[V1] Prompt logprobs + APC compatibility; prompt logprobs reqs cannot fill APC (vllm-project#13949)

Author: afeldman-nm

tests/v1/core/test_prefix_caching.py

@@ -1,6 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 """Compare the with and without prefix caching."""
 
+from typing import Optional
+
 import pytest
 
 from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
@@ -15,7 +17,8 @@
 def make_request(request_id,
                  prompt_token_ids,
                  mm_positions=None,
-                 mm_hashes=None):
+                 mm_hashes=None,
+                 prompt_logprobs: Optional[int] = None):
     if mm_positions is None:
         multi_modal_inputs = None
     else:
@@ -28,7 +31,8 @@ def make_request(request_id,
         multi_modal_inputs=multi_modal_inputs,
         multi_modal_hashes=mm_hashes,
         multi_modal_placeholders=mm_positions,
-        sampling_params=SamplingParams(max_tokens=17),
+        sampling_params=SamplingParams(max_tokens=17,
+                                       prompt_logprobs=prompt_logprobs),
         eos_token_id=100,
         arrival_time=0,
         lora_request=None,
@@ -144,6 +148,110 @@ def test_prefill():
     assert manager.block_pool.free_block_queue.free_list_tail is None
 
 
+def test_prefill_plp():
+    '''Test prefill with APC and some prompt logprobs (plp) requests.
+
+    1. Schedule plp request and validate APC block allocation
+    2. Schedule non-plp request and validate blocks
+    3. Schedule plp request; no hit should occur; validate blocks
+    '''
+    manager = KVCacheManager(
+        block_size=16,
+        num_gpu_blocks=10,
+        max_model_len=8192,
+        sliding_window=None,
+        enable_caching=True,
+        num_preallocate_tokens=16,
+    )
+
+    # Complete 3 blocks (48 tokens)
+    common_token_ids = [i for i in range(3) for _ in range(16)]
+
+    # Request #0 is a prompt logprobs request
+    # Fully cache miss
+    # Incomplete 1 block (7 tokens)
+    unique_token_ids = [3] * 7
+    all_token_ids = common_token_ids + unique_token_ids
+    req0 = make_request("0", all_token_ids, prompt_logprobs=5)
+    computed_blocks, num_computed_tokens = manager.get_computed_blocks(req0)
+    assert len(manager.req_to_block_hashes[req0.request_id]) == 3
+    assert not computed_blocks
+    assert num_computed_tokens == 0
+    blocks = manager.allocate_slots(req0, 55, computed_blocks)
+    assert [b.block_id for b in blocks] == [0, 1, 2, 3, 4]
+    req0_block_hashes = [b.block_hash for b in blocks]
+
+    # Check full block metadata
+    parent_block_hash = None
+    for block_id in (0, 1, 2):
+        block_tokens = tuple(all_token_ids[block_id * 16:(block_id + 1) * 16])
+        block_hash = hash_block_tokens(parent_block_hash, block_tokens)
+        assert manager.block_pool.blocks[block_id].block_hash == block_hash
+        assert manager.block_pool.blocks[block_id].ref_cnt == 1
+        parent_block_hash = block_hash.hash_value
+
+    # Check partial/preallocated block metadata
+    for block_id in (3, 4):
+        assert manager.block_pool.blocks[block_id].block_hash is None
+        assert manager.block_pool.blocks[block_id].ref_cnt == 1
+
+    # Request #1 is a non-prompt-logprobs request:
+    # Cache hit in the common prefix when the original block is still in use.
+    # Incomplete 1 block (5 tokens)
+    unique_token_ids = [3] * 5
+    req1 = make_request("1", common_token_ids + unique_token_ids)
+    computed_blocks, num_computed_tokens = manager.get_computed_blocks(req1)
+    assert len(manager.req_to_block_hashes[req1.request_id]) == 3
+    assert [b.block_id for b in computed_blocks] == [0, 1, 2]
+    assert num_computed_tokens == 3 * 16
+    num_new_tokens = 53 - 3 * 16
+    blocks = manager.allocate_slots(req1, num_new_tokens, computed_blocks)
+    assert [b.block_id for b in blocks] == [5, 6]
+    for block in computed_blocks:
+        assert block.ref_cnt == 2
+
+    # At this point, we should have 3 free blocks left.
+    assert manager.block_pool.free_block_queue.num_free_blocks == 3
+
+    manager.free(req0)
+    manager.free(req1)
+
+    # All blocks should be available.
+    assert manager.block_pool.free_block_queue.num_free_blocks == 10
+    # The order should be
+    # [unallocated (7, 8, 9)]
+    # [unique_req0 (4, 3)]
+    # [unique_req1 (6, 5)]
+    # [common (2, 1, 0)]
+    assert [
+        b.block_id
+        for b in manager.block_pool.free_block_queue.get_all_free_blocks()
+    ] == [7, 8, 9, 4, 3, 6, 5, 2, 1, 0]
+
+    # Request #2 is a prompt-logprobs request:
+    # NO cache hit in the common prefix; duplicates request #0 cached blocks
+    unique_token_ids = [3] * 6
+    req2 = make_request("2",
+                        common_token_ids + unique_token_ids,
+                        prompt_logprobs=5)
+    computed_blocks, num_computed_tokens = manager.get_computed_blocks(req2)
+    assert len(manager.req_to_block_hashes[req2.request_id]) == 3
+    assert not computed_blocks
+    assert num_computed_tokens == 0
+    blocks = manager.allocate_slots(req2, 55, computed_blocks)
+    block_ids = [b.block_id for b in blocks]
+    # Duplicate cached blocks have different ids but same hashes vs request #0
+    assert [b.block_hash for b in blocks] == req0_block_hashes
+    assert block_ids != [0, 1, 2, 3, 4]
+
+    # Request #2 block hashes are valid since request #0 hashes are.
+    # Check block reference counts.
+    for block_id in block_ids:
+        assert manager.block_pool.blocks[block_id].ref_cnt == 1
+
+    manager.free(req2)
+
+
 def test_decode():
     manager = KVCacheManager(
         block_size=16,

