feat: spec decode with draft models (#24322)

Signed-off-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>

examples/offline_inference/spec_decode.py

@@ -54,7 +54,7 @@ def parse_args():
         "--method",
         type=str,
         default="eagle",
-        choices=["ngram", "eagle", "eagle3", "mtp"],
+        choices=["ngram", "eagle", "eagle3", "mtp", "draft_model"],
     )
     parser.add_argument("--num-spec-tokens", type=int, default=2)
     parser.add_argument("--prompt-lookup-max", type=int, default=5)
@@ -70,7 +70,11 @@ def parse_args():
     parser.add_argument("--output-len", type=int, default=256)
     parser.add_argument("--model-dir", type=str, default=None)
     parser.add_argument("--eagle-dir", type=str, default=None)
+    parser.add_argument("--draft-model", type=str, default=None)
     parser.add_argument("--custom-mm-prompts", action="store_true")
+    parser.add_argument("--gpu-memory-utilization", type=float, default=0.9)
+    parser.add_argument("--disable-padded-drafter-batch", action="store_true")
+    parser.add_argument("--max-num-seqs", type=int, default=None)
     return parser.parse_args()
 
 
@@ -111,6 +115,7 @@ def main(args):
             "method": args.method,
             "model": eagle_dir,
             "num_speculative_tokens": args.num_spec_tokens,
+            "disable_padded_drafter_batch": args.disable_padded_drafter_batch,
         }
     elif args.method == "ngram":
         speculative_config = {
@@ -119,6 +124,15 @@ def main(args):
             "prompt_lookup_max": args.prompt_lookup_max,
             "prompt_lookup_min": args.prompt_lookup_min,
         }
+    elif args.method == "draft_model":
+        assert args.draft_model is not None and args.draft_model != ""
+        speculative_config = {
+            "method": args.method,
+            "model": args.draft_model,
+            "num_speculative_tokens": args.num_spec_tokens,
+            "enforce_eager": args.enforce_eager,
+            "max_model_len": args.max_model_len,
+        }
     elif args.method == "mtp":
         speculative_config = {
             "method": "mtp",
@@ -133,12 +147,13 @@ def main(args):
         tensor_parallel_size=args.tp,
         enable_chunked_prefill=args.enable_chunked_prefill,
         enforce_eager=args.enforce_eager,
-        gpu_memory_utilization=0.9,
+        gpu_memory_utilization=args.gpu_memory_utilization,
         speculative_config=speculative_config,
         disable_log_stats=False,
         max_model_len=args.max_model_len,
         limit_mm_per_prompt={"image": 5},
         disable_chunked_mm_input=True,
+        max_num_seqs=args.max_num_seqs,
     )
 
     sampling_params = SamplingParams(temperature=args.temp, max_tokens=args.output_len)

examples/online_serving/disaggregated_serving/moriio_toy_proxy_server.py

@@ -4,13 +4,13 @@ import asyncio
 import copy
 import logging
 import os
-import re
 import socket
 import threading
 import uuid
 
 import aiohttp
 import msgpack
+import regex as re
 import zmq
 from quart import Quart, make_response, request
 

tests/v1/e2e/test_spec_decode.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import random
+from collections.abc import Iterable
+from dataclasses import dataclass
 from typing import Any
 
 import pytest
@@ -10,32 +12,45 @@ from tests.utils import get_attn_backend_list_based_on_platform, large_gpu_mark
 from vllm import LLM, SamplingParams
 from vllm.assets.base import VLLM_S3_BUCKET_URL
 from vllm.assets.image import VLM_IMAGES_DIR
+from vllm.benchmarks.datasets import InstructCoderDataset
+from vllm.config.vllm import VllmConfig
 from vllm.distributed import cleanup_dist_env_and_memory
+from vllm.engine.arg_utils import EngineArgs
 from vllm.platforms import current_platform
+from vllm.v1.metrics.reader import Metric
+from vllm.v1.spec_decode.draft_model import (
+    create_vllm_config_for_draft_model,
+    merge_toks_kernel,
+)
 
 MTP_SIMILARITY_RATE = 0.8
 
 
 def _skip_if_insufficient_gpus_for_tp(tp_size: int):
     """Skip test if available GPUs < tp_size on ROCm."""
-    if current_platform.is_rocm():
-        available_gpus = torch.cuda.device_count()
-        if available_gpus < tp_size:
-            pytest.skip(
-                f"Test requires {tp_size} GPUs, but only {available_gpus} available"
-            )
+    available_gpus = torch.cuda.device_count()
+    if available_gpus < tp_size:
+        pytest.skip(
+            f"Test requires {tp_size} GPUs, but only {available_gpus} available"
+        )
 
 
-def get_test_prompts(mm_enabled: bool):
+Messages = list[dict[str, Any]]
+
+
+def get_test_prompts(
+    mm_enabled: bool, quiet: bool = False, num_prompts: int = 100
+) -> list[Messages]:
     prompt_types = ["repeat", "sentence"]
     if mm_enabled:
         prompt_types.append("mm")
-    num_prompts = 100
     prompts = []
 
     random.seed(0)
     random_prompt_type_choices = random.choices(prompt_types, k=num_prompts)
-    print(f"Prompt types: {random_prompt_type_choices}")
+
+    if not quiet:
+        print(f"Prompt types: {random_prompt_type_choices}")
 
     # Generate a mixed batch of prompts, some of which can be easily
     # predicted by n-gram matching and some which likely cannot.
@@ -75,11 +90,27 @@ def get_test_prompts(mm_enabled: bool):
     return prompts
 
 
+def get_instruct_coder_messages(n: int) -> list[Messages]:
+    dataset = InstructCoderDataset(
+        dataset_path="likaixin/InstructCoder", dataset_split="train"
+    )
+    prompts: Iterable[str] = dataset.sample_prompts(n=n)
+    return [[{"role": "user", "content": prompt}] for prompt in prompts]
+
+
 @pytest.fixture
 def sampling_config():
+    return greedy_sampling()
+
+
+def greedy_sampling() -> SamplingParams:
     return SamplingParams(temperature=0, max_tokens=10, ignore_eos=False)
 
