[PERF] Qwen3-next MTP speedup (change bool mask indexing to index_select / index_copy to reduce d2h) (#26437)

Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>

vllm/model_executor/layers/fla/ops/utils.py

@@ -45,7 +45,7 @@ def tensor_cache(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]
     """
 
     cache_entries: tuple[tuple | None, dict | None, Any] = []
-    cache_size = 4
+    cache_size = 8
 
     @functools.wraps(fn)
     def wrapper(*args: Any, **kwargs: Any) -> Any: