Merge db4a5a84fc into cc2aa81513

tools/server/server-context.cpp

@@ -1415,6 +1415,7 @@ private:
 
         res->n_decoded           = slot.n_decoded;
         res->n_prompt_tokens     = slot.task->n_tokens();
+        res->n_cache_read_input_tokens = slot.n_prompt_tokens_cache;
         res->post_sampling_probs = slot.task->params.post_sampling_probs;
 
         res->verbose           = slot.task->params.verbose;
@@ -1457,6 +1458,7 @@ private:
         res->truncated           = slot.truncated;
         res->n_decoded           = slot.n_decoded;
         res->n_prompt_tokens     = slot.task->n_tokens();
+        res->n_cache_read_input_tokens = slot.n_prompt_tokens_cache;
         res->n_tokens_cached     = slot.prompt.n_tokens();
         res->has_new_line        = slot.has_new_line;
         res->stopping_word       = slot.stopping_word;

tools/server/server-task.cpp

@@ -1010,6 +1010,7 @@ json server_task_result_cmpl_final::to_json_anthropic() {
     if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
         stop_reason = oaicompat_msg.tool_calls.empty() ? "end_turn" : "tool_use";
     }
+    const int32_t cache_read_input_tokens = n_cache_read_input_tokens >= 0 ? n_cache_read_input_tokens : 0;
 
     json content_blocks = json::array();
 
@@ -1063,6 +1064,7 @@ json server_task_result_cmpl_final::to_json_anthropic() {
         {"stop_sequence", stopping_word.empty() ? nullptr : json(stopping_word)},
         {"usage", {
             {"input_tokens", n_prompt_tokens},
+            {"cache_read_input_tokens", cache_read_input_tokens},
             {"output_tokens", n_decoded}
         }}
     };
@@ -1077,6 +1079,7 @@ json server_task_result_cmpl_final::to_json_anthropic_stream() {
     if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
         stop_reason = oaicompat_msg.tool_calls.empty() ? "end_turn" : "tool_use";
     }
+    const int32_t cache_read_input_tokens = n_cache_read_input_tokens >= 0 ? n_cache_read_input_tokens : 0;
 
     bool has_thinking = !oaicompat_msg.reasoning_content.empty();
     bool has_text     = !oaicompat_msg.content.empty();
@@ -1555,6 +1558,7 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
 json server_task_result_cmpl_partial::to_json_anthropic() {
     json events = json::array();
     bool first = (n_decoded == 1);
+    const int32_t cache_read_input_tokens = n_cache_read_input_tokens >= 0 ? n_cache_read_input_tokens : 0;
     // use member variables to track block state across streaming calls
     // (anthropic_thinking_block_started, anthropic_text_block_started)
 
@@ -1573,6 +1577,7 @@ json server_task_result_cmpl_partial::to_json_anthropic() {
                     {"stop_sequence", nullptr},
                     {"usage", {
                         {"input_tokens", n_prompt_tokens},
+                        {"cache_read_input_tokens", cache_read_input_tokens},
                         {"output_tokens", 0}
                     }}
                 }}

tools/server/server-task.h

@@ -342,6 +342,7 @@ struct server_task_result_cmpl_final : server_task_result {
     bool truncated;
     int32_t n_decoded;
     int32_t n_prompt_tokens;
+    int32_t n_cache_read_input_tokens = 0;
     int32_t n_tokens_cached;
     bool has_new_line;
     std::string stopping_word;
@@ -406,6 +407,7 @@ struct server_task_result_cmpl_partial : server_task_result {
 
     int32_t n_decoded;
     int32_t n_prompt_tokens;
+    int32_t n_cache_read_input_tokens = 0;
 
     bool post_sampling_probs;
     bool is_progress = false;

tools/server/tests/unit/test_compat_anthropic.py

@@ -64,9 +64,12 @@ def test_anthropic_messages_basic():
     assert res.body["stop_reason"] in ["end_turn", "max_tokens"], f"Invalid stop_reason: {res.body.get('stop_reason')}"
     assert "usage" in res.body, "Missing 'usage' field"
     assert "input_tokens" in res.body["usage"], "Missing usage.input_tokens"
+    assert "cache_read_input_tokens" in res.body["usage"], "Missing usage.cache_read_input_tokens"
     assert "output_tokens" in res.body["usage"], "Missing usage.output_tokens"
     assert isinstance(res.body["usage"]["input_tokens"], int), "input_tokens should be integer"
+    assert isinstance(res.body["usage"]["cache_read_input_tokens"], int), "cache_read_input_tokens should be integer"
     assert isinstance(res.body["usage"]["output_tokens"], int), "output_tokens should be integer"
+    assert res.body["usage"]["cache_read_input_tokens"] >= 0, "cache_read_input_tokens should be >= 0"
     assert res.body["usage"]["output_tokens"] > 0, "Should have generated some tokens"
     # Anthropic API should NOT include timings
     assert "timings" not in res.body, "Anthropic API should not include timings field"
@@ -168,6 +171,9 @@ def test_anthropic_messages_streaming():
     assert message_start["message"]["content"] == []
     assert "usage" in message_start["message"]
     assert message_start["message"]["usage"]["input_tokens"] > 0
+    assert "cache_read_input_tokens" in message_start["message"]["usage"]
+    assert isinstance(message_start["message"]["usage"]["cache_read_input_tokens"], int)
+    assert message_start["message"]["usage"]["cache_read_input_tokens"] >= 0
 
     # Check content_block_start
     block_start = next(e for e in events if e["type"] == "content_block_start")