readme : remove obsolete Zig instructions (#7471)

compare-commits.sh : hide stdout, use -oe to print markdown

.gitignore

@@ -34,9 +34,11 @@ ggml-metal-embed.metal
 lcov-report/
 gcovr-report/
 
+tags
 build*
 !build.zig
 cmake-build-*
+android-ndk-*
 out/
 tmp/
 

CMakeLists.txt

@@ -557,12 +557,17 @@ if (LLAMA_VULKAN)
 endif()
 
 if (LLAMA_HIPBLAS)
-    if ($ENV{ROCM_PATH})
-        set(ROCM_PATH $ENV{ROCM_PATH})
+    if (NOT EXISTS $ENV{ROCM_PATH})
+        if (NOT EXISTS /opt/rocm)
+            set(ROCM_PATH /usr)
+        else()
+            set(ROCM_PATH /opt/rocm)
+        endif()
     else()
-        set(ROCM_PATH /opt/rocm)
+        set(ROCM_PATH $ENV{ROCM_PATH})
     endif()
     list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})
+    list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
 
     # CMake on Windows doesn't support the HIP language yet
     if(WIN32)

Makefile

@@ -69,6 +69,10 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
+ifdef LLAMA_RPC
+	BUILD_TARGETS += rpc-server
+endif
+
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -429,6 +433,11 @@ ifdef LLAMA_BLIS
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
 endif # LLAMA_BLIS
 
+ifdef LLAMA_RPC
+	MK_CPPFLAGS   += -DGGML_USE_RPC
+	OBJS          += ggml-rpc.o
+endif # LLAMA_RPC
+
 ifdef LLAMA_CUBLAS
 # LLAMA_CUBLAS is deprecated and will be removed in the future
 	LLAMA_CUDA := 1
@@ -654,11 +663,26 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 endif
 endif # LLAMA_METAL
 
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
+
 ifndef LLAMA_NO_LLAMAFILE
 sgemm.o: sgemm.cpp sgemm.h ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif
 
+ifdef LLAMA_RPC
+ggml-rpc.o: ggml-rpc.cpp ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+rpc-server.o: examples/rpc/rpc-server.cpp ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+rpc-server: rpc-server.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif # LLAMA_RPC
+
 GF_CC := $(CC)
 include scripts/get-flags.mk
 
@@ -738,14 +762,9 @@ unicode.o: unicode.cpp unicode.h
 unicode-data.o: unicode-data.cpp unicode-data.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
-
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
-
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 

README.md

@@ -364,17 +364,6 @@ In order to build llama.cpp you have four different options.
       cmake --build build --config Debug
       ```
 
-- Using `Zig` (version 0.11 or later):
-
-    Building for optimization levels and CPU features can be accomplished using standard build arguments, for example AVX2, FMA, F16C,
-    it's also possible to cross compile for other operating systems and architectures:
-
-    ```bash
-    zig build -Doptimize=ReleaseFast -Dtarget=x86_64-windows-gnu -Dcpu=x86_64+avx2+fma+f16c
-    ```
-
-    The `zig targets` command will give you valid options to use.
-
 -   Using `gmake` (FreeBSD):
 
     1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)

cmake/arm64-windows-llvm.cmake

@@ -9,7 +9,7 @@ set( CMAKE_CXX_COMPILER  clang++ )
 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
 
-set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast" )
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
 
 set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )

examples/llama-bench/llama-bench.cpp

@@ -140,10 +140,11 @@ static std::string get_gpu_info() {
 }
 
 // command line params
-enum output_formats {CSV, JSON, MARKDOWN, SQL};
+enum output_formats {NONE, CSV, JSON, MARKDOWN, SQL};
 
 static const char * output_format_str(output_formats format) {
     switch (format) {
+        case NONE:     return "none";
         case CSV:      return "csv";
         case JSON:     return "json";
         case MARKDOWN: return "md";
@@ -152,6 +153,23 @@ static const char * output_format_str(output_formats format) {
     }
 }
 
