From 4ff09cec45230f020028c70e498d6d3d6a285748 Mon Sep 17 00:00:00 2001
From: wuli666 <421774554@qq.com>
Date: Wed, 7 Oct 2026 15:15:26 +0800
Subject: [PATCH] fix: restore the device token when an async request resumes
 in a new slot

On CUDA async, PRELUDE reads its input from AsyncState.tokens[slot], not
from request.input_token_tensor. Preemption snapshots and restores the
request's tensor, but release() returns the slot, and ensure_slot() only
imported the hidden state into the new one. A victim preempted while queued
for PRELUDE could resume in a slot holding another request's sample.

Import the token alongside the hidden state when a slot is assigned.
---
 tests/test_prefix_growth.py    | 47 ++++++++++++++++++++++++++++++++++
 vllm_rlt/worker/async_state.py | 11 +++++---
 2 files changed, 54 insertions(+), 4 deletions(-)

diff --git a/tests/test_prefix_growth.py b/tests/test_prefix_growth.py
index 672d12f6..a82c2aad 100644
--- a/tests/test_prefix_growth.py
+++ b/tests/test_prefix_growth.py
@@ -271,3 +271,50 @@ def test_async_pressure_preemption_with_resident_state():
         if blocks == 24:
             assert e.preemption.preemptions > 0
     assert results[0] == results[1]
+
+
+@pytest.mark.gpu
+def test_async_preemption_restores_prelude_token_in_a_new_slot():
+    # A preempted request releases its async slot. A later arrival takes it
+    # (LIFO), so the victim resumes in another slot whose device token belongs
+    # to a different request. Distinct prompts make a stale token visible.
+    from vllm_rlt import ExecutionConfig, ExitConfig
+    from vllm_rlt.request import Stage
+
+    torch.manual_seed(72)
+    m = OuroForCausalLM(tiny_ouro_config(head_dim=64)).to(device="cuda", dtype=torch.bfloat16)
+    results, victim_stages = [], []
+    for blocks in [128, 24]:
+        e = LLMEngine(
+            m,
+            attention_backend="triton",
+            cache_config=CacheConfig(blocks, 2, incremental_allocation=True),
+            scheduler_config=SchedulerConfig(
+                max_num_seqs=2,
+                max_num_batched_tokens=2,
+                prefill_chunk_size=2,
+                enable_preemption=True,
+            ),
+            exit_config=ExitConfig("trace", depths_by_request={"t": [4, 1, 2, 3, 4, 1, 2, 3]}),
+            execution_config=ExecutionConfig(async_scheduling=True),
+        )
+        select = e.preemption._select_preemption_victim
+
+        def record_victim(*args, **kwargs):
+            victim = select(*args, **kwargs)
+            if victim is not None:
+                victim_stages.append(victim.stage)
+            return victim
+
+        e.preemption._select_preemption_victim = record_victim
+        for i in range(4):
+            e.add_request(
+                str(i),
+                [2 * i + 1, 2 * i + 2],
+                SamplingParams(max_tokens=8, min_loops=1, ignore_eos=True),
+                trace_id="t",
+            )
+        results.append({rid: out.token_ids for rid, out in finish(e).items()})
+        assert not e.preemption.snapshots and e.cache_manager.num_used_blocks == 0
+    assert Stage.PRELUDE in victim_stages
+    assert results[0] == results[1]
diff --git a/vllm_rlt/worker/async_state.py b/vllm_rlt/worker/async_state.py
index 90156370..2b289d90 100644
--- a/vllm_rlt/worker/async_state.py
+++ b/vllm_rlt/worker/async_state.py
@@ -48,9 +48,9 @@ def transfer(self, batch=None):
 
         for slot, table in self.uploads:
             self.owner.tables[slot, :, : table.shape[1]].copy_(table, non_blocking=True)
-        for slot, hidden in self.imports:
-            hidden.record_stream(torch.cuda.current_stream(self.owner.cache.device))
-            self.owner.hidden[slot].copy_(hidden)
+        for pool, slot, value in self.imports:
+            value.record_stream(torch.cuda.current_stream(self.owner.cache.device))
+            pool[slot].copy_(value)
         self.descriptor = (
             self.host
             if self.owner.use_uva
@@ -162,8 +162,11 @@ def ensure_slot(self, request, bank):
         bank.uploads.append((slot, table))
         self.table_versions[rid] = allocation.block_tables
         if request.hidden_state is not None:
-            bank.imports.append((slot, request.hidden_state))
+            bank.imports.append((self.hidden, slot, request.hidden_state))
             request.hidden_state = self.hidden[slot]
+        # A resumed PRELUDE reads its input from the slot, not the request.
+        if request.input_token_tensor is not None:
+            bank.imports.append((self.tokens, slot, request.input_token_tensor))
         return slot
 
     def prepare(self, requests, depths, positions, size, *, recurrent=False, finalize=False):