gpu_model_runner.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import gc
import itertools
import time
from collections import defaultdict
from collections.abc import Iterator
from contextlib import contextmanager
from copy import deepcopy
from typing import TYPE_CHECKING, Any, Optional, Union, cast

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
from tqdm import tqdm

import vllm.envs as envs
from vllm.attention import Attention, AttentionType
from vllm.attention.backends.abstract import AttentionBackend
from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
from vllm.compilation.counter import compilation_counter
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
from vllm.config import (CompilationLevel, CUDAGraphMode, VllmConfig,
                         get_layers_from_vllm_config, update_config)
from vllm.distributed.eplb.eplb_state import EplbState
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group)
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
from vllm.distributed.parallel_state import (
    get_pp_group, get_tp_group, graph_capture, is_global_first_rank,
    prepare_communication_buffer_for_model)
from vllm.forward_context import (BatchDescriptor, DPMetadata,
                                  set_forward_context)
from vllm.logger import init_logger
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
from vllm.model_executor.models.interfaces import (is_mixture_of_experts,
                                                   supports_eagle3,
                                                   supports_transcription)
from vllm.model_executor.models.interfaces_base import (
    VllmModelForPooling, is_pooling_model, is_text_generation_model)
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (BatchedTensorInputs, MultiModalKwargsItem,
                                    PlaceholderRange)
from vllm.multimodal.utils import group_mm_kwargs_by_modality
from vllm.pooling_params import PoolingParams
from vllm.sampling_params import SamplingType
from vllm.sequence import IntermediateTensors, PoolerOutput
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                        GiB_bytes, LazyLoader, cdiv, check_use_alibi,
                        get_dtype_size, is_pin_memory_available, round_up,
                        supports_dynamo)
from vllm.v1.attention.backends.utils import (
    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
    create_fast_prefill_custom_backend,
    reorder_batch_to_split_decodes_and_prefills)
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
from vllm.v1.kv_cache_interface import (AttentionSpec,
                                        ChunkedLocalAttentionSpec,
                                        EncoderOnlyAttentionSpec,
                                        FullAttentionSpec, KVCacheConfig,
                                        KVCacheGroupSpec, KVCacheSpec,
                                        MambaSpec, SlidingWindowSpec)
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, AsyncModelRunnerOutput,
                             DraftTokenIds, LogprobsLists, LogprobsTensors,
                             ModelRunnerOutput, SamplerOutput)
from vllm.v1.pool.metadata import PoolingMetadata
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
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.eagle import EagleProposer
from vllm.v1.spec_decode.medusa import MedusaProposer
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
from vllm.v1.worker.kv_connector_model_runner_mixin import (
    KVConnectorModelRunnerMixin, KVConnectorOutput)
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin

from .utils import (AttentionGroup, MultiModalBudget,
                    add_kv_sharing_layers_to_kv_cache_groups, bind_kv_cache,
                    gather_mm_placeholders, sanity_check_mm_encoder_outputs,
                    scatter_mm_placeholders)

if TYPE_CHECKING:
    import xgrammar as xgr

    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
    from vllm.v1.core.sched.output import SchedulerOutput
else:
    xgr = LazyLoader("xgr", globals(), "xgrammar")

logger = init_logger(__name__)


# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):

    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        self._model_runner_output = model_runner_output
        self._invalid_req_indices = invalid_req_indices

        # Event on the copy stream so we can synchronize the non-blocking copy.
        self._async_copy_ready_event = torch.cuda.Event()

        # Keep a reference to the device tensor to avoid it being
        # deallocated until we finish copying it to the host.
        self._sampled_token_ids = sampled_token_ids

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
            self._sampled_token_ids_cpu = self._sampled_token_ids.to(
                'cpu', non_blocking=True)
            self._async_copy_ready_event.record()

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
        
        This function blocks until the copy is finished.
        """
        self._async_copy_ready_event.synchronize()

        # Release the device tensor once the copy has completed
        del self._sampled_token_ids

        valid_sampled_token_ids = self._sampled_token_ids_cpu.tolist()
        for i in self._invalid_req_indices:
            valid_sampled_token_ids[i].clear()

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
        return output


class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):