 
+def stochastic_sampling() -> SamplingParams:
+    return SamplingParams(temperature=1.0, max_tokens=10, ignore_eos=False)
+
+
 @pytest.fixture
 def model_name():
     return "meta-llama/Llama-3.1-8B-Instruct"
@@ -583,3 +614,269 @@ def test_mtp_correctness(
         del spec_llm
         torch.cuda.empty_cache()
         cleanup_dist_env_and_memory()
+
+
+@dataclass
+class ArgsTest:
+    target_model: str
+    draft_model: str
+    sampling_config: SamplingParams
+    num_speculative_tokens: int
+    expected_acceptance_rate: float
+    expected_acceptance_len: float
+    # Defaults
+    target_tensor_parallel_size: int = 1
+    draft_tensor_parallel_size: int = 1
+    max_model_len: int = 1024
+    gpu_memory_utilization: float = 0.5
+    dataset: str = "test_prompts"
+    num_prompts: int = 100
+
+
+cases = [
+    # Same model for draft and target, greedy sampling.
+    ArgsTest(
+        target_model="Qwen/Qwen3-0.6B",
+        draft_model="Qwen/Qwen3-0.6B",
+        sampling_config=greedy_sampling(),
+        num_speculative_tokens=3,  # K
+        expected_acceptance_len=3 + 1,  # K + 1
+        expected_acceptance_rate=1.0,
+    ),
+    # Smaller draft model, stochastic sampling.
+    ArgsTest(
+        target_model="Qwen/Qwen3-1.7B",
+        draft_model="Qwen/Qwen3-0.6B",
+        sampling_config=stochastic_sampling(),
+        num_speculative_tokens=3,
+        expected_acceptance_len=2.8 + 1,
+        expected_acceptance_rate=0.9,
+    ),
+]
+
+
+@pytest.mark.parametrize("args", cases)
+@pytest.mark.parametrize("enforce_eager", [True, False])
+def test_draft_model_correctness(args: ArgsTest, enforce_eager: bool):
+    assert_draft_model_correctness(args, enforce_eager)
+
+
+def test_draft_model_realistic_example():
+    args = ArgsTest(
+        target_model="Qwen/Qwen3-1.7B",
+        draft_model="Qwen/Qwen3-0.6B",
+        dataset="likaixin/InstructCoder",
+        num_speculative_tokens=3,
+        sampling_config=greedy_sampling(),
+        # values below are not derived, but just prevent a regression
+        expected_acceptance_len=2.8,
+        expected_acceptance_rate=0.55,
+    )
+    assert_draft_model_correctness(args, enforce_eager=False)
+
+
+@pytest.mark.parametrize(
+    "models",
+    [
+        # target_model,         draft_model
+        ("Qwen/Qwen3-1.7B-FP8", "Qwen/Qwen3-0.6B"),  # target quantized
+        ("Qwen/Qwen3-1.7B", "Qwen/Qwen3-0.6B-FP8"),  # draft quantized
+    ],
+    ids=["target_quantized", "draft_quantized"],
+)
+@pytest.mark.parametrize("enforce_eager", [True, False])
+def test_draft_model_quantization(models: tuple[str, str], enforce_eager: bool):
+    tgt_model, draft_model = models
+    sd_case = ArgsTest(
+        target_model=tgt_model,
+        draft_model=draft_model,
+        **some_high_acceptance_metrics(),
+    )
+    assert_draft_model_correctness(sd_case, enforce_eager)
+
+
+def test_draft_model_tensor_parallelism():
+    """Ensure spec decode works when running with TP > 1."""
+    _skip_if_insufficient_gpus_for_tp(2)
+    sd_case = ArgsTest(
+        target_model="Qwen/Qwen3-1.7B",
+        target_tensor_parallel_size=2,
+        draft_model="Qwen/Qwen3-0.6B",
+        draft_tensor_parallel_size=2,
+        **some_high_acceptance_metrics(),
+    )
+    assert_draft_model_correctness(sd_case, enforce_eager=False)
+
+
+def test_draft_model_engine_args_tensor_parallelism():
+    """Ensure the vllm_config for the draft model is created correctly,
+    and independently of the target model (quantization, TP, etc.)"""
+    _skip_if_insufficient_gpus_for_tp(2)
+
+    engine_args = EngineArgs(
+        model="Qwen/Qwen3-1.7B-FP8",  # <<< tgt quantized
+        tensor_parallel_size=2,
+        speculative_config={
+            "model": "Qwen/Qwen3-0.6B",  # <<< draft not quantized
+            "method": "draft_model",
+            "num_speculative_tokens": 3,
+            "draft_tensor_parallel_size": 1,  # <<< valid arg name
+        },
+    )
+    tgt_vllm_config: VllmConfig = engine_args.create_engine_config()
+    assert tgt_vllm_config.parallel_config.tensor_parallel_size == 2
+    assert tgt_vllm_config.quant_config.get_name() == "fp8"
+
+    draft_vllm_config: VllmConfig = create_vllm_config_for_draft_model(tgt_vllm_config)
+    assert draft_vllm_config.parallel_config.tensor_parallel_size == 1
+    assert draft_vllm_config.quant_config is None
+
+
+def test_draft_model_engine_args_rejects_invalid_tp_argname():
+    """The user should pass "draft_tensor_parallel_size" rather than
+    "tensor_parallel_size". We enforce this with validation."""
+
+    engine_args = EngineArgs(
+        model="Qwen/Qwen3-1.7B",
+        tensor_parallel_size=1,
+        speculative_config={
+            "model": "Qwen/Qwen3-0.6B",
+            "method": "draft_model",
+            "num_speculative_tokens": 3,
+            "tensor_parallel_size": 1,  # <<< invalid arg name
+        },
+    )
+    with pytest.raises(ValueError):
+        engine_args.create_engine_config()
+
+
+def assert_draft_model_correctness(args: ArgsTest, enforce_eager: bool):
+    """Compare the outputs using and not using speculative decoding.
+    In the greedy decoding case, the outputs must match EXACTLY."""
+    test_prompts: list[Messages] = get_messages(
+        dataset=args.dataset, n=args.num_prompts
+    )
+
+    spec_llm = LLM(
+        model=args.target_model,
+        speculative_config={
+            "model": args.draft_model,
+            "method": "draft_model",
+            "num_speculative_tokens": args.num_speculative_tokens,
+            "max_model_len": args.max_model_len,
+            "enforce_eager": enforce_eager,
+            "draft_tensor_parallel_size": args.draft_tensor_parallel_size,
+            "max_num_seqs": 100,  # limit cudagraph capture runtime
+        },
+        max_model_len=args.max_model_len,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        tensor_parallel_size=args.target_tensor_parallel_size,
+        enforce_eager=enforce_eager,
+        disable_log_stats=False,  # enables get_metrics()
+    )
+    # we don't check the outputs, only check the metrics
+    spec_llm.chat(test_prompts, args.sampling_config)
+    metrics = spec_llm.get_metrics()
+
+    acceptance_rate: float = compute_acceptance_rate(metrics)
+    acceptance_len: float = compute_acceptance_len(metrics)
+    del spec_llm  # CLEANUP
+    torch.cuda.empty_cache()
+    cleanup_dist_env_and_memory()
+
+    print(
+        f"spec-decode: target={args.target_model}, draft={args.draft_model}, "
+        f"temperature={args.sampling_config.temperature:.2f}, "
+        f"acceptance_rate={acceptance_rate:.2f}, "
+        f"acceptance_len={acceptance_len:.2f}, "
+    )
+
+    assert acceptance_rate >= args.expected_acceptance_rate
+    assert acceptance_len >= args.expected_acceptance_len
+
+
+def get_messages(dataset: str, n: int) -> list[Messages]:
+    if dataset == "test_prompts":
+        return get_test_prompts(mm_enabled=False, quiet=True, num_prompts=n)
+    elif dataset == "likaixin/InstructCoder":
+        return get_instruct_coder_messages(n=n)
+    else:
+        raise NotImplementedError(f"Dataset '{dataset}' not implemented")
+
+
+def some_high_acceptance_metrics() -> dict:
+    return {
+        "sampling_config": greedy_sampling(),
+        "num_speculative_tokens": 3,
+        "expected_acceptance_len": 2.90 + 1,
+        "expected_acceptance_rate": 0.90,
+    }
+
+
+def test_merge_toks_kernel():
+    device = "cuda"
+    merged_len = 5 + 2  # len(target_toks) = 5, batch_size = 2
+    merged = torch.full((merged_len,), -100, device=device)  # -100 is arbitrary
+    is_rejected_tok = torch.full((merged_len,), True, device=device)
+    grid = (2,)
+    merge_toks_kernel[grid](
+        target_toks_ptr=torch.tensor([0, 1, 2, 0, 1], device=device),
+        next_toks_ptr=torch.tensor([3, 2], device=device),
+        query_start_locs_ptr=torch.tensor([0, 3], device=device),
+        query_end_locs_ptr=torch.tensor([2, 4], device=device),
+        out_ptr_merged_toks=merged,
+        out_ptr_is_rejected_tok=is_rejected_tok,
+        target_toks_size=5,
+        rejected_tok_fill=-1,
+    )
+    expected_merged = torch.tensor([0, 1, 2, 3, 0, 1, 2], device=device)
+    assert torch.allclose(merged, expected_merged)
+
+    expected_rejected_toks = torch.tensor([False] * merged_len, device=device)
+    assert torch.allclose(is_rejected_tok, expected_rejected_toks)
+
+
+def test_merge_toks_kernel_with_rejected_tokens():
+    device = "cuda"
+    merged_size = 9 + 2  # len(target_toks) = 9, batch_size = 2
+    merged = torch.full((merged_size,), -100, device=device)
+    is_rejected_tok = torch.full((merged_size,), True, device=device)
+    grid = (2,)
+    merge_toks_kernel[grid](
+        #                                       rejected tokens
+        #                                       ↓   ↓   ↓         ↓
+        target_toks_ptr=torch.tensor([0, 1, 2, 13, 14, 15, 0, 1, 22], device=device),
+        next_toks_ptr=torch.tensor([3, 2], device=device),
+        query_start_locs_ptr=torch.tensor([0, 6], device=device),
+        query_end_locs_ptr=torch.tensor([2, 7], device=device),
+        out_ptr_merged_toks=merged,
+        out_ptr_is_rejected_tok=is_rejected_tok,
+        target_toks_size=9,
+        rejected_tok_fill=-1,
+    )
+    expected_merged = torch.tensor([0, 1, 2, 3, -1, -1, -1, 0, 1, 2, -1], device=device)
+    assert torch.allclose(merged, expected_merged)
+
+    expected_rejected_toks = torch.tensor(
+        [False, False, False, False, True, True, True, False, False, False, True],
+        device=device,
+    )
+    assert torch.allclose(is_rejected_tok, expected_rejected_toks)
+
+
+def compute_acceptance_rate(metrics: list[Metric]) -> float:
+    name2metric = {metric.name: metric for metric in metrics}
+    n_draft_toks = name2metric["vllm:spec_decode_num_draft_tokens"].value  # type: ignore
+    if n_draft_toks == 0:
+        return float("nan")
+    n_accepted_toks = name2metric["vllm:spec_decode_num_accepted_tokens"].value  # type: ignore
+    return n_accepted_toks / n_draft_toks
+
+
+def compute_acceptance_len(metrics: list[Metric]) -> float:
+    name2metric = {metric.name: metric for metric in metrics}
+    n_drafts = name2metric["vllm:spec_decode_num_drafts"].value  # type: ignore
+    n_accepted_toks = name2metric["vllm:spec_decode_num_accepted_tokens"].value  # type: ignore
+    if n_drafts == 0:
+        return 1
+    return 1 + (n_accepted_toks / n_drafts)

