model_runner.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
NOTE: Coding style guide for this file:
This model runner is shared by all models: text and multimodal, generative
and embedding, public and private. As a result, this file must only contain
code that is common to every model. Model-specific behavior belongs in the
appropriate model-specific files.

In other words:
* Be paranoid about changing this file. It should remain stable.
* Be even more paranoid about adding new lines. It should remain minimal.

Even for shared features (for example, different parallelism modes), keep the
complexity out of this path. The less common the feature, the more it should be
hidden. Prefer utility functions defined elsewhere and call them from here,
instead of embedding feature-specific logic directly.
"""

import functools
import gc
import time
from copy import deepcopy
from typing import Any, NamedTuple

import numpy as np
import torch
import torch.nn as nn

from vllm.config import VllmConfig
from vllm.config.compilation import CUDAGraphMode
from vllm.distributed.parallel_state import (
    get_dcp_group,
    get_pp_group,
    prepare_communication_buffer_for_model,
)
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.logger import init_logger
from vllm.model_executor.model_loader import get_model_loader
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.sequence import IntermediateTensors
from vllm.tasks import SupportedTask
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.kv_cache_interface import KVCacheConfig
from vllm.v1.outputs import DraftTokenIds, KVConnectorOutput, ModelRunnerOutput
from vllm.v1.worker.cp_utils import check_attention_cp_compatibility
from vllm.v1.worker.gpu.async_utils import AsyncOutput, AsyncPoolingOutput
from vllm.v1.worker.gpu.attn_utils import (
    build_slot_mappings_by_layer,
    get_kv_cache_spec,
    init_attn_backend,
    init_kv_cache,
)
from vllm.v1.worker.gpu.block_table import BlockTables
from vllm.v1.worker.gpu.buffer_utils import async_copy_to_gpu
from vllm.v1.worker.gpu.cp_utils import prepare_dcp_local_seq_lens
from vllm.v1.worker.gpu.cudagraph_utils import (
    BatchExecutionDescriptor,
    ModelCudaGraphManager,
    get_uniform_token_count,
)
from vllm.v1.worker.gpu.dp_utils import sync_cudagraph_and_dp_padding
from vllm.v1.worker.gpu.input_batch import (
    InputBatch,
    InputBuffers,
    combine_sampled_and_draft_tokens,
    expand_idx_mapping,
    get_num_sampled_and_rejected,
    post_update,
    post_update_pool,
    prepare_pos_seq_lens,
    prepare_prefill_inputs,
)
from vllm.v1.worker.gpu.kv_connector import (
    NO_OP_KV_CONNECTOR,
    KVConnector,
    get_kv_connector,
)
from vllm.v1.worker.gpu.lora_utils import LoraState
from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
from vllm.v1.worker.gpu.model_states import init_model_state
from vllm.v1.worker.gpu.pool.pooling_runner import PoolingRunner
from vllm.v1.worker.gpu.pp_utils import pp_broadcast, pp_receive
from vllm.v1.worker.gpu.sample.output import SamplerOutput
from vllm.v1.worker.gpu.sample.prompt_logprob import PromptLogprobsWorker
from vllm.v1.worker.gpu.sample.sampler import Sampler
from vllm.v1.worker.gpu.spec_decode import init_speculator
from vllm.v1.worker.gpu.spec_decode.eagle.eagle3_utils import (
    set_eagle3_aux_hidden_state_layers,
)
from vllm.v1.worker.gpu.spec_decode.rejection_sampler import RejectionSampler
from vllm.v1.worker.gpu.spec_decode.utils import DraftTokensHandler
from vllm.v1.worker.gpu.states import RequestState
from vllm.v1.worker.gpu.structured_outputs import StructuredOutputsWorker
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin

logger = init_logger(__name__)


class GPUModelRunner(LoRAModelRunnerMixin):
    def __init__(self, vllm_config: VllmConfig, device: torch.device):
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        self.compilation_config = vllm_config.compilation_config
        self.lora_config = vllm_config.lora_config
        self.load_config = vllm_config.load_config
        self.parallel_config = vllm_config.parallel_config
        self.scheduler_config = vllm_config.scheduler_config
        self.speculative_config = vllm_config.speculative_config
        self.observability_config = vllm_config.observability_config

        self.device = device
        self.dtype = self.model_config.dtype
        self.kv_cache_dtype = self.dtype
        if self.cache_config.cache_dtype != "auto":
            # Quantized KV cache.
            self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                self.cache_config.cache_dtype
            ]

        self.vocab_size = self.model_config.get_vocab_size()
        self.max_model_len = self.model_config.max_model_len
        self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
        self.max_num_reqs = self.scheduler_config.max_num_seqs
        self.is_encoder_decoder = self.model_config.is_encoder_decoder

        self.use_async_scheduling = self.scheduler_config.async_scheduling
        self.output_copy_stream = torch.cuda.Stream(self.device)
        self.output_copy_event = torch.cuda.Event()

        # Pipeline parallelism.
        self.pp_size = self.parallel_config.pipeline_parallel_size
        self.use_pp = self.pp_size > 1
        if self.use_pp:
            self.is_first_pp_rank = get_pp_group().is_first_rank
            self.is_last_pp_rank = get_pp_group().is_last_rank
        else:
            self.is_first_pp_rank = True
            self.is_last_pp_rank = True

        # Data parallelism.
        self.dp_size = self.parallel_config.data_parallel_size
        self.dp_rank = self.parallel_config.data_parallel_rank

        # Decode context parallelism.
        self.dcp_size = self.parallel_config.decode_context_parallel_size
        self.use_dcp = self.dcp_size > 1
        self.dcp_rank = get_dcp_group().rank_in_group if self.use_dcp else 0
        self.cp_interleave = self.parallel_config.cp_kv_cache_interleave_size

