diff --git a/vllm/v1/sample/ops/topk_topp_sampler.py b/vllm/v1/sample/ops/topk_topp_sampler.py
index 363b113f0a4f..cc3120233814 100644
--- a/vllm/v1/sample/ops/topk_topp_sampler.py
+++ b/vllm/v1/sample/ops/topk_topp_sampler.py
@@ -289,6 +289,11 @@ def forward_xpu(
         torch.ops.vllm.xpu_topk_topp_sampler(
             random_sampled, logits_to_return, logits, k, p, self.logprobs_mode, seeds
         )
+        # The custom XPU sampler kernel consumes RNG values internally, so advance
+        # the default generator's offset to keep future draws deterministic.
+        offset += logits.numel()
+        state.view(torch.int64)[1] = offset
+        generator.set_state(state)
         return random_sampled, logits_to_return