+static bool output_format_from_str(const std::string & s, output_formats & format) {
+    if (s == "none") {
+        format = NONE;
+    } else if (s == "csv") {
+        format = CSV;
+    } else if (s == "json") {
+        format = JSON;
+    } else if (s == "md") {
+        format = MARKDOWN;
+    } else if (s == "sql") {
+        format = SQL;
+    } else {
+        return false;
+    }
+    return true;
+}
+
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
         case LLAMA_SPLIT_MODE_NONE:  return "none";
@@ -190,31 +208,33 @@ struct cmd_params {
     int reps;
     bool verbose;
     output_formats output_format;
+    output_formats output_format_stderr;
 };
 
 static const cmd_params cmd_params_defaults = {
-    /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
-    /* n_prompt      */ {512},
-    /* n_gen         */ {128},
-    /* n_pg          */ {},
-    /* n_batch       */ {2048},
-    /* n_ubatch      */ {512},
-    /* type_k        */ {GGML_TYPE_F16},
-    /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {cpu_get_num_math()},
-    /* n_gpu_layers  */ {99},
-    /* rpc_servers   */ {""},
-    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
-    /* main_gpu      */ {0},
-    /* no_kv_offload */ {false},
-    /* flash_attn    */ {false},
-    /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
-    /* use_mmap      */ {true},
-    /* embeddings    */ {false},
-    /* numa          */ GGML_NUMA_STRATEGY_DISABLED,
-    /* reps          */ 5,
-    /* verbose       */ false,
-    /* output_format */ MARKDOWN
+    /* model                */ {"models/7B/ggml-model-q4_0.gguf"},
+    /* n_prompt             */ {512},
+    /* n_gen                */ {128},
+    /* n_pg                 */ {},
+    /* n_batch              */ {2048},
+    /* n_ubatch             */ {512},
+    /* type_k               */ {GGML_TYPE_F16},
+    /* type_v               */ {GGML_TYPE_F16},
+    /* n_threads            */ {cpu_get_num_math()},
+    /* n_gpu_layers         */ {99},
+    /* rpc_servers          */ {""},
+    /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
+    /* main_gpu             */ {0},
+    /* no_kv_offload        */ {false},
+    /* flash_attn           */ {false},
+    /* tensor_split         */ {std::vector<float>(llama_max_devices(), 0.0f)},
+    /* use_mmap             */ {true},
+    /* embeddings           */ {false},
+    /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
+    /* reps                 */ 5,
+    /* verbose              */ false,
+    /* output_format        */ MARKDOWN,
+    /* output_format_stderr */ NONE,
 };
 
 static void print_usage(int /* argc */, char ** argv) {
@@ -243,6 +263,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
     printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err <csv|json|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
@@ -284,6 +305,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
 
     params.verbose = cmd_params_defaults.verbose;
     params.output_format = cmd_params_defaults.output_format;
+    params.output_format_stderr = cmd_params_defaults.output_format_stderr;
     params.reps = cmd_params_defaults.reps;
 
     for (int i = 1; i < argc; i++) {
@@ -493,18 +515,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            if (argv[i] == std::string("csv")) {
-                params.output_format = CSV;
-            } else if (argv[i] == std::string("json")) {
-                params.output_format = JSON;
-            } else if (argv[i] == std::string("md")) {
-                params.output_format = MARKDOWN;
-            } else if (argv[i] == std::string("sql")) {
-                params.output_format = SQL;
-            } else {
+            invalid_param = !output_format_from_str(argv[i], params.output_format);
+        } else if (arg == "-oe" || arg == "--output-err") {
+            if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
+            invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
         } else if (arg == "-v" || arg == "--verbose") {
             params.verbose = true;
         } else {
@@ -1278,6 +1295,22 @@ static void llama_null_log_callback(enum ggml_log_level level, const char * text
     (void) user_data;
 }
 
+static std::unique_ptr<printer> create_printer(output_formats format) {
+    switch (format) {
+        case NONE:
+            return nullptr;
+        case CSV:
+            return std::unique_ptr<printer>(new csv_printer());
+        case JSON:
+            return std::unique_ptr<printer>(new json_printer());
+        case MARKDOWN:
+            return std::unique_ptr<printer>(new markdown_printer());
+        case SQL:
+            return std::unique_ptr<printer>(new sql_printer());
+    }
+    GGML_ASSERT(false);
+}
+
 int main(int argc, char ** argv) {
     // try to set locale for unicode characters in markdown
     setlocale(LC_CTYPE, ".UTF-8");
@@ -1304,26 +1337,18 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // initialize printer
-    std::unique_ptr<printer> p;
-    switch (params.output_format) {
-        case CSV:
-            p.reset(new csv_printer());
-            break;
-        case JSON:
-            p.reset(new json_printer());
-            break;
-        case MARKDOWN:
-            p.reset(new markdown_printer());
-            break;
-        case SQL:
-            p.reset(new sql_printer());
-            break;
-        default:
-            assert(false);
-            exit(1);
+    std::unique_ptr<printer> p = create_printer(params.output_format);
+    std::unique_ptr<printer> p_err = create_printer(params.output_format_stderr);
+
+    if (p) {
+        p->fout = stdout;
+        p->print_header(params);
+    }
+
+    if (p_err) {
+        p_err->fout = stderr;
+        p_err->print_header(params);
     }
-    p->fout = stdout;
-    p->print_header(params);
 