        # Multimodal
        self.mm_registry = MULTIMODAL_REGISTRY
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
            self.model_config
        )
        self.encoder_cache = None
        if self.supports_mm_inputs and self.is_first_pp_rank:
            self.encoder_cache = EncoderCache()

        # Speculative decoding.
        self.speculator = None
        self.num_speculative_steps = 0
        self.use_aux_hidden_state_outputs = False
        use_strict_rejection_sampling = False
        if self.speculative_config is not None:
            self.num_speculative_steps = self.speculative_config.num_speculative_tokens
            use_strict_rejection_sampling = (
                self.speculative_config.rejection_sample_method == "strict"
            )

            if self.is_last_pp_rank:
                self.speculator = init_speculator(self.vllm_config, self.device)

            if self.speculative_config.method == "eagle3":
                # EAGLE3 may require auxiliary hidden states from target model outputs.
                self.use_aux_hidden_state_outputs = True
                if self.pp_size > 1:
                    raise ValueError("EAGLE3 with pipeline parallel is not supported.")

        # Draft tokens propagation - for spec-dec + struct outputs.
        self.draft_tokens_handler = DraftTokensHandler(self.device)

        # Pooling models.
        self.is_pooling_model = self.model_config.runner_type == "pooling"
        self.pooling_runner: PoolingRunner | None = None

        # General request states.
        self.req_states = RequestState(
            max_num_reqs=self.max_num_reqs,
            max_model_len=self.max_model_len,
            max_num_batched_tokens=self.max_num_tokens,
            num_speculative_steps=self.num_speculative_steps,
            vocab_size=self.vocab_size,
            device=self.device,
            model_dtype=self.dtype,
            cache_draft_logits=not use_strict_rejection_sampling,
        )
        self.input_buffers = InputBuffers(
            max_num_reqs=self.max_num_reqs,
            max_num_tokens=self.max_num_tokens,
            device=self.device,
        )

        self.sampler: Sampler | None = None
        self.rejection_sampler: RejectionSampler | None = None
        self.prompt_logprobs_worker: PromptLogprobsWorker | None = None
        self.structured_outputs_worker: StructuredOutputsWorker | None = None
        if self.is_last_pp_rank and not self.is_pooling_model:
            # Initialize sampling-related workers.
            # These components are only set up on the last PP rank and
            # for generative (non-pooling) models.
            self.sampler = Sampler(
                max_num_reqs=self.max_num_reqs,
                vocab_size=self.vocab_size,
                device=self.device,
                req_states=self.req_states,
                logprobs_mode=self.model_config.logprobs_mode,
                num_speculative_tokens=self.num_speculative_steps + 1,
            )
            self.rejection_sampler = RejectionSampler(
                self.sampler,
                num_speculative_steps=self.num_speculative_steps,
                use_strict_rejection_sampling=use_strict_rejection_sampling,
            )
            self.prompt_logprobs_worker = PromptLogprobsWorker(self.max_num_reqs)
            self.structured_outputs_worker = StructuredOutputsWorker(
                max_num_logits=self.max_num_reqs * (self.num_speculative_steps + 1),
                vocab_size=self.vocab_size,
                device=self.device,
            )

        # CUDA graphs.
        self.decode_query_len = self.num_speculative_steps + 1
        self.cudagraph_manager = ModelCudaGraphManager(
            self.vllm_config,
            self.device,
            self.compilation_config.cudagraph_mode,
            decode_query_len=self.decode_query_len,
        )
        # LoRA-related workers.
        self.lora_state = LoraState(max_num_reqs=self.max_num_reqs)
        # KV Connector if configured.
        self.kv_connector: KVConnector = NO_OP_KV_CONNECTOR

        # For transferring state from execute_model to subsequent sample_tokens call.
        self.execute_model_state: ExecuteModelState | None = None

    def update_max_model_len(self, max_model_len: int) -> None:
        self.max_model_len = max_model_len
        self.req_states.max_model_len = max_model_len

    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        tasks: list[SupportedTask] = []
        if self.model_config.runner_type == "generate":
            tasks.extend(self.model_state.get_supported_generation_tasks())
        if self.is_pooling_model:
            # Do not rely on pooling_runner here, since this information is needed
            # on the first PP rank, while pooling_runner is only initialized
            # on the last PP rank.
            tasks.extend(PoolingRunner.get_supported_tasks(self.model))
        return tuple(tasks)

    def load_model(self, *args, **kwargs) -> None:
        time_before_load = time.perf_counter()
        with DeviceMemoryProfiler() as m:
            model_loader = get_model_loader(self.vllm_config.load_config)
            logger.info("Loading model from scratch...")

            self.model = model_loader.load_model(
                vllm_config=self.vllm_config,
                model_config=self.vllm_config.model_config,
            )
            if self.lora_config:
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )

            if self.use_aux_hidden_state_outputs:
                assert self.speculative_config is not None
                set_eagle3_aux_hidden_state_layers(self.model, self.speculative_config)
            if self.speculator is not None:
                self.speculator.load_model(self.model)
        time_after_load = time.perf_counter()

        self.model_memory_usage = m.consumed_memory
        logger.info(
            "Model loading took %s GiB and %.6f seconds",
            format_gib(m.consumed_memory),
            time_after_load - time_before_load,
        )

        prepare_communication_buffer_for_model(self.model)
        if self.speculator is not None:
            prepare_communication_buffer_for_model(self.speculator.model)

        # Initialize the components that require the model.
        self.model_state = init_model_state(
            self.vllm_config, self.model, self.encoder_cache, self.device
        )
        if self.is_pooling_model and self.is_last_pp_rank:
            self.pooling_runner = PoolingRunner(self.model)

    def get_model(self) -> nn.Module:
        return self.model

    @functools.cached_property
    def main_stream(self) -> torch.cuda.Stream:
        # Cache the default CUDA stream to avoid lookup overhead.
        return torch.cuda.current_stream(self.device)

    def get_kv_cache_spec(self):
        return get_kv_cache_spec(self.vllm_config)

