Minor fixes for server_args, parallel_state, and test_deterministic.py (#11159)

python/sglang/srt/distributed/parallel_state.py

@@ -1085,7 +1085,8 @@ class GroupCoordinator:
 
             comm_group = metadata_group if tensor.is_cpu else group
             work = send_func(tensor, self.ranks[dst], group=comm_group)
-            p2p_works.append(P2PWork(work, tensor))
+            if async_send:
+                p2p_works.append(P2PWork(work, tensor))
         return p2p_works
 
     def recv_tensor_dict(

python/sglang/srt/metrics/collector.py

@@ -809,13 +809,13 @@ class TokenizerMetricsCollector:
     def check_time_to_first_token_straggler(self, value: float) -> bool:
         his = self.histogram_time_to_first_token.labels(**self.labels)
         total_observations = sum(bucket._value for bucket in his._buckets)
-        if total_observations < 100:
+        if total_observations < 1000:
             return False
-        p99_threshold = total_observations * 0.99
+        p999_threshold = total_observations * 0.999
         cumulative_count = 0
         for i, bucket in enumerate(his._buckets):
             cumulative_count += bucket._value
-            if cumulative_count > p99_threshold:
+            if cumulative_count > p999_threshold:
                 return value >= his._upper_bounds[i]
         return False
 

python/sglang/srt/server_args.py

@@ -116,6 +116,8 @@ GRAMMAR_BACKEND_CHOICES = ["xgrammar", "outlines", "llguidance", "none"]
 
 DETERMINISTIC_ATTENTION_BACKEND_CHOICES = ["flashinfer", "fa3", "triton"]
 
+RADIX_EVICTION_POLICY_CHOICES = ["lru", "lfu"]
+
 
 # Allow external code to add more choices
 def add_load_format_choices(choices):
@@ -138,6 +140,14 @@ def add_grammar_backend_choices(choices):
     GRAMMAR_BACKEND_CHOICES.extend(choices)
 
 
+def add_deterministic_attention_backend_choices(choices):
+    DETERMINISTIC_ATTENTION_BACKEND_CHOICES.extend(choices)
+
+
+def add_radix_eviction_policy_choices(choices):
+    RADIX_EVICTION_POLICY_CHOICES.extend(choices)
+
+
 @dataclasses.dataclass
 class ServerArgs:
     # Model and tokenizer
@@ -2243,7 +2253,7 @@ class ServerArgs:
         parser.add_argument(
             "--radix-eviction-policy",
             type=str,
-            choices=["lru", "lfu"],
+            choices=RADIX_EVICTION_POLICY_CHOICES,
             default=ServerArgs.radix_eviction_policy,
             help="The eviction policy of radix trees. 'lru' stands for Least Recently Used, 'lfu' stands for Least Frequently Used.",
         )

python/sglang/test/test_deterministic.py

@@ -29,7 +29,7 @@ class BenchArgs:
     port: int = 30000
     batch_size: int = 1
     temperature: float = 0.0
-    sampling_seed: int = None
+    sampling_seed: int = 42
     max_new_tokens: int = 100
     frequency_penalty: float = 0.0
     presence_penalty: float = 0.0