[V1][Spec Decode] Handle draft tokens beyond max_model_len (#16087)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

tests/v1/core/test_scheduler.py

@@ -30,6 +30,7 @@ def create_scheduler(
     use_kv_connector: bool = False,
     num_blocks: int = 10000,
     block_size: int = 16,
+    max_model_len: Optional[int] = None,
 ) -> Scheduler:
     '''Create scheduler under test.
 
@@ -44,12 +45,15 @@ def create_scheduler(
     Returns:
       :class:`Scheduler` instance
     '''
+    if max_model_len is None:
+        max_model_len = max_num_batched_tokens
     scheduler_config = SchedulerConfig(
         max_num_seqs=max_num_seqs,
         max_num_batched_tokens=max_num_batched_tokens,
-        max_model_len=max_num_batched_tokens,
+        max_model_len=max_model_len,
         long_prefill_token_threshold=long_prefill_token_threshold,
         disable_chunked_mm_input=disable_chunked_mm_input,
+        enable_chunked_prefill=True,
     )
     model_config = ModelConfig(
         model=model,
@@ -296,6 +300,7 @@ def test_no_mm_input_chunking():
         model="llava-hf/llava-1.5-7b-hf",
         max_num_batched_tokens=1024,
         disable_chunked_mm_input=True,
+        max_model_len=2048,
     )
     mm_positions = [[PlaceholderRange(offset=400, length=800)]]
     requests = create_requests(num_requests=1,

tests/v1/spec_decode/test_max_len.py

@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Test whether spec decoding handles the max model length properly."""
+
+import pytest
+
+from vllm import LLM, SamplingParams
+
+_PROMPTS = [
+    "1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1",
+    "Repeat the following sentence 10 times: Consistency is key to mastering any skill.",  # noqa: E501
+    "Who won the Turing Award in 2018, and for what contribution? Describe in detail.",  # noqa: E501
+]
+
+
+@pytest.mark.parametrize("num_speculative_tokens", [1, 3, 10])
+def test_ngram_max_len(
+    monkeypatch: pytest.MonkeyPatch,
+    num_speculative_tokens: int,
+):
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+
+        llm = LLM(
+            model="facebook/opt-125m",
+            max_model_len=100,
+            enforce_eager=True,  # For faster initialization.
+            speculative_config={
+                "method": "ngram",
+                "prompt_lookup_max": 5,
+                "prompt_lookup_min": 3,
+                "num_speculative_tokens": num_speculative_tokens,
+            },
+        )
+        sampling_params = SamplingParams(max_tokens=100, ignore_eos=True)
+        llm.generate(_PROMPTS, sampling_params)
+
+
+@pytest.mark.parametrize("num_speculative_tokens", [1, 3, 10])
+def test_eagle_max_len(
+    monkeypatch: pytest.MonkeyPatch,
+    num_speculative_tokens: int,
+):
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+
+        llm = LLM(
+            model="meta-llama/Meta-Llama-3-8B-Instruct",
+            enforce_eager=True,  # For faster initialization.
+            speculative_config={
+                "method": "eagle",
+                "model": "yuhuili/EAGLE-LLaMA3-Instruct-8B",
+                "num_speculative_tokens": num_speculative_tokens,
+            },
+            max_model_len=100,
+        )
+        sampling_params = SamplingParams(max_tokens=100, ignore_eos=True)
+        llm.generate(_PROMPTS, sampling_params)

tests/v1/spec_decode/test_ngram.py

@@ -2,7 +2,7 @@
 
 import numpy as np
 
-from vllm.config import SpeculativeConfig, VllmConfig
+from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
 from vllm.v1.spec_decode.ngram_proposer import (NgramProposer,
                                                 _find_subarray_kmp,
                                                 _kmp_lps_array)
@@ -42,14 +42,24 @@ def test_find_subarray_kmp():
 def test_ngram_proposer():
 
     def ngram_proposer(min_n: int, max_n: int, k: int) -> NgramProposer:
-        return NgramProposer(vllm_config=VllmConfig(
-            speculative_config=SpeculativeConfig.from_dict(
-                {
-                    "prompt_lookup_min": min_n,
-                    "prompt_lookup_max": max_n,
-                    "num_speculative_tokens": k,
-                    "method": "ngram",
-                })))
+        # Dummy model config. Just to set max_model_len.
+        model_config = ModelConfig(model="facebook/opt-125m",
+                                   task="generate",
+                                   max_model_len=100,
+                                   tokenizer="facebook/opt-125m",
+                                   tokenizer_mode="auto",
+                                   dtype="auto",
+                                   seed=None,
+                                   trust_remote_code=False)
+        return NgramProposer(
+            vllm_config=VllmConfig(model_config=model_config,
+                                   speculative_config=SpeculativeConfig.
+                                   from_dict({
+                                       "prompt_lookup_min": min_n,
+                                       "prompt_lookup_max": max_n,
+                                       "num_speculative_tokens": k,
+                                       "method": "ngram",
+                                   })))
 
     # No match.
     result = ngram_proposer(

vllm/v1/core/sched/scheduler.py

@@ -185,6 +185,13 @@ class Scheduler(SchedulerInterface):
             num_new_tokens = min(num_new_tokens, token_budget)
             assert num_new_tokens > 0
 
