Add multipstep chunked-prefill support where prefill turns into decod…

…e after the first single step.

tests/multi_step/test_correctness_llm.py

@@ -5,6 +5,8 @@
 
 import pytest
 
+from tests.kernels.utils import override_backend_env_variable
+
 from ..models.utils import check_logprobs_close, check_outputs_equal
 
 MODELS = [
@@ -23,6 +25,7 @@
 @pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS)
 @pytest.mark.parametrize("num_prompts", NUM_PROMPTS)
 @pytest.mark.parametrize("num_logprobs", [None, 5])
+@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN", "FLASHINFER"])
 def test_multi_step_llm(
     hf_runner,
     vllm_runner,
@@ -36,6 +39,8 @@ def test_multi_step_llm(
     num_scheduler_steps: int,
     num_prompts: int,
     num_logprobs: Optional[int],
+    attention_backend: str,
+    monkeypatch,
 ) -> None:
     """Test vLLM engine with multi-step scheduling via sync LLM Engine.
 
@@ -63,6 +68,7 @@ def test_multi_step_llm(
       num_logprobs: corresponds to the `logprobs` argument to the OpenAI
                     completions endpoint; `None` -> 1 logprob returned.
     """
+    override_backend_env_variable(monkeypatch, attention_backend)
 
     prompts = example_prompts
     if len(prompts) < num_prompts:
@@ -114,6 +120,7 @@ def test_multi_step_llm(
 @pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS)
 @pytest.mark.parametrize("num_prompts", NUM_PROMPTS)
 @pytest.mark.parametrize("num_logprobs,num_prompt_logprobs", [(5, 5)])
+@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN", "FLASHINFER"])
 def test_multi_step_llm_w_prompt_logprobs(
     vllm_runner,
     example_prompts,
@@ -126,6 +133,8 @@ def test_multi_step_llm_w_prompt_logprobs(
     num_prompts: int,
     num_logprobs: Optional[int],
     num_prompt_logprobs: Optional[int],
+    attention_backend: str,
+    monkeypatch,
 ) -> None:
     """Test prompt logprobs with multi-step scheduling via sync LLM Engine.
 
@@ -155,6 +164,7 @@ def test_multi_step_llm_w_prompt_logprobs(
                            note that this argument is not supported by the
                            OpenAI completions endpoint.
     """
+    override_backend_env_variable(monkeypatch, attention_backend)
 
     prompts = example_prompts
     if len(prompts) < num_prompts:
@@ -205,6 +215,7 @@ def test_multi_step_llm_w_prompt_logprobs(
 @pytest.mark.parametrize("num_scheduler_steps", NUM_SCHEDULER_STEPS)
 @pytest.mark.parametrize("num_prompts", NUM_PROMPTS)
 @pytest.mark.parametrize("num_logprobs", [None, 5])
+@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN"])
 def test_multi_step_llm_chunked_prefill_prefix_cache(
     vllm_runner,
     example_prompts,
@@ -216,6 +227,8 @@ def test_multi_step_llm_chunked_prefill_prefix_cache(
     num_scheduler_steps: int,
     num_prompts: int,
     num_logprobs: Optional[int],
+    attention_backend: str,
+    monkeypatch,
 ) -> None:
     """Test vLLM engine with multi-step+"single-step chunked prefill"+APC.
 
@@ -278,6 +291,8 @@ def test_multi_step_llm_chunked_prefill_prefix_cache(
     #
     # The Incorrect scheduling behavior - if it occurs - will cause an exception
     # in the model runner resulting from `do_sample=False`.
+    override_backend_env_variable(monkeypatch, attention_backend)
+
     assert len(example_prompts) >= 2
     challenge_prompts = copy.deepcopy(example_prompts)
     challenge_prompts[0] = ('vLLM is a high-throughput and memory-efficient '

vllm/attention/backends/flashinfer.py

@@ -430,11 +430,25 @@ def advance_step(self,
         Update metadata in-place to advance one decode step.
         """
 
-        assert not turn_prefills_into_decodes, \
-            ("Chunked prefill is not supported with flashinfer yet."
-             "turn_prefills_into_decodes is a Multi-Step + Chunked-Prefill "
-             "specific parameter.")
-
+        if turn_prefills_into_decodes:
+            # When mutli-Step is enabled with chunked-Prefill, prefills and
+            # decodes are scheduled together. In the first step, all the
+            # prefills turn into decodes. This update reflects that
+            # conversion.
+            assert self.num_decode_tokens + self.num_prefills == num_seqs
+            # Flashinfer doesn't support speculative decoding + chunked-prefill
+            # + multi-step scheduling yet. 
+            assert self.decode_query_len == 1
+            self.num_decode_tokens += self.num_prefills
+            self.num_prefills = 0
+            self.num_prefill_tokens = 0
+            self.max_prefill_seq_len = 0
+            self.max_query_len = 1
+
+            self.slot_mapping = self.slot_mapping[:num_seqs]
+        else:
+            assert self.seq_lens_tensor is not None
+
         assert num_seqs > 0
         assert num_queries > 0
         assert model_input.attn_metadata is not None

vllm/worker/multi_step_model_runner.py

@@ -30,7 +30,7 @@
 logger = init_logger(__name__)
 
 MULTI_STEP_ATTENTION_BACKENDS = ["FLASH_ATTN", "ROCM_FLASH", "FLASHINFER"]
-MULTI_STEP_CHUNKED_PREFILL_ATTENTION_BACKENDS = ["FLASH_ATTN"]
+MULTI_STEP_CHUNKED_PREFILL_ATTENTION_BACKENDS = ["FLASH_ATTN", "FLASHINFER"]
 
 def _get_supported_attention_backends(chunked_prefill_enabled: bool) \
     -> List[str]: