Merge branch 'ggerganov:master' into master

.gitignore

@@ -19,6 +19,7 @@ models/*
 /main
 /quantize
 /result
+/perplexity
 
 arm_neon.h
 compile_commands.json

CMakeLists.txt

@@ -211,17 +211,6 @@ endif()
 # Build libraries
 #
 
-add_library(utils OBJECT
-            utils.cpp
-            utils.h)
-
-target_include_directories(utils PUBLIC .)
-target_compile_features(utils PUBLIC cxx_std_11) # don't bump
-target_link_libraries(utils PRIVATE ${LLAMA_EXTRA_LIBS})
-if (BUILD_SHARED_LIBS)
-    set_target_properties(utils PROPERTIES POSITION_INDEPENDENT_CODE ON)
-endif()
-
 add_library(ggml OBJECT
             ggml.c
             ggml.h)
@@ -239,22 +228,12 @@ add_library(llama
 
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
-target_link_libraries(llama PRIVATE utils ggml ${LLAMA_EXTRA_LIBS})
+target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
 endif()
 
-#
-# Executables
-#
-
-add_executable(main main.cpp)
-target_link_libraries(main PRIVATE llama ggml utils)
-
-add_executable(quantize quantize.cpp)
-target_link_libraries(quantize PRIVATE llama ggml utils)
-
 #
 # programs, examples and tests
 #
@@ -264,6 +243,6 @@ if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     add_subdirectory(tests)
 endif ()
 
-#if (LLAMA_BUILD_EXAMPLES)
+if (LLAMA_BUILD_EXAMPLES)
-#    add_subdirectory(examples)
+    add_subdirectory(examples)
-#endif()
+endif()

Makefile

@@ -212,7 +212,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize
+default: main quantize perplexity
 
 #
 # Build library
@@ -224,20 +224,23 @@ ggml.o: ggml.c ggml.h
 llama.o: llama.cpp llama.h
 	$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
 
-utils.o: utils.cpp utils.h
+common.o: examples/common.cpp examples/common.h
-	$(CXX) $(CXXFLAGS) -c utils.cpp -o utils.o
+	$(CXX) $(CXXFLAGS) -c examples/common.cpp -o common.o
 
 clean:
-	rm -f *.o main quantize
+	rm -vf *.o main quantize perplexity
 
-main: main.cpp ggml.o llama.o utils.o
+main: examples/main/main.cpp ggml.o llama.o common.o
-	$(CXX) $(CXXFLAGS) main.cpp ggml.o llama.o utils.o -o main $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: quantize.cpp ggml.o llama.o utils.o
+quantize: examples/quantize/quantize.cpp ggml.o llama.o
-	$(CXX) $(CXXFLAGS) quantize.cpp ggml.o llama.o utils.o -o quantize $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp ggml.o llama.o -o quantize $(LDFLAGS)
+
+perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
+	$(CXX) $(CXXFLAGS) examples/perplexity/perplexity.cpp ggml.o llama.o common.o -o perplexity $(LDFLAGS)
 
 #
 # Tests

README.md

@@ -179,7 +179,10 @@ Here is an example few-shot interaction, invoked with the command
 
 ```bash
 # default arguments using 7B model
-./chat.sh
+./examples/chat.sh
+
+# advanced chat with 13B model
+./examples/chat-13B.sh
 
 # custom arguments using 13B model
 ./main -m ./models/13B/ggml-model-q4_0.bin -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
@@ -195,7 +198,7 @@ Note the use of `--color` to distinguish between user input and generated text.
 2. Run the `main` tool like this:
 
