llama : impl

ggml-ci

Author: ggerganov

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";
-static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
-static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+namespace {
+
+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

@@ -1,19 +1,17 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "llama-impl.h"
+#include "llama-context.h"
+#include "common.h"
 
 #include <algorithm>
 #include <cassert>
 #include <cinttypes>
 #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_cpu = ggml_get_type_traits_cpu(type);
+        const auto * qfns     = ggml_get_type_traits(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),
-                        batch->embd + n_embd * ids[seq.offset + i],
+                        ubatch.embd + (n_embd * (ubatch.n_tokens + 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;