    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
        kv_cache_config = deepcopy(kv_cache_config)
        self.kv_cache_config = kv_cache_config
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
        ]

        block_table_max_model_len = self.max_model_len
        if self.is_encoder_decoder:
            # Cross-attention block tables need to index encoder tokens
            # (e.g., Whisper ~1500), which can exceed decoder max_model_len.
            block_table_max_model_len = max(
                block_table_max_model_len,
                getattr(self.model_config.hf_config, "max_source_positions", 0),
            )

        self.block_tables = BlockTables(
            block_sizes=block_sizes,
            max_num_reqs=self.max_num_reqs,
            max_num_batched_tokens=self.max_num_tokens,
            max_model_len=block_table_max_model_len,
            device=self.device,
            cp_size=self.dcp_size,
            cp_rank=self.dcp_rank,
            cp_interleave=self.cp_interleave,
        )

        self.attn_backends, self.attn_groups = init_attn_backend(
            self.kv_cache_config, self.vllm_config, self.device
        )
        check_attention_cp_compatibility(self.vllm_config)
        if self.speculator is not None:
            # HACK(woosuk)
            self.speculator.set_attn(
                self.model_state,
                self.kv_cache_config,
                self.attn_groups,
                self.block_tables,
            )

        self.kv_caches: list[torch.Tensor] = []
        kv_caches_dict = init_kv_cache(
            self.kv_caches,
            self.compilation_config.static_forward_context,
            self.kv_cache_config,
            self.attn_backends,
            self.device,
        )
        self.kv_connector = get_kv_connector(self.vllm_config, kv_caches_dict)

    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
        *args,
        skip_attn: bool = True,
        uniform_decode: bool = False,
        **kwargs,
    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
        # Create a dummy scheduler output.
        num_reqs = min(num_tokens, self.max_num_reqs)
        if uniform_decode:
            # HACK(lucas): for now since the worker is shared between MRV1 and MRV2,
            # and for spec-decode with MTP we want to make sure the dummy runs use
            # 1+num_speculative_tokens we use max here, this will likely be eventually
            # changed in the worker: https://github.com/vllm-project/vllm/pull/35243
            num_tokens = max(num_tokens, self.decode_query_len)
            num_reqs = num_tokens // self.decode_query_len
            assert num_tokens % self.decode_query_len == 0
        num_tokens_per_request = [num_tokens // num_reqs] * num_reqs
        num_tokens_per_request[-1] += num_tokens % num_reqs

        assert sum(num_tokens_per_request) == num_tokens
        num_scheduled_tokens = {
            f"_dummy_req_{i}": n for i, n in enumerate(num_tokens_per_request)
        }
        dummy_scheduler_output = SchedulerOutput.make_empty()
        dummy_scheduler_output.total_num_scheduled_tokens = num_tokens
        dummy_scheduler_output.num_scheduled_tokens = num_scheduled_tokens

        # Disable any use of KVConnector for dummy runs.
        self.kv_connector.set_disabled(True)

        # For non-first PP ranks, create dummy intermediate_tensors.
        intermediate_tensors = None
        if not self.is_first_pp_rank:
            intermediate_tensors = self.model.make_empty_intermediate_tensors(
                batch_size=num_tokens,
                dtype=self.model_config.dtype,
                device=self.device,
            )

        # Execute the model.
        self.execute_model(
            dummy_scheduler_output,
            intermediate_tensors=intermediate_tensors,
            dummy_run=True,
            skip_attn_for_dummy_run=skip_attn,
        )
        self.kv_connector.set_disabled(False)

        # Non-last PP ranks don't produce output for sampling.
        if not self.is_last_pp_rank:
            return None, None

        assert self.execute_model_state is not None
        input_batch = self.execute_model_state.input_batch
        attn_metadata = self.execute_model_state.attn_metadata
        slot_mappings_by_layer = self.execute_model_state.slot_mappings_by_layer
        hidden_states = self.execute_model_state.hidden_states
        aux_hidden_states = self.execute_model_state.aux_hidden_states
        num_tokens_across_dp = self.execute_model_state.num_tokens_across_dp
        self.execute_model_state = None

        # dummy run the eagle speculator's propose to ensure DP/EP sync.
        if self.speculator is not None:
            assert self.sampler is not None
            self.speculator.propose(
                input_batch=input_batch,
                attn_metadata=attn_metadata,
                slot_mappings=slot_mappings_by_layer,
                last_hidden_states=hidden_states,
                aux_hidden_states=aux_hidden_states,
                num_sampled=torch.ones(
                    input_batch.num_reqs, dtype=torch.int32, device=self.device
                ),
                num_rejected=torch.zeros(
                    input_batch.num_reqs, dtype=torch.int32, device=self.device
                ),
                last_sampled=self.req_states.last_sampled_tokens,
                next_prefill_tokens=self.req_states.next_prefill_tokens,
                temperature=self.sampler.sampling_states.temperature.gpu,
                seeds=self.sampler.sampling_states.seeds.gpu,
                draft_logits_out=self.req_states.draft_logits,
                num_tokens_across_dp=num_tokens_across_dp,
                dummy_run=True,
                skip_attn_for_dummy_run=skip_attn,
            )

        assert hidden_states is not None  # Last PP rank always has hidden_states
        sample_hidden_states = hidden_states[input_batch.logits_indices]
        return hidden_states, sample_hidden_states

    @torch.inference_mode()
    def _dummy_sampler_run(self, hidden_states: torch.Tensor) -> None:
        num_reqs = hidden_states.shape[0]
        logits = self.model.compute_logits(hidden_states)
        dummy_input_batch = InputBatch.make_dummy(
            num_reqs, num_reqs, self.input_buffers
        )

