fixup! lookup: evaluation tools, use corpus/previous gens

ggml-org · Mar 22, 2024 · acb5375 · acb5375
1 parent 3106005
commit acb5375
Show file tree

Hide file tree

Showing 2 changed files with 22 additions and 8 deletions.
diff --git a/common/ngram-cache.cpp b/common/ngram-cache.cpp
@@ -39,7 +39,14 @@ void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, in
                 const long eta_min  = eta_ms / (60*1000);
                 const long eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
 
-                fprintf(stderr, "%s: %ld/%ld done, ETA: %02ld:%02ld\n", __func__, n_done, n_todo, eta_min, eta_s);
+                // %02ld doesn't compile on Arm64 MacOS:
+                std::string eta_string;
+                eta_string += eta_min < 10 ? "0" : "";
+                eta_string += std::to_string(eta_min);
+                eta_string += ":";
+                eta_string += eta_s < 10 ? "0" : "";
+                eta_string += std::to_string(eta_s);
+                fprintf(stderr, "%s: %ld/%ld done, ETA: %s\n", __func__, n_done, n_todo, eta_string.c_str());
             }
         }
     }

diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp
@@ -73,7 +73,7 @@ int main(int argc, char ** argv){
     int n_drafted = 0;
     int n_accept  = 0;
 
-    const int64_t t_start_ms = ggml_time_ms();
+    const long t_start_ms = ggml_time_ms();
 
     // Iterate over input tokens in chunks of size n_ctx.
     // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
@@ -127,12 +127,19 @@ int main(int argc, char ** argv){
 
         }
         if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
-            const int64_t t_now_ms = ggml_time_ms();
-            const int64_t eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
-            const int64_t eta_min  = eta_ms / (60*1000);
-            const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
-
-            LOG_TEE("%d/%d done, ETA: %02ld:%02ld\n", i_start, n_input, eta_min, eta_s);
+            const long t_now_ms = ggml_time_ms();
+            const long eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
+            const long eta_min  = eta_ms / (60*1000);
+            const long eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+            // %02ld doesn't compile on Arm64 MacOS:
+            std::string eta_string;
+            eta_string += eta_min < 10 ? "0" : "";
+            eta_string += std::to_string(eta_min);
+            eta_string += ":";
+            eta_string += eta_s < 10 ? "0" : "";
+            eta_string += std::to_string(eta_s);
+            LOG_TEE("lookup-stats: %d/%d done, ETA: %s\n", i_start, n_input, eta_string.c_str());
         }
 
         // After each chunk, update the dynamic ngram cache with the context ngram cache: