[Core] Combined support for multi-step scheduling, chunked prefill & prefix caching (#8804)

Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>
Co-authored-by: Andrew Feldman <afeld2012@gmail.com>

tests/multi_step/test_correctness_llm.py

@@ -1,5 +1,6 @@
 # Test the LLMEngine with multi-step-decoding
 
+import copy
 from typing import Optional
 
 import pytest
@@ -196,3 +197,160 @@ def test_multi_step_llm_w_prompt_logprobs(
         name_0="hf",
         name_1="vllm",
     )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("tp_size", [1])
+@pytest.mark.parametrize("max_tokens", [5])
+@pytest.mark.parametrize("enforce_eager", [True])
+@pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS)
+@pytest.mark.parametrize("num_prompts", NUM_PROMPTS)
+@pytest.mark.parametrize("num_logprobs", [None, 5])
+def test_multi_step_llm_chunked_prefill_prefix_cache(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    tp_size: int,
+    max_tokens: int,
+    enforce_eager: int,
+    num_scheduler_steps: int,
+    num_prompts: int,
+    num_logprobs: Optional[int],
+) -> None:
+    """Test vLLM engine with multi-step+"single-step chunked prefill"+APC.
+
+    Set up contrived scenario which tests for a possible failure mode of
+    scheduling with multi-step+"single-step chunked prefill"+APC
+
+    "single-step chunked prefill" here refers to the current vLLM multi-step+
+    chunked-prefill implementation, which requires that a prefill may only
+    be scheduled in the same step as decodes if the prefill prompt fits in a
+    single chunk (note that "complete" multi-step+chunked-prefill would allow
+    a prefill to span multiple chunks & multiple steps but that is not yet
+    the case.)
+
+    "APC" is short for "automatic prefix caching".
+
+    This test creates a scenario where the scheduler must decide whether/how
+    to schedule a prefill with a prompt that exceeds the available token budget.
+    The correct behavior for multi-step+"single-step chunked prefill"+APC is to
+    put off scheduling the prefill until a future step.
+
+    Validate that:
+    * Multi-step kernels do not raise an exception due to incorrect scheduler
+      behavior
+    * Generated tokens match between
+      multi-step+"single-step chunked prefill"+APC and
+      single-step scheduling.
+    * (If logprobs are enabled) check logprobs are close enough
+
+    Args:
+      vllm_runner: vLLM model runner fixture
+      example_prompts: test fixture providing example prompts
+      model: model under test (same for single- and multi-step engines)
+      dtype: tensor datatype for engine to utilize
+      tp_size: degree of tensor-parallelism
+      max_tokens: the maximum number of tokens to generate
+      enforce_eager
+      num_scheduler_steps: for multi-step scheduling, GPU-side steps per
+                           GPU -> CPU output transfer
+      num_prompts: number of example prompts under test
+      num_logprobs: corresponds to the `logprobs` argument to the OpenAI
+                    completions endpoint; `None` -> 1 logprob returned.
+    """
+
+    # Set up contrived test for correct scheduling behavior with
+    # multi-step+"single-step chunked prefill"+APC.
+    #
+    # Assume block_size=16
+    #
+    # Assume max_num_batched_tokens=48
+    #   => Per-step token budget=48
+    #
+    # 1. Scheduler schedules 0th prompt (24 tokens)
+    #      => Remaining token budget=24
+    # 2. Scheduler attempts to schedule 1st prompt (30 tokens)
+    #    * 30 tokens exceeds 24 token remaining budget
+    #    * Correct behavior: do not schedule this prompt in this step
+    #    * Incorrect behavior: schedule prompt chunk
+    #      * `do_sample=False` for this prompt in this step
+    #      * Chunk size = (remaining tokens // block size) * block size
+    #
+    # The Incorrect scheduling behavior - if it occurs - will cause an exception
+    # in the model runner resulting from `do_sample=False`.
+    assert len(example_prompts) >= 2
+    challenge_prompts = copy.deepcopy(example_prompts)
+    challenge_prompts[0] = ('vLLM is a high-throughput and memory-efficient '
+                            'inference and serving engine for LLMs.\n'
+                            )  # 24 tok
+    challenge_prompts[1] = (
+        'Briefly describe the major milestones in the '
+        'development of artificial intelligence from 1950 to 2020.\n'
+    )  # 30 tok
+
+    # If necessary, adjust the length of `challenge_prompts` to match
+    # `num_prompts`
+    if len(challenge_prompts) < num_prompts:
+        challenge_prompts = (challenge_prompts *
+                             ((num_prompts // len(challenge_prompts)) + 1))
+    challenge_prompts = challenge_prompts[:num_prompts]
+    assert len(challenge_prompts) == num_prompts
+
+    # Single-step scheduler baseline
+    with vllm_runner(
+            model,
+            dtype=dtype,
+            enforce_eager=enforce_eager,
+            gpu_memory_utilization=0.7,
+            tensor_parallel_size=tp_size,
+            use_v2_block_manager=True,
+            num_scheduler_steps=num_scheduler_steps,
+            max_model_len=48,
+            max_num_batched_tokens=48,
+            max_num_seqs=4,
+            block_size=16,
+    ) as vllm_model:
+        outputs_baseline = (vllm_model.generate_greedy(
+            challenge_prompts, max_tokens) if num_logprobs is None else
+                            vllm_model.generate_greedy_logprobs(
+                                challenge_prompts, max_tokens, num_logprobs))
+
+    # multi-step+"single-step chunked prefill"+APC
+    with vllm_runner(
+            model,
+            dtype=dtype,
+            enforce_eager=enforce_eager,
+            gpu_memory_utilization=0.7,
+            tensor_parallel_size=tp_size,
+            use_v2_block_manager=True,
+            enable_chunked_prefill=True,
+            enable_prefix_caching=True,
+            num_scheduler_steps=num_scheduler_steps,
+            max_model_len=48,
+            max_num_batched_tokens=48,
+            max_num_seqs=4,
+            block_size=16,
+    ) as vllm_model:
+        outputs_w_features = (vllm_model.generate_greedy(
+            challenge_prompts, max_tokens) if num_logprobs is None else
+                              vllm_model.generate_greedy_logprobs(
+                                  challenge_prompts, max_tokens, num_logprobs))
+
+    if num_logprobs is None:
+        # No-logprobs test
+        check_outputs_equal(
+            outputs_0_lst=outputs_baseline,
+            outputs_1_lst=outputs_w_features,
+            name_0="multi-step",
+            name_1="multi-step+features",
+        )
+    else:
+        # Yes-logprobs test
+        check_logprobs_close(
+            outputs_0_lst=outputs_baseline,
+            outputs_1_lst=outputs_w_features,
+            name_0="multi-step",
+            name_1="multi-step+features",
+        )