tests/v1/core/test_scheduler.py

@@ -1,6 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 from typing import Optional
 
+import pytest
+
 from vllm.config import CacheConfig, ModelConfig, SchedulerConfig, VllmConfig
 from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
 from vllm.sampling_params import SamplingParams
@@ -16,7 +18,21 @@ def create_scheduler(
     model: str = "facebook/opt-125m",
     max_num_seqs: int = 16,
     max_num_batched_tokens: int = 8192,
+    enable_prefix_caching: Optional[bool] = None,
 ) -> Scheduler:
+    '''Create scheduler under test.
+    
+    Args:
+      model: model under test
+      max_num_seqs: max sequences to schedule
+      max_num_batch_tokens: max num tokens to batch
+      enable_prefix_caching: optionally force APC config
+                             (True/False) or use default
+                             (None)
+
+    Returns:
+      :class:`Scheduler` instance
+    '''
     scheduler_config = SchedulerConfig(
         max_num_seqs=max_num_seqs,
         max_num_batched_tokens=max_num_batched_tokens,
@@ -31,11 +47,16 @@ def create_scheduler(
         dtype="float16",
         seed=42,
     )
+    # Cache config, optionally force APC
+    kwargs_cache = ({} if enable_prefix_caching is None else {
+        'enable_prefix_caching': enable_prefix_caching
+    })
     cache_config = CacheConfig(
         block_size=16,
         gpu_memory_utilization=0.9,
         swap_space=0,
         cache_dtype="auto",
+        **kwargs_cache,
     )
     vllm_config = VllmConfig(
         scheduler_config=scheduler_config,
@@ -54,16 +75,16 @@ def create_scheduler(
     )
 
 
-def create_requests(
-    num_requests: int,
-    num_tokens: int = 10,
-    mm_positions: Optional[list[PlaceholderRange]] = None,
-    max_tokens: int = 16,
-    stop_token_ids: Optional[list[int]] = None,
-):
+def create_requests(num_requests: int,
+                    num_tokens: int = 10,
+                    mm_positions: Optional[list[PlaceholderRange]] = None,
+                    max_tokens: int = 16,
+                    stop_token_ids: Optional[list[int]] = None,
+                    prompt_logprobs: Optional[int] = None):
     sampling_params = SamplingParams(ignore_eos=False,
                                      max_tokens=max_tokens,
-                                     stop_token_ids=stop_token_ids)
+                                     stop_token_ids=stop_token_ids,
+                                     prompt_logprobs=prompt_logprobs)
     requests = []
     for i in range(num_requests):
         if mm_positions is not None:
@@ -122,9 +143,18 @@ def test_get_num_unfinished_requests():
         assert scheduler.get_num_unfinished_requests() == len(requests) - i - 1
 
 
-def test_schedule():
-    scheduler = create_scheduler()
-    requests = create_requests(num_requests=10)
+@pytest.mark.parametrize("enable_prefix_caching, prompt_logprobs", [
+    (None, None),
+    (True, 5),
+])
+def test_schedule(enable_prefix_caching: Optional[bool],
+                  prompt_logprobs: Optional[int]):
+    '''Test scheduling. 
+    Two cases: default APC/no prompt logprobs; APC=True + prompt logprobs
+    '''
+    scheduler = create_scheduler(enable_prefix_caching=enable_prefix_caching)
+    requests = create_requests(num_requests=10,
+                               prompt_logprobs=prompt_logprobs)
     for request in requests:
         scheduler.add_request(request)
 
@@ -427,14 +457,21 @@ def test_stop_via_update_from_output():
     assert list(requests[0].output_token_ids) == [EOS_TOKEN_ID, 10, 11]
 
 
-def test_schedule_concurrent_batches():
+@pytest.mark.parametrize("enable_prefix_caching, prompt_logprobs", [
+    (None, None),
+    (True, 5),
+])
+def test_schedule_concurrent_batches(enable_prefix_caching: Optional[bool],
+                                     prompt_logprobs: Optional[int]):
     scheduler = create_scheduler(
         max_num_batched_tokens=1024,
         max_num_seqs=2,
+        enable_prefix_caching=enable_prefix_caching,
     )
     requests = create_requests(
         num_requests=2,
         num_tokens=512,
+        prompt_logprobs=prompt_logprobs,
     )
 
     # Schedule the first request.

tests/v1/engine/test_async_llm.py

@@ -6,7 +6,6 @@
 
 import pytest
 
-from tests.v1.engine.utils import PLP_APC_UNSUPPORTED_MSG
 from vllm import SamplingParams
 from vllm.assets.image import ImageAsset
 from vllm.engine.arg_utils import AsyncEngineArgs
@@ -72,41 +71,6 @@ async def generate(engine: AsyncLLM,
     return count, request_id
 
 
-@pytest.mark.parametrize(
-    "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY])
-@pytest.mark.asyncio
-async def test_async_llm_refuses_prompt_logprobs_with_apc(
-        monkeypatch, output_kind: RequestOutputKind):
-    """Test passes if AsyncLLM raises an exception when it is configured
-    for automatic prefix caching and it receives a request with
-    prompt_logprobs enabled, which is incompatible."""
-    # TODO(rickyx): Remove monkeypatch VLLM_USE_V1 setting once we have a
-    # better way to test V1 so that in the future when we switch, we don't
-    # have to change all the tests.
-    monkeypatch.setenv("VLLM_USE_V1", "1")
-    # Create AsyncLLM engine with APC
-    apc_engine_args = AsyncEngineArgs(model="facebook/opt-125m",
-                                      enable_prefix_caching=True,
-                                      gpu_memory_utilization=0.8,
-                                      disable_log_requests=True)
-    engine = AsyncLLM.from_engine_args(apc_engine_args)
-    try:
-        with pytest.raises(ValueError) as excinfo:
-            # Issue a request with prompt logprobs enabled, which should fail
-            await asyncio.create_task(
-                generate(engine,
-                         "request-0",
-                         TEXT_PROMPT,
-                         output_kind,
-                         10,
-                         prompt_logprobs=5))
-        # Validate exception string is correct
-        assert str(excinfo.value) == PLP_APC_UNSUPPORTED_MSG
-    finally:
-        # Shut down engine
-        engine.shutdown()
-
-
 @pytest.mark.parametrize(
     "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY])
 @pytest.mark.parametrize("engine_args_and_prompt",

tests/v1/engine/test_llm_engine.py

@@ -5,7 +5,6 @@
 
 import pytest
 
-from tests.v1.engine.utils import PLP_APC_UNSUPPORTED_MSG
 from vllm import LLM, SamplingParams
 
 MODEL = "facebook/opt-125m"
@@ -98,17 +97,3 @@ def test_parallel_sampling(vllm_model, example_prompts) -> None:
             raise AssertionError(
                 f"{len(completion_counts)} unique completions; expected"
                 f" {n}. Repeats: {repeats}")
-
-
-def test_llm_engine_refuses_prompt_logprobs_with_apc(vllm_model_apc):
-    """Test passes if LLMEngine raises an exception when it is configured
-    for automatic prefix caching and it receives a request with
-    prompt_logprobs enabled, which is incompatible."""
-    model: LLM = vllm_model_apc.model
-    with pytest.raises(ValueError) as excinfo:
-        model.generate(
-            "Hello, my name is",
-            SamplingParams(temperature=0.8, top_p=0.95, prompt_logprobs=5))
-
-    # Validate exception string is correct
-    assert str(excinfo.value) == PLP_APC_UNSUPPORTED_MSG

tests/v1/engine/utils.py

@@ -30,9 +30,6 @@
 STOP_STRINGS = ["I love working on", "company by far", "brother in"]
 PROMPT_LEN = 5
 
-PLP_APC_UNSUPPORTED_MSG = ("Prefix caching with prompt logprobs not yet "
-                           "supported on VLLM V1.")
-
 random.seed(42)