    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

        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
        set_cpu_offload_max_bytes(
            int(self.cache_config.cpu_offload_gb * 1024**3))

        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
        self.device = device
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
        if cache_config.cache_dtype == "auto":
            self.kv_cache_dtype = self.dtype
        else:
            self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                cache_config.cache_dtype]

        self.is_pooling_model = (model_config.runner_type == 'pooling')
        self.is_multimodal_raw_input_only_model = (
            model_config.is_multimodal_raw_input_only_model)

        self.max_model_len = model_config.max_model_len
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
        self.max_num_reqs = scheduler_config.max_num_seqs

        # Model-related.
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
        self.hidden_size = model_config.get_hidden_size()
        self.attention_chunk_size = model_config.attention_chunk_size
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)

        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn

        # Multi-modal data support
        self.mm_registry = MULTIMODAL_REGISTRY
        self.uses_mrope = model_config.uses_mrope
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
            model_config)

        # Sampler
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)

        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

        Will be lazily initialized when the model is loaded.
        """

        # Lazy initializations
        # self.model: nn.Module  # Set after load_model
        # Initialize in initialize_kv_cache
        self.kv_caches: list[torch.Tensor] = []
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
        # self.kv_cache_config: KVCacheConfig

        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}

        self.use_aux_hidden_state_outputs = False
        # Set up speculative decoding.
        # NOTE(Jiayi): currently we put the entire draft model on
        # the last PP rank. This is not ideal if there are many
        # layers in the draft model.
        if self.speculative_config and get_pp_group().is_last_rank:
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
            elif self.speculative_config.use_eagle():
                self.drafter = EagleProposer(self.vllm_config, self.device,
                                             self)  # type: ignore
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
                    vllm_config=self.vllm_config,
                    device=self.device)  # type: ignore
            else:
                raise ValueError("Unknown speculative decoding method: "
                                 f"{self.speculative_config.method}")
            self.rejection_sampler = RejectionSampler()

        # Request states.
        self.requests: dict[str, CachedRequestState] = {}

        # Input Batch
        # NOTE(Chen): Ideally, we should initialize the input batch inside
        # `initialize_kv_cache` based on the kv cache config. However, as in
        # https://github.com/vllm-project/vllm/pull/18298, due to some unknown
        # reasons, we have to initialize the input batch before `load_model`,
        # quantization + weight offloading will fail otherwise. As a temporary
        # solution, we initialize the input batch here, and re-initialize it
        # in `initialize_kv_cache` if the block_sizes here is different from
        # the block_sizes in the kv cache config.
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
            max_model_len=self.max_model_len,
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
            vocab_size=self.model_config.get_vocab_size(),
            block_sizes=[self.cache_config.block_size],
            is_spec_decode=bool(self.vllm_config.speculative_config),
            logitsprocs=build_logitsprocs(
                self.vllm_config, self.device, self.pin_memory,
                self.is_pooling_model,
                self.vllm_config.model_config.logits_processors),
            is_pooling_model=self.is_pooling_model,
        )

        self.use_async_scheduling = self.scheduler_config.async_scheduling
        self.async_output_copy_stream = torch.cuda.Stream() if \
            self.use_async_scheduling else None

        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
        if self.compilation_config.cudagraph_capture_sizes and \
                self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
            self.cudagraph_batch_sizes = list(
                reversed(self.compilation_config.cudagraph_capture_sizes))

        # Cache the device properties.
        self._init_device_properties()

        # Persistent buffers for CUDA graphs.
        self.input_ids = self._make_buffer(self.max_num_tokens,
                                           dtype=torch.int32)
        self.positions = self._make_buffer(self.max_num_tokens,
                                           dtype=torch.int64)
        self.query_start_loc = self._make_buffer(self.max_num_reqs + 1,
                                                 dtype=torch.int32)
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
        # Because inputs_embeds may be bfloat16 and we don't need a numpy
        # version of this tensor, avoid a RuntimeError by not creating a
        # numpy buffer.
        self.inputs_embeds = self._make_buffer(self.max_num_tokens,
                                               self.hidden_size,
                                               dtype=self.dtype,
                                               numpy=False)

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
        if self.uses_mrope:
            # NOTE: `mrope_positions` is implemented with one additional dummy
            # position on purpose to make it non-contiguous so that it can work
            # with torch compile.
            # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923

