fix: report full prompt length in OpenAI usage

src/common/AutoModel/automodel.cpp

@@ -215,7 +215,11 @@ bool AutoModel::_shared_insert(chat_meta_info_t& meta_info, std::vector<int>& to
 
     auto prefill_end_time = this->profiler_list[PREFILL_TIME].stop(tokens.size());
     meta_info.prefill_duration = (uint64_t)time_utils::duration_ns(prefill_start_time, prefill_end_time).first;
-    meta_info.prompt_tokens = tokens.size();
+    // `tokens` was trimmed to the uncached suffix above, so add the prefix served
+    // from the KV cache back on: usage.prompt_tokens is the whole prompt, and the
+    // cached part is reported separately rather than subtracted.
+    meta_info.cached_prompt_tokens = static_cast<int>(skip_count);
+    meta_info.prompt_tokens = static_cast<int>(skip_count + tokens.size());
 
     if (meta_info.stop_reason == CANCEL_DETECTED) {
         return false;

src/common/AutoModel/modeling_qwen3_5_omni.cpp

@@ -551,7 +551,12 @@ bool Qwen3_5_Omni::insert(chat_meta_info_t& meta_info, lm_uniform_input_t& input
 
     auto prefill_end = this->profiler_list[PREFILL_TIME].stop(tokens.size());
     meta_info.prefill_duration = (uint64_t)time_utils::duration_ns(prefill_start, prefill_end).first;
-    meta_info.prompt_tokens = tokens.size(); 
+    // As in AutoModel::_shared_insert, `tokens` holds only the uncached suffix, so
+    // the cached prefix is added back to report the whole prompt. The relevant count
+    // is `skip_count`, which is what was erased from `tokens` above -- not the earlier
+    // `prefix_skip_count`, which only trims the multi-modal payload.
+    meta_info.cached_prompt_tokens = static_cast<int>(skip_count);
+    meta_info.prompt_tokens = static_cast<int>(skip_count + tokens.size());
 
     this->total_tokens += tokens.size();
 

src/common/AutoModel/modeling_qwen3vl.cpp

@@ -583,6 +583,9 @@ bool Qwen3VL_Flash::insert(chat_meta_info_t& meta_info, lm_uniform_input_t& inpu
 
     meta_info.prefill_duration = (uint64_t)time_utils::duration_ns(prefill_start_time, prefill_end_time).first;
     meta_info.prompt_tokens = static_cast<int>(tokens.size()); // report full prompt length to caller
+    // The pinned system prefix is already in the kv cache, so it is part of the
+    // prompt but was not prefilled on this turn.
+    meta_info.cached_prompt_tokens = skip;
 
     if (meta_info.stop_reason == CANCEL_DETECTED) {
         return false;

src/include/AutoModel/automodel.hpp

@@ -108,7 +108,8 @@ typedef enum {
 
 struct chat_meta_info_t {
 	int max_prefill_len;
-    int prompt_tokens;
+    int prompt_tokens;        // whole prompt, cached prefix included
+    int cached_prompt_tokens; // subset of prompt_tokens served from the KV cache
     int generated_tokens;
     uint64_t total_duration; // in nanoseconds
     uint64_t load_duration; // in nanoseconds
@@ -118,7 +119,7 @@ struct chat_meta_info_t {
 	bool restore_allowed;
 	tool_choice_t tool_choice;
 
-	chat_meta_info_t() : max_prefill_len(0), prompt_tokens(0), generated_tokens(0), total_duration(0), load_duration(0), prefill_duration(0), decoding_duration(0), stop_reason(EOT_DETECTED), restore_allowed(false), tool_choice(TOOL_CHOICE_AUTO) {}
+	chat_meta_info_t() : max_prefill_len(0), prompt_tokens(0), cached_prompt_tokens(0), generated_tokens(0), total_duration(0), load_duration(0), prefill_duration(0), decoding_duration(0), stop_reason(EOT_DETECTED), restore_allowed(false), tool_choice(TOOL_CHOICE_AUTO) {}
 };
 