 ```
-./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt -ins
+./examples/alpaca.sh
 ```
 
 Sample run:

chat.sh

@@ -1,6 +0,0 @@
-#!/bin/bash
-#
-# Temporary script - will be removed in the future
-#
-
-./main -m ./models/7B/ggml-model-q4_0.bin -b 128 -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt

examples/CMakeLists.txt

@@ -0,0 +1,36 @@
+# dependencies
+
+find_package(Threads REQUIRED)
+
+# third-party
+
+# ...
+
+# common
+
+set(TARGET common)
+
+add_library(${TARGET} OBJECT
+    common.h
+    common.cpp
+    )
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+target_include_directories(${TARGET} PUBLIC .)
+target_compile_features(${TARGET} PUBLIC cxx_std_11)
+target_link_libraries(${TARGET} PRIVATE llama)
+
+# examples
+
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
+if (EMSCRIPTEN)
+else()
+    add_subdirectory(main)
+    add_subdirectory(quantize)
+    add_subdirectory(perplexity)
+    add_subdirectory(embedding)
+endif()

examples/alpaca.sh

@@ -1,6 +1,10 @@
 #!/bin/bash
+
 #
 # Temporary script - will be removed in the future
 #
 
+cd `dirname $0`
+cd ..
+
 ./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7

examples/chat.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+# Important:
+#
+#   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
+#
+./main -m ./models/7B/ggml-model-q4_0.bin -c 512 -b 1024 -n 256 --keep 48 \
+    --repeat_penalty 1.0 --color -i \
+    -r "User:" -f prompts/chat-with-bob.txt

examples/common.cpp

@@ -1,6 +1,6 @@
-#include "ggml.h"
+#include "common.h"
 
-#include "utils.h"
+#include "ggml.h"
 
 #include <cassert>
 #include <cstring>
@@ -112,6 +112,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
             params.n_batch = std::stoi(argv[i]);
             params.n_batch = std::min(512, params.n_batch);
+        } else if (arg == "--keep") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_keep = std::stoi(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -134,7 +140,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.use_mlock = true;
         } else if (arg == "--mtest") {
             params.mem_test = true;
-        } else if (arg == "--verbose_prompt") {
+        } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
             if (++i >= argc) {
@@ -198,7 +204,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
     fprintf(stderr, "  -f FNAME, --file FNAME\n");
     fprintf(stderr, "                        prompt file to start generation.\n");
-    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d, -1 - infinity)\n", params.n_predict);
     fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
     fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
     fprintf(stderr, "  --repeat_last_n N     last n tokens to consider for penalize (default: %d)\n", params.repeat_last_n);
@@ -210,6 +216,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --n_parts N           number of model parts (default: -1 = determine from dimensions)\n");
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
+    fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt\n");
     if (ggml_mlock_supported()) {
         fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }

examples/common.h

@@ -21,6 +21,7 @@ struct gpt_params {
     int32_t n_parts       = -1;   // amount of model parts (-1 = determine from model dimensions)
     int32_t n_ctx         = 512;  // context size
     int32_t n_batch       = 8;    // batch size for prompt processing
+    int32_t n_keep        = 0;    // number of tokens to keep from initial prompt
 
     // sampling parameters
     int32_t top_k = 40;

examples/embedding/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET embedding)
+add_executable(${TARGET} embedding.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/embedding/README.md

@@ -0,0 +1,3 @@
+# embedding
+
+TODO

examples/embedding/embedding.cpp

@@ -0,0 +1,101 @@
+#include "common.h"
+#include "llama.h"
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    params.model = "models/llama-7B/ggml-model.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    params.embedding = true;
+
+    if (params.n_ctx > 2048) {
+        fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
+                "expect poor results\n", __func__, params.n_ctx);
+    }
+
+    if (params.seed <= 0) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_context * ctx;
+
+    // load the model
+    {
+        auto lparams = llama_context_default_params();
+
+        lparams.n_ctx      = params.n_ctx;
+        lparams.n_parts    = params.n_parts;
+        lparams.seed       = params.seed;
+        lparams.f16_kv     = params.memory_f16;
+        lparams.logits_all = params.perplexity;
+        lparams.use_mlock  = params.use_mlock;
+        lparams.embedding  = params.embedding;
+
+        ctx = llama_init_from_file(params.model.c_str(), lparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+    }
+
+    int n_past = 0;
+
+    // Add a space in front of the first character to match OG llama tokenizer behavior
+    params.prompt.insert(0, 1, ' ');
+
+    // tokenize the prompt
+    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+
+    // determine newline token
+    auto llama_token_newline = ::llama_tokenize(ctx, "\n", false);
+
+    if (params.verbose_prompt) {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]));
+        }
+        fprintf(stderr, "\n");
+    }
+
+    if (params.embedding){
+        if (embd_inp.size() > 0) {
+            if (llama_eval(ctx, embd_inp.data(), embd_inp.size(), n_past, params.n_threads)) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return 1;
+            }
+        }
+
+        const int n_embd = llama_n_embd(ctx);
+        const auto embeddings = llama_get_embeddings(ctx);
+
+        for (int i = 0; i < n_embd; i++) {
+            printf("%f ", embeddings[i]);
+        }
+        printf("\n");
+    }
+
+    llama_print_timings(ctx);
+    llama_free(ctx);
+
+    return 0;
+}

examples/main/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET main)
+add_executable(${TARGET} main.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/main/README.md

@@ -0,0 +1,3 @@
+# main
+
+TODO

examples/main/main.cpp

@@ -1,5 +1,4 @@
-#include "utils.h"
+#include "common.h"
-#include "ggml.h"
 #include "llama.h"
 
 #include <cassert>
@@ -45,8 +44,18 @@ enum console_state {
 static console_state con_st = CONSOLE_STATE_DEFAULT;
 static bool con_use_color = false;
 
-void set_console_state(console_state new_st)
+void enable_console_colors() {
-{
+#if defined (_WIN32)
+    // Enable ANSI colors on Windows 10+
+    unsigned long dwMode = 0;
+    void* hConOut = GetStdHandle((unsigned long)-11); // STD_OUTPUT_HANDLE (-11)
+    if (hConOut && hConOut != (void*)-1 && GetConsoleMode(hConOut, &dwMode) && !(dwMode & 0x4)) {
+        SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4)
+    }
+#endif
+}
+
+void set_console_state(console_state new_st) {
     if (!con_use_color) return;
     // only emit color code if state changed
     if (new_st != con_st) {
@@ -65,79 +74,6 @@ void set_console_state(console_state new_st)
     }
 }
 
-std::vector<double> softmax(const std::vector<float>& logits) {
-    std::vector<double> probs(logits.size());
-    float max_logit = logits[0];
-    for (float v : logits) max_logit = std::max(max_logit, v);
-    double sum_exp = 0.0;
-    for (size_t i = 0; i < logits.size(); i++) {
-        // Subtract the maximum logit value from the current logit value for numerical stability
-        float logit = logits[i] - max_logit;
-        double exp_logit = std::exp(logit);
-        sum_exp += exp_logit;
-        probs[i] = exp_logit;
-    }
-    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
-    return probs;
-}
-
-void perplexity(llama_context * ctx, const gpt_params & params) {
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
-    // Run `./main --perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
-    // Output: `perplexity: 13.5106 [114/114]`
-    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
-
-    int count = 0;
-    double nll = 0.0;
-    int seq_count = tokens.size() / params.n_ctx;
-
-    fprintf(stderr, "%s : calculating perplexity over %d chunks\n", __func__, seq_count);
-
-    for (int i = 0; i < seq_count; ++i) {
-        int start = i * params.n_ctx;
-        int end = start + params.n_ctx - 1;
-        std::vector<llama_token> embd(tokens.begin() + start, tokens.begin() + end);
-        auto start_t = std::chrono::high_resolution_clock::now();
-        if (llama_eval(ctx, embd.data(), embd.size(), 0, params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return;
-        }
-        auto end_t = std::chrono::high_resolution_clock::now();
-        if (i == 0) {
-            double seconds = std::chrono::duration<double>(end_t - start_t).count();
-            printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
-        }
-        // We get the logits for all the tokens in the context window (params.n_ctx)
-        // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
-        // calculate the perplexity over the last half the window (so the model always has
-        // some context to predict the token).
-        //
-        // We rely on the fact that attention in the forward pass only looks at previous
-        // tokens here, so the logits returned for each token are an accurate representation
-        // of what the model would have predicted at that point.
-        //
-        // Example, we have a context window of 512, we will compute perplexity for each of the
-        // last 256 tokens.  Then, we split the input up into context window size chunks to
-        // process the entire prompt.
-
-        auto logits = llama_get_logits(ctx);
