From fd4d803c605445cbf5e543a5a1c999930899b2c1 Mon Sep 17 00:00:00 2001
From: Sascha Rogmann <github@rogmann.org>
Date: Mon, 26 Jan 2026 00:20:05 +0100
Subject: [PATCH] common: print performance in spec decoding

---
 common/speculative.cpp | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/common/speculative.cpp b/common/speculative.cpp
index e9102678fa..06fd9fbf8e 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -50,6 +50,7 @@ struct common_speculative_config {
 // in a subclass of common_speculative_state
 struct common_speculative_state {
     const enum common_speculative_type type;
+    const bool gen_perf            = true; // whether to generate performance stats.
 
     size_t drafts_call_count       = 0; // number of times this implementation was called.
     size_t drafts_generated_count  = 0; // number of times a draft or part was generated by this implementation.
@@ -57,6 +58,8 @@ struct common_speculative_state {
     size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
     size_t drafts_accepted_tokens  = 0; // number of tokens accepted by the target model.
 
+    int64_t gen_duration_ms        = 0; // total time spent in this implementation in milliseconds.
+
     virtual ~common_speculative_state() = default;
 
     common_speculative_state(enum common_speculative_type type) : type(type) {}
@@ -737,6 +740,7 @@ llama_tokens common_speculative_gen_draft(
     // TODO: avoid dynamic casts
     for (auto & impl : spec->impls) {
         impl->drafts_call_count++;
+        const int64_t t_start_ms = impl->gen_perf ? ggml_time_ms() : 0;
 
         switch (impl->type) {
             case COMMON_SPECULATIVE_TYPE_NONE:
@@ -801,6 +805,9 @@ llama_tokens common_speculative_gen_draft(
             }
         }
 
+        const int64_t t_now_ms = impl->gen_perf ? ggml_time_ms() : 0;
+        impl->gen_duration_ms += t_now_ms - t_start_ms; // accumulate duration for this implementation
+
         if (!result.empty()) {
             LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
                     common_speculative_type_to_str(impl.get()->type).c_str(),
@@ -848,12 +855,17 @@ void common_speculative_print_stats(const struct common_speculative * spec) {
     }
 
     for (const auto & impl : spec->impls) {
-        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu\n",
+        // std::string via impl->gen_duration_ms (if >0, else "")
+        std::string performance = impl->gen_perf
+            ? ", dur = " + std::to_string(impl->gen_duration_ms) + " ms"
+            : "";
+        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
                 common_speculative_type_to_str(impl->type).c_str(),
                 impl->drafts_call_count,
                 impl->drafts_generated_count,
                 impl->drafts_accepted_count,
                 impl->drafts_generated_tokens,
-                impl->drafts_accepted_tokens);
+                impl->drafts_accepted_tokens,
+                performance.c_str());
     }
 }