+            # Make sure the input position does not exceed the max model len.
+            # This is necessary when using spec decoding.
+            num_new_tokens = min(
+                num_new_tokens,
+                self.max_model_len - request.num_computed_tokens)
+            assert num_new_tokens > 0
+
             # Schedule encoder inputs.
             if request.has_encoder_inputs:
                 (encoder_inputs_to_schedule, num_new_tokens,

vllm/v1/spec_decode/eagle.py

@@ -12,6 +12,8 @@ from vllm.model_executor.models.llama_eagle import EagleLlamaForCausalLM
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
 from vllm.v1.sample.metadata import SamplingMetadata
 
+PADDING_SLOT_ID = -1
+
 
 class EagleProposer:
 
@@ -23,6 +25,7 @@ class EagleProposer:
         self.vllm_config = vllm_config
         self.num_speculative_tokens = (
             vllm_config.speculative_config.num_speculative_tokens)
+        self.max_model_len = vllm_config.model_config.max_model_len
         self.block_size = vllm_config.cache_config.block_size
         # We need +1 here because the arange is used to set query_start_loc,
         # which has one more element than batch_size.
@@ -112,22 +115,48 @@ class EagleProposer:
             # Update the inputs.
             input_ids = draft_token_ids_list[-1]
             positions += 1
+
+            # NOTE(woosuk): We should handle the case where the draft model
+            # generates tokens beyond the max model length. Since it is complex
+            # to remove such requests from the batch, we keep them in the batch
+            # but adjust the position ids and slot mappings to avoid the
+            # out-of-range access during the model execution. The draft tokens
+            # generated with this adjustment should be ignored.
+            exceeds_max_model_len = positions >= self.max_model_len
+            # Mask out the position ids that exceed the max model length.
+            # Otherwise, we may get out-of-range error in RoPE.
+            clamped_positions = torch.where(exceeds_max_model_len, 0,
+                                            positions)
+
+            # Increment the sequence lengths.
             attn_metadata.max_seq_len += 1
             attn_metadata.seq_lens += 1
+            # Consider max model length.
+            attn_metadata.max_seq_len = min(attn_metadata.max_seq_len,
+                                            self.max_model_len)
+            # For the requests that exceed the max model length, we set the
+            # sequence length to 1 to minimize their overheads in attention.
+            attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)
+
             # Compute the slot mapping.
-            block_numbers = positions // self.block_size
+            block_numbers = clamped_positions // self.block_size
             block_ids = block_table.gather(dim=1,
                                            index=block_numbers.view(-1, 1))
             block_ids = block_ids.view(-1)
             attn_metadata.slot_mapping = (block_ids * self.block_size +
-                                          positions % self.block_size)
+                                          clamped_positions % self.block_size)
+            # Mask out the slot mappings that exceed the max model length.
+            # Otherwise, the KV cache will be inadvertently updated with the
+            # padding tokens.
+            attn_metadata.slot_mapping.masked_fill_(exceeds_max_model_len,
+                                                    PADDING_SLOT_ID)
 
             # Run the model.
             with set_forward_context(attn_metadata, self.vllm_config):
                 hidden_states = self.model(
                     input_ids=input_ids,
                     hidden_states=hidden_states,
-                    positions=positions,
+                    positions=clamped_positions,
                 )
             logits = self.model.compute_logits(hidden_states, None)
             draft_token_ids, probs = compute_probs_and_sample_next_token(

vllm/v1/spec_decode/ngram_proposer.py

@@ -18,6 +18,9 @@ class NgramProposer:
         # tokens follow the match, we will return the maximum amount of
         # tokens until the end.
         self.k = vllm_config.speculative_config.num_speculative_tokens
+        # Maximum length of the model.
+        self.max_model_len = vllm_config.model_config.max_model_len
+
         # Trigger Numba JIT compilation for N-gram proposer.
         # This usually takes less than 1 second.
         self.propose(np.zeros(1024, dtype=np.int32))
@@ -50,9 +53,14 @@ class NgramProposer:
               followed that pattern. Here we will return [4,2,3] because 
               we only have three tokens after the match.
         """
+        # Do not generate draft tokens beyond the max model length.
+        k = min(self.k, self.max_model_len - context_token_ids.shape[0])
+        if k <= 0:
+            return None
+
         # TODO(woosuk): Optimize this.
         for n in range(self.max_n, self.min_n - 1, -1):
-            result = _find_subarray_kmp(context_token_ids, n, self.k)
+            result = _find_subarray_kmp(context_token_ids, n, k)
             if result is not None:
                 return result
         return None

vllm/v1/worker/gpu_model_runner.py

@@ -1271,7 +1271,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 draft_token_ids.append([])
                 continue
 
-            # Skip requests that require top-p, top-k, etc.
+            # Skip requests that require sampling parameters that are not
+            # supported with speculative decoding.
             req_id = self.input_batch.req_ids[i]
             if not is_spec_decode_supported(req_id, self.input_batch):
                 draft_token_ids.append([])
@@ -1280,6 +1281,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             # Add sampled_token_ids to token_ids_cpu.
             start_idx = self.input_batch.num_tokens_no_spec[i]
             end_idx = start_idx + num_sampled_ids
+            if end_idx >= self.max_model_len:
+                # Skip requests that have already reached the max model length.
+                draft_token_ids.append([])
+                continue
+
             self.input_batch.token_ids_cpu[i, start_idx:end_idx] = sampled_ids
             drafter_output = self.drafter.propose(
                 self.input_batch.token_ids_cpu[i, :end_idx])