remove unecessary dual stream token threshold from the rest of models (qwen moe, kimi linear, etc.) (#14337)

python/sglang/srt/models/llada2.py

@@ -349,11 +349,9 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
         num_tokens, hidden_size = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_size)
 
-        DUAL_STREAM_TOKEN_THRESHOLD = 1024
         if (
             self.alt_stream is not None
             and hidden_states.shape[0] > 0
-            and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
             and get_is_capture_mode()
         ):
             final_hidden_states, shared_output = self.forward_normal_dual_stream(