local-llm-rn 1.0.0

package/cpp/src/llama-impl.cpp

@@ -0,0 +1,171 @@
+#include "llama-impl.h"
+
+#include "gguf.h"
+#include "llama.h"
+
+#include <cinttypes>
+#include <climits>
+#include <cstdarg>
+#include <cstring>
+#include <vector>
+#include <sstream>
+
+struct llama_logger_state {
+    ggml_log_callback log_callback = llama_log_callback_default;
+    void * log_callback_user_data = nullptr;
+};
+
+static llama_logger_state g_logger_state;
+
+time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+time_meas::~time_meas() {
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
+    }
+}
+
+void llama_log_get(ggml_log_callback * log_callback, void ** user_data) {
+    ggml_log_get(log_callback, user_data);
+}
+
+void llama_log_set(ggml_log_callback log_callback, void * user_data) {
+    ggml_log_set(log_callback, user_data);
+    g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
+    g_logger_state.log_callback_user_data = user_data;
+}
+
+static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) {
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = new char[len + 1];
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        delete[] buffer2;
+    }
+    va_end(args_copy);
+}
+
+void llama_log_internal(ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    llama_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
+void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+std::string format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
+    for (size_t i = 1; i < ne.size(); i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
+    }
+    return buf;
+}
+
+std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%6" PRId64, t->ne[0]);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, t->ne[i]);
+    }
+    return buf;
+}
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}

package/cpp/src/llama-impl.h

@@ -0,0 +1,73 @@
+#pragma once
+
+#include "ggml.h" // for ggml_log_level
+
+#include <string>
+#include <vector>
+
+#ifdef __GNUC__
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
+#else
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
+#endif
+
+//
+// logging
+//
+
+LLAMA_ATTRIBUTE_FORMAT(2, 3)
+void llama_log_internal        (ggml_log_level level, const char * format, ...);
+void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data);
+
+#define LLAMA_LOG(...)       llama_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
+#define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
+#define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
+#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LLAMA_LOG_DEBUG(...) llama_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LLAMA_LOG_CONT(...)  llama_log_internal(GGML_LOG_LEVEL_CONT , __VA_ARGS__)
+
+//
+// helpers
+//
+
+template <typename T>
+struct no_init {
+    T value;
+    no_init() = default;
+};
+
+struct time_meas {
+    time_meas(int64_t & t_acc, bool disable = false);
+    ~time_meas();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
+template <typename T>
+struct buffer_view {
+    T * data;
+    size_t size = 0;
+
+    bool has_data() const {
+        return data && size > 0;
+    }
+};
+
+void replace_all(std::string & s, const std::string & search, const std::string & replace);
+
+// TODO: rename to llama_format ?
+LLAMA_ATTRIBUTE_FORMAT(1, 2)
+std::string format(const char * fmt, ...);
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
+std::string llama_format_tensor_shape(const struct ggml_tensor * t);
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
+
+#define LLAMA_TENSOR_NAME_FATTN "__fattn__"

package/cpp/src/llama-io.cpp

@@ -0,0 +1,15 @@
+#include "llama-io.h"
+
+void llama_io_write_i::write_string(const std::string & str) {
+    uint32_t str_size = str.size();
+
+    write(&str_size,  sizeof(str_size));
+    write(str.data(), str_size);
+}
+
+void llama_io_read_i::read_string(std::string & str) {
+    uint32_t str_size;
+    read_to(&str_size, sizeof(str_size));
+
+    str.assign((const char *) read(str_size), str_size);
+}

package/cpp/src/llama-io.h

@@ -0,0 +1,35 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+struct ggml_tensor;
+
+class llama_io_write_i {
+public:
+    llama_io_write_i() = default;
+    virtual ~llama_io_write_i() = default;
+
+    virtual void write(const void * src, size_t size) = 0;
+    virtual void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) = 0;
+
+    // bytes written so far
+    virtual size_t n_bytes() = 0;
+
+    void write_string(const std::string & str);
+};
+
+class llama_io_read_i {
+public:
+    llama_io_read_i() = default;
+    virtual ~llama_io_read_i() = default;
+
+    virtual const uint8_t * read(size_t size) = 0;
+    virtual void read_to(void * dst, size_t size) = 0;
+
+    // bytes read so far
+    virtual size_t n_bytes() = 0;
+
+    void read_string(std::string & str);
+};