-        for (int j = params.n_ctx / 2; j < params.n_ctx - 1; ++j) {
-            // Calculate probability of next token, given the previous ones.
-            int n_vocab = llama_n_vocab(ctx);
-            std::vector<float> tok_logits(
-                logits + j * n_vocab,
-                logits + (j + 1) * n_vocab);
-            double prob = softmax(tok_logits)[tokens[start + j + 1]];
-            nll += -std::log(prob);
-            ++count;
-        }
-        // perplexity is e^(average negative log-likelihood)
-        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
-        fflush(stdout);
-    }
-    printf("\n");
-}
-
 static bool is_interacting = false;
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
@@ -155,9 +91,6 @@ void sigint_handler(int signo) {
 #endif
 
 int main(int argc, char ** argv) {
-    // has to be called once at the start of the program to init ggml stuff
-    ggml_time_init();
-
     gpt_params params;
     params.model = "models/llama-7B/ggml-model.bin";
 
@@ -165,6 +98,22 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    if (params.perplexity) {
+        printf("\n************\n");
+        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.embedding) {
+        printf("\n************\n");
+        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
                 "expect poor results\n", __func__, params.n_ctx);
@@ -198,9 +147,7 @@ int main(int argc, char ** argv) {
         lparams.n_parts    = params.n_parts;
         lparams.seed       = params.seed;
         lparams.f16_kv     = params.memory_f16;
-        lparams.logits_all = params.perplexity;
         lparams.use_mlock  = params.use_mlock;
-        lparams.embedding  = params.embedding;
 
         ctx = llama_init_from_file(params.model.c_str(), lparams);
 
@@ -236,13 +183,6 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    if (params.perplexity) {
-        perplexity(ctx, params);
-        exit(0);
-    }
-
-    int n_past = 0;
-
     // Add a space in front of the first character to match OG llama tokenizer behavior
     params.prompt.insert(0, 1, ' ');
 
@@ -251,7 +191,12 @@ int main(int argc, char ** argv) {
 
     const int n_ctx = llama_n_ctx(ctx);
 
-    params.n_predict = std::min(params.n_predict, n_ctx - (int) embd_inp.size());
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        fprintf(stderr, "%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    params.n_keep    = std::min(params.n_keep,    (int) embd_inp.size());
 
     // prefix & suffix for instruct mode
     const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true);
@@ -282,6 +227,13 @@ int main(int argc, char ** argv) {
         for (int i = 0; i < (int) embd_inp.size(); i++) {
             fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]));
         }
+        if (params.n_keep > 0) {
+        fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                fprintf(stderr, "%s", llama_token_to_str(ctx, embd_inp[i]));
+            }
+            fprintf(stderr, "'\n");
+        }
         fprintf(stderr, "\n");
     }
 
@@ -298,7 +250,7 @@ int main(int argc, char ** argv) {
 
         fprintf(stderr, "%s: interactive mode on.\n", __func__);
 
-        if(params.antiprompt.size()) {
+        if (params.antiprompt.size()) {
             for (auto antiprompt : params.antiprompt) {
                 fprintf(stderr, "Reverse prompt: '%s'\n", antiprompt.c_str());
             }
@@ -308,14 +260,12 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
         }
     }
-    fprintf(stderr, "sampling parameters: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
+    fprintf(stderr, "sampling: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
+    fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     fprintf(stderr, "\n\n");
 
-    std::vector<llama_token> embd;
+    // TODO: replace with ring-buffer
-
+    std::vector<llama_token> last_n_tokens(n_ctx);
-
-    int last_n_size = params.repeat_last_n;
-    std::vector<llama_token> last_n_tokens(last_n_size);
     std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
 
     if (params.interactive) {
@@ -328,48 +278,44 @@ int main(int argc, char ** argv) {
         is_interacting = params.interactive_start || params.instruct;
     }
 
-    int input_consumed = 0;
     bool input_noecho = false;
 
-    int remaining_tokens = params.n_predict;
+    int n_past     = 0;
+    int n_remain   = params.n_predict;
+    int n_consumed = 0;
 
-#if defined (_WIN32)
-  if (params.use_color) {
-        // Enable ANSI colors on Windows 10+
-        unsigned long dwMode = 0;
-        void* hConOut = GetStdHandle((unsigned long)-11); // STD_OUTPUT_HANDLE (-11)
-        if (hConOut && hConOut != (void*)-1 && GetConsoleMode(hConOut, &dwMode) && !(dwMode & 0x4)) {
-            SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4)
-        }
-    }