        # NOTE(woosuk): During the initial memory profiling, the sampler may skip
        # top_k, top_p, and logprobs, using less GPU memory than what is possible
        # during actual execution.
        assert self.sampler is not None
        self.sampler(logits, dummy_input_batch)

    @torch.inference_mode()
    def _dummy_pooler_run(self, hidden_states: torch.Tensor) -> None:
        assert self.pooling_runner is not None
        self.pooling_runner.dummy_pooler_run(hidden_states)

    @torch.inference_mode()
    def profile_run(self) -> None:
        hidden_states, sample_hidden_states = self._dummy_run(
            self.max_num_tokens, skip_attn=True
        )

        # Only run sampler/pooler on last PP rank (non-last ranks return None).
        if self.is_last_pp_rank:
            assert sample_hidden_states is not None
            if self.pooling_runner is None:
                self._dummy_sampler_run(sample_hidden_states)
            else:
                self._dummy_pooler_run(hidden_states)

        torch.accelerator.synchronize()
        del hidden_states, sample_hidden_states
        gc.collect()

    def reset_mm_cache(self) -> None:
        if self.encoder_cache is not None:
            self.encoder_cache.reset_mm_cache()

    def reset_encoder_cache(self) -> None:
        if self.encoder_cache is not None:
            self.encoder_cache.reset_encoder_cache()

    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
        # SP is not supported yet.
        return num_scheduled_tokens

    def profile_cudagraph_memory(self) -> int:
        # NOTE(woosuk): It is TBD whether we keep this API or not.
        return 0

    @torch.inference_mode()
    def capture_model(self) -> int:
        if not self.cudagraph_manager.needs_capture():
            logger.warning(
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
            return 0

        # TODO (zhanqiu): support CUDA graph for PP.
        if self.use_pp:
            logger.warning_once(
                "Skipping CUDA graph capture because pipeline parallel is "
                "enabled. Pipeline parallel is currently eager-only.",
            )
            return 0

        start_time = time.perf_counter()
        gc.collect()
        torch.accelerator.empty_cache()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

        with self.maybe_setup_dummy_loras(self.lora_config):
            self.cudagraph_manager.capture(
                self.model,
                self.model_state,
                self.input_buffers,
                self.block_tables,
                self.attn_groups,
                self.kv_cache_config,
                has_lora=self.lora_config is not None,
                use_aux_hidden_state_outputs=self.use_aux_hidden_state_outputs,
            )
            if self.speculator is not None:
                self.speculator.capture_model()

        end_time = time.perf_counter()
        end_free_gpu_memory = torch.cuda.mem_get_info()[0]
        elapsed_time = end_time - start_time
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
        # This usually takes 5~20 seconds.
        logger.info(
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
        )
        return cuda_graph_size

    def finish_requests(self, scheduler_output: SchedulerOutput) -> None:
        finished_req_ids = scheduler_output.finished_req_ids
        preempted_req_ids = scheduler_output.preempted_req_ids
        if preempted_req_ids:
            finished_req_ids = finished_req_ids.union(preempted_req_ids)
        for req_id in finished_req_ids:
            self.req_states.remove_request(req_id)
            if self.encoder_cache is not None:
                self.encoder_cache.remove_request(req_id)
            if self.prompt_logprobs_worker is not None:
                self.prompt_logprobs_worker.remove_request(req_id)
            self.lora_state.remove_request(req_id)

    def free_states(self, scheduler_output: SchedulerOutput) -> None:
        if self.encoder_cache is not None:
            for mm_hash in scheduler_output.free_encoder_mm_hashes:
                self.encoder_cache.free_encoder_cache(mm_hash)

    def add_requests(self, scheduler_output: SchedulerOutput) -> None:
        for new_req_data in scheduler_output.scheduled_new_reqs:
            assert new_req_data.prompt_token_ids is not None
            assert new_req_data.prefill_token_ids is not None
            req_id = new_req_data.req_id
            prompt_len = len(new_req_data.prompt_token_ids)
            self.req_states.add_request(
                req_id=req_id,
                prompt_len=prompt_len,
                all_token_ids=new_req_data.prefill_token_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
            )
            req_index = self.req_states.req_id_to_index[req_id]

            if self.encoder_cache is not None:
                self.encoder_cache.add_request(req_id, new_req_data.mm_features)

            self.model_state.add_request(req_index, new_req_data)
            self.block_tables.append_block_ids(
                req_index, new_req_data.block_ids, overwrite=True
            )
            self.lora_state.add_request(req_id, req_index, new_req_data.lora_request)

            if self.is_last_pp_rank and new_req_data.sampling_params is not None:
                assert self.sampler is not None
                self.sampler.add_request(
                    req_index, prompt_len, new_req_data.sampling_params
                )
                assert self.prompt_logprobs_worker is not None
                self.prompt_logprobs_worker.add_request(
                    req_id, req_index, new_req_data.sampling_params
                )

        if scheduler_output.scheduled_new_reqs:
            self.req_states.apply_staged_writes()
            self.model_state.apply_staged_writes()
        if self.sampler is not None:
            self.sampler.apply_staged_writes()

    def update_requests(self, scheduler_output: SchedulerOutput) -> None:
        # Add new blocks for the existing requests.
        reqs = scheduler_output.scheduled_cached_reqs
        for req_new_block_ids, req_id in zip(reqs.new_block_ids, reqs.req_ids):
            if req_new_block_ids is not None:
                req_index = self.req_states.req_id_to_index[req_id]
                self.block_tables.append_block_ids(
                    req_index, req_new_block_ids, overwrite=False
                )

    def prepare_inputs(
        self, scheduler_output: SchedulerOutput, batch_desc: BatchExecutionDescriptor
    ) -> InputBatch:
        num_tokens = scheduler_output.total_num_scheduled_tokens
        num_tokens_after_padding = batch_desc.num_tokens
        assert num_tokens > 0
        num_tokens_per_req = scheduler_output.num_scheduled_tokens
        num_reqs = len(num_tokens_per_req)