vllm/core/scheduler.py

@@ -1607,10 +1607,29 @@ class Scheduler:
         # in a decode phase. Do not chunk.
         if enable_chunking and len(seqs) == 1:
             remaining_token_budget = budget.remaining_token_budget()
-            if self.cache_config.enable_prefix_caching:
+            if self.scheduler_config.is_multi_step:
+                # The current multi-step + chunked prefill capability does
+                # not actually support chunking prompts.
+                #
+                # Therefore, `num_new_tokens` is computed in the same fashion
+                # for both multi-step+chunked-prefill &
+                # multi-step+chunked-prefill+APC
+                #
+                # Prompts with more tokens than the current remaining budget
+                # are postponed to future scheduler steps
+                if num_new_tokens > self._get_prompt_limit(seq_group):
+                    # If the seq_group is in prompt-stage, pass the
+                    # num_new_tokens as-is so the caller can ignore
+                    # the sequence.
+                    pass
+                else:
+                    num_new_tokens = 0 \
+                        if num_new_tokens > remaining_token_budget \
+                        else num_new_tokens
+            elif self.cache_config.enable_prefix_caching:
                 # When prefix caching is enabled, we always allocate
-                # the number of new tokens that is dividable by the block size
+                # the number of new tokens that is dividable by the block
-                # to avoid partial block matching.
+                # size to avoid partial block matching.
                 block_size = self.cache_config.block_size
                 remainder = budget.token_budget % block_size
                 if remainder != 0:
@@ -1623,16 +1642,6 @@ class Scheduler:
                 if remaining_token_budget < num_new_tokens:
                     num_new_tokens = (remaining_token_budget //
                                       block_size) * block_size
-            elif self.scheduler_config.is_multi_step:
-                if num_new_tokens > self._get_prompt_limit(seq_group):
-                    # If the seq_group is in prompt-stage, pass the
-                    # num_new_tokens as-is so the caller can ignore
-                    # the sequence.
-                    pass
-                else:
-                    num_new_tokens = 0 \
-                        if num_new_tokens > remaining_token_budget \
-                        else num_new_tokens
             else:
                 num_new_tokens = min(num_new_tokens, remaining_token_budget)
         return num_new_tokens

vllm/engine/arg_utils.py

@@ -999,10 +999,6 @@ class EngineArgs:
             if speculative_config is not None:
                 raise ValueError("Speculative decoding is not supported with "
                                  "multi-step (--num-scheduler-steps > 1)")
-            if self.enable_chunked_prefill and self.enable_prefix_caching:
-                raise ValueError("Multi-Step is not supported with "
-                                 "both Chunked-Prefill and Prefix-Caching "
-                                 "enabled together.")
             if self.enable_chunked_prefill and self.pipeline_parallel_size > 1:
                 raise ValueError("Multi-Step Chunked-Prefill is not supported "
                                  "for pipeline-parallel-size > 1")