-#endif
     // the first thing we will do is to output the prompt, so set color accordingly
+    if (params.use_color) {
+        enable_console_colors();
+    }
     set_console_state(CONSOLE_STATE_PROMPT);
 
-    if (params.embedding){
+    std::vector<llama_token> embd;
-        embd = embd_inp;
 
-        if (embd.size() > 0) {
+    while (n_remain != 0 || params.interactive) {
-            if (llama_eval(ctx, embd.data(), embd.size(), n_past, params.n_threads)) {
-                fprintf(stderr, "%s : failed to eval\n", __func__);
-                return 1;
-            }
-        }
-
-        const auto embeddings = llama_get_embeddings(ctx);
-
-        // TODO: print / use the embeddings
-
-        if (params.use_color) {
-            printf(ANSI_COLOR_RESET);
-        }
-
-        return 0;
-    }
-
-    while (remaining_tokens > 0 || params.interactive) {
         // predict
         if (embd.size() > 0) {
+            // infinite text generation via context swapping
+            // if we run out of context:
+            // - take the n_keep first tokens from the original prompt (via n_past)
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch
+            if (n_past + (int) embd.size() > n_ctx) {
+                const int n_left = n_past - params.n_keep;
+
+                n_past = params.n_keep;
+
+                // insert n_left/2 tokens at the start of embd from last_n_tokens
+                embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
+
+                //printf("\n---\n");
+                //printf("resetting: '");
+                //for (int i = 0; i < (int) embd.size(); i++) {
+                //    printf("%s", llama_token_to_str(ctx, embd[i]));
+                //}
+                //printf("'\n");
+                //printf("\n---\n");
+            }
+
             if (llama_eval(ctx, embd.data(), embd.size(), n_past, params.n_threads)) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return 1;
@@ -379,7 +325,7 @@ int main(int argc, char ** argv) {
         n_past += embd.size();
         embd.clear();
 
-        if ((int) embd_inp.size() <= input_consumed && !is_interacting) {
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // out of user input, sample next token
             const float top_k          = params.top_k;
             const float top_p          = params.top_p;
@@ -392,14 +338,12 @@ int main(int argc, char ** argv) {
                 auto logits = llama_get_logits(ctx);
 
                 if (params.ignore_eos) {
-                    // set the logit of the eos token to zero to avoid sampling it
-                    //logits[logits.size() - n_vocab + EOS_TOKEN_ID] = 0;
-                    // TODO: this does not work of params.logits_all == true
-                    assert(params.perplexity == false);
                     logits[llama_token_eos()] = 0;
                 }
 
-                id = llama_sample_top_p_top_k(ctx, last_n_tokens.data(), last_n_tokens.size(), top_k, top_p, temp, repeat_penalty);
+                id = llama_sample_top_p_top_k(ctx,
+                        last_n_tokens.data() + n_ctx - params.repeat_last_n,
+                        params.repeat_last_n, top_k, top_p, temp, repeat_penalty);
 
                 last_n_tokens.erase(last_n_tokens.begin());
                 last_n_tokens.push_back(id);
@@ -422,14 +366,14 @@ int main(int argc, char ** argv) {
             input_noecho = false;
 
             // decrement remaining sampling budget
-            --remaining_tokens;
+            --n_remain;
         } else {
             // some user input remains from prompt or interaction, forward it to processing
-            while ((int) embd_inp.size() > input_consumed) {
+            while ((int) embd_inp.size() > n_consumed) {
-                embd.push_back(embd_inp[input_consumed]);
+                embd.push_back(embd_inp[n_consumed]);
                 last_n_tokens.erase(last_n_tokens.begin());
-                last_n_tokens.push_back(embd_inp[input_consumed]);
+                last_n_tokens.push_back(embd_inp[n_consumed]);
-                ++input_consumed;
+                ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
                 }
@@ -444,13 +388,13 @@ int main(int argc, char ** argv) {
             fflush(stdout);
         }
         // reset color to default if we there is no pending user input
-        if (!input_noecho && (int)embd_inp.size() == input_consumed) {
+        if (!input_noecho && (int)embd_inp.size() == n_consumed) {
             set_console_state(CONSOLE_STATE_DEFAULT);
         }
 
         // in interactive mode, and not currently processing queued inputs;
         // check if we should prompt the user for more
-        if (params.interactive && (int) embd_inp.size() <= input_consumed) {
+        if (params.interactive && (int) embd_inp.size() <= n_consumed) {
             // check for reverse prompt