            # NOTE: When M-RoPE is enabled, position ids are 3D regardless of
            # the modality of inputs. For text-only inputs, each dimension has
            # identical position IDs, making M-RoPE functionally equivalent to
            # 1D-RoPE.
            # See page 5 of https://arxiv.org/abs/2409.12191
            self.mrope_positions = self._make_buffer(
                (3, self.max_num_tokens + 1), dtype=torch.int64)

        # CUDA event to synchronize use of reused CPU tensors between steps
        # when async scheduling is enabled.
        self.prepare_inputs_event: Optional[torch.cuda.Event] = None
        if self.use_async_scheduling:
            self.prepare_inputs_event = torch.cuda.Event()
            # Start in a completed state.
            self.prepare_inputs_event.record(torch.cuda.default_stream())

        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None

        # OPTIMIZATION: Cache the tensors rather than creating them every step.
        # Keep in int64 to avoid overflow with long context
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
                                       self.max_model_len,
                                       self.max_num_tokens),
                                   dtype=np.int64)

        # Layer pairings for cross-layer KV sharing.
        # If an Attention layer `layer_name` is in the keys of this dict, it
        # means this layer will perform attention using the keys and values
        # from the KV cache of `shared_kv_cache_layers[layer_name]`.
        self.shared_kv_cache_layers: dict[str, str] = {}
        self.kv_sharing_fast_prefill_eligible_layers: set[str] = set()

        self.kv_sharing_fast_prefill_logits_indices = None
        if self.cache_config.kv_sharing_fast_prefill:
            self.kv_sharing_fast_prefill_logits_indices = torch.zeros(
                self.max_num_tokens, dtype=torch.int32, device=self.device)

        self.uniform_decode_query_len = 1 if not self.speculative_config else \
            1 + self.speculative_config.num_speculative_tokens

        # Cudagraph dispatcher for runtime cudagraph dispatching.
        self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config)

        self.mm_budget = MultiModalBudget(
            self.model_config,
            self.scheduler_config,
            self.mm_registry,
        ) if self.supports_mm_inputs else None

        self.reorder_batch_threshold: Optional[int] = None

        # Attention layers that are only in the KVCacheConfig of the runner
        # (e.g., KV sharing, encoder-only attention), but not in the
        # KVCacheConfig of the scheduler.
        self.runner_only_attn_layers: set[str] = set()

        # Cached outputs.
        self._draft_token_ids: Optional[Union[list[list[int]],
                                              torch.Tensor]] = None
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
            (self.max_model_len, 1),
            dtype=torch.int64,
            device="cpu",
            pin_memory=self.pin_memory)

    def _make_buffer(self,
                     *size: Union[int, torch.SymInt],
                     dtype: torch.dtype,
                     numpy: bool = True) -> CpuGpuBuffer:
        # Bfloat16 torch tensors cannot be directly cast to a numpy array, so
        # if a bfloat16 buffer is needed without a corresponding numpy array,
        # don't bother instantiating the numpy array.
        return CpuGpuBuffer(*size,
                            dtype=dtype,
                            device=self.device,
                            pin_memory=self.pin_memory,
                            with_numpy=numpy)

    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

        if not self.is_pooling_model:
            return model_kwargs

        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
            if param.extra_kwargs is not None and \
            (token_types := param.extra_kwargs.get(
                "compressed_token_type_ids")) is not None:
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

        seq_lens = self.seq_lens.gpu[:num_reqs]
        token_type_ids = []

        for i in range(num_reqs):
            pos = token_type_id_requests.get(i, seq_lens[i])
            ids = (torch.arange(seq_lens[i]) >= pos).int()
            token_type_ids.append(ids)

        model_kwargs["token_type_ids"] = torch.concat(token_type_ids).to(
            device=self.device)
        return model_kwargs