     std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
 
@@ -1381,7 +1406,15 @@ int main(int argc, char ** argv) {
             t.samples_ns.push_back(t_ns);
         }
 
-        p->print_test(t);
+        if (p) {
+            p->print_test(t);
+            fflush(p->fout);
+        }
+
+        if (p_err) {
+            p_err->print_test(t);
+            fflush(p_err->fout);
+        }
 
         llama_print_timings(ctx);
 
@@ -1390,7 +1423,13 @@ int main(int argc, char ** argv) {
 
     llama_free_model(lmodel);
 
-    p->print_footer();
+    if (p) {
+        p->print_footer();
+    }
+
+    if (p_err) {
+        p_err->print_footer();
+    }
 
     llama_backend_free();
 

ggml-alloc.c

@@ -750,7 +750,7 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
+            ggml_backend_view_init(tensor);
         }
     } else {
         if (tensor->data == NULL) {
@@ -899,12 +899,12 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
             if (t->view_src == NULL) {
                 ggml_tallocr_alloc(&tallocr, t);
             } else if (t->buffer == NULL) {
-                ggml_backend_view_init(buffer, t);
+                ggml_backend_view_init(t);
             }
         } else {
             if (t->view_src != NULL && t->buffer == NULL) {
                 // view of a pre-allocated tensor
-                ggml_backend_view_init(buffer, t);
+                ggml_backend_view_init(t);
             }
         }
     }

ggml-backend.c

@@ -151,7 +151,7 @@ void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
 bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
     ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
     if (dst_buf->iface.cpy_tensor) {
-        return src->buffer->iface.cpy_tensor(dst_buf, src, dst);
+        return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
     }
     return false;
 }
@@ -1887,15 +1887,15 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
 
 // utils
 
-void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+void ggml_backend_view_init(struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->buffer == NULL);
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
     GGML_ASSERT(tensor->view_src->data != NULL);
 
-    tensor->buffer = buffer;
+    tensor->buffer = tensor->view_src->buffer;
     tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    ggml_backend_buffer_init_tensor(buffer, tensor);
+    ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
 }
 
 void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
@@ -1954,7 +1954,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
     struct ggml_tensor * dst = node_copies[id];
     if (dst->view_src != NULL) {
         graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
-        ggml_backend_view_init(dst->view_src->buffer, dst);
+        ggml_backend_view_init(dst);
     }
     else {
         ggml_backend_tensor_copy(src, dst);

ggml-backend.h

@@ -225,7 +225,7 @@ extern "C" {
 
     // Tensor initialization
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
 
 #ifdef  __cplusplus

ggml-rpc.cpp

@@ -491,7 +491,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer
     if (remote_ptr != 0) {
         ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
             ggml_backend_rpc_buffer_interface,
-            new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC"},
+            new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC[" + std::string(buft_ctx->endpoint) + "]"},
             remote_size);
         return buffer;
     } else {
@@ -692,7 +692,7 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const
 GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
         /* .endpoint  = */ endpoint,
-        /* .name      = */ "RPC",
+        /* .name      = */ "RPC[" + std::string(endpoint) + "]",
     };
 
     ggml_backend_t backend = new ggml_backend {

ggml.c

@@ -2272,6 +2272,11 @@ inline static float ggml_silu_f32(float x) {
     return x/(1.0f + expf(-x));
 }
 
+#if __FINITE_MATH_ONLY__
+#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
+#error "ref: https://github.com/ggerganov/llama.cpp/pull/7154#issuecomment-2143844461"
+#endif
+
 #if defined(__ARM_NEON) && defined(__aarch64__)
 