 typedef enum {

src/server/rest_handler.cpp

@@ -1303,13 +1303,14 @@ void RestHandler::handle_openai_chat_completion(const json& request,
                 {"choices", choices},
                 {"usage", {
                     {"prompt_tokens", meta_info.prompt_tokens},
+                    {"prompt_tokens_details", {{"cached_tokens", meta_info.cached_prompt_tokens}}},
                     {"completion_tokens", meta_info.generated_tokens},
                     {"total_tokens", meta_info.prompt_tokens + meta_info.generated_tokens},
                     {"kv_token_occupancy_rate_percentage", (float)this->auto_chat_engine->get_current_context_length() / (float)this->auto_chat_engine->get_max_length() * 100},
                     {"load_duration", static_cast<double>(meta_info.load_duration) / 1'000'000'000},
                     {"prefill_duration_ttft", static_cast<double>(meta_info.prefill_duration) / 1'000'000'000},
                     {"decoding_duration", static_cast<double>(meta_info.decoding_duration) / 1'000'000'000},
-                    {"prefill_speed_tps", static_cast<double>(meta_info.prompt_tokens) / static_cast<double>(meta_info.prefill_duration) * 1'000'000'000},
+                    {"prefill_speed_tps", static_cast<double>(meta_info.prompt_tokens - meta_info.cached_prompt_tokens) / static_cast<double>(meta_info.prefill_duration) * 1'000'000'000},
                     {"decoding_speed_tps", static_cast<double>(meta_info.generated_tokens) / static_cast<double>(meta_info.decoding_duration) * 1'000'000'000},
                 }},
                 {"service_tier", "default"}
@@ -1511,6 +1512,7 @@ void RestHandler::handle_openai_completion(const json& request,
                 })},
                 {"usage", {
                     {"prompt_tokens", meta_info.prompt_tokens},
+                    {"prompt_tokens_details", {{"cached_tokens", meta_info.cached_prompt_tokens}}},
                     {"completion_tokens", meta_info.generated_tokens},
                     {"total_tokens", meta_info.prompt_tokens + meta_info.generated_tokens}
                 }}

src/server/streaming_ostream_openai.hpp

@@ -207,6 +207,7 @@ class streaming_buf_openai : public std::streambuf {
             })},
             {"usage", {
                 {"prompt_tokens", meta_info.prompt_tokens},
+                {"prompt_tokens_details", {{"cached_tokens", meta_info.cached_prompt_tokens}}},
                 {"completion_tokens", meta_info.generated_tokens},
                 {"total_tokens", meta_info.prompt_tokens + meta_info.generated_tokens}
             }}
@@ -497,6 +498,7 @@ class streaming_buf_openai_chat : public std::streambuf {
             })},
             {"usage", {
                 {"prompt_tokens", meta_info.prompt_tokens},
+                {"prompt_tokens_details", {{"cached_tokens", meta_info.cached_prompt_tokens}}},
                 {"completion_tokens", meta_info.generated_tokens},
                 {"total_tokens", meta_info.prompt_tokens + meta_info.generated_tokens},
                 {"active_kv_tokens", this->auto_chat_engine->get_current_context_length()},
@@ -505,7 +507,7 @@ class streaming_buf_openai_chat : public std::streambuf {
                 {"load_duration", static_cast<double>(meta_info.load_duration) / 1'000'000'000},
                 {"prefill_duration_ttft", static_cast<double>(meta_info.prefill_duration) / 1'000'000'000},
                 {"decoding_duration", static_cast<double>(meta_info.decoding_duration) / 1'000'000'000},
-                {"prefill_speed_tps", static_cast<double>(meta_info.prompt_tokens) / static_cast<double>(meta_info.prefill_duration) * 1'000'000'000},
+                {"prefill_speed_tps", static_cast<double>(meta_info.prompt_tokens - meta_info.cached_prompt_tokens) / static_cast<double>(meta_info.prefill_duration) * 1'000'000'000},
                 {"decoding_speed_tps", static_cast<double>(meta_info.generated_tokens) / static_cast<double>(meta_info.decoding_duration) * 1'000'000'000},
             }}
         };