llama : impl

ggml-ci

common/common.h

@@ -638,6 +638,10 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 // Split utils
 //
 
-static const char * const LLM_KV_SPLIT_NO            = "split.no";
+namespace {
-static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+
-static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+const char * const LLM_KV_SPLIT_NO            = "split.no";
+const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+
+}

examples/gguf-split/gguf-split.cpp

@@ -2,15 +2,14 @@
 #include "common.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdlib>
 #include <fstream>
 #include <string>
 #include <vector>
-
-#include <stdio.h>
-#include <string.h>
 #include <climits>
+
+#include <cstdio>
+#include <cstring>
 #include <stdexcept>
 
 #if defined(_WIN32)

examples/quantize-stats/quantize-stats.cpp

@@ -1,7 +1,7 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "llama-impl.h"
+#include "llama-context.h"
+#include "common.h"
 
 #include <algorithm>
 #include <cassert>
@@ -9,11 +9,9 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
-#include <map>
 #include <numeric>
 #include <regex>
 #include <string>
-#include <unordered_map>
 #include <vector>
 #include <thread>
 #include <mutex>
@@ -330,13 +328,13 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const auto &tensors = llama_internal_get_tensor_map(ctx);
+    const auto & tensors = llama_internal_get_tensor_map(ctx);
 
     // check layer tensors
     int included_layers = 0;
     int64_t max_nelements = 0;
     bool is_f16 = false;
-    for (const auto& kv_tensor : tensors) {
+    for (const auto & kv_tensor : tensors) {
         if (!layer_included(params, kv_tensor.first)) {
             continue;
         }
@@ -371,8 +369,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        const auto *  qfns     = ggml_get_type_traits(type);
+        const auto * qfns     = ggml_get_type_traits(type);
-        const auto *  qfns_cpu = ggml_get_type_traits_cpu(type);
+        const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
         if (qfns_cpu->from_float && qfns->to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
@@ -382,7 +380,7 @@ int main(int argc, char ** argv) {
 
             error_stats global_stats {};
 
-            for (const auto& kv_tensor : tensors) {
+            for (const auto & kv_tensor : tensors) {
                 if (!layer_included(params, kv_tensor.first)) {
                     continue;
                 }

src/CMakeLists.txt

@@ -15,6 +15,7 @@ add_library(llama
             llama-chat.cpp
             llama-context.cpp
             llama-hparams.cpp
+            llama-impl.cpp
             llama-grammar.cpp
             llama-kv-cache.cpp
             llama-mmap.cpp

src/llama-adapter.cpp

@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <map>
 #include <cassert>
+#include <stdexcept>
 
 // vec
 

src/llama-batch.h

@@ -26,7 +26,9 @@ struct llama_ubatch {
 
 struct llama_sbatch_seq {
     int32_t n_seq_id;
+
     llama_seq_id * seq_id;
+
     size_t offset;
     size_t length;
 };
@@ -112,8 +114,8 @@ struct llama_sbatch {
             if (ubatch.equal_seqs) {
                 for (size_t i = 0; i < length; ++i) {
                     memcpy(
-                        ubatch.embd + n_embd * (ubatch.n_tokens + i),
+                        ubatch.embd + (n_embd * (ubatch.n_tokens + i)),
-                        batch->embd + n_embd * ids[seq.offset + i],
+                        batch->embd + (n_embd * ids[seq.offset + i]),
                         n_embd * sizeof(float)
                     );
                 }

src/llama-context.cpp

@@ -1,5 +1,7 @@
 #include "llama-context.h"
 
+#include <stdexcept>
+
 // deprecated
 size_t llama_get_state_size(struct llama_context * ctx) {
     return llama_state_get_size(ctx);
@@ -968,3 +970,8 @@ size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepa
     }
 }
 
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+) {
+    return ctx->model.tensors_by_name;
+}

src/llama-context.h

@@ -219,3 +219,7 @@ static void llama_output_reorder(struct llama_context * ctx) {
         out_ids.clear();
     }
 }
+
+// For internal test use
+// TODO: remove
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(struct llama_context * ctx);

src/llama-grammar.cpp

@@ -1,5 +1,6 @@
 #include "llama-grammar.h"
 
+#include "llama-impl.h"
 #include "llama-vocab.h"
 #include "llama-sampling.h"
 

src/llama-grammar.h

@@ -1,8 +1,10 @@
 #pragma once
 
-#include "llama-impl.h"
+#include "llama.h"
 
 #include <map>
+#include <string>
+#include <vector>
 
 struct llama_vocab;
 

src/llama-impl.cpp

@@ -0,0 +1,74 @@
+#include "llama-impl.h"
+
+#include "llama.h"
+
+#include <cstdarg>
+
+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 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);
+}
+
+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);
+}

src/llama-impl.h

@@ -1,10 +1,8 @@
 #pragma once
 
-#include "llama.h"
+#include "ggml.h"
 