tests/v1/worker/test_utils.py

@@ -55,3 +55,38 @@ def test_bind_kv_cache_non_attention(default_vllm_config):
 
     assert runner_kv_caches[0] is kv_cache["model.layers.20.attn"]
     assert runner_kv_caches[1] is kv_cache["model.layers.28.attn"]
+
+
+def test_bind_kv_cache_draft_model(default_vllm_config):
+    from vllm.attention.layer import Attention
+
+    layer_names = [
+        "model.layers.0.attn",
+        "model.layers.1.attn",
+        "draft_model.layers.0.attn",
+        "draft_model.layers.1.attn",
+    ]
+    ctx = {
+        layer_name: Attention(32, 128, 0.1, prefix=layer_name)
+        for layer_name in layer_names
+    }
+    kv_cache = {layer_name: torch.zeros((1,)) for layer_name in layer_names}
+    runner_kv_caches: list[torch.Tensor] = []
+    bind_kv_cache(kv_cache, ctx, runner_kv_caches)
+
+    assert ctx["model.layers.0.attn"].kv_cache[0] is kv_cache["model.layers.0.attn"]
+    assert ctx["model.layers.1.attn"].kv_cache[0] is kv_cache["model.layers.1.attn"]
+    assert (
+        ctx["draft_model.layers.0.attn"].kv_cache[0]
+        is kv_cache["draft_model.layers.0.attn"]
+    )
+    assert (
+        ctx["draft_model.layers.1.attn"].kv_cache[0]
+        is kv_cache["draft_model.layers.1.attn"]
+    )
+
+    # caches are ordered by layer_index, interleaving target and draft model
+    assert runner_kv_caches[0] is kv_cache["model.layers.0.attn"]
+    assert runner_kv_caches[1] is kv_cache["draft_model.layers.0.attn"]
+    assert runner_kv_caches[2] is kv_cache["model.layers.1.attn"]
+    assert runner_kv_caches[3] is kv_cache["draft_model.layers.1.attn"]

vllm/benchmarks/datasets.py

@@ -2593,17 +2593,10 @@ class InstructCoderDataset(HuggingFaceDataset):
         request_id_prefix: str = "",
         no_oversample: bool = False,
         **kwargs,
-    ) -> list:
+    ) -> list[SampleRequest]:
         output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
         sampled_requests = []
-        for i, item in enumerate(self.data):
-            if len(sampled_requests) >= num_requests:
-                break
-            prompt = (
-                f"{item['input']}\n\n{item['instruction']} Just output "
-                "the code, do not include any explanation."
-            )
-
+        for i, prompt in enumerate(self.sample_prompts(n=num_requests)):
             # apply template
             if not skip_chat_template:
                 prompt = tokenizer.apply_chat_template(
@@ -2626,6 +2619,14 @@ class InstructCoderDataset(HuggingFaceDataset):
         )
         return sampled_requests
 
+    def sample_prompts(self, n: int) -> Iterator[str]:
+        for item in self.data.take(n):
+            prompt = (
+                f"{item['input']}\n\n{item['instruction']} Just output "
+                "the code, do not include any explanation."
+            )
+            yield prompt
+
 
 # -----------------------------------------------------------------------------
 # MT-Bench Dataset Implementation

vllm/benchmarks/lib/ready_checker.py

@@ -8,8 +8,12 @@ import time
 import aiohttp
 from tqdm.asyncio import tqdm
 
+from vllm.logger import init_logger
+
 from .endpoint_request_func import RequestFunc, RequestFuncInput, RequestFuncOutput
 
+logger = init_logger(__name__)
+
 
 async def wait_for_endpoint(
     request_func: RequestFunc,
@@ -61,6 +65,8 @@ async def wait_for_endpoint(
                 if output.success:
                     pbar.close()
                     return output
+                else:
+                    logger.warning("Endpoint is not ready. Error='%s'", output.error)
             except aiohttp.ClientConnectorError:
                 pass
 

vllm/config/parallel.py

@@ -3,6 +3,7 @@
 
 import os
 from collections.abc import Callable
+from dataclasses import replace
 from typing import TYPE_CHECKING, Any, Literal
 
 import torch
@@ -709,3 +710,6 @@ class ParallelConfig:
             )
 
         return self
+
+    def replace(self, **kwargs) -> Self:
+        return replace(self, **kwargs)

vllm/config/speculative.py

@@ -77,6 +77,9 @@ class SpeculativeConfig:
     draft_tensor_parallel_size: int | None = Field(default=None, ge=1)
     """The degree of the tensor parallelism for the draft model. Can only be 1
     or the same as the target model's tensor parallel size."""
+    tensor_parallel_size: int | None = None
+    """Users should pass "draft_tensor_parallel_size". This parameter's purpose is to
+    warn users when they mistakenly provide the wrong argument."""
 