    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
        """
        Update the order of requests in the batch based on the attention
        backend's needs. For example, some attention backends (namely MLA) may
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
        # Attention free models have zero kv_cache_goups, however models
        # like Mamba are also attention free but use the kv_cache for
        # keeping its internal state. This is why we check the number
        # of kv_cache groups instead of solely checking
        # for self.model_config.is_attention_free.
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return

        if self.reorder_batch_threshold is not None:
            # NOTE(lucas): currently no backend supports the custom masking
            #  required for DCP with q_len > 1, so we assert here. Remove this
            #  assert once the custom mask is support is added to FA3.
            if self.dcp_world_size > 1:
                assert self.reorder_batch_threshold == 1, \
                    "DCP not support reorder_batch_threshold > 1 now."
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
                decode_threshold=self.reorder_batch_threshold)

    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
        """Initialize attributes from torch.cuda.get_device_properties
        """
        self.device_properties = torch.cuda.get_device_properties(self.device)
        self.num_sms = self.device_properties.multi_processor_count

    # Note: used for model runner override.
    def _sync_device(self) -> None:
        torch.cuda.synchronize()

    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
        """Update the cached states and the persistent batch with the scheduler
        output.

        The updated states are used by the `_prepare_inputs` function to create
        the input GPU tensors for the model.

        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
        """
        # Remove finished requests from the cached states.
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
        # Remove the finished requests from the persistent batch.
        # NOTE(woosuk): There could be an edge case where finished_req_ids and
        # scheduled_req_ids overlap. This happens when a request is aborted and
        # then resubmitted with the same ID. In this case, we treat them as two
        # distinct requests - clearing the cached states for the first request
        # and handling the second as a new request.
        for req_id in scheduler_output.finished_req_ids:
            self.input_batch.remove_request(req_id)

        # Free the cached encoder outputs.
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)

        # Remove the unscheduled requests from the persistent batch.
        # NOTE(woosuk): The unscheduled requests are either preempted requests
        # or running requests that are not scheduled in this step. We remove
        # them from the persistent batch but keep their cached states since
        # they will be scheduled again sometime in the future.
        scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys()
        cached_req_ids = self.input_batch.req_id_to_index.keys()
        unscheduled_req_ids = cached_req_ids - scheduled_req_ids
        # NOTE(woosuk): The persistent batch optimization assumes that
        # consecutive batches contain mostly the same requests. If batches
        # have low request overlap (e.g., alternating between two distinct
        # sets of requests), this optimization becomes very inefficient.
        for req_id in unscheduled_req_ids:
            self.input_batch.remove_request(req_id)

        reqs_to_add: list[CachedRequestState] = []
        # Add new requests to the cached states.
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
            pooling_params = new_req_data.pooling_params

            if sampling_params and \
                sampling_params.sampling_type == SamplingType.RANDOM_SEED:
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

            if self.is_pooling_model:
                assert pooling_params is not None
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"

                model = cast(VllmModelForPooling, self.get_model())
                to_update = model.pooler.get_pooling_updates(task)
                to_update.apply(pooling_params)

            req_state = CachedRequestState(
                req_id=req_id,
                prompt_token_ids=new_req_data.prompt_token_ids,
                mm_kwargs=new_req_data.mm_kwargs,
                mm_positions=new_req_data.mm_positions,
                mm_hashes=new_req_data.mm_hashes,
                sampling_params=sampling_params,
                pooling_params=pooling_params,
                generator=generator,
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
                output_token_ids=[],
                lora_request=new_req_data.lora_request,
            )
            self.requests[req_id] = req_state

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            if self.uses_mrope:
                self._init_mrope_positions(req_state)

            reqs_to_add.append(req_state)

        # Update the states of the running/resumed requests.
        is_last_rank = get_pp_group().is_last_rank
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
            req_state = self.requests[req_id]
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
            resumed_from_preemption = req_data.resumed_from_preemption[i]

            # Update the cached states.
            req_state.num_computed_tokens = num_computed_tokens

            if not is_last_rank:
                # When using PP, the scheduler sends the sampled tokens back,
                # because there's no direct communication between the first-
                # stage worker and the last-stage worker.
                new_token_ids = req_data.new_token_ids[i]
                # Add the sampled token(s) from the previous step (if any).
                # This doesn't include "unverified" tokens like spec tokens.
                num_new_tokens = (num_computed_tokens + len(new_token_ids) -
                                  req_state.num_tokens)
                if num_new_tokens == 1:
                    # Avoid slicing list in most common case.
                    req_state.output_token_ids.append(new_token_ids[-1])
                elif num_new_tokens > 0:
                    req_state.output_token_ids.extend(
                        new_token_ids[-num_new_tokens:])

