whisper : add integer quantization support (#540)

* whisper : add integer quantization support

* examples : add common-ggml + prepare to add "quantize" tool

* whisper : quantization tool ready

* whisper : fix F32 support

* whisper : try to fix shared lib linkage

* wasm : update quantized models to Q5

* bench.wasm : remove "medium" button

* bench.wasm : fix custom model button

* ggml : add Q5_0 and Q5_1 WASM SIMD

* wasm : add quantized models to all WASM examples

* wasm : bump DB version number to 2

* talk-llama : update example to latest llama.cpp

* node : increase test timeout to 10s

* readme : add information for model quantization

* wasm : add links to other examples

examples/talk-llama/llama-util.h

@@ -21,12 +21,17 @@
         #if defined(_POSIX_MAPPED_FILES)
             #include <sys/mman.h>
         #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
     #endif
 #endif
 
 #if defined(_WIN32)
     #define WIN32_LEAN_AND_MEAN
-    #define NOMINMAX
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
     #include <windows.h>
     #include <io.h>
     #include <stdio.h> // for _fseeki64
@@ -41,8 +46,12 @@
     } while (0)
 
 #ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
 __attribute__((format(printf, 1, 2)))
 #endif
+#endif
 static std::string format(const char * fmt, ...) {
     va_list ap, ap2;
     va_start(ap, fmt);
@@ -55,7 +64,7 @@ static std::string format(const char * fmt, ...) {
     va_end(ap2);
     va_end(ap);
     return std::string(buf.data(), size);
-};
+}
 
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
@@ -162,7 +171,7 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file) {
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
@@ -170,15 +179,16 @@ struct llama_mmap {
         flags |= MAP_POPULATE;
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        close(fd);
         if (addr == MAP_FAILED) {
             throw format("mmap failed: %s", strerror(errno));
         }
 
-        // Advise the kernel to preload the mapped memory
-        if (madvise(addr, file->size, MADV_WILLNEED)) {
-            fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
-                    strerror(errno));
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, file->size, MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
         }
     }
 
@@ -188,14 +198,13 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file) {
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
         DWORD error = GetLastError();
-        CloseHandle(hFile);
 
         if (hMapping == NULL) {
             throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
@@ -209,14 +218,20 @@ struct llama_mmap {
             throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
         }
 
-        // Advise the kernel to preload the mapped memory
-        WIN32_MEMORY_RANGE_ENTRY range;
-        range.VirtualAddress = addr;
-        range.NumberOfBytes = (SIZE_T)size;
-        if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-            fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
+        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
         }
+        #else
+        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
+        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
     }
 
     ~llama_mmap() {
@@ -291,8 +306,18 @@ struct llama_mlock {
         if (!mlock(addr, size)) {
             return true;
         } else {
-            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION,
-                    size, this->size, std::strerror(errno));
+            char* errmsg = std::strerror(errno);
+            bool suggest = (errno == ENOMEM);
+
+            // Check if the resource limit is fine after all
+            struct rlimit lock_limit;
+            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
+                suggest = false;
+            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
+                suggest = false;
+
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
             return false;
         }
     }
@@ -338,8 +363,8 @@ struct llama_mlock {
             // Hopefully a megabyte is enough overhead:
             size_t increment = size + 1048576;
             // The minimum must be <= the maximum, so we need to increase both:
-            min_ws_size += size;
-            max_ws_size += size;
+            min_ws_size += increment;
+            max_ws_size += increment;
             if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
                 fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
                         llama_format_win_err(GetLastError()).c_str());
@@ -380,4 +405,29 @@ struct llama_buffer {
         delete[] addr;
     }
 };
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+struct llama_ctx_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    void resize(size_t size) {
+        if (addr) {
+            ggml_cuda_host_free(addr);
+        }
+        addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        this->size = size;
+    }
+
+    ~llama_ctx_buffer() {
+        if (addr) {
+            ggml_cuda_host_free(addr);
+        }
+    }
+};
+#else
+typedef llama_buffer llama_ctx_buffer;
+#endif
+
 #endif

examples/talk-llama/llama.h

@@ -39,12 +39,16 @@ extern "C" {
 
     typedef struct llama_token_data {
         llama_token id;  // token id
-
+        float logit; // log-odds of the token
         float p;     // probability of the token
-        float plog;  // log probability of the token
-
     } llama_token_data;
 
+    typedef struct llama_token_data_array {
+        llama_token_data * data;
+        size_t size;
+        bool sorted;
+    } llama_token_data_array;
+
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     struct llama_context_params {
@@ -65,6 +69,20 @@ extern "C" {
         void * progress_callback_user_data;
     };
 
+    // model file types
+    enum llama_ftype {
+        LLAMA_FTYPE_ALL_F32     = 0,
+        LLAMA_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
+        // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+    };
+
     LLAMA_API struct llama_context_params llama_context_default_params();
 
     LLAMA_API bool llama_mmap_supported();
@@ -82,27 +100,46 @@ extern "C" {
 
     // TODO: not great API - very likely to change
     // Returns 0 on success
+    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-                   int   itype);
+      enum llama_ftype   ftype,
+            int          nthread);
 
-    // Returns the KV cache that will contain the context for the
-    // ongoing prediction with the model.
-    LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
-
-    // Returns the size of the KV cache
-    LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
+    // Apply a LoRA adapter to a loaded model
+    // path_base_model is the path to a higher quality model to use as a base for
+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+    // will be applied on top of the previous one
+    // Returns 0 on success
+    LLAMA_API int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads);
 
     // Returns the number of tokens in the KV cache
     LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
 
-    // Sets the KV cache containing the current context for the model
-    LLAMA_API void llama_set_kv_cache(
-            struct llama_context * ctx,
-                   const uint8_t * kv_cache,
-                          size_t   n_size,
-                             int   n_token_count);
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
+
+    // Returns the size in bytes of the state (rng, logits, embedding and kv_cache)
+    LLAMA_API size_t llama_get_state_size(struct llama_context * ctx);
+
+    // Copies the state to the specified destination address.
+    // Destination needs to have allocated enough memory.
+    // Returns the number of bytes copied
+    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dest);
+
+    // Set the state reading from the specified address
+    // Returns the number of bytes read
+    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
+
+    // Save/load session file
+    LLAMA_API size_t llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
+    LLAMA_API size_t llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
 
     // Run the llama inference to obtain the logits and probabilities for the next token.
     // tokens + n_tokens is the provided batch of new tokens to process
@@ -148,16 +185,52 @@ extern "C" {
     // Special tokens
     LLAMA_API llama_token llama_token_bos();
     LLAMA_API llama_token llama_token_eos();
+    LLAMA_API llama_token llama_token_nl();
 
-    // TODO: improve the last_n_tokens interface ?
-    LLAMA_API llama_token llama_sample_top_p_top_k(
-       struct llama_context * ctx,
-          const llama_token * last_n_tokens_data,
-                        int   last_n_tokens_size,
-                        int   top_k,
-                      float   top_p,
-                      float   temp,
-                      float   repeat_penalty);
+    // Sampling functions
+
+    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
+    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, llama_token * last_tokens, size_t last_tokens_size, float penalty);
+
+    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+
+    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
+    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep = 1);
+
+    /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
+
+    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
+    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep = 1);
+
+    /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
+    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
+    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+
+    /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+
+    /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+
+    /// @details Selects the token with the highest probability.
+    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Randomly selects a token from the candidates based on their probabilities.
+    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
 
     // Performance information
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
@@ -170,4 +243,15 @@ extern "C" {
 }
 #endif
 
+// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
+#ifdef LLAMA_API_INTERNAL
+
+#include <vector>
+#include <string>
+struct ggml_tensor;
+
+std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+
+#endif
+
 #endif // LLAMA_H

examples/talk-llama/llama_internal.h

@@ -1,12 +0,0 @@
-// Internal header to be included by llama.cpp and tests/benchmarks only.
-
-#ifndef LLAMA_INTERNAL_H
-#define LLAMA_INTERNAL_H
-
-#include <vector>
-#include <string>
-struct ggml_tensor;
-
-std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
-
-#endif // LLAMA_INTERNAL_H

examples/talk-llama/talk-llama.cpp

@@ -487,11 +487,37 @@ int main(int argc, char ** argv) {
 
                         {
                             auto logits = llama_get_logits(ctx_llama);
+                            auto n_vocab = llama_n_vocab(ctx_llama);
+
                             logits[llama_token_eos()] = 0;
 
-                            id = llama_sample_top_p_top_k(ctx_llama,
+                            std::vector<llama_token_data> candidates;
+                            candidates.reserve(n_vocab);
+                            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+                            }
+
+                            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+                            // apply repeat penalty
+                            const float nl_logit = logits[llama_token_nl()];
+
+                            llama_sample_repetition_penalty(ctx_llama, &candidates_p,
                                     embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, top_k, top_p, temp, repeat_penalty);
+                                    repeat_last_n, repeat_penalty);
+
+                            logits[llama_token_nl()] = nl_logit;
+
+                            if (temp <= 0) {
+                                // Greedy sampling
+                                id = llama_sample_token_greedy(ctx_llama, &candidates_p);
+                            } else {
+                                // Temperature sampling
+                                llama_sample_top_k(ctx_llama, &candidates_p, top_k);
+                                llama_sample_top_p(ctx_llama, &candidates_p, top_p);
+                                llama_sample_temperature(ctx_llama, &candidates_p, temp);
+                                id = llama_sample_token(ctx_llama, &candidates_p);
+                            }
                         }
 
                         if (id != llama_token_eos()) {