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

2025-12-06 19:57:26 -08:00
parent 91c9c14c28
commit 6d5d76ad97
4 changed files with 0 additions and 8 deletions
@@ -304,11 +304,9 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
        if get_moe_a2a_backend().is_deepep():
            return self._forward_deepep(hidden_states, forward_batch)

-        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(