Filter non-int token ids in benchmark and observe decode-side bootstrap/alloc metrics (#24684)

python/sglang/benchmark/datasets/common.py

@@ -66,8 +66,12 @@ def compute_random_lens(full_len: int, range_ratio: float, num: int) -> List[int
 
 @lru_cache(maxsize=1)
 def get_available_tokens(tokenizer):
-    """Get all available token ids from the tokenizer vocabulary."""
-    return list(tokenizer.get_vocab().values())
+    """Get valid token ids from the tokenizer vocabulary."""
+    return [
+        token_id
+        for token_id in tokenizer.get_vocab().values()
+        if isinstance(token_id, int)
+    ]
 
 
 def gen_prompt(tokenizer, token_num):

python/sglang/srt/observability/req_time_stats.py

@@ -936,6 +936,16 @@ class SchedulerReqTimeStats(ReqTimeStatsBase):
         )
         self.trace_slice(stage, self.decode_prealloc_queue_entry_time, ts)
 
+        if self.enable_metrics and self.bootstrap_done_time > 0:
+            bootstrap_ms = (
+                self.bootstrap_done_time - self.decode_prealloc_queue_entry_time
+            ) * 1000
+            alloc_ms = (ts - self.bootstrap_done_time) * 1000
+            self.metrics_collector.observe_kv_transfer_bootstrap(
+                bootstrap_ms=bootstrap_ms,
+                alloc_ms=alloc_ms,
+            )
+
     def set_bootstrap_done_time(self, ts=None):
         ts = ts or time.perf_counter()
         if self.bootstrap_done_time == 0.0: