
+  0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x30, 0x2e, 0x30, 0x22, 0x20, 0x6d, 0x61,
+  0x78, 0x3d, 0x22, 0x31, 0x2e, 0x30, 0x22, 0x20, 0x73, 0x74, 0x65, 0x70,
+  0x3d, 0x22, 0x30, 0x2e, 0x30, 0x31, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72,
+  0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x2e, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65,
+  0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24,
+  0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65,
+  0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x7d, 0x3c, 0x2f,
+  0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x6e, 0x50,
+  0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x3e, 0x50, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, 0x2f, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74,
+  0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65,
+  0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x6e, 0x50, 0x72, 0x65, 0x64, 0x69,
+  0x63, 0x74, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x31, 0x22, 0x20,
+  0x6d, 0x61, 0x78, 0x3d, 0x22, 0x32, 0x30, 0x34, 0x38, 0x22, 0x20, 0x73,
+  0x74, 0x65, 0x70, 0x3d, 0x22, 0x31, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3d, 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22,
+  0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e,
+  0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x7d, 0x22, 0x20, 0x6f,
+  0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64,
+  0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f,
+  0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70,
+  0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66,
+  0x6f, 0x72, 0x3d, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70,
+  0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x3e, 0x50, 0x65, 0x6e, 0x61,
+  0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x20,
+  0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x3c, 0x2f, 0x6c, 0x61,
+  0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75,
+  0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67,
+  0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61,
+  0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x20, 0x6d,
+  0x69, 0x6e, 0x3d, 0x22, 0x30, 0x2e, 0x30, 0x22, 0x20, 0x6d, 0x61, 0x78,
+  0x3d, 0x22, 0x32, 0x2e, 0x30, 0x22, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3d,
+  0x22, 0x30, 0x2e, 0x30, 0x31, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d,
+  0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61,
+  0x6c, 0x74, 0x79, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22,
+  0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65,
+  0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e,
+  0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65,
+  0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d,
+  0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e,
+  0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65,
+  0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c,
+  0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74,
+  0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x3e, 0x43, 0x6f, 0x6e,
+  0x73, 0x69, 0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65,
+  0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c,
+  0x69, 0x7a, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70,
+  0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64,
+  0x3d, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73,
+  0x74, 0x5f, 0x6e, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x30, 0x2e,
+  0x30, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x32, 0x30, 0x34, 0x38,
+  0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x72, 0x65, 0x70, 0x65,
+  0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x20, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70,
+  0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x7d, 0x22,
+  0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75,
+  0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46,
+  0x6c, 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65,
+  0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x7d, 0x3c, 0x2f,
+  0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f,
+  0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, 0x3d, 0x20, 0x28,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x64, 0x20,
+  0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x74, 0x65, 0x78,
+  0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61,
+  0x63, 0x65, 0x28, 0x2f, 0x5e, 0x23, 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20,
+  0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27,
+  0x3c, 0x68, 0x33, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27,
+  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61,
+  0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f,
+  0x29, 0x5c, 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73,
+  0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74,
+  0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x5f,
+  0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27,
+  0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f,
+  0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f,
+  0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c,
+  0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d,
+  0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70,
+  0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29,
+  0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31,
+  0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x60,
+  0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, 0x73, 0x5c, 0x53,
+  0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27,
+  0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24,
+  0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, 0x2f, 0x70, 0x72,
+  0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65,
+  0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f,
+  0x29, 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65,
+  0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63,
+  0x65, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27,
+  0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60,
+  0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72,
+  0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65,
+  0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, 0x20, 0x5f, 0x5f,
+  0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, 0x7d, 0x7d, 0x20,
+  0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x70,
+  0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
+  0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f,
+  0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, 0x64, 0x65,
+  0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63,
+  0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72,
+  0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x6d, 0x61, 0x69, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74,
+  0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20,
+  0x3a, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d,
+  0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65,
+  0x72, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, 0x74, 0x65, 0x22,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x24, 0x7b, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e,
+  0x70, 0x75, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72,
+  0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77,
+  0x65, 0x72, 0x65, 0x64, 0x20, 0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68,
+  0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f,
+  0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f,
+  0x67, 0x67, 0x65, 0x72, 0x67, 0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61,
+  0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61,
+  0x6e, 0x64, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22,
+  0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c,
+  0x2e, 0x61, 0x69, 0x22, 0x3e, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69,
+  0x3c, 0x2f, 0x61, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65,
+  0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70,
+  0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e,
+  0x62, 0x6f, 0x64, 0x79, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73,
+  0x63, 0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61,
+  0x64, 0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x3c,
+  0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74,
+  0x6d, 0x6c, 0x3e, 0x0a
+};
+unsigned int index_html_len = 10108;

examples/server/public/completion.js

@@ -0,0 +1,81 @@
+const paramDefaults = {
+  stream: true,
+  n_predict: 500,
+  temperature: 0.2,
+  stop: ["</s>"]
+};
+
+/**
+ * This function completes the input text using a llama dictionary.
+ * @param {object} params - The parameters for the completion request.
+ * @param {object} controller - an instance of AbortController if you need one, or null.
+ * @param {function} callback - The callback function to call when the completion is done.
+ * @returns {string} the completed text as a string. Ideally ignored, and you get at it via the callback.
+ */
+export const llamaComplete = async (params, controller, callback) => {
+  if (!controller) {
+    controller = new AbortController();
+  }
+  const completionParams = { ...paramDefaults, ...params };
+
+  // we use fetch directly here becasue the built in fetchEventSource does not support POST
+  const response = await fetch("/completion", {
+    method: 'POST',
+    body: JSON.stringify(completionParams),
+    headers: {
+      'Connection': 'keep-alive',
+      'Content-Type': 'application/json',
+      'Accept': 'text/event-stream'
+    },
+    signal: controller.signal,
+  });
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+
+  let content = "";
+
+  try {
+
+    let cont = true;
+
+    while (cont) {
+      const result = await reader.read();
+      if (result.done) {
+        break;
+      }
+
+      // sse answers in the form multiple lines of: value\n with data always present as a key. in our case we
+      // mainly care about the data: key here, which we expect as json
+      const text = decoder.decode(result.value);
+
+      // parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const match of text.matchAll(regex)) {
+        result[match[1]] = match[2]
+      }
+
+      // since we know this is llama.cpp, let's just decode the json in data
+      result.data = JSON.parse(result.data);
+      content += result.data.content;
+
+      // callack
+      if (callback) {
+        cont = callback(result) != false;
+      }
+
+      // if we got a stop token from server, we will break here
+      if (result.data.stop) {
+        break;
+      }
+    }
+  } catch (e) {
+    console.error("llama error: ", e);
+    throw e;
+  }
+  finally {
+    controller.abort();
+  }
+
+  return content;
+}

examples/server/public/index.html

@@ -0,0 +1,359 @@
+<html>
+
+<head>
+  <meta charset="UTF-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1" />
+  <title>llama.cpp - chat</title>
+
+  <style>
+
+    body {
+      background-color: #fff;
+      color: #000;
+      font-family: system-ui;
+      font-size: 90%;
+    }
+
+    #container {
+      margin: 0em auto;
+      display: flex;
+      flex-direction: column;
+      justify-content: space-between;
+      height: 100%;
+    }
+
+    header, footer {
+      text-align: center;
+    }
+
+    main {
+      margin: 3px;
+      display: flex;
+      flex-direction: column;
+      justify-content: space-between;
+      gap: 1em;
+
+      flex-grow: 1;
+      overflow-y: auto;
+
+      border: 1px solid #ccc;
+      border-radius: 5px;
+      padding: 0.5em;
+    }
+
+    body {
+      max-width: 600px;
+      min-width: 300px;
+      line-height: 1.2;
+      margin: 0 auto;
+      padding: 0 0.5em;
+    }
+
+    p {
+      overflow-wrap: break-word;
+      word-wrap: break-word;
+      hyphens: auto;
+      margin-top: 0.5em;
+      margin-bottom: 0.5em;
+    }
+
+    #write form {
+      margin: 1em 0 0 0;
+      display: flex;
+      flex-direction: column;
+      gap: 0.5em;
+      align-items: stretch;
+    }
+
+    .right {
+      display: flex;
+      flex-direction: row;
+      gap: 0.5em;
+      justify-content: flex-end;
+    }
+
+    fieldset {
+      border: none;
+      padding: 0;
+      margin: 0;
+    }
+
+    textarea {
+      padding: 5px;
+      flex-grow: 1;
+      width: 100%;
+    }
+
+    pre code {
+      display: block;
+      background-color: #222;
+      color: #ddd;
+    }
+    code {
+      font-family: monospace;
+      padding: 0.1em 0.3em;
+      border-radius: 3px;
+    }
+
+    fieldset label {
+      margin: 0.5em 0;
+      display: block;
+    }
+  </style>
+
+  <script type="module">
+    import {
+      html, h, signal, effect, computed, render, useSignal, useEffect, useRef
+    } from '/index.js';
+
+    import { llamaComplete } from '/completion.js';
+
+    const session = signal({
+      prompt: "This is a conversation between user and llama, a friendly chatbot. respond in markdown.",
+      template: "{{prompt}}\n\n{{history}}\n{{char}}:",
+      historyTemplate: "{{name}}: {{message}}",
+      transcript: [],
+      type: "chat",
+      char: "llama",
+      user: "User",
+    })
+
+    const transcriptUpdate = (transcript) => {
+      session.value = {
+        ...session.value,
+        transcript
+      }
+    }
+
+    const chatStarted = computed(() => session.value.transcript.length > 0)
+
+    const params = signal({
+      n_predict: 400,
+      temperature: 0.7,
+      repeat_last_n: 256,
+      repeat_penalty: 1.18,
+      top_k: 40,
+      top_p: 0.5,
+    })
+
+    const controller = signal(null)
+    const generating = computed(() => controller.value == null )
+
+    // simple template replace
+    const template = (str, extraSettings) => {
+      let settings = session.value;
+      if (extraSettings) {
+        settings = { ...settings, ...extraSettings };
+      }
+      return String(str).replaceAll(/\{\{(.*?)\}\}/g, (_, key) => template(settings[key]));
+    }
+
+    // send message to server
+    const chat = async (msg) => {
+      if (controller.value) {
+        console.log('already running...');
+        return;
+      }
+      controller.value = new AbortController();
+
+      transcriptUpdate([...session.value.transcript, ["{{user}}", msg]])
+
+      const payload = template(session.value.template, {
+        message: msg,
+        history: session.value.transcript.flatMap(([name, message]) => template(session.value.historyTemplate, {name, message})).join("\n"),
+      });
+
+      let currentMessage = '';
+      const history = session.value.transcript
+
+      const llamaParams = {
+        ...params.value,
+        prompt: payload,
+        stop: ["</s>", template("{{char}}:"), template("{{user}}:")],
+      }
+
+      await llamaComplete(llamaParams, controller.value, (message) => {
+        const data = message.data;
+        currentMessage += data.content;
+        // remove leading whitespace
+        currentMessage = currentMessage.replace(/^\s+/, "")
+
+        transcriptUpdate([...history, ["{{char}}", currentMessage]])
+
+        if (data.stop) {
+          console.log("-->", data, ' response was:', currentMessage, 'transcript state:', session.value.transcript);
+        }
+      })
+
+      controller.value = null;
+    }
+
+    function MessageInput() {
+      const message = useSignal("")
+
+      const stop = (e) => {
+        e.preventDefault();
+        if (controller.value) {
+          controller.value.abort();
+          controller.value = null;
+        }
+      }
+
+      const reset = (e) => {
+        stop(e);
+        transcriptUpdate([]);
+      }
+
+      const submit = (e) => {
+        stop(e);
+        chat(message.value);
+        message.value = "";
+      }
+
+      const enterSubmits = (event) => {
+        if (event.which === 13 && !event.shiftKey) {
+          submit(event);
+        }
+      }
+
+      return html`
+        <form onsubmit=${submit}>
+          <div>
+          <textarea type="text" rows=2 onkeypress=${enterSubmits} value="${message}" oninput=${(e) => message.value = e.target.value} placeholder="Say something..."/>
+
+          </div>
+          <div class="right">
+          <button type="submit" disabled=${!generating.value} >Send</button>
+          <button onclick=${stop} disabled=${generating}>Stop</button>
+          <button onclick=${reset}>Reset</button>
+          </div>
+        </form>
+      `
+    }
+
+    const ChatLog = (props) => {
+      const messages = session.value.transcript;
+      const container = useRef(null)
+
+      useEffect(() => {
+        // scroll to bottom (if needed)
+        if (container.current && container.current.scrollHeight <= container.current.scrollTop + container.current.offsetHeight + 300) {
+          container.current.scrollTo(0, container.current.scrollHeight)
+        }
+      }, [messages])
+
+      const chatLine = ([user, msg]) => {
+        return html`<p key=${msg}><strong>${template(user)}:</strong> <${Markdown} text=${template(msg)} /></p>`
+      };
+
+      return html`
+        <section id="chat" ref=${container}>
+          ${messages.flatMap(chatLine)}
+        </section>`;
+    };
+
+    const ConfigForm = (props) => {
+      const updateSession = (el) => session.value = { ...session.value, [el.target.name]: el.target.value }
+      const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
+      const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
+
+      return html`
+        <form>
+          <fieldset>
+            <div>
+              <label for="prompt">Prompt</label>
+              <textarea type="text" name="prompt" value="${session.value.prompt}" rows=4 oninput=${updateSession}/>
+            </div>
+
+            <div>
+              <label for="user">User name</label>
+              <input type="text" name="user" value="${session.value.user}" oninput=${updateSession} />
+            </div>
+
+            <div>
+              <label for="bot">Bot name</label>
+              <input type="text" name="char" value="${session.value.char}" oninput=${updateSession} />
+            </div>
+
+            <div>
+              <label for="template">Prompt template</label>
+              <textarea id="template" name="template" value="${session.value.template}" rows=4 oninput=${updateSession}/>
+            </div>
+
+            <div>
+              <label for="template">Chat history template</label>
+              <textarea id="template" name="historyTemplate" value="${session.value.historyTemplate}" rows=1 oninput=${updateSession}/>
+            </div>
+
+            <div>
+              <label for="temperature">Temperature</label>
+              <input type="range" id="temperature" min="0.0" max="1.0" step="0.01" name="temperature" value="${params.value.temperature}" oninput=${updateParamsFloat} />
+              <span>${params.value.temperature}</span>
+            </div>
+
+            <div>
+              <label for="nPredict">Predictions</label>
+              <input type="range" id="nPredict" min="1" max="2048" step="1" name="n_predict" value="${params.value.n_predict}" oninput=${updateParamsFloat} />
+              <span>${params.value.n_predict}</span>
+            </div>
+
+            <div>
+              <label for="repeat_penalty">Penalize repeat sequence</label>
+              <input type="range" id="repeat_penalty" min="0.0" max="2.0" step="0.01" name="repeat_penalty" value="${params.value.repeat_penalty}" oninput=${updateParamsFloat} />
+              <span>${params.value.repeat_penalty}</span>
+            </div>
+
+            <div>
+              <label for="repeat_last_n">Consider N tokens for penalize</label>
+              <input type="range" id="repeat_last_n" min="0.0" max="2048" name="repeat_last_n" value="${params.value.repeat_last_n}" oninput=${updateParamsFloat} />
+              <span>${params.value.repeat_last_n}</span>
+            </div>
+
+          </fieldset>
+        </form>
+      `
+    }
+const Markdown = (params) => {
+  const md = params.text
+    .replace(/^#{1,6} (.*)$/gim, '<h3>$1</h3>')
+    .replace(/\*\*(.*?)\*\*/g, '<strong>$1</strong>')
+    .replace(/__(.*?)__/g, '<strong>$1</strong>')
+    .replace(/\*(.*?)\*/g, '<em>$1</em>')
+    .replace(/_(.*?)_/g, '<em>$1</em>')
+    .replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
+    .replace(/`(.*?)`/g, '<code>$1</code>')
+    .replace(/\n/gim, '<br />');
+  return html`<span dangerouslySetInnerHTML=${{ __html: md }} />`;
+};
+
+    function App(props) {
+
+      return html`
+      <div id="container">
+        <header>
+          <h1>llama.cpp</h1>
+        </header>
+
+        <main id="content">
+          <${chatStarted.value ? ChatLog : ConfigForm} />
+        </main>
+
+        <footer id="write">
+          <${MessageInput} />
+        </footer>
+
+        <footer>
+          <p>Powered by <a href="https://github.com/ggerganov/llama.cpp">llama.cpp</a> and <a href="https://ggml.ai">ggml.ai</a></p>
+        </footer>
+      </div>
+      `;
+    }
+
+    render(h(App), document.body);
+  </script>
+</head>
+
+<body>
+</body>
+
+</html>

