Merge branch 'master' into concedo

# Conflicts:
#	.github/workflows/build.yml
#	README.md

Makefile

@@ -128,9 +128,9 @@ ifdef LLAMA_CUBLAS
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
-	NVCCFLAGS = --forward-unknown-to-host-linker -arch=native
+	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg

examples/CMakeLists.txt

@@ -34,4 +34,5 @@ else()
     add_subdirectory(quantize-stats)
     add_subdirectory(perplexity)
     add_subdirectory(embedding)
+    add_subdirectory(save-load-state)
 endif()

examples/common.cpp

@@ -156,10 +156,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive = true;
         } else if (arg == "--embedding") {
             params.embedding = true;
-        } else if (arg == "--interactive-start") {
-            params.interactive = true;
         } else if (arg == "--interactive-first") {
-            params.interactive_start = true;
+            params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
         } else if (arg == "--color") {

examples/common.h

@@ -43,7 +43,7 @@ struct gpt_params {
     bool interactive       = false; // interactive mode
 
     bool embedding         = false; // get only sentence embedding
-    bool interactive_start = false; // wait for user input immediately
+    bool interactive_first = false; // wait for user input immediately
 
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool ignore_eos        = false; // do not stop generating after eos

examples/main/README.md

@@ -21,12 +21,20 @@ To get started right away, run the following command, making sure to use the cor
 ./main -m models/7B/ggml-model.bin --prompt "Once upon a time"
 ```
 
+The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it):
+
+```bash
+./main -m models/7B/ggml-model.bin --ignore-eos --n_predict -1 --keep -1 --prompt "Once upon a time"
+```
+
 For an interactive experience, try this command:
 