package/cpp/src/llama-kv-cache-iswa.cpp

@@ -0,0 +1,330 @@
+#include "llama-kv-cache-iswa.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <cassert>
+
+//
+// llama_kv_cache_iswa
+//
+
+llama_kv_cache_iswa::llama_kv_cache_iswa(
+        const llama_model & model,
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   offload,
+                     bool   swa_full,
+                     bool   unified,
+                 uint32_t   kv_size,
+                 uint32_t   n_seq_max,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad,
+    const layer_filter_cb & filter,
+    const  layer_reuse_cb & reuse) : hparams(model.hparams), unified(unified) {
+
+    // chain filters
+    const layer_filter_cb filter_base = [&](int32_t il) {
+        if (filter && !filter(il)) {
+            return false;
+        }
+
+        return !model.hparams.is_swa(il);
+    };
+
+    const layer_filter_cb filter_swa  = [&](int32_t il) {
+        if (filter && !filter(il)) {
+            return false;
+        }
+
+        return  model.hparams.is_swa(il);
+    };
+
+    const uint32_t size_base = kv_size;
+
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
+
+    // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
+    if (swa_full) {
+        LLAMA_LOG_WARN("%s: using full-size SWA cache (ref: %s)\n",
+                __func__, "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
+
+        size_swa = size_base;
+    }
+
+    LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
+
+    kv_base = std::make_unique<llama_kv_cache>(
+            model, type_k, type_v,
+            v_trans, offload, unified, size_base, n_seq_max, n_pad,
+            0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
+
+    LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
+
+    kv_swa = std::make_unique<llama_kv_cache>(
+            model, type_k, type_v,
+            v_trans, offload, unified, size_swa, n_seq_max, n_pad,
+            hparams.n_swa, hparams.swa_type, filter_swa, reuse);
+}
+
+void llama_kv_cache_iswa::clear(bool data) {
+    kv_base->clear(data);
+    kv_swa ->clear(data);
+}
+
+bool llama_kv_cache_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    bool res = true;
+
+    res = res & kv_base->seq_rm(seq_id, p0, p1);
+    res = res & kv_swa ->seq_rm(seq_id, p0, p1);
+
+    return res;
+}
+
+void llama_kv_cache_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    kv_base->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    kv_swa ->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_iswa::seq_keep(llama_seq_id seq_id) {
+    kv_base->seq_keep(seq_id);
+    kv_swa ->seq_keep(seq_id);
+}
+
+void llama_kv_cache_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    kv_base->seq_add(seq_id, p0, p1, shift);
+    kv_swa ->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_kv_cache_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    kv_base->seq_div(seq_id, p0, p1, d);
+    kv_swa ->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_kv_cache_iswa::seq_pos_min(llama_seq_id seq_id) const {
+    // the base cache is a superset of the SWA cache, so we can just check the SWA cache
+    return kv_swa->seq_pos_min(seq_id);
+}
+
+llama_pos llama_kv_cache_iswa::seq_pos_max(llama_seq_id seq_id) const {
+    return kv_swa->seq_pos_max(seq_id);
+}
+
+std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_iswa::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = kv_base->memory_breakdown();
+    for (const auto & buft_size : kv_swa->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+llama_memory_context_ptr llama_kv_cache_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
+
+    // first try simple split
+    do {
+        if (!unified) {
+            // requires equal splits, so we skip the simple split
+            break;
+        }
+
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_simple(n_ubatch);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            break;
+        }
+
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
+    } while (false);
+
+    // if it fails, try equal split
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_equal(n_ubatch, !unified);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            break;
+        }
+
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
+    } while (false);
+
+    // TODO: if we fail again, we should attempt different splitting strategies
+    //       but to do that properly, we first have to refactor the batches to be more flexible
+
+    return std::make_unique<llama_kv_cache_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_kv_cache_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_iswa_context>(this);
+}
+
+llama_memory_context_ptr llama_kv_cache_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_iswa_context>(this, lctx, optimize);
+}
+
+bool llama_kv_cache_iswa::get_can_shift() const {
+    return kv_base->get_can_shift() &&
+           kv_swa->get_can_shift() &&
+           kv_base->get_size() == kv_swa->get_size();
+}
+
+void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
+        kv_base->state_write(io, seq_id, flags);
+    }
+
+    kv_swa->state_write(io, seq_id, flags);
+}
+
+void llama_kv_cache_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
+        kv_base->state_read(io, seq_id, flags);
+    }
+
+    kv_swa->state_read(io, seq_id, flags);
+}
+
+llama_kv_cache * llama_kv_cache_iswa::get_base() const {
+    return kv_base.get();
+}
+
+llama_kv_cache * llama_kv_cache_iswa::get_swa() const {
+    return kv_swa.get();
+}
+
+//
+// llama_kv_cache_iswa_context
+//
+
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(llama_memory_status status) : status(status) {}
+
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv) :
+    ctx_base(kv->get_base()->init_full()),
+    ctx_swa (kv->get_swa ()->init_full()),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv,
+        llama_context * lctx,
+        bool optimize) :
+    ctx_base(kv->get_base()->init_update(lctx, optimize)),
+    ctx_swa (kv->get_swa ()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv,
+        slot_info_vec_t sinfos_base,
+        slot_info_vec_t sinfos_swa,
+        std::vector<llama_ubatch> ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_base(new llama_kv_cache_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_iswa_context:: ~llama_kv_cache_iswa_context() = default;
+
+bool llama_kv_cache_iswa_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_base->next();
+    ctx_swa ->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_kv_cache_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_base->apply();
+    res = res & ctx_swa ->apply();
+
+    return res;
+}
+
+llama_memory_status llama_kv_cache_iswa_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_kv_cache_iswa_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_context * llama_kv_cache_iswa_context::get_base() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_context *>(ctx_base.get());
+}
+
+const llama_kv_cache_context * llama_kv_cache_iswa_context::get_swa()  const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_context *>(ctx_swa.get());
+}