 // adapted from arm limited optimized routine

llama.cpp

@@ -2149,12 +2149,12 @@ struct llama_control_vector {
 struct llama_vocab {
     using id    = int32_t;
     using token = std::string;
-    using ttype = llama_token_type;
+    using tattr = llama_token_attr;
 
     struct token_data {
         token text;
         float score;
-        ttype type;
+        tattr attr;
     };
 
     enum llama_vocab_type     type     = LLAMA_VOCAB_TYPE_SPM;
@@ -2371,13 +2371,34 @@ struct llama_context {
     struct llama_control_vector cvec;
 };
 
+static size_t llama_get_device_count(const llama_model & model) {
+    size_t count = 1;
+#if defined(GGML_USE_CUDA)
+    count = ggml_backend_cuda_get_device_count();
+#elif defined(GGML_USE_SYCL)
+    count = ggml_backend_sycl_get_device_count();
+#elif defined(GGML_USE_VULKAN)
+    count = ggml_backend_vk_get_device_count();
+#endif
+#if defined(GGML_USE_RPC)
+    count += model.rpc_servers.size();
+#endif
+    return count;
+    GGML_UNUSED(model);
+}
+
 static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) {
     ggml_backend_buffer_type_t buft = nullptr;
 
-#ifdef GGML_USE_RPC
-    std::string endpoint = model.rpc_servers[gpu];
-    buft = ggml_backend_rpc_buffer_type(endpoint.c_str());
-#elif defined(GGML_USE_METAL)
+#if defined(GGML_USE_RPC)
+    int dev_count = (int)llama_get_device_count(model);
+    int rpc_count = (int)model.rpc_servers.size();
+    if (gpu >= dev_count - rpc_count) {
+        const char * endpoint = model.rpc_servers[gpu - dev_count + rpc_count].c_str();
+        return ggml_backend_rpc_buffer_type(endpoint);
+    }
+#endif
+#if defined(GGML_USE_METAL)
     buft = ggml_backend_metal_buffer_type();
 #elif defined(GGML_USE_CUDA)
     buft = ggml_backend_cuda_buffer_type(gpu);
@@ -2425,29 +2446,19 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo
     GGML_UNUSED(tensor_split);
 }
 
-static size_t llama_get_device_count(const llama_model & model) {
-#if defined(GGML_USE_RPC)
-    return model.rpc_servers.size();
-#elif defined(GGML_USE_CUDA)
-    return ggml_backend_cuda_get_device_count();
-#elif defined(GGML_USE_SYCL)
-    return ggml_backend_sycl_get_device_count();
-#elif defined(GGML_USE_VULKAN)
-    return ggml_backend_vk_get_device_count();
-#else
-    return 1;
-#endif
-    GGML_UNUSED(model);
-}
-
 static size_t llama_get_device_memory(const llama_model & model, int device) {
 #if defined(GGML_USE_RPC)
-    size_t total;
-    size_t free;
-    std::string endpoint = model.rpc_servers[device];
-    ggml_backend_rpc_get_device_memory(endpoint.c_str(), &free, &total);
-    return free;
-#elif defined(GGML_USE_CUDA)
+    int dev_count = (int)llama_get_device_count(model);
+    int rpc_count = (int)model.rpc_servers.size();
+    if (device >= dev_count - rpc_count) {
+        size_t total;
+        size_t free;
+        const char * endpoint = model.rpc_servers[device - dev_count + rpc_count].c_str();
+        ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
+        return free;
+    }
+#endif
+#if defined(GGML_USE_CUDA)
     size_t total;
     size_t free;
     ggml_backend_cuda_get_device_memory(device, &free, &total);
@@ -4739,7 +4750,20 @@ static void llm_load_vocab(
         auto & token_data = vocab.id_to_token[i];
         token_data.text  = std::move(word);
         token_data.score = scores ? scores[i] : 0.0f;
-        token_data.type  = toktypes ? (llama_token_type) toktypes[i] : LLAMA_TOKEN_TYPE_NORMAL;
+        token_data.attr  = LLAMA_TOKEN_ATTR_NORMAL;
+
+        if (toktypes) {  //TODO: remove, required until per token attributes are available from GGUF file
+            switch(toktypes[i]) {
+                case LLAMA_TOKEN_TYPE_UNKNOWN:      token_data.attr = LLAMA_TOKEN_ATTR_UNKNOWN;      break;
+                case LLAMA_TOKEN_TYPE_UNUSED:       token_data.attr = LLAMA_TOKEN_ATTR_UNUSED;       break;
+                case LLAMA_TOKEN_TYPE_NORMAL:       token_data.attr = LLAMA_TOKEN_ATTR_NORMAL;       break;
+                case LLAMA_TOKEN_TYPE_CONTROL:      token_data.attr = LLAMA_TOKEN_ATTR_CONTROL;      break;
+                case LLAMA_TOKEN_TYPE_USER_DEFINED: token_data.attr = LLAMA_TOKEN_ATTR_USER_DEFINED; break;
+                case LLAMA_TOKEN_TYPE_BYTE:         token_data.attr = LLAMA_TOKEN_ATTR_BYTE;         break;
+                case LLAMA_TOKEN_TYPE_UNDEFINED:    token_data.attr = LLAMA_TOKEN_ATTR_UNDEFINED;    break;
+                default:                            token_data.attr = LLAMA_TOKEN_ATTR_UNDEFINED;    break;
+            }
+        }
     }
     GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
 