     # Draft model configuration
     quantization: me_quant.QuantizationMethods | None = None
@@ -397,13 +400,11 @@ class SpeculativeConfig:
                             "one layer. Might need some code changes "
                             "to support multiple layers."
                         )
+                elif self.method == "draft_model":
+                    pass
                 else:
-                    self.method = "draft_model"
                     raise NotImplementedError(
-                        "Speculative decoding with draft model is not "
-                        "supported yet. Please consider using other "
-                        "speculative decoding methods such as ngram, medusa, "
-                        "eagle, or mtp."
+                        f"Unsupported speculative method: '{self.method}'"
                     )
 
                 # Replace hf_config for EAGLE draft_model
@@ -631,6 +632,12 @@ class SpeculativeConfig:
 
     @model_validator(mode="after")
     def _verify_args(self) -> Self:
+        if self.tensor_parallel_size is not None:
+            raise ValueError(
+                "'tensor_parallel_size' is not a valid argument in the "
+                "speculative_config. Please pass 'draft_tensor_parallel_size' instead."
+            )
+
         if self.num_speculative_tokens is None:
             raise ValueError(
                 "num_speculative_tokens must be provided with "
@@ -669,12 +676,32 @@ class SpeculativeConfig:
                 f"Eagle3 is only supported for {eagle3_target_supported} models. "  # noqa: E501
                 f"Got {self.target_model_config.hf_text_config.model_type=}"
             )
-
+        self.verify_equal_vocab_size_if_draft_model()
         return self
 
+    def verify_equal_vocab_size_if_draft_model(self):
+        if (
+            self.method == "draft_model"
+            and self.target_model_config is not None
+            and self.draft_model_config is not None
+        ):
+            target_vocab_size = self.target_model_config.get_vocab_size()
+            draft_vocab_size = self.draft_model_config.get_vocab_size()
+            if target_vocab_size != draft_vocab_size:
+                raise ValueError(
+                    f"Target and draft model should have the same vocabulary size. "
+                    f"Target model vocab_size={target_vocab_size}. "
+                    f"Draft model vocab_size={draft_vocab_size}. "
+                    f"Using models with different tokenizers can cause out-of-bounds "
+                    f"errors during speculative decoding."
+                )
+
     def use_eagle(self) -> bool:
         return self.method in ("eagle", "eagle3", "mtp")
 
+    def uses_draft_model(self) -> bool:
+        return self.method == "draft_model"
+
     def __repr__(self) -> str:
         method = self.method
         model = None if method in ("ngram", "suffix") else self.draft_model_config.model

vllm/config/vllm.py

@@ -1214,10 +1214,19 @@ class VllmConfig:
         compilation_config = self.compilation_config
         computed_compile_ranges_split_points = []
 
-        # The upper bound of the compile ranges is the max_num_batched_tokens
-        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
-        if max_num_batched_tokens is not None:
-            computed_compile_ranges_split_points.append(max_num_batched_tokens)
+        # The upper bound of the compile ranges is the max_num_batched_tokens.
+        # For speculative decoding with draft model, the compile range must be extended
+        # by 1 for each sequence.
+        compile_range_end = self.scheduler_config.max_num_batched_tokens
+        if compile_range_end is not None:
+            do_extend: bool = (
+                self.speculative_config is not None
+                and self.speculative_config.uses_draft_model()
+            )
+            if do_extend:
+                compile_range_end += self.scheduler_config.max_num_seqs
+
+            computed_compile_ranges_split_points.append(compile_range_end)
 
         # Add the compile ranges for flashinfer
         if compilation_config.pass_config.fuse_allreduce_rms:
@@ -1228,10 +1237,7 @@ class VllmConfig:
                     self.model_config.get_hidden_size()
                     * self.model_config.dtype.itemsize
                 )
-                if (
-                    max_num_batched_tokens is not None
-                    and max_token_num < max_num_batched_tokens
-                ):
+                if compile_range_end is not None and max_token_num < compile_range_end:
                     computed_compile_ranges_split_points.append(max_token_num)
                 else:
                     logger.debug(
@@ -1243,11 +1249,7 @@ class VllmConfig:
             for x in compilation_config.compile_ranges_split_points:
                 assert isinstance(x, int)
                 assert x > 0, f"Invalid compile range split point: {x}"
-                if (
-                    max_num_batched_tokens is not None
-                    and x < max_num_batched_tokens
-                    and x > 1
-                ):
+                if compile_range_end is not None and x < compile_range_end and x > 1:
                     computed_compile_ranges_split_points.append(x)
         compilation_config.compile_ranges_split_points = sorted(
             computed_compile_ranges_split_points
@@ -1316,6 +1318,14 @@ class VllmConfig:
         path = self.compilation_config.debug_dump_path / append_path
         return path
 
+    def replace(self, **kwargs):
+        """
+        Replace attributes of the config, and 'recompute' the config.
+        dataclass.replace() calls __init__() and __post_init__(), source:
+        https://docs.python.org/3/library/dataclasses.html#dataclasses.replace
+        """
+        return replace(self, **kwargs)
+
     def __str__(self):
         return (
             f"model={self.model_config.model!r}, "

vllm/engine/arg_utils.py

@@ -1776,21 +1776,6 @@ class EngineArgs:
         ):
             _raise_unsupported_error(feature_name="Concurrent Partial Prefill")
 
-        # N-gram, Medusa, and Eagle are supported for speculative decoding.
-        if self.speculative_config is not None:
-            # speculative_config could still be a dict at this point
-            if isinstance(self.speculative_config, dict):
-                method = self.speculative_config.get("method", None)
-            else:
-                method = self.speculative_config.method
-
-            if method == "draft_model":
-                raise NotImplementedError(
-                    "Draft model speculative decoding is not supported yet. "
-                    "Please consider using other speculative decoding methods "
-                    "such as ngram, medusa, eagle, or mtp."
-                )
-
         if self.pipeline_parallel_size > 1:
             supports_pp = getattr(
                 self.distributed_executor_backend, "supports_pp", False

vllm/model_executor/model_loader/__init__.py

@@ -124,12 +124,17 @@ def get_model_loader(load_config: LoadConfig) -> BaseModelLoader:
 
 
 def get_model(
-    *, vllm_config: VllmConfig, model_config: ModelConfig | None = None
+    *,
+    vllm_config: VllmConfig,
+    model_config: ModelConfig | None = None,
+    prefix: str = "",
 ) -> nn.Module:
     loader = get_model_loader(vllm_config.load_config)
     if model_config is None:
         model_config = vllm_config.model_config
-    return loader.load_model(vllm_config=vllm_config, model_config=model_config)
+    return loader.load_model(
+        vllm_config=vllm_config, model_config=model_config, prefix=prefix
+    )
 
 
 __all__ = [

vllm/model_executor/model_loader/base_loader.py

@@ -36,7 +36,7 @@ class BaseModelLoader(ABC):
         raise NotImplementedError
 
     def load_model(
-        self, vllm_config: VllmConfig, model_config: ModelConfig
+        self, vllm_config: VllmConfig, model_config: ModelConfig, prefix: str = ""
     ) -> nn.Module:
         """Load a model with the given configurations."""
         device_config = vllm_config.device_config
@@ -48,7 +48,7 @@ class BaseModelLoader(ABC):
         with set_default_torch_dtype(model_config.dtype):
             with target_device:
                 model = initialize_model(
-                    vllm_config=vllm_config, model_config=model_config
+                    vllm_config=vllm_config, model_config=model_config, prefix=prefix
                 )
 
             log_model_inspection(model)

vllm/model_executor/model_loader/gguf_loader.py

@@ -335,7 +335,7 @@ class GGUFModelLoader(BaseModelLoader):
         )
 
     def load_model(
-        self, vllm_config: VllmConfig, model_config: ModelConfig
+        self, vllm_config: VllmConfig, model_config: ModelConfig, prefix: str = ""
     ) -> nn.Module:
         device_config = vllm_config.device_config
         local_model_path = self._prepare_weights(model_config)
@@ -364,7 +364,7 @@ class GGUFModelLoader(BaseModelLoader):
         target_device = torch.device(device_config.device)
         with set_default_torch_dtype(model_config.dtype):
             with target_device:
-                model = initialize_model(vllm_config=vllm_config)
+                model = initialize_model(vllm_config=vllm_config, prefix=prefix)
             self.load_weights(model, model_config)
 
             process_weights_after_loading(model, model_config, target_device)

vllm/model_executor/model_loader/tensorizer_loader.py

@@ -68,6 +68,7 @@ class TensorizerLoader(BaseModelLoader):
     def _load_model_serialized_cpu(
         self,
         vllm_config: VllmConfig,
+        prefix: str = "",
     ) -> nn.Module:
         """Load a serialized model with tensorizer to the CPU.
 
