Update for compat with quantize-stats

Author: comex

examples/quantize-stats/quantize-stats.cpp

@@ -266,15 +266,13 @@ int main(int argc, char ** argv) {
         }
     }
 
-    // Sort tensors for consistent output
-    const auto tensors = llama_internal_get_tensor_map(ctx);
-    std::map<std::string, struct ggml_tensor *> tensors_sorted { tensors.begin(), tensors.end() };
+    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_sorted) {
+    for (const auto& kv_tensor : tensors) {
         if (!layer_included(params, kv_tensor.first)) {
             continue;
         }
@@ -315,7 +313,7 @@ int main(int argc, char ** argv) {
 
             error_stats global_stats {};
 
-            for (const auto& kv_tensor : tensors_sorted) {
+            for (const auto& kv_tensor : tensors) {
                 if (!layer_included(params, kv_tensor.first)) {
                     continue;
                 }

llama.cpp

@@ -1,3 +1,4 @@
+// XXX: CHECK REGEX
 #include "llama_util.h"
 #include "llama.h"
 
@@ -148,6 +149,9 @@ struct llama_model {
     llama_mlock mlock_buf;
     llama_mlock mlock_mmap;
 
+    // for quantize-stats only
+    std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
+
     ~llama_model() {
         if (ctx) {
             ggml_free(ctx);
@@ -634,10 +638,13 @@ struct llama_model_loader {
         return tensor;
     }
 
-    void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
+    void done_getting_tensors() {
         if (num_ggml_tensors_created != tensors_map.tensors.size()) {
             throw std::string("llama.cpp: file contained more tensors than expected");
         }
+    }
+
+    void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
@@ -928,6 +935,13 @@ static void llama_model_load_internal(
         }
     }
 
+    ml->done_getting_tensors();
+
+    // populate `tensors_by_name`
+    for (llama_load_tensor & lt : ml->tensors_map.tensors) {
+        model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor);
+    }
+
     ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL);
 
     model.mapping = std::move(ml->mapping);
@@ -1887,6 +1901,6 @@ const char * llama_print_system_info(void) {
 }
 
 // For internal test use
-std::unordered_map<std::string, struct ggml_tensor *>& llama_internal_get_tensor_map(struct llama_context * ctx) {
-    return ctx->model.tensors;
+std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
+    return ctx->model.tensors_by_name;
 }

llama.h

@@ -174,7 +174,7 @@ extern "C" {
 //
 // Internal function exposed for tests and benchmarks
 //
-std::unordered_map<std::string, struct ggml_tensor *>& llama_internal_get_tensor_map(struct llama_context * ctx);
+std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
 #endif
 
 #endif