        # Decode first, then prefill.
        # batch_idx -> req_id
        req_ids = sorted(num_tokens_per_req, key=num_tokens_per_req.get)  # type: ignore[arg-type]
        numtoks_iter = map(num_tokens_per_req.get, req_ids)
        num_scheduled_tokens = np.fromiter(numtoks_iter, dtype=np.int32, count=num_reqs)

        idx_mapping_iter = map(self.req_states.req_id_to_index.get, req_ids)
        idx_mapping_np = np.fromiter(idx_mapping_iter, dtype=np.int32, count=num_reqs)
        idx_mapping = async_copy_to_gpu(idx_mapping_np, device=self.device)

        # Get the number of draft tokens for each request.
        draft_tokens = scheduler_output.scheduled_spec_decode_tokens
        if not draft_tokens:
            # No draft token scheduled (common case).
            total_num_draft_tokens = 0
            total_num_logits = num_reqs
            cu_num_logits_np = np.arange(num_reqs + 1, dtype=np.int32)
            cu_num_logits = torch.arange(
                num_reqs + 1, device=self.device, dtype=torch.int32
            )
            expanded_idx_mapping = idx_mapping
            expanded_local_pos = torch.zeros(
                num_reqs, dtype=torch.int32, device=self.device
            )
        else:
            num_draft_tokens = np.fromiter(
                (len(draft_tokens.get(req_id, ())) for req_id in req_ids),
                dtype=np.int32,
                count=num_reqs,
            )
            total_num_draft_tokens = int(num_draft_tokens.sum())
            total_num_logits = num_reqs + total_num_draft_tokens

            num_logits = num_draft_tokens + 1
            cu_num_logits_np = np.empty(num_reqs + 1, dtype=np.int32)
            cu_num_logits_np[0] = 0
            np.cumsum(num_logits, out=cu_num_logits_np[1:])
            cu_num_logits = async_copy_to_gpu(cu_num_logits_np, device=self.device)

            max_expand_len = self.num_speculative_steps + 1
            expanded_idx_mapping, expanded_local_pos = expand_idx_mapping(
                idx_mapping, total_num_logits, cu_num_logits, max_expand_len
            )

        # Get query_start_loc.
        # num_reqs_padded is None for PIECEWISE graphs (no request padding needed)
        num_reqs_padded = batch_desc.num_reqs or num_reqs
        query_start_loc_np = np.empty(self.max_num_reqs + 1, dtype=np.int32)
        query_start_loc_np[0] = 0
        np.cumsum(num_scheduled_tokens, out=query_start_loc_np[1 : num_reqs + 1])
        # Pad for full CUDA graph mode.
        # Some attention backends like FA3 require query_start_loc to be non-decreasing.
        query_start_loc_np[num_reqs + 1 :] = num_tokens
        async_copy_to_gpu(query_start_loc_np, out=self.input_buffers.query_start_loc)
        query_start_loc_np = query_start_loc_np[: num_reqs_padded + 1]
        query_start_loc = self.input_buffers.query_start_loc[: num_reqs_padded + 1]

        # Get prefill tokens if any.
        if self.req_states.any_prefills(idx_mapping_np):
            prepare_prefill_inputs(
                self.input_buffers.input_ids,
                self.req_states.next_prefill_tokens,
                idx_mapping,
                query_start_loc,
                self.req_states.all_token_ids.gpu,
                self.req_states.prefill_len.gpu,
                self.req_states.num_computed_tokens.gpu,
            )

        # Prepare positions and seq_lens.
        prepare_pos_seq_lens(
            idx_mapping,
            query_start_loc,
            self.req_states.num_computed_tokens.gpu,
            self.input_buffers.positions,
            self.input_buffers.seq_lens,
        )
        seq_lens = self.input_buffers.seq_lens[:num_reqs_padded]

        dcp_local_seq_lens = None
        if self.use_dcp:
            # Prepare dcp local seq_lens.
            prepare_dcp_local_seq_lens(
                self.input_buffers.dcp_local_seq_lens,
                self.input_buffers.seq_lens,
                num_reqs,
                self.dcp_size,
                self.dcp_rank,
                self.cp_interleave,
            )
            dcp_local_seq_lens = self.input_buffers.dcp_local_seq_lens[:num_reqs_padded]

        # Some input token ids are directly read from the last sampled tokens
        # and draft tokens. Also, get the logits indices to sample tokens from.
        logits_indices = combine_sampled_and_draft_tokens(
            self.input_buffers.input_ids,
            idx_mapping,
            self.req_states.last_sampled_tokens,
            query_start_loc,
            seq_lens,
            self.req_states.prefill_len.gpu,
            self.req_states.draft_tokens,
            cu_num_logits,
            total_num_logits,
        )

        return InputBatch(
            req_ids=req_ids,
            num_reqs=num_reqs,
            num_reqs_after_padding=num_reqs_padded,
            idx_mapping=idx_mapping,
            idx_mapping_np=idx_mapping_np,
            expanded_idx_mapping=expanded_idx_mapping,
            expanded_local_pos=expanded_local_pos,
            num_scheduled_tokens=num_scheduled_tokens,
            num_tokens=num_tokens,
            num_tokens_after_padding=num_tokens_after_padding,
            num_draft_tokens=total_num_draft_tokens,
            query_start_loc=query_start_loc,
            query_start_loc_np=query_start_loc_np,
            seq_lens=seq_lens,
            dcp_local_seq_lens=dcp_local_seq_lens,
            input_ids=self.input_buffers.input_ids[:num_tokens_after_padding],
            positions=self.input_buffers.positions[:num_tokens_after_padding],
            logits_indices=logits_indices,
            cu_num_logits=cu_num_logits,
            cu_num_logits_np=cu_num_logits_np,
            has_structured_output_reqs=scheduler_output.has_structured_output_requests,
        )