@@ -80,7 +81,7 @@ class TensorizerLoader(BaseModelLoader):
         model_config = vllm_config.model_config
         with set_default_torch_dtype(model_config.dtype):
             with torch.device(device_config.device):
-                model = initialize_model(vllm_config=vllm_config)
+                model = initialize_model(vllm_config=vllm_config, prefix=prefix)
 
             model.load_weights(self._get_weights_iterator())
         return model.eval()
@@ -112,7 +113,7 @@ class TensorizerLoader(BaseModelLoader):
             model.load_weights(self._get_weights_iterator())
 
     def load_model(
-        self, vllm_config: VllmConfig, model_config: ModelConfig
+        self, vllm_config: VllmConfig, model_config: ModelConfig, prefix: str = ""
     ) -> nn.Module:
         parallel_config = vllm_config.parallel_config
         self._verify_config(model_config, parallel_config)
@@ -134,7 +135,7 @@ class TensorizerLoader(BaseModelLoader):
                     )
             self.load_weights(model, model_config)
             return model
-        return self._load_model_serialized_cpu(vllm_config=vllm_config)
+        return self._load_model_serialized_cpu(vllm_config=vllm_config, prefix=prefix)
 
     @staticmethod
     def save_model(

vllm/v1/attention/backend.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from abc import ABC, abstractmethod
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
 from enum import Enum
 from typing import TYPE_CHECKING, Any, ClassVar, Generic, Protocol, TypeVar, get_args
 
@@ -329,6 +329,16 @@ class CommonAttentionMetadata:
 
     _num_computed_tokens_cache: torch.Tensor | None = None
 
+    def batch_size(self) -> int:
+        return self.seq_lens.shape[0]
+
+    def naive_query_lens(self) -> torch.Tensor:
+        """Naive because it assumes that query ends where the next query starts."""
+        return self.query_start_loc[1:] - self.query_start_loc[:-1]
+
+    def replace(self, **kwargs) -> "CommonAttentionMetadata":
+        return replace(self, **kwargs)
+
     @property
     @deprecated(
         """

vllm/v1/attention/backends/utils.py

@@ -818,3 +818,35 @@ def get_dcp_local_seq_lens(
     )
     dcp_local_seq_lens = base + remainder
     return dcp_local_seq_lens.squeeze(1)
+
+
+def extend_all_queries_by_1(
+    common_attn_metadata: CommonAttentionMetadata,
+    arange: torch.Tensor,
+    new_slot_mapping: torch.Tensor,
+) -> CommonAttentionMetadata:
+    """
+    Creates a new CommonAttentionMetadata with all query lengths increased by 1.
+    Also all seq lens are increased by 1.
+    This is useful e.g. in speculative decoding with draft models, where we
+    extend each sequence by 1 token.
+    The slot mapping is computed externally, as it requires more information.
+    """
+    cad = common_attn_metadata
+    # query start loc must be increased by [+0, +1, +2, ..., +batch_size]
+    new_query_start_loc = cad.query_start_loc + arange[: len(cad.query_start_loc)]
+    new_query_start_loc_cpu = cad.query_start_loc_cpu + torch.arange(
+        len(cad.query_start_loc_cpu), dtype=torch.int32
+    )
+    new_cad = cad.replace(
+        query_start_loc=new_query_start_loc,
+        query_start_loc_cpu=new_query_start_loc_cpu,
+        seq_lens=cad.seq_lens + 1,
+        # each request is extended by 1 token -> batch_size tokens are added
+        num_actual_tokens=cad.num_actual_tokens + cad.batch_size(),
+        # All query lens increase by 1, so max query len increases by 1
+        max_query_len=cad.max_query_len + 1,
+        max_seq_len=cad.max_seq_len + 1,
+        slot_mapping=new_slot_mapping,
+    )
+    return new_cad

vllm/v1/core/sched/scheduler.py

@@ -208,6 +208,8 @@ class Scheduler(SchedulerInterface):
             if speculative_config.use_eagle():
                 self.use_eagle = True
                 self.num_lookahead_tokens = self.num_spec_tokens
+            if speculative_config.uses_draft_model():
+                self.num_lookahead_tokens = self.num_spec_tokens
 
