[V1][Spec Decoding] Log accumulated metrics after system goes idle (#17913)

Signed-off-by: Mark McLoughlin <[email protected]>

Author: markmc

vllm/v1/metrics/loggers.py

@@ -128,9 +128,7 @@ def log(self):
             scheduler_stats.gpu_cache_usage * 100,
             self.prefix_caching_metrics.hit_rate * 100,
         )
-
-        if scheduler_stats.spec_decoding_stats is not None:
-            self.spec_decoding_logging.log(log_fn=log_fn)
+        self.spec_decoding_logging.log(log_fn=log_fn)
 
     def log_engine_initialized(self):
         logger.info(

vllm/v1/spec_decode/metrics.py

@@ -67,6 +67,8 @@ def observe(self, spec_decoding_stats: SpecDecodingStats):
             spec_decoding_stats.num_accepted_tokens_per_pos)
 
     def log(self, log_fn=logger.info):
+        if not self.num_drafts:
+            return
         num_drafts = np.sum(self.num_drafts)
         num_draft_tokens = np.sum(self.num_draft_tokens)
         num_accepted_tokens = np.sum(self.num_accepted_tokens)