@@ -4830,7 +4854,7 @@ static void llm_load_vocab(
     // build special tokens cache
     {
         for (llama_vocab::id id = 0; id < (llama_vocab::id)n_vocab; ++id) {
-            if (vocab.id_to_token[id].type != LLAMA_TOKEN_TYPE_NORMAL) {
+            if (!(vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL)) {
                 vocab.cache_special_tokens.push_back(id);
             }
         }
@@ -4860,6 +4884,59 @@ static void llm_load_vocab(
 
         LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0);
     }
+
+    // Handle per token attributes
+    //NOTE: Each model customizes per token attributes.
+    //NOTE: Per token attributes are missing from the GGUF file.
+    //TODO: Extract attributes from GGUF file.
+    {
+        auto _contains_any = [] (const std::string &str, const std::vector<std::string> &substrs) -> bool {
+            for (auto substr : substrs) {
+                if (str.find(substr) < std::string::npos) {
+                    return true;
+                }
+            }
+            return false;
+        };
+
+        auto _set_tokenid_attr = [&] (const llama_vocab::id id, llama_token_attr attr, bool value) {
+            uint32_t current = vocab.id_to_token.at(id).attr;
+            current = value ? (current | attr) : (current & ~attr);
+            vocab.id_to_token[id].attr = (llama_token_attr) current;
+        };
+
+        auto _set_token_attr = [&] (const std::string & token, llama_token_attr attr, bool value) {
+            _set_tokenid_attr(vocab.token_to_id.at(token), attr, value);
+        };
+
+        std::string model_name;
+        std::string tokenizer_pre;
+
+        ml.get_key(LLM_KV_GENERAL_NAME, model_name, false);
+        ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false);
+
+        // model name to lowercase
+        std::transform(model_name.begin(), model_name.end(), model_name.begin(),
+            [] (const std::string::value_type x) {
+                return std::tolower(x);
+            }
+        );
+
+        // set attributes by model/tokenizer name
+        if (_contains_any(tokenizer_pre, {"jina-v2-es", "jina-v2-de"})) {
+            _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
+        } else if (_contains_any(model_name, {"phi-3", "phi3"})) {
+            for (auto id : vocab.cache_special_tokens) {
+                _set_tokenid_attr(id, LLAMA_TOKEN_ATTR_RSTRIP, true);
+            }
+            for (auto token : {"</s>"}) {
+                _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, true);
+            }
+            for (auto token : {"<unk>", "<s>", "<|endoftext|>"}) {
+                _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, false);
+            }
+        }
+    }
 }
 
 static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
@@ -12609,27 +12686,27 @@ static enum llama_vocab_type llama_vocab_get_type(const llama_vocab & vocab) {
 
 static bool llama_is_normal_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL;
 }
 
 static bool llama_is_unknown_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_UNKNOWN;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNKNOWN;
 }
 
 static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_CONTROL;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_CONTROL;
 }
 
 static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_BYTE;
 }
 
 static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_USER_DEFINED;
 }
 
 static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
@@ -13247,7 +13324,8 @@ struct fragment_buffer_variant {
 static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer) {
     // for each special token
     for (const llama_vocab::id special_id : vocab.cache_special_tokens) {
-        const auto & special_token = vocab.id_to_token[special_id].text;
+        const auto & data = vocab.id_to_token[special_id];
+        const auto & special_token = data.text;
 