 ```bash
 ./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " --prompt $'User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:'
 ```
 
+Note that the newline characters in the prompt string above only work on Linux. On Windows, you will have to use the ``--file`` option (see below) to load a multi-line prompt from file instead.
+
 ## Common Options
 
 In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:
@@ -84,6 +92,8 @@ Instruction mode is particularly useful when working with Alpaca models, which a
 
 -   `-ins, --instruct`: Enable instruction mode to leverage the capabilities of Alpaca models in completing tasks based on user-provided instructions.
 
+Technical detail: the user's input is internally prefixed with the reverse prompt (or ``### Instruction:`` as the default), and followed by ``### Response:`` (except if you just press Return without any input, to keep generating a longer response).
+
 By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
 
 ## Context Management
@@ -114,7 +124,7 @@ The following options are related to controlling the text generation process, in
 
 The `--n_predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
 
-It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n_predict` value.
+It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n_predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the ``--ignore-eos`` parameter.
 
 ### RNG Seed
 
@@ -126,7 +136,7 @@ The RNG seed is used to initialize the random number generator that influences t
 
 -   `--temp N`: Adjust the randomness of the generated text (default: 0.8).
 
-Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism.
+Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.
 
 Example usage: `--temp 0.8`
 

examples/main/main.cpp

@@ -178,12 +178,12 @@ int main(int argc, char ** argv) {
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
-        params.interactive_start = true;
+        params.interactive_first = true;
         params.antiprompt.push_back("### Instruction:\n\n");
     }
 
     // enable interactive mode if reverse prompt or interactive start is specified
-    if (params.antiprompt.size() != 0 || params.interactive_start) {
+    if (params.antiprompt.size() != 0 || params.interactive_first) {
         params.interactive = true;
     }
 
@@ -246,7 +246,7 @@ int main(int argc, char ** argv) {
 #endif
                " - Press Return to return control to LLaMa.\n"
                " - If you want to submit another line, end your input in '\\'.\n\n");
-        is_interacting = params.interactive_start;
+        is_interacting = params.interactive_first;
     }
 
     bool is_antiprompt = false;

examples/save-load-state/CMakeLists.txt

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

examples/save-load-state/save-load-state.cpp

@@ -0,0 +1,128 @@
+#include <vector>
+#include <cstdio>
+#include <chrono>
+
+#include "common.h"
+#include "llama.h"
+#include "llama.cpp"
+
+using namespace std;
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    params.model = "models/llama-7B/ggml-model.bin";
+    params.seed = 42;
+    params.n_threads = 4;
+    params.repeat_last_n = 64;
+    params.prompt = "The quick brown fox";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    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.use_mmap   = params.use_mmap;
+    lparams.use_mlock  = params.use_mlock;
+
+    auto n_past = 0;
+    auto last_n_tokens_data = vector<llama_token>(params.repeat_last_n, 0);
+
+    // init
+    auto ctx = llama_init_from_file(params.model.c_str(), lparams);
+    auto tokens = vector<llama_token>(params.n_ctx);
+    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true);
+
+    if (n_prompt_tokens < 1) {
+        fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
+        return 1;
+    }
+
+    // evaluate prompt
+
+    llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads);
+
+    last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
+    n_past += n_prompt_tokens;
+
+    // Save state (rng, logits, embedding and kv_cache) to file
+    FILE *fp_write = fopen("dump_state.bin", "wb");
+    auto state_size = llama_get_state_size(ctx);
+    auto state_mem = new uint8_t[state_size];
+    llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file
+    fwrite(state_mem, 1, state_size, fp_write);
+    fclose(fp_write);
+
+    // save state (last tokens)
+    auto last_n_tokens_data_saved = vector<llama_token>(last_n_tokens_data);
+    auto n_past_saved = n_past;
+
+    // first run
+    printf("\n%s", params.prompt.c_str());
+    for (auto i = 0; i < params.n_predict; i++) {
+        auto next_token = llama_sample_top_p_top_k(
+            ctx,
+            &last_n_tokens_data.back() - params.repeat_last_n,
+            params.repeat_last_n,
+            40,
+            1.0,
+            1.0,
+            1.1);
+        auto next_token_str = llama_token_to_str(ctx, next_token);
+        last_n_tokens_data.push_back(next_token);
+        printf("%s", next_token_str);
+        if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
+            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            return 1;
+        }
+        n_past += 1;
+    }
+    printf("\n\n");
+
+    // free old model
+    llama_free(ctx);
+
+    // load new model
+
+    auto ctx2 = llama_init_from_file(params.model.c_str(), lparams);
+
+    // Load state (rng, logits, embedding and kv_cache) from file
+    FILE *fp_read = fopen("dump_state.bin", "rb");
+    auto state_size2 = llama_get_state_size(ctx2);
+    if (state_size != state_size2) {
+        fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
+    }
+    fread(state_mem, 1, state_size, fp_read);
+    llama_set_state_data(ctx2, state_mem);  // could also read directly from memory mapped file
+    fclose(fp_read);
+
+    // restore state (last tokens)
+    last_n_tokens_data = last_n_tokens_data_saved;
+    n_past = n_past_saved;
+
+    // second run
+    for (auto i = 0; i < params.n_predict; i++) {
+        auto next_token = llama_sample_top_p_top_k(
+            ctx2,
+            &last_n_tokens_data.back() - params.repeat_last_n,
+            params.repeat_last_n,
+            40,
+            1.0,
+            1.0,
+            1.1);
+        auto next_token_str = llama_token_to_str(ctx2, next_token);
+        last_n_tokens_data.push_back(next_token);
+        printf("%s", next_token_str);
+        if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
+            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            return 1;
+        }
+        n_past += 1;
+    }
+    printf("\n\n");
+    return 0;
+}

ggml.c

@@ -438,7 +438,7 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi)
 {
     // Load 8 bytes from memory
-    __m128i tmp = _mm_loadu_si64( ( const __m128i* )rsi );
+    __m128i tmp = _mm_loadl_epi64( ( const __m128i* )rsi );
 
     // Expand bytes into uint16_t values
     __m128i bytes = _mm_cvtepu8_epi16( tmp );
@@ -6781,15 +6781,20 @@ static void ggml_compute_forward_sum_f32(
     const size_t nb02 = src0->nb[2];
     const size_t nb03 = src0->nb[3];
 
+    ggml_float sum     = 0;
+    float      row_sum = 0;
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = 0; i01 < ne01; i01++) {
                 ggml_vec_sum_f32(ne00,
-                        (float *) (dst->data),
+                        &row_sum,
                         (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03));
+                sum += row_sum;
             }
         }
     }
+    ((float *) dst->data)[0] = sum;
 }
 
 static void ggml_compute_forward_sum(

llama.cpp

@@ -54,7 +54,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
         { MODEL_30B,   512ull * MB },
-        { MODEL_65B,   512ull * MB },
+        { MODEL_65B,  1024ull * MB },
     };
     return _MEM_REQ_SCRATCH0;
 }
@@ -65,7 +65,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
         { MODEL_30B,   512ull * MB },
-        { MODEL_65B,   512ull * MB },
+        { MODEL_65B,  1024ull * MB },
     };
     return _MEM_REQ_SCRATCH1;
 }