examples/server/server.cpp

@@ -2,8 +2,6 @@
 #include "llama.h"
 #include "build-info.h"
 
-// single thread
-#define CPPHTTPLIB_THREAD_POOL_COUNT 1
 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
 #define CPPHTTPLIB_NO_EXCEPTIONS 1
@@ -12,6 +10,11 @@
 #include "httplib.h"
 #include "json.hpp"
 
+// auto generated files (update with ./deps.sh)
+#include "index.html.hpp"
+#include "index.js.hpp"
+#include "completion.js.hpp"
+
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
 #endif
@@ -21,11 +24,23 @@ using json = nlohmann::json;
 
 struct server_params {
     std::string hostname = "127.0.0.1";
+    std::string public_path = "examples/server/public";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
 };
 
+// completion token output with probabilities
+struct completion_token_output {
+    struct token_prob {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+};
+
 static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
     size_t i;
     for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
@@ -86,6 +101,40 @@ static void server_log(const char * level, const char * function, int line,
     fflush(stdout);
 }
 
+// format incomplete utf-8 multibyte character for output
+static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
+    std::string out = token == -1 ? "" : llama_token_to_str(ctx, token);
+    // if first bit is 1, meaning it's a partial character
+    if (out.size() > 0 && (out[0] & 0x80) == 0x80) {
+        std::stringstream ss;
+        ss<< std::hex << (out[0] & 0xff);
+        std::string res ( ss.str() );
+        out = "byte: \\x" + res;
+    }
+    return out;
+}
+
+// convert a vector of completion_token_output to json
+static json probs_vector_to_json(const llama_context * ctx, const std::vector<completion_token_output> probs) {
+    json out = json::array();
+    for (const auto & prob : probs) {
+        json probs_for_token = json::array();
+        for (const auto & p : prob.probs) {
+            std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
+            probs_for_token.push_back(json {
+                { "tok_str", tok_str },
+                { "prob", p.prob },
+            });
+        }
+        std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
+        out.push_back(json {
+            {"content", tok_str},
+            {"probs", probs_for_token},
+        });
+    }
+    return out;
+}
+
 static bool server_verbose = false;
 
 #if SERVER_VERBOSE != 1
@@ -107,7 +156,9 @@ struct llama_server_context {
     bool stream = false;
     bool has_next_token = false;
     std::string generated_text;
+    std::vector<completion_token_output> generated_token_probs;
 
+    size_t num_prompt_tokens = 0;
     size_t num_tokens_predicted = 0;
     size_t n_past = 0;
     size_t n_remain = 0;
@@ -126,6 +177,12 @@ struct llama_server_context {
     std::string stopping_word;
     int32_t multibyte_pending = 0;
 
+    std::mutex mutex;
+
+    std::unique_lock<std::mutex> lock() {
+        return std::unique_lock<std::mutex>(mutex);
+    }
+
     ~llama_server_context() {
         if (ctx) {
             llama_free(ctx);
@@ -139,9 +196,11 @@ struct llama_server_context {
 
     void rewind() {
         params.antiprompt.clear();
+        num_prompt_tokens = 0;
         num_tokens_predicted = 0;
         generated_text = "";
         generated_text.reserve(params.n_ctx);
+        generated_token_probs.clear();
         truncated = false;
         stopped_eos = false;
         stopped_word = false;
@@ -169,17 +228,18 @@ struct llama_server_context {
     void loadPrompt() {
         params.prompt.insert(0, 1, ' '); // always add a first space
         std::vector<llama_token> prompt_tokens = ::llama_tokenize(ctx, params.prompt, true);
+        num_prompt_tokens = prompt_tokens.size();
 
         if (params.n_keep < 0) {
-            params.n_keep = (int)prompt_tokens.size();
+            params.n_keep = (int)num_prompt_tokens;
         }
         params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
 
         // if input prompt is too big, truncate like normal
-        if (prompt_tokens.size() >= (size_t)params.n_ctx) {
+        if (num_prompt_tokens>= (size_t)params.n_ctx) {
             const int n_left = (params.n_ctx - params.n_keep) / 2;
             std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
-            const int erased_blocks = (prompt_tokens.size() - params.n_keep - n_left - 1) / n_left;
+            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
             new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
             std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
 
@@ -193,7 +253,7 @@ struct llama_server_context {
             truncated = true;
             prompt_tokens = new_tokens;
         } else {
-            const size_t ps = prompt_tokens.size();
+            const size_t ps = num_prompt_tokens;
             std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
             std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
         }
@@ -201,7 +261,7 @@ struct llama_server_context {
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
         embd = prompt_tokens;
-        if (n_past == prompt_tokens.size()) {
+        if (n_past == num_prompt_tokens) {
             // we have to evaluate at least 1 token to generate logits.
             n_past--;
         }
@@ -221,8 +281,9 @@ struct llama_server_context {
         llama_set_rng_seed(ctx, params.seed);
     }
 
-    llama_token nextToken() {
-        llama_token result = -1;
+    completion_token_output nextToken() {
+        completion_token_output result;
+        result.tok = -1;
 
         if (embd.size() >= (size_t)params.n_ctx) {
             // Reset context
@@ -261,7 +322,8 @@ struct llama_server_context {
 
         if (params.n_predict == 0) {
             has_next_token = false;
-            return llama_token_eos();
+            result.tok = llama_token_eos();
+            return result;
         }
 
         // out of user input, sample next token
@@ -278,7 +340,7 @@ struct llama_server_context {
         const float mirostat_tau = params.mirostat_tau;
         const float mirostat_eta = params.mirostat_eta;
         const bool penalize_nl = params.penalize_nl;
-        llama_token id = 0;
+        const int32_t n_probs = params.n_probs;
 
         {
             auto * logits = llama_get_logits(ctx);
@@ -312,35 +374,42 @@ struct llama_server_context {
 
             if (temp <= 0) {
                 // Greedy sampling
-                id = llama_sample_token_greedy(ctx, &candidates_p);
+                result.tok = llama_sample_token_greedy(ctx, &candidates_p);
+                if (n_probs > 0) {
+                    llama_sample_softmax(ctx, &candidates_p);
+                }
             } else {
                 if (mirostat == 1) {
                     static float mirostat_mu = 2.0f * mirostat_tau;
                     const int mirostat_m = 100;
                     llama_sample_temperature(ctx, &candidates_p, temp);
-                    id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+                    result.tok = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
                 } else if (mirostat == 2) {
                     static float mirostat_mu = 2.0f * mirostat_tau;
                     llama_sample_temperature(ctx, &candidates_p, temp);
-                    id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+                    result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                 } else {
                     // Temperature sampling
-                    llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
-                    llama_sample_typical(ctx, &candidates_p, typical_p, 1);
-                    llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+                    size_t min_keep = std::max(1, n_probs);
+                    llama_sample_top_k(ctx, &candidates_p, top_k, min_keep);
+                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, min_keep);
+                    llama_sample_typical(ctx, &candidates_p, typical_p, min_keep);
+                    llama_sample_top_p(ctx, &candidates_p, top_p, min_keep);
                     llama_sample_temperature(ctx, &candidates_p, temp);
-                    id = llama_sample_token(ctx, &candidates_p);
+                    result.tok = llama_sample_token(ctx, &candidates_p);
                 }
             }
+
+            for (size_t i = 0; i < std::min(candidates_p.size, (size_t) n_probs); ++i) {
+                result.probs.push_back({candidates_p.data[i].id, candidates_p.data[i].p});
+            }
             last_n_tokens.erase(last_n_tokens.begin());
-            last_n_tokens.push_back(id);
+            last_n_tokens.push_back(result.tok);
             num_tokens_predicted++;
         }
 
         // add it to the context
-        embd.push_back(id);
-        result = id;
+        embd.push_back(result.tok);
         // decrement remaining sampling budget
         --n_remain;
 
@@ -382,12 +451,16 @@ struct llama_server_context {
         return stop_pos;
     }
 
-    std::string doCompletion() {
-        const llama_token token = nextToken();
+    completion_token_output doCompletion() {
+        const completion_token_output token_with_probs = nextToken();
 
-        const std::string token_text = token == -1 ? "" : llama_token_to_str(ctx, token);
+        const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(ctx, token_with_probs.tok);
         generated_text += token_text;
 
+        if (params.n_probs > 0) {
+            generated_token_probs.push_back(token_with_probs);
+        }
+
         if (multibyte_pending > 0) {
             multibyte_pending -= token_text.size();
         } else if (token_text.size() == 1) {
@@ -416,8 +489,8 @@ struct llama_server_context {
         }
 
         LOG_VERBOSE("next token", {
-            { "token", token },
-            { "token_text", llama_token_to_str(ctx, token) },
+            { "token", token_with_probs.tok },
+            { "token_text", tokens_to_output_formatted_string(ctx, token_with_probs.tok) },
             { "has_next_token", has_next_token },
             { "n_remain", n_remain },
             { "num_tokens_predicted", num_tokens_predicted },
@@ -427,7 +500,7 @@ struct llama_server_context {
             { "stopping_word", stopping_word },
         });
 
-        return token_text;
+        return token_with_probs;
     }
 
     std::vector<float> getEmbedding() {
@@ -479,6 +552,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params,
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
     fprintf(stderr, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    fprintf(stderr, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
     fprintf(stderr, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     fprintf(stderr, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     fprintf(stderr, "\n");
@@ -505,6 +579,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 break;
             }
             sparams.hostname = argv[i];
+        } else if (arg == "--path") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.public_path = argv[i];
         } else if (arg == "--timeout" || arg == "-to") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -669,6 +749,7 @@ static json format_generation_settings(llama_server_context & llama) {
         { "ignore_eos", ignore_eos },
         { "stream", llama.stream },
         { "logit_bias", llama.params.logit_bias },
+        { "n_probs", llama.params.n_probs },
     };
 }
 
@@ -678,12 +759,14 @@ static json format_embedding_response(llama_server_context & llama) {
     };
 }
 
-static json format_final_response(llama_server_context & llama, const std::string & content) {
-    return json {
+static json format_final_response(llama_server_context & llama, const std::string & content, const std::vector<completion_token_output> & probs) {
+
+    json res = json {
         { "content", content },
         { "stop", true },
         { "model", llama.params.model_alias },
         { "tokens_predicted", llama.num_tokens_predicted },
+        { "tokens_evaluated", llama.num_prompt_tokens },
         { "generation_settings", format_generation_settings(llama) },
         { "prompt", llama.params.prompt },
         { "truncated", llama.truncated },
@@ -692,13 +775,25 @@ static json format_final_response(llama_server_context & llama, const std::strin
         { "stopped_limit", llama.stopped_limit },
         { "stopping_word", llama.stopping_word },
     };
+
+    if (llama.params.n_probs > 0) {
+        res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
+    }
+
+    return res;
 }
 
-static json format_partial_response(const std::string & content) {
-    return json {
+static json format_partial_response(llama_server_context & llama, const std::string & content, const std::vector<completion_token_output> & probs) {
+    json res = json {
         { "content", content },
         { "stop", false },
     };
+
+    if (llama.params.n_probs > 0) {
+        res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
+    }
+
+    return res;
 }
 
 static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
@@ -728,6 +823,7 @@ static void parse_options_completion(const json & body, llama_server_context & l
     llama.params.n_keep = body.value("n_keep", default_params.n_keep);
     llama.params.seed = body.value("seed", default_params.seed);
     llama.params.prompt = body.value("prompt", default_params.prompt);
+    llama.params.n_probs = body.value("n_probs", default_params.n_probs);
 
     llama.params.logit_bias.clear();
     if (body.value("ignore_eos", false)) {
@@ -764,17 +860,24 @@ static void parse_options_completion(const json & body, llama_server_context & l
     LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
 }
 
+
 static void log_server_request(const Request & req, const Response & res) {
     LOG_INFO("request", {
         { "remote_addr", req.remote_addr },
         { "remote_port", req.remote_port },
         { "status", res.status },
+        { "method", req.method },
         { "path", req.path },
+        { "params", req.params },
+    });
+
+    LOG_VERBOSE("request", {
         { "request", req.body },
         { "response", res.body },
     });
 }
 
+
 int main(int argc, char ** argv) {
     // own arguments required by this example
     gpt_params params;
@@ -809,16 +912,34 @@ int main(int argc, char ** argv) {
     Server svr;
 
     svr.set_default_headers({
+        { "Server", "llama.cpp" },
         { "Access-Control-Allow-Origin", "*" },
         { "Access-Control-Allow-Headers", "content-type" }
     });
 
+    // this is only called if no index.js is found in the public --path
+    svr.Get("/index.js", [](const Request &, Response & res) {
+        res.set_content(reinterpret_cast<const char *>(&index_js), index_js_len, "text/javascript");
+        return false;
+    });
+
+    // this is only called if no index.html is found in the public --path
     svr.Get("/", [](const Request &, Response & res) {
-        res.set_content("<h1>llama.cpp server works</h1>", "text/html");
+        res.set_content(reinterpret_cast<const char*>(&index_html), index_html_len, "text/html");
+        return false;
+    });
+
+    // this is only called if no index.html is found in the public --path
+    svr.Get("/completion.js", [](const Request &, Response & res) {
+        res.set_content(reinterpret_cast<const char*>(&completion_js), completion_js_len, "application/javascript");
+        return false;
     });
 
     svr.Post("/completion", [&llama](const Request & req, Response & res) {
+        auto lock = llama.lock();
+
         llama.rewind();
+
         llama_reset_timings(llama.ctx);
 
         parse_options_completion(json::parse(req.body), llama);
@@ -830,7 +951,8 @@ int main(int argc, char ** argv) {
             size_t stop_pos = std::string::npos;
 
             while (llama.has_next_token) {
-                const std::string token_text = llama.doCompletion();
+                const completion_token_output token_with_probs = llama.doCompletion();
+                const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
 
                 stop_pos = llama.findStoppingStrings(llama.generated_text,
                     token_text.size(), STOP_FULL);
@@ -844,7 +966,7 @@ int main(int argc, char ** argv) {
                     llama.generated_text.end());
             }
 
-            const json data = format_final_response(llama, llama.generated_text);
+            const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
 
             llama_print_timings(llama.ctx);
 
@@ -853,9 +975,11 @@ int main(int argc, char ** argv) {
         } else {
             const auto chunked_content_provider = [&](size_t, DataSink & sink) {
                 size_t sent_count = 0;
+                size_t sent_token_probs_index = 0;
 
                 while (llama.has_next_token) {
-                    const std::string token_text = llama.doCompletion();
+                    const completion_token_output token_with_probs = llama.doCompletion();
+                    const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
                     if (llama.multibyte_pending > 0) {
                         continue;
                     }
@@ -878,10 +1002,22 @@ int main(int argc, char ** argv) {
                     const std::string to_send = llama.generated_text.substr(pos, stop_pos);
                     sent_count += to_send.size();
 
+                    std::vector<completion_token_output> probs_output = {};
+
+                    if (llama.params.n_probs > 0) {
+                        const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                        size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                        size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                        if (probs_pos < probs_stop_pos) {
+                            probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                        }
+                        sent_token_probs_index = probs_stop_pos;
+                    }
+
                     const json data = llama.has_next_token
-                                          ? format_partial_response(to_send)
+                                          ? format_partial_response(llama, to_send, probs_output)
                                           // Generation is done, send extra information.
-                                          : format_final_response(llama, to_send);
+                                          : format_final_response(llama, to_send, llama.generated_token_probs);
 
                     const std::string str =
                         "data: " +
@@ -912,6 +1048,8 @@ int main(int argc, char ** argv) {
     });
 
     svr.Post("/tokenize", [&llama](const Request & req, Response & res) {
+        auto lock = llama.lock();
+
         const json body = json::parse(req.body);
         const std::string content = body.value("content", "");
         const std::vector<llama_token> tokens = llama_tokenize(llama.ctx, content, false);
@@ -920,6 +1058,8 @@ int main(int argc, char ** argv) {
     });
 
     svr.Post("/embedding", [&llama](const Request & req, Response & res) {
+        auto lock = llama.lock();
+
         const json body = json::parse(req.body);
 
         llama.rewind();
@@ -950,18 +1090,27 @@ int main(int argc, char ** argv) {
         res.status = 500;
     });
 
+    svr.set_error_handler([](const Request &, Response & res) {
+        res.set_content("File Not Found", "text/plain");
+        res.status = 404;
+    });
+
+
     // set timeouts and change hostname and port
     svr.set_read_timeout(sparams.read_timeout);
     svr.set_write_timeout(sparams.write_timeout);
 
     if (!svr.bind_to_port(sparams.hostname, sparams.port)) {
-        LOG_ERROR("couldn't bind to server socket", {
-            { "hostname", sparams.hostname },
-            { "port", sparams.port },
-        });
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
         return 1;
     }
 
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
+
+    // to make it ctrl+clickable:
+    fprintf(stdout, "\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
     LOG_INFO("HTTP server listening", {
         { "hostname", sparams.hostname },
         { "port", sparams.port },

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -2671,7 +2671,8 @@ struct train_params {
     const char * fn_checkpoint_out;
     const char * fn_model_out;
 
-    int seed;
+    uint32_t seed;
+
     int n_ctx;
     int n_embd;
     int n_mult;
@@ -2768,7 +2769,7 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p
     fprintf(stderr, "  --checkpoint-in FNAME      path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in);
     fprintf(stderr, "  --checkpoint-out FNAME     path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out);
     fprintf(stderr, "  --model-out FNAME          path to save ggml model (default '%s')\n", params->fn_model_out);
-    fprintf(stderr, "  -s SEED, --seed SEED       RNG seed (default: -1, use random seed for < 0)\n");
+    fprintf(stderr, "  -s SEED, --seed SEED       RNG seed (default: -1, use random seed for -1)\n");
     fprintf(stderr, "  -c N, --ctx N              Context size used during training (default %d)\n", params->n_ctx);
     fprintf(stderr, "  --embd N                   Embedding size used for new models (default %d)\n", params->n_embd);
     fprintf(stderr, "  --mult N                   Mult size used for new models, influences feedforward size. (default %d)\n", params->n_mult);
@@ -3034,10 +3035,10 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.seed < 0) {
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
-    printf("%s: seed: %d\n", __func__, params.seed);
+    printf("%s: seed: %u\n", __func__, params.seed);
     srand(params.seed);
 
     struct llama_context_params llama_params = llama_context_default_params();

ggml-cuda.cu

@@ -70,9 +70,11 @@ typedef void (*ggml_cuda_op_t)(
 
 // QK = number of values after dequantization
 // QR = QK / number of values before dequantization
+// QI = number of 32 bit integers before dequantization
 
 #define QK4_0 32
 #define QR4_0 2
+#define QI4_0 4
 typedef struct {
     half    d;              // delta
     uint8_t qs[QK4_0 / 2];  // nibbles / quants
@@ -81,6 +83,7 @@ static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0
 
 #define QK4_1 32
 #define QR4_1 2
+#define QI4_1 4
 typedef struct {
     half    d;              // delta
     half    m;              // min
@@ -90,6 +93,7 @@ static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong
 
 #define QK5_0 32
 #define QR5_0 2
+#define QI5_0 4
 typedef struct {
     half d;                 // delta
     uint8_t qh[4];          // 5-th bit of quants
@@ -99,6 +103,7 @@ static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5
 
 #define QK5_1 32
 #define QR5_1 2
+#define QI5_1 4
 typedef struct {
     half d;                 // delta
     half m;                 // min
@@ -109,12 +114,25 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) +
 
 #define QK8_0 32
 #define QR8_0 1
+#define QI8_0 8
 typedef struct {
     half    d;              // delta
     int8_t  qs[QK8_0];      // quants
 } block_q8_0;
 static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
 
+#define QK8_1 32
+#define QR8_1 1
+#define QI8_1 8
+typedef struct {
+    half    d;              // delta
+    half    s;              // unquantized sum
+    int8_t  qs[QK8_0];      // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
+
+typedef float (*vec_dot_q_cuda_t)(const void * vbq, const block_q8_1 * bq8_1, const int iqs);
+
 //================================= k-quants
 
 #ifdef GGML_QKK_64
@@ -198,14 +216,15 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define CUDA_SCALE_BLOCK_SIZE 256
 #define CUDA_ROPE_BLOCK_SIZE 256
 #define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
+#define CUDA_QUANTIZE_BLOCK_SIZE 256
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_CUDA_DMMV_X
 #define GGML_CUDA_DMMV_X 32
 #endif
-#ifndef GGML_CUDA_DMMV_Y
-#define GGML_CUDA_DMMV_Y 1
+#ifndef GGML_CUDA_MMV_Y
+#define GGML_CUDA_MMV_Y 1
 #endif
 
 #ifndef K_QUANTS_PER_ITERATION
@@ -214,6 +233,11 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
 #endif
 
+struct ggml_tensor_extra_gpu {
+    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
+    cudaEvent_t events[GGML_CUDA_MAX_DEVICES]; // events for synchronizing multiple GPUs
+};
+
 static __global__ void add_f32(const float * x, const float * y, float * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -265,7 +289,6 @@ static __global__ void rms_norm_f32(const float * x, float * dst, const int ncol
     }
 
     // sum up partial sums
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -709,7 +732,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * vx, const float
 #endif
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -814,7 +836,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float
 #endif
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -918,7 +939,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * vx, const float
 #endif
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1023,7 +1043,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * vx, const float
 #endif
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1134,7 +1153,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * vx, const float
 #endif
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1153,6 +1171,41 @@ static __device__ void convert_f16(const void * vx, const int ib, const int iqs,
     v.y = x[ib + iqs + 1];
 }
 
+static __global__ void quantize_q8_1(const float * x, void * vy, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    block_q8_1 * y = (block_q8_1 *) vy;
+
+    const int ib = i / QK8_0; // block index
+    const int iqs = i % QK8_0; // quant index
+
+    const float xi = x[i];
+    float amax = fabsf(xi);
+    float sum = xi;
+
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xffffffff, amax, mask, 32));
+        sum += __shfl_xor_sync(0xffffffff, sum, mask, 32);
+    }
+
+    const float d = amax / 127;
+    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
+
+    y[ib].qs[iqs] = q;
+
+    if (iqs > 0) {
+        return;
+    }
+
+    y[ib].d = d;
+    y[ib].s = sum;
+}
+
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
 static __global__ void dequantize_block(const void * vx, float * y, const int k) {
     const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;
@@ -1174,6 +1227,182 @@ static __global__ void dequantize_block(const void * vx, float * y, const int k)
     y[iybs + iqs + y_offset] = v.y;
 }
 
+static __device__ __forceinline__ float vec_dot_q4_0_q8_1(const void * vbq, const block_q8_1 * bq8_1, const int iqs) {
+#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
+    const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
+
+    int vi;
+    memcpy(&vi,  &bq4_0->qs[sizeof(int) * (iqs + 0)], sizeof(int));
+    const int ui0 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + 0)]);
+    const int ui1 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + QI4_0)]);
+
+    const float d = __half2float(bq4_0->d) * __half2float(bq8_1->d);
+
+    // subtract 8 from each quantized value
+    const int vi0 = __vsub4((vi >> 0) & 0x0F0F0F0F, 0x08080808);
+    const int vi1 = __vsub4((vi >> 4) & 0x0F0F0F0F, 0x08080808);
+
+    // SIMD dot product of quantized values
+    int sumi = __dp4a(vi0, ui0, 0);
+    sumi     = __dp4a(vi1, ui1, sumi);
+
+    return sumi*d;
+#else
+    return 0.0f; // only to satisfy the compiler
+#endif // __CUDA_ARCH__ >= 600
+}
+
+static __device__ __forceinline__ float vec_dot_q4_1_q8_1(const void * vbq, const block_q8_1 * bq8_1, const int iqs) {
+#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
+    const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
+
+    const int vi  = *((int *) &bq4_1->qs[sizeof(int) * (iqs + 0)]);
+    const int ui0 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + 0)]);
+    const int ui1 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + QI4_1)]);
+
+    const float d = __half2float(bq4_1->d) * __half2float(bq8_1->d);
+    const float m = bq4_1->m;
+    const float s = bq8_1->s;
+
+    const int vi0 = (vi >> 0) & 0x0F0F0F0F;
+    const int vi1 = (vi >> 4) & 0x0F0F0F0F;
+
+    // SIMD dot product of quantized values
+    int sumi = __dp4a(vi0, ui0, 0);
+    sumi     = __dp4a(vi1, ui1, sumi);
+
+    return sumi*d + m*s / QI4_1; // scale sum by QI4_1 because there are QI4_1 threads working on this block
+#else
+    return 0.0f; // only to satisfy the compiler
+#endif // __CUDA_ARCH__ >= 600
+}
+
+static __device__ __forceinline__ float vec_dot_q5_0_q8_1(const void * vbq, const block_q8_1 * bq8_1, const int iqs) {
+#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
+    const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
+
+    int qs;
+    memcpy(&qs, &bq5_0->qs[sizeof(int) * (iqs + 0)], sizeof(int));
+    const int qh0 = bq5_0->qh[iqs/2 + 0] >> 4*(iqs%2);
+    const int qh1 = bq5_0->qh[iqs/2 + 2] >> 4*(iqs%2);
+    const int ui0 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + 0)]);
+    const int ui1 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + QI5_0)]);
+
+    const float d = __half2float(bq5_0->d) * __half2float(bq8_1->d);
+
+    int vi0 = (qs  >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh0 as 5th bits
+    vi0    |= (qh0 <<  4) & 0x00000010; // 1 ->  5
+    vi0    |= (qh0 << 11) & 0x00001000; // 2 -> 13
+    vi0    |= (qh0 << 18) & 0x00100000; // 3 -> 21
+    vi0    |= (qh0 << 25) & 0x10000000; // 4 -> 29
+    vi0     = __vsub4(vi0,  0x10101010); // subtract 16 from quantized values
+    int sumi = __dp4a(vi0, ui0, 0); // SIMD dot product of quantized values
+
+    int vi1 = (qs  >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh1 as 5th bits
+    vi1    |= (qh1 <<  4) & 0x00000010; // 1 ->  5
+    vi1    |= (qh1 << 11) & 0x00001000; // 2 -> 13
+    vi1    |= (qh1 << 18) & 0x00100000; // 3 -> 21
+    vi1    |= (qh1 << 25) & 0x10000000; // 4 -> 29
+    vi1     = __vsub4(vi1,  0x10101010); // subtract 16 from quantized values
+    sumi = __dp4a(vi1, ui1, sumi); // SIMD dot product of quantized values
+
+    return sumi*d;
+#else
+    return 0.0f; // only to satisfy the compiler
+#endif // __CUDA_ARCH__ >= 600
+}
+
+static __device__ __forceinline__ float vec_dot_q5_1_q8_1(const void * vbq, const block_q8_1 * bq8_1, const int iqs) {
+#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
+    const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
+
+    const int qs  = *((int *) &bq5_1->qs[sizeof(int) * (iqs + 0)]);
+    const int qh0 = bq5_1->qh[iqs/2 + 0] >> 4*(iqs%2);
+    const int qh1 = bq5_1->qh[iqs/2 + 2] >> 4*(iqs%2);
+    const int ui0 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + 0)]);
+    const int ui1 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + QI5_1)]);
+
+    const float d = __half2float(bq5_1->d) * __half2float(bq8_1->d);
+    const float m = bq5_1->m;
+    const float s = bq8_1->s;
+
+    int vi0 = (qs  >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh0 as 5th bits
+    vi0    |= (qh0 <<  4) & 0x00000010; // 1 ->  5
+    vi0    |= (qh0 << 11) & 0x00001000; // 2 -> 13
+    vi0    |= (qh0 << 18) & 0x00100000; // 3 -> 21
+    vi0    |= (qh0 << 25) & 0x10000000; // 4 -> 29
+    int sumi = __dp4a(vi0, ui0, 0); // SIMD dot product of quantized values
+
+    int vi1 = (qs  >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh1 as 5th bits
+    vi1    |= (qh1 <<  4) & 0x00000010; // 1 ->  5
+    vi1    |= (qh1 << 11) & 0x00001000; // 2 -> 13
+    vi1    |= (qh1 << 18) & 0x00100000; // 3 -> 21
+    vi1    |= (qh1 << 25) & 0x10000000; // 4 -> 29
+    sumi = __dp4a(vi1, ui1, sumi); // SIMD dot product of quantized values
+
+    return sumi*d + m*s / QI5_1; // scale sum by QI5_1 because there are QI5_1 threads working on this block
+#else
+    return 0.0f; // only to satisfy the compiler
+#endif // __CUDA_ARCH__ >= 600
+}
+
+static __device__ __forceinline__ float vec_dot_q8_0_q8_1(const void * vbq, const block_q8_1 * bq8_1, const int iqs) {
+#if __CUDA_ARCH__ >= 600 // lowest compute capability for integer intrinsics
+    const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
+
+    int vi;
+    memcpy(&vi,  &bq8_0->qs[sizeof(int) * (iqs + 0)], sizeof(int));
+    const int ui = *((int *) &bq8_1->qs[sizeof(int) * (iqs + 0)]);
+
+    const float d = __half2float(bq8_0->d) * __half2float(bq8_1->d);
+
+    // SIMD dot product of quantized values
+    int sumi = __dp4a(vi, ui, 0);
+
+    return sumi*d;
+#else
+    return 0.0f; // only to satisfy the compiler
+#endif // __CUDA_ARCH__ >= 600
+}
+
+template <int qk, int qi, typename block_q_t, vec_dot_q_cuda_t vec_dot_q_cuda>
+static __global__ void mul_mat_vec_q(const void * vx, const void * vy, float * dst, const int ncols, const int nrows) {
+    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = 0; i < blocks_per_row; i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i + threadIdx.x / qi; // x block index
+
+        const int iby = i + threadIdx.x / qi; // y block index
+
+        const int iqs  = threadIdx.x % qi; // x block quant index when casting the quants to int
+
+        tmp += vec_dot_q_cuda(&x[ibx], &y[iby], iqs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
+    }
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
 static __global__ void dequantize_mul_mat_vec(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows) {
     // qk = quantized weights per x block
@@ -1228,7 +1457,6 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const dfloat * y,
     }
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1279,7 +1507,6 @@ static __global__ void mul_mat_p021_f16_f32(const void * vx, const float * y, fl
     const int idst = channel*nrows_dst + row_dst;
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1292,7 +1519,7 @@ static __global__ void mul_mat_p021_f16_f32(const void * vx, const float * y, fl
 
 static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int nchannels_x, const int channel_stride_x) {
+    const int row_stride_x, const int channel_stride_x) {
 
     const half * x = (const half *) vx;
 
@@ -1325,7 +1552,6 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     }
 
     // sum up partial sums and write back result
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1435,7 +1661,6 @@ static __global__ void soft_max_f32(const float * x, float * dst, const int ncol
     }
 
     // sum up partial sums
-    __syncthreads();
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
@@ -1489,6 +1714,11 @@ static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, con
     rms_norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
 }
 
+static void quantize_row_q8_1_cuda(const float * x, void * vy, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    quantize_q8_1<<<num_blocks, CUDA_QUANTIZE_BLOCK_SIZE, 0, stream>>>(x, vy, k);
+}
+
 static void dequantize_row_q4_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<QK4_0, QR4_0, dequantize_q4_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
@@ -1557,45 +1787,45 @@ static void dequantize_row_q6_K_cuda(const void * vx, float * y, const int k, cu
 
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
@@ -1642,6 +1872,51 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
     dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
+static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, vec_dot_q4_0_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
+static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, vec_dot_q4_1_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
+static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, vec_dot_q5_0_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
+static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, vec_dot_q5_1_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
+static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, vec_dot_q8_0_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
 static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<1, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
@@ -1649,9 +1924,9 @@ static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, c
 
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     dequantize_mul_mat_vec<1, 1, convert_f16>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
@@ -1698,7 +1973,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_cuda(
     const dim3 block_nums(1, nrows_x, nchannels_x);
     const dim3 block_dims(WARP_SIZE, 1, 1);
     mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, nchannels_x, channel_stride_x);
+        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x);
 }
 
 static void ggml_cpy_f32_f32_cuda(
@@ -1817,6 +2092,7 @@ static size_t g_scratch_offset = 0;
 
 static int g_device_count = -1;
 static int g_main_device = 0;
+static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
@@ -1834,9 +2110,12 @@ void ggml_init_cublas() {
         for (int id = 0; id < g_device_count; ++id) {
             cudaDeviceProp prop;
             CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-            fprintf(stderr, "  Device %d: %s\n", id, prop.name);
+            fprintf(stderr, "  Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
+
             g_tensor_split[id] = total_vram;
             total_vram += prop.totalGlobalMem;
+
+            g_compute_capabilities[id] = 100*prop.major + 10*prop.minor;
         }
         for (int id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
@@ -1970,7 +2249,6 @@ inline void ggml_cuda_op_add(
     } else {
         GGML_ASSERT(false);
     }
-    CUDA_CHECK(cudaGetLastError());
 
     (void) src1;
     (void) dst;
@@ -2002,7 +2280,6 @@ inline void ggml_cuda_op_mul(
 
         // compute
         mul_f32_cuda(src0_ddf_i01, src1_ddf_i01, dst_ddf_i01, ne00, ne10, cudaStream_main);
-        CUDA_CHECK(cudaGetLastError());
     }
 
     (void) dst;
@@ -2023,7 +2300,6 @@ inline void ggml_cuda_op_silu(
 
     // compute
     silu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main);
-    CUDA_CHECK(cudaGetLastError());
 
     (void) src1;
     (void) dst;
@@ -2046,7 +2322,6 @@ inline void ggml_cuda_op_rms_norm(
 
     // compute
     rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
-    CUDA_CHECK(cudaGetLastError());
 
     (void) src1;
     (void) dst;
@@ -2056,7 +2331,7 @@ inline void ggml_cuda_op_rms_norm(
     (void) i1;
 }
 
-inline void ggml_cuda_op_dequantize_mul_mat_vec(
+inline void ggml_cuda_op_mul_mat_vec(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
     float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
     cudaStream_t & cudaStream_main){
@@ -2068,70 +2343,116 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
     const int64_t ne00 = src0->ne[0];
     const int64_t nrows = i01_high - i01_low;
 
