[Core][KVConnector] Propagate all tokens on resumed preemptions (vllm-project#24926)

Signed-off-by: Qier Li <kevin44036@gmail.com>
Co-authored-by: Qier Li <qier@fb.com>
Signed-off-by: xuebwang-amd <xuebwang@amd.com>

Author: 2 people

tests/v1/core/test_scheduler.py

@@ -1950,7 +1950,7 @@ def test_schedule_skip_tokenizer_init_structured_output_request():
     assert len(scheduler.waiting) == 1
 
 
-def test_priority_scheduling_preemption_when_out_of_kv():
+def test_priority_scheduling_preemption_and_resumption_when_out_of_kv():
     """Test that priority scheduling preempts lower priority requests
     when out of KV cache space."""
     # Create scheduler with very limited memory to force preemption
@@ -1959,6 +1959,7 @@ def test_priority_scheduling_preemption_when_out_of_kv():
         max_num_batched_tokens=200,
         num_blocks=5,  # Can hold 64 tokens (first block is null)
         block_size=16,  # Standard block size
+        use_kv_connector=True,
     )
 
     # Create a request and schedule it
@@ -1970,12 +1971,13 @@ def test_priority_scheduling_preemption_when_out_of_kv():
         starting_idx=0,
     )[0]
     scheduler.add_request(request_low)
+    # 1st schedule
     output = scheduler.schedule()
     assert len(output.scheduled_new_reqs) == 1
     assert len(scheduler.waiting) == 0
     assert len(scheduler.running) == 1
 
-    # Simulate model execution
+    # Simulate model execution - 1st decode
     model_output = ModelRunnerOutput(
         req_ids=[request_low.request_id],
         req_id_to_index={request_low.request_id: 0},
@@ -1996,6 +1998,7 @@ def test_priority_scheduling_preemption_when_out_of_kv():
         starting_idx=1,
     )[0]
     scheduler.add_request(request_high)
+    # 2nd schedule
     output = scheduler.schedule()
     # KV cache should be full at this point
     assert scheduler.kv_cache_manager.block_pool.get_num_free_blocks() == 0
@@ -2004,7 +2007,7 @@ def test_priority_scheduling_preemption_when_out_of_kv():
     assert len(scheduler.waiting) == 0
     assert len(scheduler.running) == 2
 
-    # Simulate model execution
+    # Simulate model execution - 2nd decode
     requests = [request_low, request_high]
     model_output = ModelRunnerOutput(
         req_ids=[req.request_id for req in requests],
@@ -2017,7 +2020,7 @@ def test_priority_scheduling_preemption_when_out_of_kv():
     )
     scheduler.update_from_output(output, model_output)
 
-    # Schedule again - this should trigger preemption
+    # 3rd schedule - this should trigger preemption
     # req_low needs 32 tokens = 2 blocks
     # req_high needs 33 tokens = 3 blocks
     # so doesn't fit in 4 blocks.
@@ -2027,9 +2030,44 @@ def test_priority_scheduling_preemption_when_out_of_kv():
     assert len(output.scheduled_new_reqs) == 0
     assert output.scheduled_cached_reqs.num_reqs == 1
     assert output.scheduled_cached_reqs.req_ids[0] == request_high.request_id
+    assert scheduler.requests[request_low.request_id].status == RequestStatus.PREEMPTED
     assert len(scheduler.waiting) == 1
     assert len(scheduler.running) == 1
 
+    # Simulate model execution - 3rd decode
+    model_output = ModelRunnerOutput(
+        req_ids=[req.request_id for req in requests],
+        req_id_to_index={req.request_id: i for i, req in enumerate(requests)},
+        sampled_token_ids=[[], [100]],
+        # spec_token_ids=None,
+        logprobs=None,
+        prompt_logprobs_dict={},
+        pooler_output=[],
+    )
+    # Finish the requests to make room for the preempted requests to resume
+    scheduler.update_from_output(output, model_output)
+    scheduler.finish_requests(request_high.request_id, RequestStatus.FINISHED_STOPPED)
+
+    # 4th Schedule - this should trigger the resumption
+    output = scheduler.schedule()
+    scheduled_cached_reqs = output.scheduled_cached_reqs
+    resumed_from_preemption = scheduled_cached_reqs.resumed_from_preemption
+
+    assert len(output.scheduled_new_reqs) == 0
+    assert scheduled_cached_reqs.num_reqs == 1
+    assert len(scheduler.waiting) == 0
+    assert len(scheduler.running) == 1
+
+    # Preempted request resumed in scheduled_cached_reqs
+    assert len(resumed_from_preemption) == 1
+    assert len(scheduled_cached_reqs.resumed_req_token_ids) == 1
+    assert resumed_from_preemption[0]
+    assert scheduled_cached_reqs.req_ids[0] == request_low.request_id
+    assert scheduled_cached_reqs.resumed_req_token_ids[0] is not None
+    # Resumed tokens include 30 prompt tokens and 2 decoded tokens
+    assert len(scheduled_cached_reqs.resumed_req_token_ids[0]) == 32
+    assert scheduled_cached_reqs.resumed_req_token_ids[0][31] == 100
+
 