         # Create the KV cache manager.
         self.kv_cache_manager = KVCacheManager(

vllm/v1/spec_decode/draft_model.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any
+
+import torch
+
+from vllm.attention.layer import Attention
+from vllm.config import VllmConfig, get_layers_from_vllm_config
+from vllm.config.speculative import SpeculativeConfig
+from vllm.logger import init_logger
+from vllm.model_executor.model_loader import get_model
+from vllm.triton_utils import tl, triton
+from vllm.v1.attention.backends.utils import (
+    CommonAttentionMetadata,
+    extend_all_queries_by_1,
+)
+from vllm.v1.spec_decode.eagle import PADDING_SLOT_ID, SpecDecodeBaseProposer
+
+logger = init_logger(__name__)
+
+
+class DraftModelProposer(SpecDecodeBaseProposer):
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        device: torch.device,
+        runner=None,
+    ):
+        super().__init__(
+            vllm_config=vllm_config,
+            device=device,
+            pass_hidden_states_to_model=False,
+            runner=runner,
+        )
+        self._raise_if_multimodal()
+        self._raise_if_mrope()
+        self._raise_if_padded_drafter_batch_disabled()
+        self._raise_if_vocab_size_mismatch()
+        self._raise_if_draft_tp_mismatch()
+
+    def _block_size(self) -> int:
+        builder = self._get_attention_metadata_builder()
+        return builder.kv_cache_spec.block_size
+
+    def _raise_if_multimodal(self):
+        if self.supports_mm_inputs:
+            raise NotImplementedError(
+                "Speculative Decoding with draft models "
+                "does not support multimodal models yet"
+            )
+
+    def _raise_if_mrope(self):
+        if self.draft_model_config.uses_mrope:
+            raise NotImplementedError(
+                "Speculative Decoding with draft models does not support M-RoPE yet"
+            )
+
+    def _raise_if_padded_drafter_batch_disabled(self):
+        if self.vllm_config.speculative_config.disable_padded_drafter_batch:
+            raise NotImplementedError(
+                "Speculative Decoding with draft models only supports "
+                "padded drafter batch. Please don't pass --disable-padded-drafter-batch"
+                " in the speculative_config."
+            )
+
+    def _raise_if_vocab_size_mismatch(self):
+        self.vllm_config.speculative_config.verify_equal_vocab_size_if_draft_model()
+
+    def _raise_if_draft_tp_mismatch(self):
+        # Note(Tomas Ruiz) If we run the target model with TP > 1 and
+        # the draft model with TP = 1, then the different TP ranks collide.
+        # Specifically when all ranks compile the draft model on rank 0
+        # (because TP=1), then the torch compile cache is overwritten and corrupted.
+        # We need a mechanism like this: https://github.com/vllm-project/vllm/pull/5414
+        # To prevent this error, we assert that both TP sizes must be the same.
+        spec_cfg: SpeculativeConfig = self.vllm_config.speculative_config
+        tgt_tp = spec_cfg.target_parallel_config.tensor_parallel_size
+        draft_tp = spec_cfg.draft_parallel_config.tensor_parallel_size
+        if draft_tp != tgt_tp:
+            raise ValueError(
+                f"Currently, 'draft_tensor_parallel_size' and 'tensor_parallel_size' "
+                f"must be the same. Got {draft_tp} and {tgt_tp}. "
+                "Please pass 'draft_tensor_parallel_size' in the speculative_config."
+            )
+
+    def set_inputs_first_pass(
+        self,
+        target_token_ids: torch.Tensor,
+        next_token_ids: torch.Tensor,
+        target_positions: torch.Tensor,
+        last_token_indices: torch.Tensor | None,
+        cad: CommonAttentionMetadata,
+        num_rejected_tokens_gpu: torch.Tensor | None,
+    ) -> tuple[int, torch.Tensor, CommonAttentionMetadata]:
+        batch_size = cad.batch_size()
+        grid = (batch_size,)
+        start_locs = cad.query_start_loc[:-1]
+        end_locs = cad.query_start_loc[1:] - 1
+        if num_rejected_tokens_gpu is not None:
+            end_locs -= num_rejected_tokens_gpu
+
+        num_tokens = target_token_ids.shape[0] + batch_size
+        is_rejected_tok = torch.empty(
+            (num_tokens,), device=self.input_ids.device, dtype=torch.bool
+        )
+        merge_toks_kernel[grid](
+            target_toks_ptr=target_token_ids,
+            next_toks_ptr=next_token_ids,
+            query_start_locs_ptr=start_locs,
+            query_end_locs_ptr=end_locs,
+            out_ptr_merged_toks=self.input_ids,
+            out_ptr_is_rejected_tok=is_rejected_tok,
+            target_toks_size=target_token_ids.shape[0],
+            # passing a negative rejected_tok_fill value will raise an error
+            # when the value is used to index into embeddings.
+            # Therefore, we pass a valid integer, e.g. 0.
+            rejected_tok_fill=0,
+        )
+        merge_toks_kernel[grid](
+            target_toks_ptr=target_positions,
+            next_toks_ptr=target_positions[end_locs] + 1,
+            query_start_locs_ptr=start_locs,
+            query_end_locs_ptr=end_locs,
+            out_ptr_merged_toks=self.positions,
+            out_ptr_is_rejected_tok=is_rejected_tok,
+            target_toks_size=target_positions.shape[0],
+            rejected_tok_fill=0,
+        )
+
+        # recompute slot mapping
+        new_slot_mapping = compute_new_slot_mapping(
+            cad=cad,
+            new_positions=self.positions[:num_tokens],
+            is_rejected_token_mask=is_rejected_tok,
+            block_size=self._block_size(),
+            max_model_len=self.max_model_len,
+        )
+        # update common_attn_metadata
+        new_cad: CommonAttentionMetadata = extend_all_queries_by_1(
+            cad,
+            arange=self.arange,
+            new_slot_mapping=new_slot_mapping,
+        )
+
+        new_last_token_indices = new_cad.query_start_loc[1:] - 1
+        if num_rejected_tokens_gpu is not None:
+            new_last_token_indices -= num_rejected_tokens_gpu
+
+        return num_tokens, new_last_token_indices, new_cad
+
+    def load_model(self, target_model: Any) -> None:
+        """Takes target_model to satisfy the type checker."""
+
+        # This must be computed before loading the draft model
+        # because that mutates the forward_context of the vllm_config
+        target_attn_layer_names = set(
+            get_layers_from_vllm_config(self.vllm_config, Attention).keys()
+        )
+
+        from vllm.compilation.backends import set_model_tag
+
+        draft_vllm_config: VllmConfig = create_vllm_config_for_draft_model(
+            target_model_vllm_config=self.vllm_config
+        )
+        logger.info(
+            "Starting to load draft model %s. TP=%d, rank=%d",
+            draft_vllm_config.model_config.model,
+            draft_vllm_config.parallel_config.tensor_parallel_size,
+            draft_vllm_config.parallel_config.rank,
+        )
+        with set_model_tag("draft_model"):
+            self.model = get_model(vllm_config=draft_vllm_config, prefix="draft_model")
+
+        # This must be computed after loading the draft model
+        # because that mutates the forward_context of the vllm_config
+        draft_attn_layer_names = (
+            get_layers_from_vllm_config(self.vllm_config, Attention).keys()
+            - target_attn_layer_names
+        )
+        self.attn_layer_names = list(draft_attn_layer_names)
+
+
+def create_vllm_config_for_draft_model(
+    target_model_vllm_config: VllmConfig,
+) -> VllmConfig:
+    """The vllm_config is configured for the target model, e.g.
+    its quant_config and parallel_config. But the draft model is potentially
+    quantized differently, and has potentially different tensor_parallel_size.
+    This function creates a new vllm_config configured for the draft model.
+    The vllm_config is useful when loading the draft model with get_model().
+    """
+    old = target_model_vllm_config
+    new_parallel_config = old.speculative_config.draft_parallel_config.replace(
+        rank=old.parallel_config.rank
+    )
+    new: VllmConfig = old.replace(
+        quant_config=None,  # quant_config is recomputed in __init__()
+        model_config=old.speculative_config.draft_model_config,
+        parallel_config=new_parallel_config,
+    )
+    return new
+
+
+def compute_new_slot_mapping(
+    cad: CommonAttentionMetadata,
+    new_positions: torch.Tensor,
+    is_rejected_token_mask: torch.Tensor,
+    block_size: int,
+    max_model_len: int,
+):
+    batch_size, n_blocks_per_req = cad.block_table_tensor.shape
+    req_indices = torch.arange(batch_size, device=cad.query_start_loc.device)
+    req_indices = torch.repeat_interleave(
+        req_indices, cad.naive_query_lens() + 1, output_size=len(new_positions)
+    )
+    # Clamp the positions to prevent an out-of-bounds error when indexing
+    # into block_table_tensor.
+    clamped_positions = torch.clamp(new_positions, max=max_model_len - 1)
+    block_table_indices = (
+        req_indices * n_blocks_per_req + clamped_positions // block_size
+    )
+    block_nums = cad.block_table_tensor.view(-1)[block_table_indices]
+    block_offsets = clamped_positions % block_size
+    new_slot_mapping = block_nums * block_size + block_offsets
+    # Mask out the position ids that exceed the max model length.
+    exceeds_max_model_len = new_positions >= max_model_len
+    new_slot_mapping.masked_fill_(exceeds_max_model_len, PADDING_SLOT_ID)
+    # Mask out rejected tokens to prevent saves to the KV cache.
+    new_slot_mapping.masked_fill_(is_rejected_token_mask, PADDING_SLOT_ID)
+    return new_slot_mapping
+
+
+@triton.jit
+def merge_toks_kernel(
+    target_toks_ptr,
+    next_toks_ptr,
+    query_start_locs_ptr,
+    query_end_locs_ptr,
+    out_ptr_merged_toks,
+    out_ptr_is_rejected_tok,
+    target_toks_size,
+    rejected_tok_fill,
+):
+    """
+    Merges the `target_toks_ptr` and the `next_toks_ptr` into a new tensor
+    called `out_ptr_merged_toks`. Rejected tokens are those after the
+    `query_end_locs_ptr` and before the next `query_start_locs_ptr`. Fills the
+    rejected tokens positions with the value `rejected_tok_fill`. Also fills a mask
+    of the rejected tokens in `out_ptr_is_rejected_tok`.
+    """
+    pid = tl.program_id(0)
+    start_loc = tl.load(query_start_locs_ptr + pid)
+    is_last_program = pid == tl.num_programs(0) - 1
+    if is_last_program:
+        next_start_loc = target_toks_size.to(tl.int32)
+    else:
+        next_start_loc = tl.load(query_start_locs_ptr + pid + 1).to(tl.int32)
+
+    end_loc = tl.load(query_end_locs_ptr + pid)
+    new_val = tl.load(next_toks_ptr + pid)
+    for i in range(start_loc, next_start_loc + 1):
+        if i <= end_loc:  # copy existing tokens
+            old_val = tl.load(target_toks_ptr + i)
+            tl.store(out_ptr_merged_toks + pid + i, old_val)
+            tl.store(out_ptr_is_rejected_tok + pid + i, False)
+        elif i == end_loc + 1:  # copy bonus token
+            tl.store(out_ptr_merged_toks + pid + i, new_val)
+            tl.store(out_ptr_is_rejected_tok + pid + i, False)
+        else:  # fill rejected tokens
+            tl.store(out_ptr_merged_toks + pid + i, rejected_tok_fill)
+            tl.store(out_ptr_is_rejected_tok + pid + i, True)