-// on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
-#ifdef GGML_CUDA_DMMV_F16
-    size_t ash;
-    dfloat * src1_dfloat = nullptr; // dfloat == half
-
-    bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
-        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
-        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
-
-    if (src1_convert_f16) {
-        src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash);
-        ggml_cpy_f32_f16_cuda((char *) src1_ddf_i, (char *) src1_dfloat, ne00,
-                                ne00, 1, sizeof(float), 0, 0,
-                                ne00, 1, sizeof(half),  0, 0, cudaStream_main);
-    }
+#ifdef GGML_CUDA_FORCE_DMMV
+    const bool use_mul_mat_vec_q = false;
 #else
-    dfloat * src1_dfloat = src1_ddf_i; // dfloat == float, no conversion
+    int id;
+    CUDA_CHECK(cudaGetDevice(&id));
+
+    const bool mul_mat_vec_q_implemented = src0->type == GGML_TYPE_Q4_0 ||
+        src0->type == GGML_TYPE_Q4_1 ||
+        src0->type == GGML_TYPE_Q5_0 ||
+        src0->type == GGML_TYPE_Q5_1 ||
+        src0->type == GGML_TYPE_Q8_0;
+
+    // The integer intrinsics used in mul_mat_vec_q are available with compute capability 6.
+    // However, they have bad performance with Pascal cards.
+    // Therefore, in a multi GPU setting decide at runtime which GPUs should use mul_mat_vec_q.
+    const bool use_mul_mat_vec_q = g_compute_capabilities[id] >= 700 && mul_mat_vec_q_implemented;
+#endif
+
+    if (use_mul_mat_vec_q) {
+        size_t as;
+        void * src1_q8_1 = ggml_cuda_pool_malloc(ne00*sizeof(block_q8_1)/QK8_1, &as);
+        quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne00, cudaStream_main);
+
+        switch (src0->type) {
+            case GGML_TYPE_Q4_0:
+                mul_mat_vec_q4_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q4_1:
+                mul_mat_vec_q4_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q5_0:
+                mul_mat_vec_q5_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q5_1:
+                mul_mat_vec_q5_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q8_0:
+                mul_mat_vec_q8_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+
+        ggml_cuda_pool_free(src1_q8_1, as);
+    } else {
+        // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
+#ifdef GGML_CUDA_DMMV_F16
+        size_t ash;
+        dfloat * src1_dfloat = nullptr; // dfloat == half
+
+        bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+            src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
+            src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
+
+        if (src1_convert_f16) {
+            src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash);
+            ggml_cpy_f32_f16_cuda((char *) src1_ddf_i, (char *) src1_dfloat, ne00,
+                                    ne00, 1, sizeof(float), 0, 0,
+                                    ne00, 1, sizeof(half),  0, 0, cudaStream_main);
+        }
+#else
+        dfloat * src1_dfloat = src1_ddf_i; // dfloat == float, no conversion
 #endif // GGML_CUDA_DMMV_F16
 
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q4_1:
-            dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q5_0:
-            dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q5_1:
-            dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q8_0:
-            dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q2_K:
-            dequantize_mul_mat_vec_q2_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q3_K:
-            dequantize_mul_mat_vec_q3_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q4_K:
-            dequantize_mul_mat_vec_q4_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q5_K:
-            dequantize_mul_mat_vec_q5_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_Q6_K:
-            dequantize_mul_mat_vec_q6_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        case GGML_TYPE_F16:
-            convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-    CUDA_CHECK(cudaGetLastError());
+        switch (src0->type) {
+            case GGML_TYPE_Q4_0:
+                dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q4_1:
+                dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q5_0:
+                dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q5_1:
+                dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q8_0:
+                dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q2_K:
+                dequantize_mul_mat_vec_q2_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q3_K:
+                dequantize_mul_mat_vec_q3_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q4_K:
+                dequantize_mul_mat_vec_q4_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q5_K:
+                dequantize_mul_mat_vec_q5_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_Q6_K:
+                dequantize_mul_mat_vec_q6_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            case GGML_TYPE_F16:
+                convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
 
 #ifdef GGML_CUDA_DMMV_F16
-    if (src1_convert_f16) {
-        ggml_cuda_pool_free(src1_dfloat, ash);
-    }
+        if (src1_convert_f16) {
+            ggml_cuda_pool_free(src1_dfloat, ash);
+        }
 #endif // GGML_CUDA_DMMV_F16
+    }
 
     (void) src1;
     (void) dst;
@@ -2202,7 +2523,6 @@ inline void ggml_cuda_op_rope(
 
     // compute
     rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p, theta_scale, cudaStream_main);
-    CUDA_CHECK(cudaGetLastError());
 
     (void) dst;
     (void) src0_ddq_i;
@@ -2226,7 +2546,6 @@ inline void ggml_cuda_op_diag_mask_inf(
 
     // compute
     diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main);
-    CUDA_CHECK(cudaGetLastError());
 
     (void) dst;
     (void) src0_ddq_i;
@@ -2248,7 +2567,6 @@ inline void ggml_cuda_op_soft_max(
 
     // compute
     soft_max_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
-    CUDA_CHECK(cudaGetLastError());
 
     (void) src1;
     (void) dst;
@@ -2344,10 +2662,11 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0};
     size_t  dst_asf[GGML_CUDA_MAX_DEVICES] = {0};
 
-    // if multiple GPUs are used they need to wait for the main GPU to finish
+    // if multiple devices are used they need to wait for the main device
+    // here an event is recorded that signifies that the main device has finished calculating the input data
     if (split && g_device_count > 1) {
         CUDA_CHECK(cudaSetDevice(g_main_device));
-        CUDA_CHECK(cudaDeviceSynchronize());
+        CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device], g_cudaStreams_main[g_main_device]));
     }
 
     for (int id = 0; id < g_device_count; ++id) {
@@ -2373,6 +2692,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
         int64_t row_diff = row_high - row_low;
 
         cudaSetDevice(id);
+        cudaStream_t cudaStream_main = g_cudaStreams_main[id];
+
+        // wait for main GPU data if necessary
+        if (split && id != g_main_device) {
+            CUDA_CHECK(cudaStreamWaitEvent(cudaStream_main, src0_extra->events[g_main_device]));
+        }
 
         if (src0_on_device && src0_is_contiguous) {
             if (src0_is_f32) {
@@ -2448,8 +2773,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                 }
                 const int64_t i11 = i13*ne12 + i12;
 
-                cudaStream_t cudaStream_main = g_cudaStreams_main[id];
-
                 // for split tensors the data begins at i0 == i0_offset_low
                 char  * src0_ddq_i = src0_ddq[id] + (i0 - i0_offset_low)*src0_stride*src0_ts/src0_bs;
                 float * src0_ddf_i = src0_ddf[id] + (i0 - i0_offset_low)*src0_stride;
@@ -2509,6 +2832,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
 
                 // do the computation
                 op(src0, src1, dst, src0_ddq_i, src0_ddf_i, src1_ddf_i, dst_ddf_i, i02, i01_low, i01_high, i11, cudaStream_main);
+                CUDA_CHECK(cudaGetLastError());
 
                 // copy dst to host or other device if necessary
                 if (!dst_on_device) {
@@ -2538,6 +2862,11 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                         CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i, dst_stride*sizeof(float), kind, cudaStream_main));
                     }
                 }
+
+                // signify to main device that other device is done
+                if (split && g_device_count > 1 && id != g_main_device) {
+                    CUDA_CHECK(cudaEventRecord(src0_extra->events[id], cudaStream_main));
+                }
             }
         }
     }
@@ -2549,7 +2878,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
         }
 
         CUDA_CHECK(cudaSetDevice(id));
-        CUDA_CHECK(cudaDeviceSynchronize());
 
         if (src0_asq[id] > 0) {
             ggml_cuda_pool_free(src0_ddq[id], src0_asq[id]);
@@ -2564,6 +2892,21 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
             ggml_cuda_pool_free(dst_ddf[id], dst_asf[id]);
         }
     }
+
+    // main device waits for all other devices to be finished
+    if (split && g_device_count > 1) {
+        CUDA_CHECK(cudaSetDevice(g_main_device));
+        for (int id = 0; id < g_device_count; ++id) {
+            if (id != g_main_device) {
+                CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams_main[g_main_device], src0_extra->events[id]));
+            }
+        }
+    }
+
+    if (dst->backend == GGML_BACKEND_CPU) {
+        CUDA_CHECK(cudaSetDevice(g_main_device));
+        CUDA_CHECK(cudaDeviceSynchronize());
+    }
 }
 
 void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -2679,8 +3022,8 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_
     }else if (src0->type == GGML_TYPE_F32) {
         ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false);
     } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
-        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src0->ne[1] % GGML_CUDA_DMMV_Y == 0) {
-            ggml_cuda_op(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false, false);
+        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) {
+            ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_vec, false, false);
         } else {
             ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false);
         }
@@ -2803,25 +3146,32 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
         cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice);
 
         extra->data_device[id] = buf;
+
+        if (backend == GGML_BACKEND_GPU_SPLIT) {
+            CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id], cudaEventDisableTiming));
+        }
     }
 
     tensor->extra = extra;
 }
 
 void ggml_cuda_free_data(struct ggml_tensor * tensor) {
-    if (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) {
+    if (!tensor || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
         return;
     }
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
     for (int id = 0; id < g_device_count; ++id) {
-        if (extra->data_device[id] == nullptr) {
-            continue;
+        if (extra->data_device[id] != nullptr) {
+            CUDA_CHECK(cudaSetDevice(id));
+            CUDA_CHECK(cudaFree(extra->data_device[id]));
         }
 
-        CUDA_CHECK(cudaSetDevice(id));
-        CUDA_CHECK(cudaFree(extra->data_device[id]));
+        if (extra->events[id] != nullptr) {
+            CUDA_CHECK(cudaSetDevice(id));
+            CUDA_CHECK(cudaEventDestroy(extra->events[id]));
+        }
     }
 
     delete extra;

ggml-cuda.h

@@ -8,10 +8,6 @@ extern "C" {
 
 #define GGML_CUDA_MAX_DEVICES       16
 
-struct ggml_tensor_extra_gpu {
-    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
-};
-
 void   ggml_init_cublas(void);
 void   ggml_cuda_set_tensor_split(const float * tensor_split);
 

ggml-metal.m

@@ -202,7 +202,9 @@ struct ggml_metal_context * ggml_metal_init(void) {
 
 void ggml_metal_free(struct ggml_metal_context * ctx) {
     fprintf(stderr, "%s: deallocating\n", __func__);
-
+    for (int i = 0; i < ctx->n_buffers; ++i) {
+        [ctx->buffers[i].metal release];
+    }
     free(ctx);
 }
 

ggml-opencl.cpp

@@ -21,11 +21,19 @@
 
 #define CL_DMMV_BLOCK_SIZE 32
 
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 1
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
 #define MULTILINE_QUOTE(...) #__VA_ARGS__
 static std::string program_source = MULTILINE_QUOTE(
 
 typedef char int8_t;
 typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
 typedef int int32_t;
 typedef uint uint32_t;
 
@@ -175,7 +183,9 @@ void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float
     *v0 = vload_half(0, &x[ib + 0]);
     *v1 = vload_half(0, &x[ib + 1]);
 }
+);
 
+static std::string k_quants_source = MULTILINE_QUOTE(
 inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m)
 {
     if (j < 4)
@@ -199,7 +209,7 @@ __kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __globa
     const int is = 8 * n + l / 16;
 
     const uint8_t q = x[i].qs[32 * n + l];
-    __global float *y = yy + i * 256 + 128 * n;
+    __global float *y = yy + i * QK_K + 128 * n;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -231,7 +241,7 @@ __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __globa
     float d_all = vload_half(0, &x[i].d);
     float dl = d_all * (us - 32);
 
-    __global float *y = yy + i * 256 + 128 * n + 32 * j;
+    __global float *y = yy + i * QK_K + 128 * n + 32 * j;
     const __global uint8_t *q = x[i].qs + 32 * n;
     const __global uint8_t *hm = x[i].hmask;
 
@@ -248,7 +258,7 @@ __kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __globa
     const int is = 2 * il;
     const int n = 4;
 
-    __global float *y = yy + i * 256 + 64 * il + n * ir;
+    __global float *y = yy + i * QK_K + 64 * il + n * ir;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -277,7 +287,7 @@ __kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __globa
     const int ir = tid % 16;
     const int is = 2 * il;
 
-    __global float *y = yy + i * 256 + 64 * il + 2 * ir;
+    __global float *y = yy + i * QK_K + 64 * il + 2 * ir;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -309,7 +319,7 @@ __kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __globa
     const int il = tid - 32 * ip;
     const int is = 8 * ip + il / 16;
 
-    __global float *y = yy + i * 256 + 128 * ip + il;
+    __global float *y = yy + i * QK_K + 128 * ip + il;
 
     const float d = vload_half(0, &x[i].d);
 
@@ -323,161 +333,383 @@ __kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __globa
     y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
 }
 
+__kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
 
-void vec_dot_q2_K(__global const struct block_q2_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+    const int row = get_group_id(0);
 
-    int n = iqs / 128;
-    int r = iqs - 128 * n;
-    int l = r / 8;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
 
-    __global const float *y = yy + 128 * n + l;
-    __global const uint8_t *q = x[ib].qs + 32 * n + l;
-    __global const uint8_t *s = x[ib].scales + 8 * n;
+    __global const struct block_q2_K * x = xx + ib0;
 
-    const float dall = vload_half(0, &x[ib].d);
-    const float dmin = vload_half(0, &x[ib].dmin);
+    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
+    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
-    float sum = y[  0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4))
-              + y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4))
-              + y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4))
-              + y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4))
-              + y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4))
-              + y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4))
-              + y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4))
-              + y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4));
+    const int step = 16/K_QUANTS_PER_ITERATION;
 
-    *result = sum;
-}
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
 
-void vec_dot_q3_K(__global const struct block_q3_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
+    const int q_offset = 32*im + l0;
+    const int s_offset = 8*im;
+    const int y_offset = 128*im + l0;
 
-    const uint32_t kmask1 = 0x03030303;
-    const uint32_t kmask2 = 0x0f0f0f0f;
+    tmp[16 * ix + tid] = 0;
 
-    uint32_t aux[3];
-    uint32_t utmp[4];
+    uint32_t aux[4];
+    const uint8_t * d = (const uint8_t *)aux;
+    const uint8_t * m = (const uint8_t *)(aux + 2);
 
-    int n = iqs/128;
-    int r = iqs - 128*n;
-    int l = r/8;
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
 
-    __global const float   * y = yy + 128*n + l;
-    __global const uint8_t * q = x[ib].qs + 32*n + l;
-    __global const uint8_t * hm = x[ib].hmask + l;
-    const int8_t * s = (const int8_t *)utmp + 8*n;
+        __global const float   * y = yy + i * QK_K + y_offset;
+        __global const uint8_t * q = x[i].qs + q_offset;
 
-    aux[0] = x[ib].scales[0] | x[ib].scales[1] << 8 | x[ib].scales[2] << 16 | x[ib].scales[3] << 24;
-    aux[1] = x[ib].scales[4] | x[ib].scales[5] << 8 | x[ib].scales[6] << 16 | x[ib].scales[7] << 24;
-    aux[2] = x[ib].scales[8] | x[ib].scales[9] << 8 | x[ib].scales[10] << 16 | x[ib].scales[11] << 24;
+        const float dall = vload_half(0, &x[i].d);
+        const float dmin = vload_half(0, &x[i].dmin);
 
-    utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-    utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-    utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-    utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+        __global const uint32_t * a = (__global const uint32_t *)(x[i].scales + s_offset);
+        aux[0] = a[0] & 0x0f0f0f0f;
+        aux[1] = a[1] & 0x0f0f0f0f;
+        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
+        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
 
-    const float dall = vload_half(0, &x[ib].d);
-    const uint8_t m = 1 << (4*n);
+        float sum1 = 0, sum2 = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
+            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
 
-    float sum = y[  0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4))
-              + y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4))
-              + y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4))
-              + y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4))
-              + y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4))
-              + y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4))
-              + y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4))
-              + y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4));
+        }
+        tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
 
-    *result = sum * dall;
-
-}
-
-void vec_dot_q4_K(__global const struct block_q4_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
-
-    const int j  = iqs / 64;        // j  is in 0...3
-    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
-    const int is = 2*j;             // is is in 0...6 in steps of 2
-
-    __global const float   * y = yy + 64*j + ir;
-    __global const uint8_t * q = x[ib].qs + 32*j + ir;
-
-    const float dall = vload_half(0, &x[ib].d);
-    const float dmin = vload_half(0, &x[ib].dmin);
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[ib].scales, &sc, &m);
-    const float d1 = dall * sc;
-    const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[ib].scales, &sc, &m);
-    const float d2 = dall * sc;
-    const float m2 = dmin * m;
-
-    float sum = 0;
-    for (int k = 0; k < 4; ++k) {
-        sum += y[k +  0] * (d1 * (q[k] & 0xF) - m1);
-        sum += y[k + 32] * (d2 * (q[k] >>  4) - m2);
     }
 
-    *result = sum;
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=16; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
 }
 
-void vec_dot_q5_K(__global const struct block_q5_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+__kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
 
-    const int j  = iqs / 64;
-    const int ir = (iqs - 64*j)/2;
-    const int is = 2*j;
+    const int row = get_group_id(0);
 
-    __global const float   * y  = yy + 64*j + ir;
-    __global const uint8_t * ql = x[ib].qs + 32*j + ir;
-    __global const uint8_t * qh = x[ib].qh + ir;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
 
-    const float dall = vload_half(0, &x[ib].d);
-    const float dmin = vload_half(0, &x[ib].dmin);
+    __global const struct block_q3_K * x = xx + ib0;
 
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[ib].scales, &sc, &m);
-    const float d1 = dall * sc;
-    const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[ib].scales, &sc, &m);
-    const float d2 = dall * sc;
-    const float m2 = dmin * m;
+    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
+    const int step = 16/K_QUANTS_PER_ITERATION;
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0....15 or 0...7
+
+    const uint8_t m = 1 << (4*im);
+
+    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
+    const int q_offset =  32*im + l0;
+    const int y_offset = 128*im + l0;
+
+    uint16_t utmp[4];
+    const int8_t * s = (const int8_t *)utmp;
+
+    const uint16_t s_shift = 4*im;
+
+    tmp[16 * ix + tid] = 0;
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        __global const float   * y  = yy + i * QK_K + y_offset;
+        __global const uint8_t * q = x[i].qs + q_offset;
+        __global const uint8_t * h = x[i].hmask + l0;
+
+        __global const uint16_t * a = (__global const uint16_t *)x[i].scales;
+        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
+        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
+
+        const float d = vload_half(0, &x[i].d);
+
+        float sum = 0;
+        for (int l = 0; l < n; ++l) {
+            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
+            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
+        }
+        tmp[16 * ix + tid] += d * sum;
 
-    uint8_t hm  = 1 << is;
-    float sum = 0;
-    for (int k = 0; k < 4; ++k) {
-        sum += y[k +  0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1);
     }
-    hm <<= 1;
-    for (int k = 0; k < 4; ++k) {
-        sum += y[k + 32] * (d2 * ((ql[k] >>  4) + (qh[k] & hm ? 16 : 0)) - m2);
-    }
-    *result = sum;
 
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=16; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
 }
 
-void vec_dot_q6_K(__global const struct block_q6_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+__kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
 
+    //to rename it later, just to test now
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
 
-    const int ip = iqs / 128;        // 0 or 1
-    const int il = (iqs - 128*ip)/8; // 0...15
-    const int is = 8*ip;
+    const int row = get_group_id(0);
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
 
-    __global const float * y = yy + 128*ip + il;
+    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...15
+    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;
 
-    const float d = vload_half(0, &x[ib].d);
+    const int step = 8/K_QUANTS_PER_ITERATION;
 
-    __global const uint8_t * ql = x[ib].ql + 64*ip + il;
-    __global const uint8_t * qh = x[ib].qh + 32*ip + il;
-    __global const int8_t  * sc = x[ib].scales + is;
+    const int il  = tid/step;     // 0...3
+    const int ir  = tid - step*il;// 0...3
+    const int n   = 2*K_QUANTS_PER_ITERATION;
 
-    *result = y[  0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32)
-           + y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32)
-           + y[ 64] * d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32)
-           + y[ 96] * d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32)
-           + y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32)
-           + y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32)
-           + y[ 80] * d * sc[5] * ((int8_t)((ql[16]  >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32)
-           + y[112] * d * sc[7] * ((int8_t)((ql[48]  >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32);
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
 
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+    __global const struct block_q4_K * x = xx + ib0;
+
+    tmp[16 * ix + tid] = 0;
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        __global const uint8_t * q1 = x[i].qs + q_offset;
+        __global const uint8_t * q2 = q1 + 64;
+        __global const float   * y1 = yy + i*QK_K + y_offset;
+        __global const float   * y2 = y1 + 128;
+
+        const float dall = vload_half(0, &x[i].d);
+        const float dmin = vload_half(0, &x[i].dmin);
+
+        __global const uint16_t * a = (__global const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+        float4 s = (float4)(0.f);
+        float smin = 0;
+        for (int l = 0; l < n; ++l) {
+            s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
+            s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp[16 * ix + tid] += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
+
+    }
+
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=16; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
+}
+
+__kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int row = get_group_id(0);
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const int tid = get_local_id(0)/2;  // 0...15
+    const int ix  = get_local_id(0)%2;
+
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
+    const int n   = 2;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    const uint8_t hm1  = 1 << (2*im);
+    const uint8_t hm2  = hm1 << 4;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+    __global const struct block_q5_K * x = xx + ib0;
+
+    tmp[16 * ix + tid] = 0;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2) {
+
+        __global const uint8_t * ql1 = x[i].qs + q_offset;
+        __global const uint8_t * ql2 = ql1 + 64;
+        __global const uint8_t * qh  = x[i].qh + l0;
+        __global const float   * y1  = yy + i*QK_K + y_offset;
+        __global const float   * y2  = y1 + 128;
+
+        const float dall = vload_half(0, &x[i].d);
+        const float dmin = vload_half(0, &x[i].dmin);
+
+        __global const uint16_t * a = (__global const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+        float4 sum = (float4)(0.f);
+        float smin = 0;
+        for (int l = 0; l < n; ++l) {
+            sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+                   + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
+            sum.y += y1[l+32] * ((ql1[l+ 0] >>  4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+                   + y1[l+48] * ((ql1[l+16] >>  4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
+            sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+                   + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
+            sum.w += y2[l+32] * ((ql2[l+ 0] >>  4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+                   + y2[l+48] * ((ql2[l+16] >>  4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
+            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
+        }
+        tmp[16 * ix + tid] += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
+
+    }
+
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=16; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
+}
+
+__kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, __local float* tmp, __global const float * yy, __global float * dst, const int ncols) {
+
+    const int row = get_group_id(0);
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    __global const struct block_q6_K * x = xx + ib0;
+
+    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
+    const int is = 0;
+#else
+    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
+    const int is = in / 4;
+#endif
+    const int ql_offset = 64*im + l0;
+    const int qh_offset = 32*im + l0;
+    const int s_offset  =  8*im + is;
+    const int y_offset = 128*im + l0;
+
+    tmp[16 * ix + tid] = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        __global const float   * y  = yy + i * QK_K + y_offset;
+        __global const uint8_t * ql = x[i].ql + ql_offset;
+        __global const uint8_t * qh = x[i].qh + qh_offset;
+        __global const int8_t  * s  = x[i].scales + s_offset;
+
+        const float d = vload_half(0, &x[i].d);
+
+#if K_QUANTS_PER_ITERATION == 1
+        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
+        tmp[16 * ix + tid] += sum;
+#else
+        float sum = 0;
+        for (int l = 0; l < 4; ++l) {
+            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
+        }
+        tmp[16 * ix + tid] += sum;
+#endif
+
+    }
+
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=16; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
 }
 
 );
@@ -549,44 +781,6 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
 }
 );
 
-std::string dequant_mul_mat_vec_k_template = MULTILINE_QUOTE(
-__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
-    const int block_size = get_local_size(0);
-    const int row = get_group_id(0);
-    const int tid = get_local_id(0);
-
-    const int iter_stride = 256;
-    const int vals_per_iter = iter_stride / block_size;
-    const int num_blocks_per_row = ncols / 256;
-    const int ib0 = row*num_blocks_per_row;
-
-    tmp[tid] = 0;
-
-    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i + vals_per_iter*tid;
-        const int ib = ib0 + col/256; // x block index
-        const int iqs = col%256; // x quant index
-        const int iybs = col - col%256; // y block start index
-
-        // dequantize
-        float v;
-        DOT_KERNEL(x, ib, iqs, y + iybs, &v);
-        tmp[tid] += v;
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=block_size/2; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-);
 
 std::string mul_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
@@ -649,18 +843,6 @@ std::array<std::string, 2> mul_str_values = {
     "mul_f32", "float"
 };
 
-std::array<std::string, 3> dmmv_k_str_keys = {
-    "KERNEL_NAME", "X_TYPE", "DOT_KERNEL"
-};
-
-std::array<std::string, 15> dmmv_k_str_values = {
-    "dequantize_mul_mat_vec_q2_K", "struct block_q2_K", "vec_dot_q2_K",
-    "dequantize_mul_mat_vec_q3_K", "struct block_q3_K", "vec_dot_q3_K",
-    "dequantize_mul_mat_vec_q4_K", "struct block_q4_K", "vec_dot_q4_K",
-    "dequantize_mul_mat_vec_q5_K", "struct block_q5_K", "vec_dot_q5_K",
-    "dequantize_mul_mat_vec_q6_K", "struct block_q6_K", "vec_dot_q6_K",
-};
-
 std::string& replace(std::string& s, const std::string& from, const std::string& to) {
     size_t pos = 0;
     while ((pos = s.find(from, pos)) != std::string::npos) {
@@ -673,6 +855,7 @@ std::string& replace(std::string& s, const std::string& from, const std::string&
 std::string generate_kernels() {
     std::stringstream src;
     src << program_source << '\n';
+    src << k_quants_source << '\n';
     for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) {
         std::string dequant_kernel = dequant_template;
         std::string dmmv_kernel = dequant_mul_mat_vec_template;
@@ -690,13 +873,6 @@ std::string generate_kernels() {
         }
         src << mul_kernel << '\n';
     }
-    for (size_t i = 0; i < dmmv_k_str_values.size(); i += dmmv_k_str_keys.size()) {
-        std::string dmmv_k_kernel = dequant_mul_mat_vec_k_template;
-        for (size_t j = 0; j < dmmv_k_str_keys.size(); j++) {
-            replace(dmmv_k_kernel, dmmv_k_str_keys[j], dmmv_k_str_values[i + j]);
-        }
-        src << dmmv_k_kernel << '\n';
-    }
 
     return src.str();
 }
@@ -729,10 +905,11 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co
         exit(1);
     }
 
-    const char* compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
-                               "-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1";
+    std::string compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
+                               "-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1 "
+                               "-DQK_K=256 -DK_QUANTS_PER_ITERATION=" + std::to_string(K_QUANTS_PER_ITERATION);
 
-    err = clBuildProgram(p, 0, NULL, compile_opts, NULL, NULL);
+    err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL);
     if(err < 0) {
 
         clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
@@ -1199,7 +1376,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
     const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];

ggml.h

@@ -201,6 +201,8 @@
 #define GGML_MAX_NAME          48
 #define GGML_DEFAULT_N_THREADS 4
 
+#define GGML_UNUSED(x) (void)(x)
+
 #define GGML_ASSERT(x) \
     do { \
         if (!(x)) { \
@@ -209,6 +211,30 @@
         } \
     } while (0)
 
+// used to copy the number of elements and stride in bytes of tensors into local variables.
+// main purpose is to reduce code duplication and improve readability.
+//
+// example:
+//
+//    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
+//    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb);
+//
+#define GGML_TENSOR_LOCALS_1(type, prefix, pointer, array) \
+    const type prefix##0 = (pointer)->array[0]; \
+    GGML_UNUSED(prefix##0);
+#define GGML_TENSOR_LOCALS_2(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_1    (type, prefix, pointer, array) \
+    const type prefix##1 = (pointer)->array[1]; \
+    GGML_UNUSED(prefix##1);
+#define GGML_TENSOR_LOCALS_3(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_2    (type, prefix, pointer, array) \
+    const type prefix##2 = (pointer)->array[2]; \
+    GGML_UNUSED(prefix##2);
+#define GGML_TENSOR_LOCALS(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_3  (type, prefix, pointer, array) \
+    const type prefix##3 = (pointer)->array[3]; \
+    GGML_UNUSED(prefix##3);
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -224,8 +250,8 @@ extern "C" {
     GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
     GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
 
-    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n);
-    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n);
+    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
+    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);
 
     struct ggml_object;
     struct ggml_context;
@@ -295,12 +321,15 @@ extern "C" {
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
+        GGML_OP_ARGMAX,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
         GGML_OP_ABS,
         GGML_OP_SGN,
         GGML_OP_NEG,
         GGML_OP_STEP,
+        GGML_OP_TANH,
+        GGML_OP_ELU,
         GGML_OP_RELU,
         GGML_OP_GELU,
         GGML_OP_GELU_QUICK,
@@ -332,9 +361,8 @@ extern "C" {
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
-        GGML_OP_CONV_1D_S1_PH,
-        GGML_OP_CONV_1D_S2_PH,
-        GGML_OP_CONV_2D_SK_P0,
+        GGML_OP_CONV_1D,
+        GGML_OP_CONV_2D,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
@@ -444,6 +472,9 @@ extern "C" {
 
 
     // compute types
+
+    // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
+    // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
     enum ggml_task_type {
         GGML_TASK_INIT = 0,
         GGML_TASK_COMPUTE,
@@ -687,6 +718,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // argmax along rows
+    GGML_API struct ggml_tensor * ggml_argmax(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // if a is the same shape as b, and a is not parameter, return a
     // otherwise, return a new tensor: repeat(a) to fit in b
     GGML_API struct ggml_tensor * ggml_repeat(
@@ -731,6 +767,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_tanh(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_tanh_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_relu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1081,58 +1133,33 @@ extern "C" {
             float                 min,
             float                 max);
 
-    // TODO: implement general-purpose convolutions
-    // GGML_API struct ggml_tensor * ggml_conv_1d(
-    //        struct ggml_context * ctx,
-    //        struct ggml_tensor  * a,
-    //        struct ggml_tensor  * b,
-    //        int                   s0
-    //        int                   p0,
-    //        int                   d0);
-    //
-    // GGML_API struct ggml_tensor * ggml_conv_2d(
-    //        struct ggml_context * ctx,
-    //        struct ggml_tensor  * a,
-    //        struct ggml_tensor  * b,
-    //        int                   s0,
-    //        int                   s1,
-    //        int                   p0,
-    //        int                   p1,
-    //        int                   d0,
-    //        int                   d1);
-
-    // padding = half
-    // TODO: we don't support extra parameters for now
-    //       that's why we are hard-coding the stride, padding, and dilation
-    //       not great ..
-    // example:
-    // a:      3   80  768    1
-    // b:   3000   80    1    1
-    // res: 3000  768    1    1
-    // used in whisper
-    GGML_API struct ggml_tensor * ggml_conv_1d_s1_ph(
+    GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
 
-    // used in whisper
-    GGML_API struct ggml_tensor * ggml_conv_1d_s2_ph(
+    GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   s0,
+            int                   s1,
+            int                   p0,
+            int                   p1,
+            int                   d0,
+            int                   d1);
 
-    // kernel size is a->ne[0] x a->ne[1]
-    // stride is equal to kernel size
-    // padding is zero
-    // example:
-    // a:     16   16    3  768
-    // b:   1024 1024    3    1
-    // res:   64   64  768    1
-    // used in sam
-    GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
+    // conv_1d with padding = half
+    // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
+    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   s,
+            int                   d);
 
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
@@ -1487,26 +1514,19 @@ extern "C" {
     // Internal types and functions exposed for tests and benchmarks
     //
 
-#ifdef  __cplusplus
-    // restrict not standard in C++
-#define GGML_RESTRICT
-#else
-#define GGML_RESTRICT restrict
-#endif
-    typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-    typedef void (*quantize_row_q_t)  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-    typedef void (*vec_dot_q_t)       (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+    typedef void (*ggml_to_float_t)(const void * x, float * y, int k);
+    typedef void (*ggml_from_float_t)(const float * x, void * y, int k);
+    typedef void (*ggml_vec_dot_t)(const int n, float * s, const void * x, const void * y);
 
     typedef struct {
-        dequantize_row_q_t dequantize_row_q;
-        quantize_row_q_t   quantize_row_q;
-        quantize_row_q_t   quantize_row_q_reference;
-        quantize_row_q_t   quantize_row_q_dot;
-        vec_dot_q_t        vec_dot_q;
-        enum ggml_type     vec_dot_type;
-    } quantize_fns_t;
+        ggml_to_float_t   to_float;
+        ggml_from_float_t from_float;
+        ggml_from_float_t from_float_reference;
+        ggml_vec_dot_t    vec_dot;
+        enum ggml_type    vec_dot_type;
+    } ggml_type_traits_t;
 
-    quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
+    ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type i);
 
 #ifdef  __cplusplus
 }

llama.cpp

@@ -66,6 +66,7 @@ enum e_model {
     MODEL_65B,
 };
 
+static const size_t kB = 1024;
 static const size_t MB = 1024*1024;
 
 // computed for n_ctx == 2048
@@ -129,6 +130,34 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
     return k_sizes;
 }
 
+// amount of VRAM needed per batch size to hold temporary results
+// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
+{
+    static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,   512ull * kB },
+        { MODEL_7B,   512ull * kB },
+        { MODEL_13B,  640ull * kB },
+        { MODEL_30B,  768ull * kB },
+        { MODEL_65B, 1536ull * kB },
+    };
+    return k_sizes;
+}
+
+// amount of VRAM needed per batch size and context to hold temporary results
+// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
+{
+    static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,  128ull },
+        { MODEL_7B,  128ull },
+        { MODEL_13B, 160ull },
+        { MODEL_30B, 208ull },
+        { MODEL_65B, 416ull },
+    };
+    return k_sizes;
+}
+
 // default hparams (LLaMA 7B)
 struct llama_hparams {
     uint32_t n_vocab = 32000;
@@ -165,8 +194,8 @@ struct llama_layer {
 };
 
 struct llama_kv_cache {
-    struct ggml_tensor * k;
-    struct ggml_tensor * v;
+    struct ggml_tensor * k = NULL;
+    struct ggml_tensor * v = NULL;
 
     struct ggml_context * ctx = NULL;
 
@@ -253,7 +282,13 @@ struct llama_model {
 
 struct llama_context {
     llama_context(const llama_model & model, const llama_vocab & vocab) : model(model), vocab(vocab), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {}
-
+#ifdef GGML_USE_METAL
+    ~llama_context() {
+        if (ctx_metal) {
+            ggml_metal_free(ctx_metal);
+        }
+    }
+#endif
     std::mt19937 rng;
 
     bool has_evaluated_once = false;
@@ -446,9 +481,7 @@ struct llama_file_loader {
             std::string word = file.read_string(len);
 
             float score = 0.0f;
-            if (file_version >= LLAMA_FILE_VERSION_GGMF_V1) {
-                file.read_raw(&score, sizeof(score));
-            }
+            file.read_raw(&score, sizeof(score));
 
             vocab.token_to_id[word] = i;
 
@@ -777,7 +810,7 @@ static bool kv_cache_init(
 
 struct llama_context_params llama_context_default_params() {
     struct llama_context_params result = {
-        /*.seed                        =*/ -1,
+        /*.seed                        =*/ LLAMA_DEFAULT_SEED,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
         /*.gpu_layers                  =*/ 0,
@@ -1112,14 +1145,18 @@ static void llama_model_load_internal(
             fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
             ggml_cuda_set_scratch_size(0); // disable scratch
         } else {
-            vram_scratch = n_batch * MB;
+            const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type);
+            const size_t vram_scratch_per_context = VRAM_REQ_SCRATCH_PER_CONTEXT().at(model.type);
+            vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context);
             ggml_cuda_set_scratch_size(vram_scratch);
             if (n_gpu_layers > 0) {
-                fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n",
-                        __func__, vram_scratch / MB);
+                fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n",
+                        __func__, vram_scratch_base / kB, vram_scratch_per_context,
+                        (vram_scratch + MB - 1) / MB); // round up
             }
         }
 #endif // GGML_USE_CUBLAS
+
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
@@ -1128,6 +1165,10 @@ static void llama_model_load_internal(
             fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
         }
         size_t vram_kv_cache = 0;
+
+#ifdef GGML_USE_CUBLAS
+        const int max_backend_supported_layers = hparams.n_layer + 3;
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
         if (n_gpu_layers > (int) hparams.n_layer + 1) {
             if (low_vram) {
                 fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
@@ -1144,14 +1185,18 @@ static void llama_model_load_internal(
                 vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
             }
         }
-        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+#elif defined(GGML_USE_CLBLAST)
+        const int max_backend_supported_layers = hparams.n_layer + 1;
+        const int max_offloadable_layers = hparams.n_layer + 1;
+#endif // GGML_USE_CUBLAS
+
         fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
-                __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3);
+                __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
         fprintf(stderr, "%s: total VRAM used: %zu MB\n",
                 __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
-#endif
+#endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
     }
 
     // populate `tensors_by_name`
@@ -1860,10 +1905,10 @@ void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * can
         return;
     }
 
-    const int64_t t_start_sample_us = ggml_time_us();
-
     llama_sample_softmax(ctx, candidates);
 
+    const int64_t t_start_sample_us = ggml_time_us();
+
     // Compute the cumulative probabilities
     float cum_sum = 0.0f;
     size_t last_idx = candidates->size;
@@ -1892,9 +1937,8 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array *
         return;
     }
 
-    const int64_t t_start_sample_us = ggml_time_us();
-
     llama_sample_softmax(nullptr, candidates);
+    const int64_t t_start_sample_us = ggml_time_us();
 
     // Compute the first and second derivatives
     std::vector<float> first_derivatives(candidates->size - 1);
@@ -1946,11 +1990,11 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
         return;
     }
 
-    const int64_t t_start_sample_us = ggml_time_us();
-
     // Compute the softmax of logits and calculate entropy
     llama_sample_softmax(nullptr, candidates);
 
+    const int64_t t_start_sample_us = ggml_time_us();
+
     float entropy = 0.0f;
     for (size_t i = 0; i < candidates->size; ++i) {
         entropy += -candidates->data[i].p * logf(candidates->data[i].p);
@@ -2119,13 +2163,11 @@ llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_
 
     if (ctx) {
         ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-        ctx->n_sample++;
     }
     return X;
 }
 
 llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu) {
-    assert(ctx);
     int64_t t_start_sample_us;
     t_start_sample_us = ggml_time_us();
 
@@ -2140,13 +2182,14 @@ llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_tok
         candidates->size = 1;
     }
 
+    if (ctx) {
+        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    }
+
     // Normalize the probabilities of the remaining words
     llama_sample_softmax(ctx, candidates);
 
     // Sample the next word X from the remaining words
-    if (ctx) {
-        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-    }
     llama_token X = llama_sample_token(ctx, candidates);
     t_start_sample_us = ggml_time_us();
 
@@ -2214,10 +2257,10 @@ static void llama_convert_tensor_internal(const llama_load_tensor & tensor, llam
     }
     float * f32_output = (float *) output.addr;
 
