fixed Status field semantic mismatch and Timing left partially error

Author: abhisekupadhyaya

sdk/runanywhere-commons/include/rac/backends/rac_llm_llamacpp.h

@@ -177,7 +177,12 @@ RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_generate_stream(
  * @param options Generation options
  * @param callback Callback for each token
  * @param user_data User context passed to callback
- * @param timing_out Output: Benchmark timing (can be NULL for no timing)
+ * @param timing_out Output: Benchmark timing struct, caller-allocated.
+ *                   Must remain valid for the duration of the call.
+ *                   Caller should initialize via rac_benchmark_timing_init() before passing.
+ *                   On success, all t2/t3/t5 fields are populated.
+ *                   On failure, status is set but timing fields may be partial.
+ *                   Pass NULL to skip timing (zero overhead).
  * @return RAC_SUCCESS or error code
  */
 RAC_LLAMACPP_API rac_result_t rac_llm_llamacpp_generate_stream_with_timing(

sdk/runanywhere-commons/include/rac/core/rac_benchmark.h

@@ -67,9 +67,12 @@ typedef struct rac_benchmark_timing {
     int32_t output_tokens;
 
     /**
-     * Status of the request:
-     * - 0: Success
-     * - Non-zero: Error code (from rac_result_t)
+     * Status of the benchmark request.
+     * Uses RAC_BENCHMARK_STATUS_* codes:
+     * - RAC_BENCHMARK_STATUS_SUCCESS (0): Completed successfully
+     * - RAC_BENCHMARK_STATUS_ERROR (1): Failed
+     * - RAC_BENCHMARK_STATUS_TIMEOUT (2): Timed out
+     * - RAC_BENCHMARK_STATUS_CANCELLED (3): Cancelled
      */
     int32_t status;
 

sdk/runanywhere-commons/include/rac/features/llm/rac_llm_component.h

@@ -217,7 +217,13 @@ RAC_API rac_result_t rac_llm_component_generate_stream(
  * @param complete_callback Called when generation completes
  * @param error_callback Called on error
  * @param user_data User context passed to callbacks
- * @param timing_out Output: Benchmark timing (can be NULL for no timing)
+ * @param timing_out Output: Benchmark timing struct, caller-allocated.
+ *                   Must remain valid for the duration of the call.
+ *                   Caller should initialize via rac_benchmark_timing_init() before passing.
+ *                   Component fills t0/t4/t6, backend fills t2/t3/t5.
+ *                   On success, all timing fields are populated.
+ *                   On failure, status is set but timing fields may be partial.
+ *                   Pass NULL to skip timing (zero overhead).
  * @return RAC_SUCCESS or error code
  */
 RAC_API rac_result_t rac_llm_component_generate_stream_with_timing(

sdk/runanywhere-commons/src/backends/llamacpp/llamacpp_backend.cpp

@@ -883,6 +883,11 @@ bool LlamaCppTextGeneration::generate_stream_with_timing(const TextGenerationReq
 
     if (llama_decode(context_, batch) != 0) {
         LOGE("llama_decode failed for prompt");
+        if (timing_out != nullptr) {
+            int64_t now = rac_monotonic_now_ms();
+            timing_out->t3_prefill_end_ms = now;
+            timing_out->t5_last_token_ms = now;
+        }
         llama_batch_free(batch);
         return false;
     }