[DSA] Fix IMA in fused top-k v2: write all output slots on tie overflow (#30512)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

python/sglang/jit_kernel/include/sgl_kernel/deepseek_v4/topk_impl.cuh

@@ -209,6 +209,15 @@ struct TopKConfig {
 
     if (num_ties <= topk) {
       if (tx < num_ties) problem.emit(base + tx, tie_buffer[tx].idx);
+      // Fewer tie candidates than remaining slots (ties beyond kMaxNumTie are
+      // dropped at collect): pad [num_ties, topk) with -1 ("no token"). The
+      // transform pass reads all `topk` output slots, and any slot left
+      // unwritten holds uninitialized staging memory whose page-table
+      // translation yields a garbage KV index (-> illegal memory access in
+      // the downstream sparse attention kernel).
+      for (uint32_t t = num_ties + tx; t < topk; t += kBlockSize) {
+        problem.emit(base + t, -1u);
+      }
     } else if (num_ties <= kWarpSize) {
       if (lane_id >= num_ties || warp_id >= num_ties) return;  // some threads are idle
       /// NOTE: use long long to avoid mask overflow when num_tie == 32