 @pytest.mark.parametrize(
     ("enable_chunked_prefill", "is_encoder_decoder", "expect_enabled"),

tests/v1/worker/test_gpu_model_runner.py

@@ -257,6 +257,7 @@ def test_update_states_request_resumed(model_runner, dist_init):
         req_ids=[req_id],
         resumed_from_preemption=[False],
         new_token_ids=[[]],
+        resumed_req_token_ids=[None],
         new_block_ids=([[0]],),
         num_computed_tokens=[0],
         num_output_tokens=[0],

vllm/v1/core/sched/output.py

@@ -98,6 +98,9 @@ class CachedRequestData:
     # NOTE(woosuk): new_token_ids is only used for pipeline parallelism.
     # When PP is not used, new_token_ids will be empty.
     new_token_ids: list[list[int]]
+    # If resumed_from_preemption is True, propogate the token ids to the
+    # connector, otherwise will be empty.
+    resumed_req_token_ids: list[list[int] | None]
     new_block_ids: list[tuple[list[int], ...] | None]
     num_computed_tokens: list[int]
     num_output_tokens: list[int]
@@ -112,6 +115,7 @@ def make_empty(cls) -> CachedRequestData:
             req_ids=[],
             resumed_from_preemption=[],
             new_token_ids=[],
+            resumed_req_token_ids=[],
             new_block_ids=[],
             num_computed_tokens=[],
             num_output_tokens=[],

vllm/v1/core/sched/scheduler.py

@@ -709,10 +709,15 @@ def _make_cached_request_data(
         req_ids: list[str] = []
         new_token_ids: list[list[int]] = []
         new_block_ids: list[tuple[list[int], ...] | None] = []
+        resumed_req_token_ids: list[list[int] | None] = []
         num_computed_tokens: list[int] = []
         num_output_tokens: list[int] = []
 
-        for req in itertools.chain(running_reqs, resumed_reqs):
+        # Because resumed_reqs is usually empty, it is more efficient to do
+        # in-place appending so that we don't need to allocate a new list.
+        resumed_from_preemption = [False] * len(running_reqs)
+        resumed_from_preemption += [True] * len(resumed_reqs)
+        for idx, req in enumerate(itertools.chain(running_reqs, resumed_reqs)):
             req_id = req.request_id
             req_ids.append(req_id)
             num_tokens = num_scheduled_tokens[req_id] - len(
@@ -728,20 +733,23 @@ def _make_cached_request_data(
                     req.num_computed_tokens : req.num_computed_tokens + num_tokens
                 ]
                 new_token_ids.append(token_ids)
+            resumed_token_ids = None
+            if resumed_from_preemption[idx]:
+                resumed_token_ids = req.all_token_ids[
+                    : req.num_computed_tokens + num_tokens
+                ]
+            resumed_req_token_ids.append(resumed_token_ids)
             new_block_ids.append(
                 req_to_new_blocks[req_id].get_block_ids(allow_none=True)
             )
             num_computed_tokens.append(req.num_computed_tokens)
             num_output_tokens.append(req.num_output_tokens)
-        # Because resumed_reqs is usually empty, it is more efficient to do
-        # in-place appending so that we don't need to allocate a new list.
-        resumed_from_preemption = [False] * len(running_reqs)
-        resumed_from_preemption += [True] * len(resumed_reqs)
 
         return CachedRequestData(
             req_ids=req_ids,
             resumed_from_preemption=resumed_from_preemption,
             new_token_ids=new_token_ids,
+            resumed_req_token_ids=resumed_req_token_ids,
             new_block_ids=new_block_ids,
             num_computed_tokens=num_computed_tokens,
             num_output_tokens=num_output_tokens,