             std::string last_output;
             for (auto id : last_n_tokens) {
@@ -472,7 +416,7 @@ int main(int argc, char ** argv) {
                 set_console_state(CONSOLE_STATE_USER_INPUT);
 
                 if (params.instruct) {
-                    input_consumed = embd_inp.size();
+                    n_consumed = embd_inp.size();
                     embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
 
                     printf("\n> ");
@@ -506,7 +450,7 @@ int main(int argc, char ** argv) {
                     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
                 }
 
-                remaining_tokens -= line_inp.size();
+                n_remain -= line_inp.size();
 
                 input_noecho = true; // do not echo this again
             }
@@ -527,8 +471,8 @@ int main(int argc, char ** argv) {
         }
 
         // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
-        if (params.interactive && remaining_tokens <= 0) {
+        if (params.interactive && n_remain <= 0) {
-            remaining_tokens = params.n_predict;
+            n_remain = params.n_predict;
             is_interacting = true;
         }
     }

examples/perplexity/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET perplexity)
+add_executable(${TARGET} perplexity.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/perplexity/README.md

@@ -0,0 +1,3 @@
+# perplexity
+
+TODO

examples/perplexity/perplexity.cpp

@@ -0,0 +1,138 @@
+#include "common.h"
+#include "llama.h"
+
+std::vector<double> softmax(const std::vector<float>& logits) {
+    std::vector<double> probs(logits.size());
+    float max_logit = logits[0];
+    for (float v : logits) max_logit = std::max(max_logit, v);
+    double sum_exp = 0.0;
+    for (size_t i = 0; i < logits.size(); i++) {
+        // Subtract the maximum logit value from the current logit value for numerical stability
+        float logit = logits[i] - max_logit;
+        double exp_logit = std::exp(logit);
+        sum_exp += exp_logit;
+        probs[i] = exp_logit;
+    }
+    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
+    return probs;
+}
+
+void perplexity(llama_context * ctx, const gpt_params & params) {
+    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Run `./main --perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+    // Output: `perplexity: 13.5106 [114/114]`
+    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
+
+    int count = 0;
+    double nll = 0.0;
+    int seq_count = tokens.size() / params.n_ctx;
+
+    fprintf(stderr, "%s : calculating perplexity over %d chunks\n", __func__, seq_count);
+
+    for (int i = 0; i < seq_count; ++i) {
+        int start = i * params.n_ctx;
+        int end = start + params.n_ctx - 1;
+        std::vector<llama_token> embd(tokens.begin() + start, tokens.begin() + end);
+        auto start_t = std::chrono::high_resolution_clock::now();
+        if (llama_eval(ctx, embd.data(), embd.size(), 0, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return;
+        }
+        auto end_t = std::chrono::high_resolution_clock::now();
+        if (i == 0) {
+            double seconds = std::chrono::duration<double>(end_t - start_t).count();
+            printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
+        }
+        // We get the logits for all the tokens in the context window (params.n_ctx)
+        // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
+        // calculate the perplexity over the last half the window (so the model always has
+        // some context to predict the token).
+        //
+        // We rely on the fact that attention in the forward pass only looks at previous
+        // tokens here, so the logits returned for each token are an accurate representation
+        // of what the model would have predicted at that point.
+        //
+        // Example, we have a context window of 512, we will compute perplexity for each of the
+        // last 256 tokens.  Then, we split the input up into context window size chunks to
+        // process the entire prompt.
+
+        auto logits = llama_get_logits(ctx);
+        for (int j = params.n_ctx / 2; j < params.n_ctx - 1; ++j) {
+            // Calculate probability of next token, given the previous ones.
+            int n_vocab = llama_n_vocab(ctx);
+            std::vector<float> tok_logits(
+                logits + j * n_vocab,
+                logits + (j + 1) * n_vocab);
+            double prob = softmax(tok_logits)[tokens[start + j + 1]];
+            nll += -std::log(prob);
+            ++count;
+        }
+        // perplexity is e^(average negative log-likelihood)
+        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        fflush(stdout);
+    }
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    params.model = "models/llama-7B/ggml-model.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    params.perplexity = true;