         // for each text fragment
         std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin();
@@ -13284,13 +13362,22 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
                     if (match > raw_text_base_offset) {
                         // left
                         const int64_t left_reminder_offset = raw_text_base_offset + 0;
-                        const int64_t left_reminder_length = match - raw_text_base_offset;
-                        buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length);
+                        int64_t left_reminder_length = match - raw_text_base_offset;
+
+                        if (data.attr & LLAMA_TOKEN_ATTR_LSTRIP) {
+                            while (left_reminder_length > 0 && isspace(raw_text[left_reminder_offset + left_reminder_length - 1])) {
+                                left_reminder_length--;
+                            }
+                        }
+
+                        if (left_reminder_length > 0) {
+                            buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length);
+                            it++;
+                        }
 
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str());
 #endif
-                        it++;
                     }
 
                     // special token
@@ -13299,16 +13386,25 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
                     // right
                     if (match + special_token.length() < raw_text_base_offset + raw_text_base_length) {
-                        const int64_t right_reminder_offset = match + special_token.length();
-                        const int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length());
-                        buffer.emplace_after(it, raw_text, right_reminder_offset, right_reminder_length);
+                        int64_t right_reminder_offset = match + special_token.length();
+                        int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length());
+
+                        if (data.attr & LLAMA_TOKEN_ATTR_RSTRIP) {
+                            while (right_reminder_length > 0 && isspace(raw_text[right_reminder_offset])) {
+                                right_reminder_offset++;
+                                right_reminder_length--;
+                            }
+                        }
+
+                        if (right_reminder_length > 0) {
+                            buffer.emplace_after(it, raw_text, right_reminder_offset, right_reminder_length);
+                            it++;
+                        }
 
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("FR: (%ld %ld) '%s'\n", right_reminder_offset, right_reminder_length, raw_text->substr(right_reminder_offset, right_reminder_length).c_str());
 #endif
 
-                        it++;
-
                         if (source == 0) {
                             buffer.erase_after(buffer.before_begin());
                         } else {
@@ -13354,9 +13450,7 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                 // tokenizer.encode('', add_special_tokens=True)  returns [1]
                 // tokenizer.encode('', add_special_tokens=False) returns []
 
-                static const bool rtrim = true;  //TODO: as param
                 bool is_prev_special = false;
-                bool special_token_rtrim = false;
 
                 if (add_special && vocab.special_add_bos != 0) {
                     GGML_ASSERT(vocab.special_bos_id != -1);
@@ -13366,25 +13460,8 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
-                        // without adding this leading whitespace, we do not get the same results as the original tokenizer
-
-                        // TODO: It's likely possible to get rid of this string copy entirely
-                        //  by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer
-                        //  and passing 'add space prefix' as bool argument
-                        //
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
 
-                        if (special_token_rtrim) {
-                            size_t num_whitespaces = 0;
-                            while (isspace(raw_text[num_whitespaces])) {
-                                num_whitespaces++;
-                            }
-                            if (num_whitespaces == raw_text.size()) {
-                                continue; // skip if all whitespaces
-                            }
-                            raw_text = raw_text.substr(num_whitespaces);
-                        }
-
                         if (vocab.add_space_prefix) {
                             if (!output.size() || is_prev_special) {  // prefix with space if first token
                                 raw_text = " " + raw_text;
@@ -13400,11 +13477,6 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                         is_prev_special = true;
-                        // phi-3 special tokens without rtrim, works fine for llama-spm too
-                        special_token_rtrim = rtrim
-                            && fragment.token != vocab.special_bos_id
-                            && fragment.token != vocab.special_unk_id
-                            && fragment.token != vocab.special_eos_id;
                     }
                 }
 
@@ -16160,7 +16232,7 @@ struct llama_model * llama_load_model_from_file(
             return true;
         };
     }
-    if (params.rpc_servers != nullptr) {
+    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
         // split the servers set them into model->rpc_servers
         std::string servers(params.rpc_servers);
         size_t pos = 0;
@@ -16323,17 +16395,7 @@ struct llama_context * llama_new_context_with_model(
 
     if (!hparams.vocab_only) {
         // initialize backends
-#if defined(GGML_USE_RPC)
-        for (auto & server : model->rpc_servers) {
-            ggml_backend_t backend = ggml_backend_rpc_init(server.c_str());
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to connect RPC backend to %s\n", __func__, server.c_str());
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(backend);
-        }
-#elif defined(GGML_USE_METAL)
+#if defined(GGML_USE_METAL)
         if (model->n_gpu_layers > 0) {
             ctx->backend_metal = ggml_backend_metal_init();
             if (ctx->backend_metal == nullptr) {
@@ -16425,6 +16487,19 @@ struct llama_context * llama_new_context_with_model(
             }
             ctx->backends.push_back(backend);
         }
+#endif
+#if defined(GGML_USE_RPC)
+        if (model->n_gpu_layers > 0) {
+            for (const auto & endpoint : model->rpc_servers) {
+                ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
 #endif
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {
@@ -18207,9 +18282,9 @@ float llama_token_get_score(const struct llama_model * model, llama_token token)
     return model->vocab.id_to_token[token].score;
 }
 
-llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) {
+llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token) {
     GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return model->vocab.id_to_token[token].type;
+    return model->vocab.id_to_token[token].attr;
 }
 
 bool llama_token_is_eog(const struct llama_model * model, llama_token token) {

llama.h

@@ -97,7 +97,7 @@ extern "C" {
         LLAMA_ROPE_TYPE_GLM  =  4,
     };
 
-    enum llama_token_type {
+    enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
         LLAMA_TOKEN_TYPE_UNDEFINED    = 0,
         LLAMA_TOKEN_TYPE_NORMAL       = 1,
         LLAMA_TOKEN_TYPE_UNKNOWN      = 2,
@@ -107,6 +107,20 @@ extern "C" {
         LLAMA_TOKEN_TYPE_BYTE         = 6,
     };
 
+    enum llama_token_attr {
+        LLAMA_TOKEN_ATTR_UNDEFINED    = 0,
+        LLAMA_TOKEN_ATTR_UNKNOWN      = 1 <<  1,
+        LLAMA_TOKEN_ATTR_UNUSED       = 1 <<  2,
+        LLAMA_TOKEN_ATTR_NORMAL       = 1 <<  3,
+        LLAMA_TOKEN_ATTR_CONTROL      = 1 <<  4,  // SPECIAL?
+        LLAMA_TOKEN_ATTR_USER_DEFINED = 1 <<  5,
+        LLAMA_TOKEN_ATTR_BYTE         = 1 <<  6,
+        LLAMA_TOKEN_ATTR_NORMALIZED   = 1 <<  7,
+        LLAMA_TOKEN_ATTR_LSTRIP       = 1 <<  8,
+        LLAMA_TOKEN_ATTR_RSTRIP       = 1 <<  9,
+        LLAMA_TOKEN_ATTR_SINGLE_WORD  = 1 << 10,
+    };
+
     // model file types
     enum llama_ftype {
         LLAMA_FTYPE_ALL_F32              = 0,
@@ -821,7 +835,7 @@ extern "C" {
 
     LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
 
-    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
+    LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token);
 
     // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
     LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);

scripts/compare-commits.sh

@@ -10,16 +10,18 @@ set -x
 
 bench_args="${@:3}"
 
-rm -f llama-bench.sqlite
+rm -f llama-bench.sqlite > /dev/null
 
 # to test a backend, call the script with the corresponding environment variable (e.g. LLAMA_CUDA=1 ./scripts/compare-commits.sh ...)
 
-git checkout $1
-make clean && make -j32 $make_opts llama-bench
-./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
+git checkout $1 > /dev/null
+make clean > /dev/null
+make -j$(nproc) $make_opts llama-bench > /dev/null
+./llama-bench -o sql -oe md $bench_args | sqlite3 llama-bench.sqlite
 
-git checkout $2
-make clean && make -j32 $make_opts llama-bench
-./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
+git checkout $2 > /dev/null
+make clean > /dev/null
+make -j$(nproc) $make_opts llama-bench > /dev/null
+./llama-bench -o sql -oe md $bench_args | sqlite3 llama-bench.sqlite
 
 ./scripts/compare-llama-bench.py -b $1 -c $2

tests/test-tokenizer-random.py

@@ -156,17 +156,39 @@ def generator_custom_text_edge_cases() -> Iterator[str]:
         '<s>a',       # Phi-3 fail
         '<unk><|endoftext|><s>',  # Phi-3 fail
         'a\na',       # TODO: Bert fail
+        'a </s> b',   # rstrip phi-3
+        'a <mask> b', # lstrip jina-v2
     ]
 
 
-def generator_random_special_tokens(tokenizer, iterations=100) -> Iterator[str]:
-    special_tokens = set(tokenizer.all_special_tokens)
-    special_tokens.update([" ", "\n", "\t", "-", "!", "one", "1", "<s>", "</s>"])
-    special_tokens = list(sorted(special_tokens))
+def generator_vocab_words(vocab: list[str]) -> Iterator[str]:
+    """Brute force check all vocab words"""
+    yield from vocab
+
+
+def generator_added_lr_strip(tokenizer) -> Iterator[str]:
+    WHITESPACES = ["", " ", "  ", "    "]
+    special_tokens = list(tokenizer.all_special_tokens)
+    added_tokens   = list(tokenizer.added_tokens_encoder)
+    all_tokens     = list(sorted(set(special_tokens + added_tokens)))
+    for token in all_tokens:
+        for lstrip in WHITESPACES:
+            for rstrip in WHITESPACES:
+                yield lstrip + token + rstrip
+                yield "a" + lstrip + token + rstrip
+                yield lstrip + token + rstrip + "z"
+                yield "a" + lstrip + token + rstrip + "z"
+
+
+def generator_random_added_tokens(tokenizer, iterations=100) -> Iterator[str]:
+    special_tokens = list(tokenizer.all_special_tokens)
+    added_tokens   = list(tokenizer.added_tokens_encoder)
+    separations    = [" ", "\n", "\t", "-", "!", "one", "1", "<s>", "</s>"]
+    all_tokens     = list(sorted(set(special_tokens + added_tokens + separations)))
     rand = random.Random()
     for m in range(iterations):
         rand.seed(m)
-        words = rand.choices(special_tokens, k=500)
+        words = rand.choices(all_tokens, k=500)
         if words[0] == tokenizer.bos_token:  # skip spam warning of double BOS
             while len(words) > 1 and words[1] == tokenizer.bos_token:  # leave one starting BOS
                 words.pop(0)
@@ -175,11 +197,6 @@ def generator_random_special_tokens(tokenizer, iterations=100) -> Iterator[str]:
         yield "".join(words)
 
 
-def generator_vocab_words(vocab: list[str]) -> Iterator[str]:
-    """Brute force check all vocab words"""
-    yield from vocab
-
-
 def generator_random_chars(iterations=100) -> Iterator[str]:
     """Brute force random text with simple characters"""
 
@@ -274,8 +291,8 @@ def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, g
         ids2 = func_tokenize2(text)
         if ids1 != ids2:
             i = find_first_mismatch(ids1, ids2)
-            ids1 = list(ids1)[max(0, i - 2) : i + 2 + 1]
-            ids2 = list(ids2)[max(0, i - 2) : i + 2 + 1]
+            ids1 = list(ids1)[max(0, i - 2) : i + 5 + 1]
+            ids2 = list(ids2)[max(0, i - 2) : i + 5 + 1]
             logger.info(" TokenIDs: " + str(ids1))
             logger.info(" Expected: " + str(ids2))
             raise Exception()
@@ -309,8 +326,9 @@ def main(argv: list[str] = None):
     vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_special_tokens(tokenizer, 10_000))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_added_lr_strip(tokenizer))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_added_tokens(tokenizer, 10_000))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000))
@@ -322,14 +340,14 @@ def main(argv: list[str] = None):
 if __name__ == "__main__":
     # main()
 
-    path_tokenizers = "./models/tokenizers/"
+    path_tokenizers   = "./models/tokenizers/"
     path_vocab_format = "./models/ggml-vocab-%s.gguf"
 
     # import os
     # tokenizers = os.listdir(path_tokenizers)
     tokenizers = [
-        # "llama-spm",   # SPM
-        # "phi-3",       # SPM
+        "llama-spm",   # SPM
+        "phi-3",       # SPM
         "jina-v2-en",  # WPM
         "bert-bge",    # WPM
     ]