vllm/v1/spec_decode/eagle.py

@@ -53,11 +53,12 @@ logger = init_logger(__name__)
 PADDING_SLOT_ID = -1
 
 
-class EagleProposer:
+class SpecDecodeBaseProposer:
     def __init__(
         self,
         vllm_config: VllmConfig,
         device: torch.device,
+        pass_hidden_states_to_model: bool,
         runner=None,
     ):
         self.vllm_config = vllm_config
@@ -65,6 +66,7 @@ class EagleProposer:
         assert self.speculative_config is not None
         self.draft_model_config = self.speculative_config.draft_model_config
         self.method = self.speculative_config.method
+        self.pass_hidden_states_to_model = pass_hidden_states_to_model
 
         self.runner = runner
         self.device = device
@@ -72,7 +74,11 @@ class EagleProposer:
         self.max_model_len = vllm_config.model_config.max_model_len
         self.dp_rank = vllm_config.parallel_config.data_parallel_rank
         self.num_speculative_tokens = self.speculative_config.num_speculative_tokens
-        self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens
+        # The drafter can get longer sequences than the target model.
+        max_batch_size = vllm_config.scheduler_config.max_num_seqs
+        self.max_num_tokens = (
+            vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size
+        )
         self.token_arange_np = np.arange(self.max_num_tokens)
         # We need to get the hidden size from the draft model config because
         # the draft model's hidden size can be different from the target model's
@@ -143,7 +149,6 @@ class EagleProposer:
 
         # We need +1 here because the arange is used to set query_start_loc,
         # which has one more element than batch_size.
-        max_batch_size = vllm_config.scheduler_config.max_num_seqs
         max_num_slots_for_arange = max(max_batch_size + 1, self.max_num_tokens)
         self.arange = torch.arange(
             max_num_slots_for_arange, device=device, dtype=torch.int32
@@ -245,11 +250,7 @@ class EagleProposer:
         mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None,
         num_rejected_tokens_gpu: torch.Tensor | None = None,
     ) -> torch.Tensor:
-        num_tokens = target_token_ids.shape[0]
-        batch_size = next_token_ids.shape[0]
-
-        if last_token_indices is None:
-            last_token_indices = common_attn_metadata.query_start_loc[1:] - 1
+        batch_size = common_attn_metadata.batch_size()
 
         if self.method == "eagle3":
             assert isinstance(self.model, Eagle3LlamaForCausalLM)
@@ -257,12 +258,17 @@ class EagleProposer:
                 target_hidden_states
             )
             assert target_hidden_states.shape[-1] == self.hidden_size
-        # Shift the input ids by one token.
-        # E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
-        self.input_ids[: num_tokens - 1] = target_token_ids[1:]
-        # Replace the last token with the next token.
-        # E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
-        self.input_ids[last_token_indices] = next_token_ids
+
+        num_tokens, last_token_indices, common_attn_metadata = (
+            self.set_inputs_first_pass(
+                target_token_ids=target_token_ids,
+                next_token_ids=next_token_ids,
+                target_positions=target_positions,
+                last_token_indices=last_token_indices,
+                cad=common_attn_metadata,
+                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
+            )
+        )
 
         assert self.runner is not None
 
@@ -311,9 +317,10 @@ class EagleProposer:
         if num_tokens_across_dp is not None:
             num_tokens_across_dp[self.dp_rank] = num_input_tokens
 
-        # copy inputs to buffer for cudagraph
-        self._set_positions(num_tokens, target_positions)
-        self.hidden_states[:num_tokens] = target_hidden_states
+        if self.pass_hidden_states_to_model:
+            # target_hidden_states and self.hidden_states can have different
+            # hidden dims. E.g. large target model and small draft model.
+            self.hidden_states[:num_tokens] = target_hidden_states
 
         if self.supports_mm_inputs:
             mm_embeds, is_mm_embed = mm_embed_inputs or (None, None)
@@ -330,6 +337,14 @@ class EagleProposer:
             input_ids = self.input_ids[:num_input_tokens]
             inputs_embeds = None
 
+        model_kwargs = {
+            "input_ids": input_ids,
+            "positions": self._get_positions(num_input_tokens),
+            "inputs_embeds": inputs_embeds,
+        }
+        if self.pass_hidden_states_to_model:
+            model_kwargs["hidden_states"] = self.hidden_states[:num_input_tokens]
+
         with set_forward_context(
             per_layer_attn_metadata,
             self.vllm_config,
@@ -337,17 +352,13 @@ class EagleProposer:
             num_tokens_across_dp=num_tokens_across_dp,
             cudagraph_runtime_mode=cudagraph_runtime_mode,
         ):
-            ret_hidden_states = self.model(
-                input_ids=input_ids,
-                positions=self._get_positions(num_input_tokens),
-                hidden_states=self.hidden_states[:num_input_tokens],
-                inputs_embeds=inputs_embeds,
-            )
-            if self.method == "mtp":
+            ret_hidden_states = self.model(**model_kwargs)
+            if not self.model_returns_tuple():
                 last_hidden_states = ret_hidden_states
                 hidden_states = last_hidden_states
             else:
                 last_hidden_states, hidden_states = ret_hidden_states
+
         sample_hidden_states = last_hidden_states[last_token_indices]
         logits = self.model.compute_logits(sample_hidden_states)
 
@@ -357,9 +368,9 @@ class EagleProposer:
             return draft_token_ids.view(-1, 1)
 
         if self.uses_mrope:
-            positions = target_positions[:, last_token_indices]
+            positions = self.positions[:, last_token_indices]
         else:
-            positions = target_positions[last_token_indices]
+            positions = self.positions[last_token_indices]
         if self.method in (
             "deepseek_mtp",
             "ernie_mtp",
@@ -527,6 +538,14 @@ class EagleProposer:
                 inputs_embeds = None
 
             # Run the model.
+            model_kwargs = {
+                "input_ids": input_ids,
+                "positions": self._get_positions(input_batch_size),
+                "inputs_embeds": inputs_embeds,
+            }
+            if self.pass_hidden_states_to_model:
+                model_kwargs["hidden_states"] = self.hidden_states[:input_batch_size]
+
             with set_forward_context(
                 per_layer_attn_metadata,
                 self.vllm_config,
@@ -534,17 +553,13 @@ class EagleProposer:
                 num_tokens_across_dp=batch_size_across_dp,
                 cudagraph_runtime_mode=cudagraph_runtime_mode,
             ):
-                ret_hidden_states = self.model(
-                    input_ids=input_ids,
-                    positions=self._get_positions(input_batch_size),
-                    hidden_states=self.hidden_states[:input_batch_size],
-                    inputs_embeds=inputs_embeds,
-                )
-                if self.method == "mtp":
+                ret_hidden_states = self.model(**model_kwargs)
+                if not self.model_returns_tuple():
                     last_hidden_states = ret_hidden_states
                     hidden_states = ret_hidden_states
                 else:
                     last_hidden_states, hidden_states = ret_hidden_states
+
             hidden_states = hidden_states[:batch_size]
             logits = self.model.compute_logits(last_hidden_states[:batch_size])
             draft_token_ids = logits.argmax(dim=-1)
@@ -554,6 +569,34 @@ class EagleProposer:
         draft_token_ids = torch.stack(draft_token_ids_list, dim=1)
         return draft_token_ids
 
