Fix using ended span

neuron/neuron/submission_tester/testing.py

@@ -150,7 +150,7 @@ def compare_checkpoints(
     vram_used = max(output.vram_used for output in outputs) - start_vram
     watts_used = max(output.watts_used for output in outputs)
 
-    with tracer.start_span("compare_outputs") as compare_span:
+    with tracer.start_span("compare_outputs"):
         with CURRENT_CONTEST.output_comparator() as output_comparator:
             def calculate_similarity(comparator: OutputComparator, baseline_output: GenerationOutput, optimized_output: GenerationOutput):
                 try:
@@ -191,18 +191,6 @@ def calculate_similarity(comparator: OutputComparator, baseline_output: Generati
         min_similarity=min_similarity,
     )
 
-    compare_span.set_attributes({
-        "benchmark.samples": len(inputs),
-        "benchmark.score": benchmark.calculate_score(baseline.metric_data),
-        "benchmark.average_similarity": average_similarity,
-        "benchmark.min_similarity": min_similarity,
-        "benchmark.average_generation_time": average_time,
-        "benchmark.model_size": size,
-        "benchmark.model_load_time": sandbox.load_time,
-        "benchmark.vram_usage": vram_used,
-        "benchmark.power_usage": watts_used,
-    })
-
     logger.info(
         f"Tested {len(inputs)} Samples\n"
         f"Score: {benchmark.calculate_score(baseline.metric_data)}\n"

validator/weight_setting/validator.py

@@ -815,6 +815,7 @@ def get_uid(hotkey: Key) -> Uid | None:
 
         self.sleep_for_blocks(now, epoch_length / 4, "Benchmarking in progress")
 
+    @tracer.start_as_current_span("sleep_for_blocks")
     def sleep_for_blocks(self, now: datetime, blocks: int, reason: str):
         next_noon = datetime.combine(now.date() + timedelta(days=int(now.hour >= 12)), time(12), tzinfo=TIMEZONE)
         blocks_to_sleep = min(blocks, ceil((next_noon - now).total_seconds() / 12))