            # Update the block IDs.
            if not resumed_from_preemption:
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
                    for block_ids, new_ids in zip(req_state.block_ids,
                                                  new_block_ids):
                        block_ids.extend(new_ids)
            else:
                assert new_block_ids is not None
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
                req_state.block_ids = new_block_ids

            req_index = self.input_batch.req_id_to_index.get(req_id)
            if req_index is None:
                # The request is not in the persistent batch.
                # The request was either preempted and resumed later, or was not
                # scheduled in the previous step and needs to be added again.
                reqs_to_add.append(req_state)
                continue

            # Update the persistent batch.
            self.input_batch.num_computed_tokens_cpu[req_index] = (
                num_computed_tokens)
            if new_block_ids is not None:
                self.input_batch.block_table.append_row(
                    new_block_ids, req_index)

            # For the last rank, we don't need to update the token_ids_cpu
            # because the sampled tokens are already cached.
            if not is_last_rank:
                # Add new_token_ids to token_ids_cpu.
                start_token_index = num_computed_tokens
                end_token_index = num_computed_tokens + len(new_token_ids)
                self.input_batch.token_ids_cpu[
                    req_index,
                    start_token_index:end_token_index] = new_token_ids
                self.input_batch.num_tokens_no_spec[
                    req_index] = end_token_index
                self.input_batch.num_tokens[req_index] = end_token_index

            # Add spec_token_ids to token_ids_cpu.
            spec_token_ids = (
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, ()))
            if spec_token_ids:
                num_spec_tokens = len(spec_token_ids)
                start_index = self.input_batch.num_tokens_no_spec[req_index]
                end_token_index = start_index + num_spec_tokens
                self.input_batch.token_ids_cpu[
                    req_index, start_index:end_token_index] = spec_token_ids
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens

        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
        for request in reqs_to_add:
            self.input_batch.add_request(request)

        # Condense the batched states if there are gaps left by removed requests
        self.input_batch.condense()
        # Allow attention backend to reorder the batch, potentially
        self._may_reorder_batch(scheduler_output)
        # Refresh batch metadata with any pending updates.
        self.input_batch.refresh_metadata()

    def _init_mrope_positions(self, req_state: CachedRequestState):
        image_grid_thw = []
        video_grid_thw = []
        second_per_grid_ts = []
        audio_feature_lengths = []
        use_audio_in_video = False
        for mm_item in req_state.mm_kwargs:
            mm_input = mm_item.get_data()
            if (t := mm_input.get("image_grid_thw")) is not None:
                image_grid_thw.append(t.tolist())
            if (t := mm_input.get("video_grid_thw")) is not None:
                video_grid_thw.append(t.tolist())
            if (t := mm_input.get("second_per_grid_ts")) is not None:
                second_per_grid_ts.append(t)
            if (t := mm_input.get("audio_feature_lengths")) is not None:
                audio_feature_lengths.append(t)
            if mm_input.get("use_audio_in_video") is True:
                use_audio_in_video = True

        req_state.mrope_positions, req_state.mrope_position_delta = \
            MRotaryEmbedding.get_input_positions_tensor(
                req_state.prompt_token_ids,
                hf_config=self.model_config.hf_config,
                image_grid_thw=image_grid_thw,
                video_grid_thw=video_grid_thw,
                second_per_grid_ts=second_per_grid_ts,
                audio_feature_lengths=audio_feature_lengths,
                use_audio_in_video=use_audio_in_video,
            )

    def _extract_mm_kwargs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
            return {}

        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
            mm_kwargs.extend(req.mm_kwargs)

        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                mm_kwargs,
                device=self.device,
                pin_memory=self.pin_memory,
        ):
            mm_kwargs_combined.update(mm_kwargs_group)

        return mm_kwargs_combined

    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
        if not self.is_multimodal_raw_input_only_model:
            return {}

        mm_budget = self.mm_budget
        assert mm_budget is not None

        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)

    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
        cumsum_dtype: Optional[np.dtype] = None,
    ) -> tuple[np.ndarray, np.ndarray]:
        """Get the cumulative sum and batched arange of the given array.
        # E.g., [2, 5, 3] -> ([2, 7, 10], [0, 1, 0, 1, 2, 3, 4, 0, 1, 2])
        # Equivalent to but faster than:
        # np.concatenate([np.arange(n) for n in num_tokens])
        """
        # Step 1. [2, 5, 3] -> [2, 7, 10]
        cu_num_tokens = np.cumsum(num_tokens, dtype=cumsum_dtype)
        total_num_tokens = cu_num_tokens[-1]
        # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
        cumsums_offsets = np.repeat(cu_num_tokens - num_tokens, num_tokens)
        # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        arange = self.arange_np[:total_num_tokens] - cumsums_offsets

        return cu_num_tokens, arange

    def _prepare_input_ids(self, total_num_scheduled_tokens: int,
                           cu_num_tokens: np.ndarray) -> None:
        """Prepare the input IDs for the current batch.
        
        Carefully handles the `prev_sampled_token_ids` which can be cached
        from the previous engine iteration, in which case those tokens on the
        GPU need to be copied into the corresponding slots into input_ids."""

        if self.input_batch.prev_sampled_token_ids is None:
            # Normal scheduling case
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
            return

        # Async scheduling case, where some decode requests from the previous
        # iteration won't have entries in input_ids_cpu and need to be copied
        # on the GPU from prev_sampled_token_ids.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        assert prev_req_id_to_index is not None
        flattened_indices = []
        prev_common_req_indices = []
        indices_match = True
        max_flattened_index = -1
        for req_id, cur_index in self.input_batch.req_id_to_index.items():
            if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
                prev_common_req_indices.append(prev_index)
                # We need to compute the flattened input_ids index of the
                # last token in each common request.
                flattened_index = cu_num_tokens[cur_index].item() - 1
                flattened_indices.append(flattened_index)
                indices_match &= (prev_index == flattened_index)
                max_flattened_index = max(max_flattened_index, flattened_index)
        num_commmon_tokens = len(flattened_indices)
        if num_commmon_tokens < total_num_scheduled_tokens:
            # If not all requests are decodes from the last iteration,
            # We need to copy the input_ids_cpu to the GPU first.
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
            # So input_ids_cpu will have all the input ids.
            return
        if indices_match and max_flattened_index == (num_commmon_tokens - 1):
            # Common-case optimization: the batch is unchanged
            # and no reordering happened.
            # The indices are both the same permutation of 0..N-1 so
            # we can copy directly using a single slice.
            self.input_ids.gpu[:num_commmon_tokens].copy_(
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens,
                                                        0],
                non_blocking=True)
            return
        # Upload the index tensors asynchronously
        # so the scatter can be non-blocking.
        input_ids_index_tensor = torch.tensor(flattened_indices,
                                              dtype=torch.int64,
                                              pin_memory=self.pin_memory).to(
                                                  self.device,
                                                  non_blocking=True)
        prev_common_req_indices_tensor = torch.tensor(
            prev_common_req_indices,
            dtype=torch.int64,
            pin_memory=self.pin_memory).to(self.device, non_blocking=True)
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
                prev_common_req_indices_tensor, 0])

    def _prepare_inputs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> tuple[dict[str, Any], torch.Tensor, Optional[SpecDecodeMetadata],
               np.ndarray, Optional[CommonAttentionMetadata], int]:
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
            logits_indices, spec_decode_metadata
        ]
        """
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        assert total_num_scheduled_tokens > 0
        num_reqs = self.input_batch.num_reqs
        assert num_reqs > 0

        # OPTIMIZATION: Start copying the block table first.
        # This way, we can overlap the copy with the following CPU operations.
        self.input_batch.block_table.commit_block_table(num_reqs)

        # Get the number of scheduled tokens for each request.
        req_ids = self.input_batch.req_ids
        tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
        num_scheduled_tokens = np.array(tokens, dtype=np.int32)
        max_num_scheduled_tokens = max(tokens)

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)

        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        cu_num_tokens, arange = self._get_cumsum_and_arange(
            num_scheduled_tokens)

        # Get positions.
        positions_np = self.positions.np[:total_num_scheduled_tokens]
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
        if self.uses_mrope:
            self._calc_mrope_positions(scheduler_output)