    def prepare_attn(
        self, input_batch: InputBatch
    ) -> tuple[tuple[torch.Tensor, ...], torch.Tensor]:
        # Block tables: num_kv_cache_groups x [num_reqs_padded, max_num_blocks].
        block_tables = self.block_tables.gather_block_tables(
            input_batch.idx_mapping,
            num_reqs_padded=input_batch.num_reqs_after_padding,
        )
        # Slot mappings: [num_kv_cache_groups, num_tokens_padded].
        # Kernel pads beyond num_tokens with PAD_SLOT_ID.
        slot_mappings = self.block_tables.compute_slot_mappings(
            input_batch.idx_mapping,
            input_batch.query_start_loc,
            input_batch.positions,
            num_tokens_padded=input_batch.num_tokens_after_padding,
        )
        return block_tables, slot_mappings

    def prepare_dummy_attn(
        self, input_batch: InputBatch
    ) -> tuple[tuple[torch.Tensor, ...], torch.Tensor]:
        block_tables = self.block_tables.get_dummy_block_tables(input_batch.num_reqs)
        slot_mappings = self.block_tables.get_dummy_slot_mappings(
            input_batch.num_tokens
        )
        return block_tables, slot_mappings

    def sample(
        self,
        hidden_states: torch.Tensor,
        input_batch: InputBatch,
        grammar_output: GrammarOutput | None,
    ) -> tuple[SamplerOutput, torch.Tensor, torch.Tensor]:
        sample_hidden_states = hidden_states[input_batch.logits_indices]
        logits = self.model.compute_logits(sample_hidden_states)
        if grammar_output is not None:
            # Apply grammar bitmask to the logits in-place.
            assert self.structured_outputs_worker is not None
            self.structured_outputs_worker.apply_grammar_bitmask(
                logits,
                input_batch,
                grammar_output.structured_output_request_ids,
                grammar_output.grammar_bitmask,
            )

        if input_batch.num_draft_tokens == 0:
            # No draft tokens (common case).
            assert self.sampler is not None
            sampler_output = self.sampler(logits, input_batch)
        else:
            # Rejection sampling for spec decoding.
            assert self.rejection_sampler is not None
            sampler_output = self.rejection_sampler(
                logits,
                input_batch,
                # Draft logits are needed for probabilistic rejection sampling.
                self.req_states.draft_logits[input_batch.idx_mapping]
                if self.req_states.draft_logits is not None
                else None,
            )

        # Get the number of sampled and rejected tokens.
        # For chunked prefills, num_sampled and num_rejected are both 0.
        num_sampled, num_rejected = get_num_sampled_and_rejected(
            sampler_output.num_sampled,
            input_batch.seq_lens,
            input_batch.cu_num_logits,
            input_batch.idx_mapping,
            self.req_states.prefill_len.gpu,
        )
        return sampler_output, num_sampled, num_rejected

    def postprocess(
        self,
        input_batch: InputBatch,
        sampled_tokens: torch.Tensor,
        num_sampled: torch.Tensor,
        num_rejected: torch.Tensor,
    ) -> None:
        # Update the number of computed tokens.
        if self.is_last_pp_rank:
            assert self.sampler is not None
            output_bin_counts = self.sampler.penalties_state.output_bin_counts
        else:
            output_bin_counts = None
        post_update(
            input_batch.idx_mapping,
            self.req_states.num_computed_tokens.gpu,
            self.req_states.last_sampled_tokens,
            output_bin_counts,
            sampled_tokens,
            num_sampled,
            num_rejected,
            input_batch.query_start_loc,
            self.req_states.all_token_ids.gpu,
            self.req_states.total_len.gpu,
        )

        # Update the number of computed prefill tokens.
        idx_mapping_np = input_batch.idx_mapping_np
        computed_prefill = self.req_states.num_computed_prefill_tokens
        computed_prefill[idx_mapping_np] += input_batch.num_scheduled_tokens
        np.minimum(
            computed_prefill, self.req_states.prefill_len.np, out=computed_prefill
        )

    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: SchedulerOutput,
        intermediate_tensors: IntermediateTensors | None = None,
        dummy_run: bool = False,
        skip_attn_for_dummy_run: bool = False,
    ) -> ModelRunnerOutput | IntermediateTensors | None:
        if not dummy_run:
            # Update the request states.
            self.finish_requests(scheduler_output)
            self.free_states(scheduler_output)
            self.add_requests(scheduler_output)
            self.update_requests(scheduler_output)
            self.block_tables.apply_staged_writes()
            if scheduler_output.total_num_scheduled_tokens == 0:
                # No need to run the model.
                empty_output = self.kv_connector.no_forward(scheduler_output)
                return empty_output

        # Get batch descriptor and sync across DP ranks.
        num_reqs = len(scheduler_output.num_scheduled_tokens)
        num_toks = scheduler_output.total_num_scheduled_tokens
        max_query_len = max(scheduler_output.num_scheduled_tokens.values())
        uniform_tok_count = get_uniform_token_count(num_reqs, num_toks, max_query_len)

        batch_desc = self.cudagraph_manager.dispatch(
            num_reqs, num_toks, uniform_tok_count
        )
        num_tokens_across_dp = None

        skip_compiled = False
        if self.is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
            # Encoder-decoder models such as Whisper should run eager/non-compiled
            # when encoder inputs are scheduled, because this step updates
            # cross-attention cache with dynamic encoder outputs.
            # Override batch_desc to NONE.
            skip_compiled = True
            batch_desc = BatchExecutionDescriptor(
                cg_mode=CUDAGraphMode.NONE,
                num_tokens=num_toks,
                num_reqs=num_reqs,
            )

        if self.dp_size > 1:
            batch_desc, num_tokens_across_dp = sync_cudagraph_and_dp_padding(
                self.cudagraph_manager,
                batch_desc,
                num_toks,
                num_reqs,
                uniform_tok_count,
                self.dp_size,
                self.dp_rank,
            )

        if batch_desc.num_tokens == 0:
            # All DP ranks have zero tokens to run.
            empty_output = self.kv_connector.no_forward(scheduler_output)
            return empty_output

        if not dummy_run:
            # Common case.
            # Prepare all the inputs and copy to the input buffers.
            input_batch = self.prepare_inputs(scheduler_output, batch_desc)
            block_tables, slot_mappings = self.prepare_attn(input_batch)

            if self.lora_config:
                # Activate LoRA adapters.
                lora_inputs = self.lora_state.make_lora_inputs(
                    input_batch.req_ids,
                    input_batch.idx_mapping_np,
                    input_batch.num_scheduled_tokens,
                )
                self._set_active_loras(*lora_inputs)
        else:
            # No actual tokens to run. A dummy run for DP or memory profiling.
            input_batch = InputBatch.make_dummy(
                batch_desc.num_reqs or num_reqs,
                batch_desc.num_tokens,
                self.input_buffers,
            )
            if not skip_attn_for_dummy_run:
                block_tables, slot_mappings = self.prepare_dummy_attn(input_batch)
            else:
                block_tables = None
                slot_mappings = None
            # FIXME(woosuk): Fix warmup for LoRA.

        attn_metadata = None
        slot_mappings_by_layer = None
        if not (dummy_run and skip_attn_for_dummy_run):
            assert slot_mappings is not None
            slot_mappings_by_layer = build_slot_mappings_by_layer(
                slot_mappings, self.kv_cache_config
            )
            assert block_tables is not None
            attn_metadata = self.model_state.prepare_attn(
                input_batch,
                batch_desc.cg_mode,
                block_tables,
                slot_mappings,
                self.attn_groups,
                self.kv_cache_config,
            )

        inputs_embeds = None
        if self.supports_mm_inputs and self.is_first_pp_rank:
            # Run MM encoder (if needed) and get multimodal embeddings.
            # Only first PP rank prepares multimodal embeddings.
            # NOTE(woosuk): We must call get_mm_embeddings even during dummy runs
            # to obtain inputs_embeds, because the compiled model expects this input.
            inputs_embeds = self.model_state.get_mm_embeddings(
                scheduler_output.scheduled_encoder_inputs,
                input_batch,
                self.req_states,
            )

        model_inputs = {
            "input_ids": input_batch.input_ids,
            "positions": input_batch.positions,
            "inputs_embeds": inputs_embeds,
            "intermediate_tensors": intermediate_tensors,
            # NOTE: Values returned by `prepare_inputs` will override the default
            # values above.
            **self.model_state.prepare_inputs(input_batch, self.req_states),
        }
        if not self.is_first_pp_rank:
            # Update for non-first PP ranks.
            model_inputs["input_ids"] = None
            model_inputs["inputs_embeds"] = None
            assert intermediate_tensors is not None

        # Run model.
        if batch_desc.cg_mode == CUDAGraphMode.FULL:
            # Use explicit cudagraph replay for FULL mode.
            # NOTE(woosuk): Here, we don't need to pass the input tensors,
            # because they are already copied to the CUDA graph input buffers.
            self.kv_connector.pre_forward(scheduler_output)
            model_output = self.cudagraph_manager.run_fullgraph(batch_desc)
            if self.use_aux_hidden_state_outputs:
                hidden_states, aux_hidden_states = model_output
            else:
                hidden_states = model_output
                aux_hidden_states = None
        else:
            # For piecewise and eager mode, just call model().
            batch_descriptor = BatchDescriptor(
                num_tokens=input_batch.num_tokens_after_padding,
                has_lora=self.lora_config is not None,
            )

            with set_forward_context(
                attn_metadata,
                self.vllm_config,
                num_tokens=input_batch.num_tokens_after_padding,
                cudagraph_runtime_mode=batch_desc.cg_mode,
                num_tokens_across_dp=num_tokens_across_dp,
                batch_descriptor=batch_descriptor,
                slot_mapping=slot_mappings_by_layer,
                skip_compiled=skip_compiled,
            ):
                self.kv_connector.pre_forward(scheduler_output)
                model_output = self.model(**model_inputs)
                if self.use_aux_hidden_state_outputs:
                    hidden_states, aux_hidden_states = model_output
                else:
                    hidden_states = model_output
                    aux_hidden_states = None

        kv_connector_output = self.kv_connector.post_forward(scheduler_output)
        self.execute_model_state = ExecuteModelState(
            input_batch=input_batch,
            attn_metadata=attn_metadata,
            slot_mappings_by_layer=slot_mappings_by_layer,
            hidden_states=hidden_states,
            aux_hidden_states=aux_hidden_states,
            kv_connector_output=kv_connector_output,
            num_tokens_across_dp=num_tokens_across_dp,
        )

        if not self.is_last_pp_rank:
            # Non-last PP rank: return IntermediateTensors for sending.
            assert isinstance(hidden_states, IntermediateTensors)
            hidden_states.kv_connector_output = kv_connector_output
            return hidden_states
        # Last rank (or no PP): hidden_states is a tensor for sampling.
        assert isinstance(hidden_states, torch.Tensor)
        return None

    @torch.inference_mode()
    def sample_tokens(
        self, grammar_output: GrammarOutput | None
    ) -> AsyncOutput | ModelRunnerOutput | None:
        if self.execute_model_state is None:
            # The prior execute_model call must have failed.
            return None

        input_batch = self.execute_model_state.input_batch
        attn_metadata = self.execute_model_state.attn_metadata
        slot_mappings_by_layer = self.execute_model_state.slot_mappings_by_layer
        hidden_states = self.execute_model_state.hidden_states
        aux_hidden_states = self.execute_model_state.aux_hidden_states
        kv_connector_output = self.execute_model_state.kv_connector_output
        num_tokens_across_dp = self.execute_model_state.num_tokens_across_dp
        self.execute_model_state = None

        if not self.is_last_pp_rank:
            # Non-last PP rank: hidden_states is None because this rank produced
            # IntermediateTensors instead of final hidden states. Receive the
            # sampled tokens broadcast from the last rank and update local state.
            sampled, num_sampled, num_rejected = pp_receive(
                input_batch.num_reqs, max_sample_len=self.num_speculative_steps + 1
            )
            self.postprocess(input_batch, sampled, num_sampled, num_rejected)
            return None

