Opt tp: tp attn support tp reduce scattered input (#10568)

python/sglang/srt/server_args.py

@@ -524,6 +524,7 @@ class ServerArgs:
     numa_node: Optional[List[int]] = None
     enable_deterministic_inference: bool = False
     rl_on_policy_target: Optional[str] = None
+    enable_attn_tp_input_scattered: bool = False
 
     # Dynamic batch tokenizer
     enable_dynamic_batch_tokenizer: bool = False
@@ -3481,6 +3482,11 @@ class ServerArgs:
             choices=RL_ON_POLICY_TARGET_CHOICES,
             help="The training system that SGLang needs to match for true on-policy.",
         )
+        parser.add_argument(
+            "--enable-attn-tp-input-scattered",
+            action="store_true",
+            help="Allow input of attention to be scattered when only using tensor parallelism, to reduce the computational load of operations such as qkv latent.",
+        )
 
         # Dynamic batch tokenizer
         parser.add_argument(