Merge pull request #5 from l3utterfly/master

merged upstream

CMakeLists.txt

@@ -118,6 +118,7 @@ option(LLAMA_SYCL                            "llama: use SYCL"
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
 option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
+set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeline parallelism")
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
@@ -147,6 +148,8 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -1085,6 +1088,8 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
+            unicode.h
+            unicode.cpp
             )
 
 target_include_directories(llama PUBLIC .)

Makefile

@@ -167,6 +167,10 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
+ifdef LLAMA_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
+endif
+
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
@@ -633,9 +637,12 @@ ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
+unicode.o: unicode.cpp unicode.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o
 
-llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
+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

Package.swift

@@ -31,6 +31,7 @@ let package = Package(
             sources: [
                 "ggml.c",
                 "llama.cpp",
+                "unicode.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",

README.md

@@ -10,12 +10,14 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Recent API changes
 
-- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
+- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
+- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
 - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
 
 ### Hot topics
 
+- Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
 - Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
 - Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
 - Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
@@ -902,6 +904,9 @@ First, install the essential packages for termux:
 pkg install clang wget git cmake
 ```
 Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
+
+You can execute the following commands on your computer to avoid downloading the NDK to your mobile. Of course, you can also do this in Termux.
+
 ```
 $ mkdir build-android
 $ cd build-android
@@ -910,7 +915,28 @@ $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROI
 $ make
 ```
 Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
-Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+```
+
+Here is a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 

build.zig

@@ -115,6 +115,7 @@ pub fn build(b: *std.build.Builder) !void {
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
     const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
+    const unicode = make.obj("unicode", "unicode.cpp");
     const llama = make.obj("llama", "llama.cpp");
     const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
@@ -125,14 +126,14 @@ pub fn build(b: *std.build.Builder) !void {
     const clip = make.obj("clip", "examples/llava/clip.cpp");
     const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

common/common.cpp

@@ -483,6 +483,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_batch = std::stoi(argv[i]);
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_ubatch = std::stoi(argv[i]);
         } else if (arg == "--keep") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -977,7 +983,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        binary file containing multiple choice tasks.\n");
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
-    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  -b N, --batch-size N  logical maximum batch size (default: %d)\n", params.n_batch);
+    printf("  -ub N, --ubatch-size N\n");
+    printf("                        physical maximum batch size (default: %d)\n", params.n_ubatch);
     printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
     printf("                        (default: %s)\n", sampler_type_names.c_str());
     printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
@@ -1287,8 +1295,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     auto cparams = llama_context_default_params();
 
     cparams.n_ctx             = params.n_ctx;
+    cparams.n_seq_max         = params.n_parallel;
     cparams.n_batch           = params.n_batch;
-    cparams.n_parallel        = params.n_parallel;
+    cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.n_threads;
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
     cparams.seed              = params.seed;
@@ -1379,6 +1388,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
         llama_kv_cache_clear(lctx);
+        llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }
 
@@ -1786,17 +1796,17 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
     static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
 
     printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
 
     llama_kv_cache_view_cell * c_curr = view.cells;
     llama_seq_id * cs_curr = view.cells_sequences;
 
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
         if (i % row_size == 0) {
             printf("\n%5d: ", i);
         }
         int seq_count = 0;
-        for (int j = 0; j < view.n_max_seq; j++) {
+        for (int j = 0; j < view.n_seq_max; j++) {
             if (cs_curr[j] >= 0) { seq_count++; }
         }
         putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
@@ -1809,14 +1819,14 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
     static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
 
     printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
 
     std::unordered_map<llama_seq_id, size_t> seqs;
     llama_kv_cache_view_cell * c_curr = view.cells;
     llama_seq_id * cs_curr = view.cells_sequences;
 
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
-        for (int j = 0; j < view.n_max_seq; j++) {
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        for (int j = 0; j < view.n_seq_max; j++) {
             if (cs_curr[j] < 0) { continue; }
             if (seqs.find(cs_curr[j]) == seqs.end()) {
                 if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
@@ -1835,11 +1845,11 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
 
     c_curr = view.cells;
     cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
         if (i % row_size == 0) {
             printf("\n%5d: ", i);
         }
-        for (int j = 0; j < view.n_max_seq; j++) {
+        for (int j = 0; j < view.n_seq_max; j++) {
             if (cs_curr[j] >= 0) {
                 const auto & it = seqs.find(cs_curr[j]);
                 putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');

common/common.h

@@ -51,7 +51,8 @@ struct gpt_params {
     int32_t n_threads_batch_draft = -1;
     int32_t n_predict             = -1;    // new tokens to predict
     int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
     int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
     int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)

common/sampling.cpp

@@ -17,6 +17,13 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
             return nullptr;
         }
 
+        // Ensure that there is a "root" node.
+        if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
+            fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
+            delete result;
+            return nullptr;
+        }
+
         std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
 
         result->grammar = llama_grammar_init(

examples/batched-bench/batched-bench.cpp

@@ -106,7 +106,7 @@ int main(int argc, char ** argv) {
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
     // ensure enough sequences are available
-    ctx_params.n_parallel = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -138,6 +138,8 @@ int main(int argc, char ** argv) {
                 LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
             }
+
+            llama_synchronize(ctx);
         }
 
         return true;

examples/batched/batched.cpp

@@ -80,7 +80,7 @@ int main(int argc, char ** argv) {
     ctx_params.seed  = 1234;
     ctx_params.n_ctx = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_parallel      = n_parallel;
+    ctx_params.n_seq_max       = n_parallel;
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 

examples/embedding/embedding.cpp

@@ -107,7 +107,7 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch == params.n_ctx);
+    GGML_ASSERT(params.n_batch >= params.n_ctx);
 
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;