-    quantize_fns_t qtype;
+    ggml_type_traits_t qtype;
     if (ggml_is_quantized(tensor.type)) {
-        qtype = ggml_internal_get_quantize_fn(tensor.type);
-        if (qtype.dequantize_row_q == NULL) {
+        qtype = ggml_internal_get_type_traits(tensor.type);
+        if (qtype.to_float == NULL) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor.type)));
         }
     } else if (tensor.type != GGML_TYPE_F16) {
@@ -2228,7 +2271,7 @@ static void llama_convert_tensor_internal(const llama_load_tensor & tensor, llam
         if (tensor.type == GGML_TYPE_F16) {
             ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor.data, f32_output, nelements);
         } else if (ggml_is_quantized(tensor.type)) {
-            qtype.dequantize_row_q(tensor.data, f32_output, nelements);
+            qtype.to_float(tensor.data, f32_output, nelements);
         } else {
             LLAMA_ASSERT(false); // unreachable
         }
@@ -2253,7 +2296,7 @@ static void llama_convert_tensor_internal(const llama_load_tensor & tensor, llam
             if (typ == GGML_TYPE_F16) {
                 ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
             } else {
-                qtype.dequantize_row_q(inbuf, outbuf, nels);
+                qtype.to_float(inbuf, outbuf, nels);
             }
         };
         workers.push_back(std::thread(compute, tensor.type, tensor.data + in_buff_offs, f32_output + out_buff_offs, thr_elems));
@@ -2541,7 +2584,7 @@ struct llama_context * llama_new_context_with_model(
 
     llama_context * ctx = new llama_context(*model, model->vocab);
 
-    if (params.seed < 0) {
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
@@ -2723,7 +2766,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
 
     // create a name -> tensor map of the model to accelerate lookups
     std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
-    for (auto & kv: model.tensors_by_name) {
+    for (const auto & kv: model.tensors_by_name) {
         model_tensors.insert(kv);
     }
 
@@ -2974,8 +3017,8 @@ int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
 
 #define LLAMA_MAX_RNG_STATE (64*1024)
 
-void llama_set_rng_seed(struct llama_context * ctx, int seed) {
-    if (seed < 0) {
+void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) {
+    if (seed == LLAMA_DEFAULT_SEED) {
         seed = time(NULL);
     }
     ctx->rng.seed(seed);
@@ -3219,7 +3262,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
     return nread;
 }
 
-bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+static bool llama_load_session_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
     llama_file file(path_session, "rb");
 
     // sanity checks
@@ -3273,6 +3316,15 @@ bool llama_load_session_file(struct llama_context * ctx, const char * path_sessi
     return true;
 }
 
+bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    try {
+        return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
+    } catch (const std::exception & err) {
+        fprintf(stderr, "error loading session file: %s\n", err.what());
+        return false;
+    }
+}
+
 bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
     llama_file file(path_session, "wb");
 

llama.h

@@ -46,6 +46,8 @@
 #define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
+#define LLAMA_DEFAULT_SEED           0xFFFFFFFF
+
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
@@ -81,11 +83,11 @@ extern "C" {
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
    struct llama_context_params {
-        int seed;                              // RNG seed, -1 for random
-        int n_ctx;                             // text context
-        int n_batch;                           // prompt processing batch size
-        int n_gpu_layers;                      // number of layers to store in VRAM
-        int main_gpu;                          // the GPU that is used for scratch and small tensors
+        uint32_t seed;                         // RNG seed, -1 for random
+        int32_t  n_ctx;                        // text context
+        int32_t  n_batch;                      // prompt processing batch size
+        int32_t  n_gpu_layers;                 // number of layers to store in VRAM
+        int32_t  main_gpu;                     // the GPU that is used for scratch and small tensors
         float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
         // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
@@ -196,7 +198,7 @@ extern "C" {
     LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
 
     // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
 
     // Returns the maximum size in bytes of the state (rng, logits, embedding
     // and kv_cache) - will often be smaller after compacting tokens

pocs/vdot/q8dot.cpp

@@ -136,7 +136,7 @@ int main(int argc, char** argv) {
 
     auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
 
-    auto funcs = ggml_internal_get_quantize_fn(ggml_type);
+    auto funcs = ggml_internal_get_type_traits(ggml_type);
 
     Stat simple, ggml;
 
@@ -156,8 +156,8 @@ int main(int argc, char** argv) {
 
         t1 = std::chrono::high_resolution_clock::now();
         float fs;
-        if (type == 0) funcs.vec_dot_q(kVecSize * QK4_1, &fs, x40.data(), y.data());
-        else funcs.vec_dot_q(kVecSize * QK4_1, &fs, x41.data(), y.data());
+        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, x40.data(), y.data());
+        else funcs.vec_dot(kVecSize * QK4_1, &fs, x41.data(), y.data());
         t2 = std::chrono::high_resolution_clock::now();
         t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
         if (iloop > 3) ggml.addResult(fs, t);

pocs/vdot/vdot.cpp

@@ -235,7 +235,7 @@ int main(int argc, char** argv) {
     int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
     int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
 
-    auto funcs = useQ4_1 ? ggml_internal_get_quantize_fn(GGML_TYPE_Q4_1) : ggml_internal_get_quantize_fn(GGML_TYPE_Q4_0);
+    auto funcs = useQ4_1 ? ggml_internal_get_type_traits(GGML_TYPE_Q4_1) : ggml_internal_get_type_traits(GGML_TYPE_Q4_0);
 
     std::vector<block_q4_0> q40;
     std::vector<block_q4_1> q41;
@@ -261,9 +261,9 @@ int main(int argc, char** argv) {
         // Note, we do not include this in the timing as in practical application
         // we already have the quantized model weights.
         if (useQ4_1) {
-            funcs.quantize_row_q(x1.data(), q41.data(), kVecSize);
+            funcs.from_float(x1.data(), q41.data(), kVecSize);
         } else {
-            funcs.quantize_row_q(x1.data(), q40.data(), kVecSize);
+            funcs.from_float(x1.data(), q40.data(), kVecSize);
         }
 
         // Now measure time the dot product needs using the "scalar" version above
@@ -282,9 +282,10 @@ int main(int argc, char** argv) {
             dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
         }
         else {
-            funcs.quantize_row_q_dot(y1.data(), q8.data(), kVecSize);
-            if (useQ4_1) funcs.vec_dot_q(kVecSize, &result, q41.data(), q8.data());
-            else funcs.vec_dot_q(kVecSize, &result, q40.data(), q8.data());
+            auto vdot = ggml_internal_get_type_traits(funcs.vec_dot_type);
+            vdot.from_float(y1.data(), q8.data(), kVecSize);
+            if (useQ4_1) funcs.vec_dot(kVecSize, &result, q41.data(), q8.data());
+            else funcs.vec_dot(kVecSize, &result, q40.data(), q8.data());
         }
         sumq += result;
         t2 = std::chrono::high_resolution_clock::now();

scripts/sync-ggml.sh

@@ -1,6 +1,14 @@
 #!/bin/bash
 
-cp -rpv ../ggml/src/ggml.c          ./ggml.c
-cp -rpv ../ggml/src/ggml-cuda.cu    ./ggml-cuda.cu
-cp -rpv ../ggml/src/ggml-cuda.h     ./ggml-cuda.h
-cp -rpv ../ggml/include/ggml/ggml.h ./ggml.h
+cp -rpv ../ggml/src/ggml.c           ./ggml.c
+cp -rpv ../ggml/src/ggml-cuda.h      ./ggml-cuda.h
+cp -rpv ../ggml/src/ggml-cuda.cu     ./ggml-cuda.cu
+cp -rpv ../ggml/src/ggml-opencl.h    ./ggml-opencl.h
+cp -rpv ../ggml/src/ggml-opencl.cpp  ./ggml-opencl.cpp
+cp -rpv ../ggml/src/ggml-metal.h     ./ggml-metal.h
+cp -rpv ../ggml/src/ggml-metal.m     ./ggml-metal.m
+cp -rpv ../ggml/src/ggml-metal.metal ./ggml-metal.metal
+cp -rpv ../ggml/include/ggml/ggml.h  ./ggml.h
+
+cp -rpv ../ggml/tests/test-opt.c    ./tests/test-opt.c
+cp -rpv ../ggml/tests/test-grad0.c  ./tests/test-grad0.c

tests/test-grad0.c

@@ -1154,7 +1154,7 @@ int main(int argc, const char ** argv) {
                             continue;
                         }
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
 
                         GGML_PRINT_DEBUG("rope: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                         check_gradient("rope", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);

tests/test-quantize-fns.cpp

@@ -40,26 +40,26 @@ float array_rmse(const float * a1, const float * a2, size_t n) {
 }
 
 // Total quantization error on test data
-float total_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
+float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
 
-    qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
-    qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
+    qfns.from_float(test_data, tmp_q.data(), test_size);
+    qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
     return array_rmse(test_data, tmp_out.data(), test_size);
 }
 
 // Total quantization error on test data
-float reference_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
+float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
     std::vector<float> tmp_out_ref(test_size);
 
-    qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
-    qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
+    qfns.from_float(test_data, tmp_q.data(), test_size);
+    qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
 
-    qfns.quantize_row_q_reference(test_data, tmp_q.data(), test_size);
-    qfns.dequantize_row_q(tmp_q.data(), tmp_out_ref.data(), test_size);
+    qfns.from_float_reference(test_data, tmp_q.data(), test_size);
+    qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
 
     return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
 }
@@ -73,15 +73,17 @@ float dot_product(const float * a1, const float * a2, size_t test_size) {
 }
 
 // Total dot product error
-float dot_product_error(quantize_fns_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
+float dot_product_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
     std::vector<uint8_t> tmp_q1(2*test_size);
     std::vector<uint8_t> tmp_q2(2*test_size);
 
-    qfns.quantize_row_q    (test_data1, tmp_q1.data(), test_size);
-    qfns.quantize_row_q_dot(test_data2, tmp_q2.data(), test_size);
+    auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
+
+    qfns.from_float(test_data1, tmp_q1.data(), test_size);
+    vdot.from_float(test_data2, tmp_q2.data(), test_size);
 
     float result = INFINITY;
-    qfns.vec_dot_q(test_size, &result, tmp_q1.data(), tmp_q2.data());
+    qfns.vec_dot(test_size, &result, tmp_q1.data(), tmp_q2.data());
 
     const float dot_ref = dot_product(test_data1, test_data2, test_size);
 
@@ -123,9 +125,9 @@ int main(int argc, char * argv[]) {
 
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
-        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
 
-        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+        if (qfns.from_float && qfns.to_float) {
             const float total_error = total_quantization_error(qfns, test_size, test_data.data());
             const float max_quantization_error =
                 type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :

tests/test-quantize-perf.cpp

@@ -123,9 +123,9 @@ void usage(char * argv[]) {
     printf("  --type TYPE           set test type as");
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
-        quantize_fns_t qfns = ggml_internal_get_quantize_fn(type);
+        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
         if (ggml_type_name(type) != NULL) {
-            if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+            if (qfns.from_float && qfns.to_float) {
                 printf(" %s", ggml_type_name(type));
             }
         }
@@ -271,12 +271,12 @@ int main(int argc, char * argv[]) {
 
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
-        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
         if (!params.include_types.empty() && ggml_type_name(type) && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
             continue;
         }
 
-        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+        if (qfns.from_float && qfns.to_float) {
             printf("%s\n", ggml_type_name(type));
 
             if (params.op_quantize_row_q_reference) {
@@ -284,7 +284,7 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void ) {
-                        qfns.quantize_row_q_reference(test_data1, test_q1, size);
+                        qfns.from_float_reference(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
@@ -298,7 +298,7 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void ) {
-                        qfns.quantize_row_q(test_data1, test_q1, size);
+                        qfns.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
@@ -309,11 +309,11 @@ int main(int argc, char * argv[]) {
 
             if (params.op_dequantize_row_q) {
                 printf("  dequantize_row_q\n");
-                qfns.quantize_row_q(test_data1, test_q1, largest);
+                qfns.from_float(test_data1, test_q1, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void ) {
-                        qfns.dequantize_row_q(test_q1, test_out, size);
+                        qfns.to_float(test_q1, test_out, size);
                         return test_out[0];
                     };
                     size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
@@ -327,7 +327,8 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void ) {
-                        qfns.quantize_row_q_dot(test_data1, test_q1, size);
+                        auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
+                        vdot.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
@@ -338,13 +339,13 @@ int main(int argc, char * argv[]) {
 
             if (params.op_vec_dot_q) {
                 printf("  vec_dot_q\n");
-                qfns.quantize_row_q(test_data1, test_q1, largest);
-                qfns.quantize_row_q(test_data2, test_q2, largest);
+                qfns.from_float(test_data1, test_q1, largest);
+                qfns.from_float(test_data2, test_q2, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void ) {
                         float result;
-                        qfns.vec_dot_q(size, &result, test_q1, test_q2);
+                        qfns.vec_dot(size, &result, test_q1, test_q2);
                         return result;
                     };
                     size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);