        # Last rank: sample tokens
        sampler_output, num_sampled, num_rejected = self.sample(
            hidden_states, input_batch, grammar_output
        )

        if self.use_pp:
            # Broadcast to non-last PP ranks (handles spec decode multi-token).
            pp_broadcast(sampler_output.sampled_token_ids, num_sampled, num_rejected)

        assert self.prompt_logprobs_worker is not None
        prompt_logprobs_dict = self.prompt_logprobs_worker.compute_prompt_logprobs(
            self.model.compute_logits,
            hidden_states,
            input_batch,
            self.req_states.all_token_ids.gpu,
            self.req_states.num_computed_tokens.gpu,
            self.req_states.prompt_len.np,
            self.req_states.prefill_len.np,
            self.req_states.num_computed_prefill_tokens,
        )

        # Prepare the model runner output.
        model_runner_output = ModelRunnerOutput(
            req_ids=input_batch.req_ids,
            # NOTE(woosuk): req_id_to_index is unused in this model runner.
            # Only for compatibility with the existing model runner and scheduler.
            req_id_to_index={req_id: i for i, req_id in enumerate(input_batch.req_ids)},
            sampled_token_ids=None,  # type: ignore
            prompt_logprobs_dict=prompt_logprobs_dict,  # type: ignore[arg-type]
            kv_connector_output=kv_connector_output,
        )
        async_output = AsyncOutput(
            model_runner_output=model_runner_output,
            sampler_output=sampler_output,
            num_sampled_tokens=num_sampled,
            main_stream=self.main_stream,
            copy_stream=self.output_copy_stream,
            copy_event=self.output_copy_event,
        )

        # Postprocess results and update request states.
        # NOTE: This is intentionally done after creating the AsyncOutput,
        # ensuring that `copy_event` is recorded before calling postprocess.
        # This sequencing may slightly reduce latency as async D2H copy does not
        # need to wait for the postprocess to finish.
        self.postprocess(
            input_batch, sampler_output.sampled_token_ids, num_sampled, num_rejected
        )
        if self.speculator is not None:
            assert self.sampler is not None
            draft_tokens = self.speculator.propose(
                input_batch,
                attn_metadata,
                slot_mappings_by_layer,
                hidden_states,
                aux_hidden_states,
                num_sampled,
                num_rejected,
                self.req_states.last_sampled_tokens,
                self.req_states.next_prefill_tokens,
                self.sampler.sampling_states.temperature.gpu,
                self.sampler.sampling_states.seeds.gpu,
                self.req_states.draft_logits,
                num_tokens_across_dp=num_tokens_across_dp,
            )
            self.req_states.draft_tokens[input_batch.idx_mapping] = draft_tokens
            self.draft_tokens_handler.set_draft_tokens(input_batch, draft_tokens)

        if self.use_async_scheduling:
            return async_output
        return async_output.get_output()

    def take_draft_token_ids(self) -> DraftTokenIds | None:
        return self.draft_tokens_handler.get_draft_tokens()

    @torch.inference_mode()
    def pool(self) -> AsyncPoolingOutput | ModelRunnerOutput | None:
        if self.execute_model_state is None:
            # The prior execute_model call must have failed.
            return None

        input_batch = self.execute_model_state.input_batch
        hidden_states = self.execute_model_state.hidden_states
        kv_connector_output = self.execute_model_state.kv_connector_output
        self.execute_model_state = None

        if not self.is_last_pp_rank:
            self.postprocess_pool(input_batch)
            return None

        assert self.pooling_runner is not None
        pooler_output, is_valid = self.pooling_runner.pool(
            hidden_states, input_batch, self.req_states
        )
        self.postprocess_pool(input_batch)

        # Build the model runner output.
        model_runner_output = ModelRunnerOutput(
            req_ids=input_batch.req_ids,
            req_id_to_index={req_id: i for i, req_id in enumerate(input_batch.req_ids)},
            kv_connector_output=kv_connector_output,
        )
        async_output = AsyncPoolingOutput(
            model_runner_output=model_runner_output,
            pooler_output=pooler_output,
            is_valid=is_valid,
            main_stream=self.main_stream,
            copy_stream=self.output_copy_stream,
            copy_event=self.output_copy_event,
        )
        if self.use_async_scheduling:
            return async_output
        return async_output.get_output()

    def postprocess_pool(self, input_batch: InputBatch) -> None:
        # Update the number of computed tokens.
        post_update_pool(
            input_batch.idx_mapping,
            self.req_states.num_computed_tokens.gpu,
            input_batch.query_start_loc,
        )

        # Update the number of computed prefill tokens.
        idx_mapping_np = input_batch.idx_mapping_np
        computed_prefill = self.req_states.num_computed_prefill_tokens
        computed_prefill[idx_mapping_np] += input_batch.num_scheduled_tokens
        np.minimum(
            computed_prefill, self.req_states.prefill_len.np, out=computed_prefill
        )


class ExecuteModelState(NamedTuple):
    input_batch: InputBatch
    attn_metadata: dict[str, Any] | None
    slot_mappings_by_layer: dict[str, torch.Tensor] | None
    hidden_states: torch.Tensor | IntermediateTensors
    aux_hidden_states: list[torch.Tensor] | None
    kv_connector_output: KVConnectorOutput | None
    num_tokens_across_dp: torch.Tensor | None