+
+    if (params.n_ctx > 2048) {
+        fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
+                "expect poor results\n", __func__, params.n_ctx);
+    }
+
+    if (params.seed <= 0) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_context * ctx;
+
+    // load the model
+    {
+        auto lparams = llama_context_default_params();
+
+        lparams.n_ctx      = params.n_ctx;
+        lparams.n_parts    = params.n_parts;
+        lparams.seed       = params.seed;
+        lparams.f16_kv     = params.memory_f16;
+        lparams.logits_all = params.perplexity;
+        lparams.use_mlock  = params.use_mlock;
+        lparams.embedding  = params.embedding;
+
+        ctx = llama_init_from_file(params.model.c_str(), lparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+    }
+
+    perplexity(ctx, params);
+
+    llama_print_timings(ctx);
+    llama_free(ctx);
+
+    return 0;
+}

examples/quantize/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET quantize)
+add_executable(${TARGET} quantize.cpp)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize/README.md

@@ -0,0 +1,3 @@
+# quantize
+
+TODO

llama.cpp

@@ -1261,10 +1261,10 @@ static llama_vocab::id llama_sample_top_p_top_k(
         double repeat_penalty) {
     auto & rng = lctx.rng;
 
-    const auto & vocab = lctx.vocab;
+    const int n_logits = lctx.model.hparams.n_vocab;
-    const auto & logits = lctx.logits;
 
-    int n_logits = vocab.id_to_token.size();
+    const auto & logits = lctx.logits;
+    const auto * plogits = logits.data() + logits.size() - n_logits;
 
     std::vector<std::pair<double, llama_vocab::id>> logits_id;
     logits_id.reserve(n_logits);
@@ -1276,13 +1276,13 @@ static llama_vocab::id llama_sample_top_p_top_k(
             // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
             if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) {
                 // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
-                if (logits[i] < 0.0) {
+                if (plogits[i] < 0.0) {
-                    logits_id.push_back(std::make_pair(logits[i]*scale*repeat_penalty, i));
+                    logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
                 } else {
-                    logits_id.push_back(std::make_pair(logits[i]*scale/repeat_penalty, i));
+                    logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
                 }
             } else {
-                logits_id.push_back(std::make_pair(logits[i]*scale, i));
+                logits_id.push_back(std::make_pair(plogits[i]*scale, i));
             }
         }
     }
@@ -1677,6 +1677,8 @@ struct llama_context * llama_init_from_file(
         }
 
         const auto & hparams = ctx->model.hparams;
+
+        // resized during inference
         if (params.logits_all) {
             ctx->logits.reserve(hparams.n_ctx*hparams.n_vocab);
         } else {
@@ -1684,7 +1686,7 @@ struct llama_context * llama_init_from_file(
         }
 
         if (params.embedding){
-            ctx->embedding.reserve(hparams.n_embd);
+            ctx->embedding.resize(hparams.n_embd);
         }
 
         ctx->buf_compute.resize(MEM_REQ_EVAL.at(ctx->model.type));
@@ -1761,6 +1763,10 @@ int llama_n_ctx(struct llama_context * ctx) {
     return ctx->model.hparams.n_ctx;
 }
 
+int llama_n_embd(struct llama_context * ctx) {
+    return ctx->model.hparams.n_embd;
+}
+
 float * llama_get_logits(struct llama_context * ctx) {
     return ctx->logits.data();
 }

llama.h

@@ -109,6 +109,7 @@ extern "C" {
 
     LLAMA_API int llama_n_vocab(struct llama_context * ctx);
     LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (struct llama_context * ctx);
 
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row

tests/CMakeLists.txt

@@ -1,7 +1,7 @@
 function(llama_add_test source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
     add_executable(${TEST_TARGET} ${source})
-    target_link_libraries(${TEST_TARGET} PRIVATE llama ggml utils)
+    target_link_libraries(${TEST_TARGET} PRIVATE llama)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
 endfunction()
 

tests/test-tokenizer-0.cpp

@@ -1,9 +1,9 @@
-#include "utils.h"
 #include "llama.h"
 
 #include <cstdio>
 #include <string>
 #include <map>
+#include <vector>
 
 static const std::map<std::string, std::vector<llama_token>> k_tests = {
     { "Hello World",        { 1,  10994,   2787, }, },
@@ -48,7 +48,9 @@ int main(int argc, char **argv) {
     }
 
     for (const auto & test_kv : k_tests) {
-        const auto res = ::llama_tokenize(ctx, test_kv.first, true);
+        std::vector<llama_token> res(test_kv.first.size());
+        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true);
+        res.resize(n);
 
         bool correct = res.size() == test_kv.second.size();