+    def set_inputs_first_pass(
+        self,
+        target_token_ids: torch.Tensor,
+        next_token_ids: torch.Tensor,
+        target_positions: torch.Tensor,
+        last_token_indices: torch.Tensor | None,
+        cad: CommonAttentionMetadata,
+        num_rejected_tokens_gpu: torch.Tensor | None,
+    ) -> tuple[int, torch.Tensor, CommonAttentionMetadata]:
+        if last_token_indices is None:
+            last_token_indices = cad.query_start_loc[1:] - 1
+
+        num_tokens = target_token_ids.shape[0]
+        # Shift the input ids by one token.
+        # E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
+        self.input_ids[: num_tokens - 1] = target_token_ids[1:]
+        # Replace the last token with the next token.
+        # E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
+        self.input_ids[last_token_indices] = next_token_ids
+
+        # copy inputs to buffer for cudagraph
+        self._set_positions(num_tokens, target_positions)
+
+        return num_tokens, last_token_indices, cad
+
+    def model_returns_tuple(self) -> bool:
+        return self.method not in ("mtp", "draft_model")
+
     def prepare_next_token_ids_cpu(
         self,
         sampled_token_ids: list[list[int]],
@@ -1214,12 +1257,14 @@ class EagleProposer:
                     input_ids = self.input_ids[:num_input_tokens]
                     inputs_embeds = None
 
-                self.model(
+                kwargs = dict(
                     input_ids=input_ids,
                     positions=self._get_positions(num_input_tokens),
-                    hidden_states=self.hidden_states[:num_input_tokens],
                     inputs_embeds=inputs_embeds,
                 )
+                if self.pass_hidden_states_to_model:
+                    kwargs["hidden_states"] = self.hidden_states[:num_input_tokens]
+                self.model(**kwargs)
 
     def _get_attention_metadata_builder(self) -> AttentionMetadataBuilder:
         """Find and return the attention metadata builders for EAGLE layers.
@@ -1264,8 +1309,8 @@ class EagleProposer:
 
     def validate_same_kv_cache_group(self, kv_cache_config: KVCacheConfig) -> None:
         """
-        Validate that all eagle layers belong to the same KVCacheGroup.
-        Need this assumption to ensure all eagle layers can use the
+        Validate that all drafting layers belong to the same KVCacheGroup.
+        Need this assumption to ensure all drafting layers can use the
         same AttentionMetadata.
         May extend to multiple AttentionMetadata in the future.
         """
@@ -1283,7 +1328,7 @@ class EagleProposer:
                 )
             )
             == 1
-        ), "All eagle layers should belong to the same kv cache group"
+        ), "All drafting layers should belong to the same kv cache group"
 
     def _pad_batch_across_dp(
         self,
@@ -1308,6 +1353,21 @@ class EagleProposer:
         return num_tokens_dp_padded, num_toks_across_dp
 
 
+class EagleProposer(SpecDecodeBaseProposer):
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        device: torch.device,
+        runner=None,
+    ):
+        super().__init__(
+            vllm_config,
+            device,
+            pass_hidden_states_to_model=True,
+            runner=runner,
+        )
+
+
 # NOTE(woosuk): Currently, the below code is not used and we always use argmax
 # to sample the draft tokens. We will use this after we find a way to manage
 # the draft prob tensor.

vllm/v1/worker/gpu_model_runner.py

@@ -145,6 +145,7 @@ from vllm.v1.sample.logits_processor.interface import LogitsProcessor
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.sample.rejection_sampler import RejectionSampler
 from vllm.v1.sample.sampler import Sampler
+from vllm.v1.spec_decode.draft_model import DraftModelProposer
 from vllm.v1.spec_decode.eagle import EagleProposer
 from vllm.v1.spec_decode.medusa import MedusaProposer
 from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
@@ -432,10 +433,20 @@ class GPUModelRunner(
         # layers in the draft model.
         if self.speculative_config and get_pp_group().is_last_rank:
             self.drafter: (
-                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
+                NgramProposer
+                | SuffixDecodingProposer
+                | EagleProposer
+                | DraftModelProposer
+                | MedusaProposer
             )
             if self.speculative_config.method == "ngram":
                 self.drafter = NgramProposer(self.vllm_config)
+            elif self.speculative_config.uses_draft_model():
+                self.drafter = DraftModelProposer(
+                    vllm_config=self.vllm_config,
+                    device=self.device,
+                    runner=self,
+                )
             elif self.speculative_config.method == "suffix":
                 self.drafter = SuffixDecodingProposer(self.vllm_config)
             elif self.speculative_config.use_eagle():
@@ -3443,10 +3454,13 @@ class GPUModelRunner(
                 spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
                 <= self.effective_drafter_max_model_len
             )
-            if spec_config.use_eagle() and not spec_config.disable_padded_drafter_batch:
-                # EAGLE speculative decoding can use the GPU sampled tokens
+            use_gpu_toks = (
+                spec_config.use_eagle() or spec_config.uses_draft_model()
+            ) and not spec_config.disable_padded_drafter_batch
+            if use_gpu_toks:
+                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
                 # as inputs, and does not need to wait for bookkeeping to finish.
-                assert isinstance(self.drafter, EagleProposer)
+                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
                 sampled_token_ids = sampler_output.sampled_token_ids
                 if input_fits_in_drafter:
                     propose_draft_token_ids(sampled_token_ids)
@@ -3679,8 +3693,8 @@ class GPUModelRunner(
                 target_hidden_states=hidden_states,
                 sampling_metadata=sampling_metadata,
             )
-        elif spec_config.use_eagle():
-            assert isinstance(self.drafter, EagleProposer)
+        elif spec_config.use_eagle() or spec_config.uses_draft_model():
+            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
 
             if spec_config.disable_padded_drafter_batch:
                 # When padded-batch is disabled, the sampled_token_ids should be
@@ -4475,8 +4489,12 @@ class GPUModelRunner(
             else:
                 hidden_states = outputs
 
-            if self.speculative_config and self.speculative_config.use_eagle():
-                assert isinstance(self.drafter, EagleProposer)
+            if self.speculative_config and (
+                self.speculative_config.use_eagle()
+                or self.speculative_config.uses_draft_model()
+            ):
+                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
+                assert self.speculative_config is not None
                 # Eagle currently only supports PIECEWISE cudagraphs.
                 # Therefore only use cudagraphs if the main model uses PIECEWISE
                 # NOTE(lucas): this is a hack, need to clean up.
@@ -5652,8 +5670,11 @@ class GPUModelRunner(
             kv_cache_config, kernel_block_sizes
         )
 
-        if self.speculative_config and self.speculative_config.use_eagle():
-            assert isinstance(self.drafter, EagleProposer)
+        if self.speculative_config and (
+            self.speculative_config.use_eagle()
+            or self.speculative_config.uses_draft_model()
+        ):
+            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
             # validate all draft model layers belong to the same kv cache
             # group
             self.drafter.validate_same_kv_cache_group(kv_cache_config)