 #include <string>
-#include <vector>
-#include <stdexcept>
 
 #ifdef __GNUC__
 #ifdef __MINGW32__
@@ -40,146 +38,12 @@ std::string format(const char * fmt, ...);
 //
 
 struct time_meas {
-    time_meas(int64_t & t_acc, bool disable = false) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+    time_meas(int64_t & t_acc, bool disable = false);
-
+    ~time_meas();
-    ~time_meas() {
-        if (t_start_us >= 0) {
-            t_acc += ggml_time_us() - t_start_us;
-        }
-    }
 
     const int64_t t_start_us;
 
     int64_t & t_acc;
 };
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+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);
-}
-
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
-    struct llama_context * ctx
-);
-
-// the ring buffer works similarly to std::deque, but with a fixed capacity
-template<typename T>
-struct ring_buffer {
-    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
-
-    T & front() {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[first];
-    }
-
-    const T & front() const {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[first];
-    }
-
-    T & back() {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[pos];
-    }
-
-    const T & back() const {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[pos];
-    }
-
-    void push_back(const T & value) {
-        if (capacity == 0) {
-            throw std::runtime_error("ring buffer: capacity is zero");
-        }
-
-        if (sz == capacity) {
-            // advance the start when buffer is full
-            first = (first + 1) % capacity;
-        } else {
-            sz++;
-        }
-        data[pos] = value;
-        pos = (pos + 1) % capacity;
-    }
-
-    T pop_front() {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        T value = data[first];
-        first = (first + 1) % capacity;
-        sz--;
-        return value;
-    }
-
-    //T & operator[](size_t i) {
-    //    if (i >= sz) {
-    //        throw std::runtime_error("ring buffer: index out of bounds");
-    //    }
-    //    return data[(first + i) % capacity];
-    //}
-
-    //const T & at(size_t i) const {
-    //    if (i >= sz) {
-    //        throw std::runtime_error("ring buffer: index out of bounds");
-    //    }
-    //    return data[(first + i) % capacity];
-    //}
-
-    const T & rat(size_t i) const {
-        if (i >= sz) {
-            throw std::runtime_error("ring buffer: index out of bounds");
-        }
-        return data[(first + sz - i - 1) % capacity];
-    }
-
-    std::vector<T> to_vector() const {
-        std::vector<T> result;
-        result.reserve(sz);
-        for (size_t i = 0; i < sz; i++) {
-            result.push_back(data[(first + i) % capacity]);
-        }
-        return result;
-    }
-
-    void clear() {
-        // here only reset the status of the buffer
-        sz = 0;
-        first = 0;
-        pos = 0;
-    }
-
-    bool empty() const {
-        return sz == 0;
-    }
-
-    size_t size() const {
-        return sz;
-    }
-
-    size_t capacity = 0;
-    size_t sz = 0;
-    size_t first = 0;
-    size_t pos = 0;
-    std::vector<T> data;
-};

src/llama-mmap.cpp

@@ -6,6 +6,7 @@
 
 #include <cstring>
 #include <climits>
+#include <stdexcept>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)

src/llama-model.cpp

@@ -4,6 +4,7 @@
 
 #include <algorithm>
 #include <cassert>
+#include <stdexcept>
 
 const char * llm_type_name(llm_type type) {
     switch (type) {

src/llama-sampling.cpp

@@ -1,5 +1,6 @@
 #include "llama-sampling.h"
 
+#include "llama-impl.h"
 #include "llama-vocab.h"
 #include "llama-grammar.h"
 
@@ -14,6 +15,117 @@
 #include <numeric>
 #include <random>
 #include <unordered_map>
+#include <stdexcept>
+
+// the ring buffer works similarly to std::deque, but with a fixed capacity
+template<typename T>
+struct ring_buffer {
+    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
+
+    T & front() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[first];
+    }
+
+    const T & front() const {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[first];
+    }
+
+    T & back() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[pos];
+    }
+
+    const T & back() const {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[pos];
+    }
+
+    void push_back(const T & value) {
+        if (capacity == 0) {
+            throw std::runtime_error("ring buffer: capacity is zero");
+        }
+
+        if (sz == capacity) {
+            // advance the start when buffer is full
+            first = (first + 1) % capacity;
+        } else {
+            sz++;
+        }
+        data[pos] = value;
+        pos = (pos + 1) % capacity;
+    }
+
+    T pop_front() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        T value = data[first];
+        first = (first + 1) % capacity;
+        sz--;
+        return value;
+    }
+
+    //T & operator[](size_t i) {
+    //    if (i >= sz) {
+    //        throw std::runtime_error("ring buffer: index out of bounds");
+    //    }
+    //    return data[(first + i) % capacity];
+    //}
+
+    //const T & at(size_t i) const {
+    //    if (i >= sz) {
+    //        throw std::runtime_error("ring buffer: index out of bounds");
+    //    }
+    //    return data[(first + i) % capacity];
+    //}
+
+    const T & rat(size_t i) const {
+        if (i >= sz) {
+            throw std::runtime_error("ring buffer: index out of bounds");
+        }
+        return data[(first + sz - i - 1) % capacity];
+    }
+
+    std::vector<T> to_vector() const {
+        std::vector<T> result;
+        result.reserve(sz);
+        for (size_t i = 0; i < sz; i++) {
+            result.push_back(data[(first + i) % capacity]);
+        }
+        return result;
+    }
+
+    void clear() {
+        // here only reset the status of the buffer
+        sz = 0;
+        first = 0;
+        pos = 0;
+    }
+
+    bool empty() const {
+        return sz == 0;
+    }
+
+    size_t size() const {
+        return sz;
+    }
+
+    size_t capacity = 0;
+    size_t sz = 0;
+    size_t first = 0;
+    size_t pos = 0;
+    std::vector<T> data;
+};
 
 static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng) {
     // iterator for the probabilities

src/llama.cpp

@@ -202,13 +202,6 @@ static std::string llama_token_to_piece(const struct llama_model * model, llama_
 // globals
 //
 
-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;
-
 static const size_t kiB = 1024;
 static const size_t MiB = 1024*kiB;
 static const size_t GiB = 1024*MiB;
@@ -17188,46 +17181,3 @@ void llama_perf_context_reset(struct llama_context * ctx) {
     ctx->t_p_eval_us = ctx->n_p_eval = 0;
 }
 
-// For internal test use
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
-    struct llama_context * ctx
-) {
-    return ctx->model.tensors_by_name;
-}
-
-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);
-}