        # Get token indices.
        # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        # -> [0, 1, M, M + 1, M + 2, M + 3, M + 4, 2 * M, 2 * M + 1, 2 * M + 2]
        # where M is the max_model_len.
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])

        # NOTE(woosuk): We use torch.index_select instead of np.take here
        # because torch.index_select is much faster than np.take for large
        # tensors.
        torch.index_select(self.input_batch.token_ids_cpu_tensor.flatten(),
                           0,
                           torch.from_numpy(token_indices),
                           out=self.input_ids.cpu[:total_num_scheduled_tokens])

        self.input_batch.block_table.compute_slot_mapping(
            req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(
            total_num_scheduled_tokens)

        # Prepare the attention metadata.
        self.query_start_loc.np[0] = 0
        self.query_start_loc.np[1:num_reqs + 1] = cu_num_tokens
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
        self.query_start_loc.np[num_reqs + 1:].fill(cu_num_tokens[-1])
        self.query_start_loc.copy_to_gpu()
        query_start_loc = self.query_start_loc.gpu[:num_reqs + 1]

        self.seq_lens.np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
        # Fill unused with 0 for full cuda graph mode.
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
        seq_lens = self.seq_lens.gpu[:num_reqs]
        max_seq_len = self.seq_lens.np[:num_reqs].max().item()

        # Copy the tensors to the GPU.
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

        if self.uses_mrope:
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
                non_blocking=True)
        else:
            # Common case (1D positions)
            self.positions.copy_to_gpu(total_num_scheduled_tokens)

        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
        if not use_spec_decode:
            # NOTE(woosuk): Due to chunked prefills, the batch may contain
            # partial requests. While we should not sample any token
            # from these partial requests, we do so for simplicity.
            # We will ignore the sampled tokens from the partial requests.
            # TODO: Support prompt logprobs.
            logits_indices = query_start_loc[1:] - 1
            spec_decode_metadata = None
        else:
            # Get the number of draft tokens for each request.
            # Iterate over the dictionary rather than all requests since not all
            # requests have draft tokens.
            num_draft_tokens = np.zeros(num_reqs, dtype=np.int32)
            for req_id, draft_token_ids in (
                    scheduler_output.scheduled_spec_decode_tokens.items()):
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)

            spec_decode_metadata = self._calc_spec_decode_metadata(
                num_draft_tokens, cu_num_tokens)
            logits_indices = spec_decode_metadata.logits_indices

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
                logits_indices)

        attn_metadata: dict[str, Any] = {}

        # Used in the below loop.
        query_start_loc_cpu = self.query_start_loc.cpu[:num_reqs + 1]
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
        num_computed_tokens_cpu = (
            self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs])
        spec_decode_common_attn_metadata = None

        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
        for kv_cache_group_id, kv_cache_group_spec in enumerate(
                self.kv_cache_config.kv_cache_groups):

            if isinstance(kv_cache_group_spec.kv_cache_spec,
                          EncoderOnlyAttentionSpec):
                # Encoder-only layers do not have KV cache, so we need to
                # create a dummy block table and slot mapping for them.
                blk_table_tensor = torch.zeros(
                    (num_reqs, 1),
                    dtype=torch.int32,
                    device=self.device,
                )
                slot_mapping = torch.zeros(
                    (total_num_scheduled_tokens, ),
                    dtype=torch.int64,
                    device=self.device,
                )
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
                blk_table_tensor = blk_table.get_device_tensor()[:num_reqs]
                slot_mapping = blk_table.slot_mapping[:
                                                      total_num_scheduled_tokens]

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
                blk_table.slot_mapping[total_num_scheduled_tokens:].fill_(-1)
                num_common_prefix_blocks = (
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id])

            common_attn_metadata = CommonAttentionMetadata(
                query_start_loc=query_start_loc,
                query_start_loc_cpu=query_start_loc_cpu,
                seq_lens=seq_lens,
                seq_lens_cpu=seq_lens_cpu,
                num_computed_tokens_cpu=num_computed_tokens_cpu,
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
                max_seq_len=max_seq_len,
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
                causal=True,
            )

            if self.speculative_config and \
                spec_decode_common_attn_metadata is None:
                spec_decode_common_attn_metadata = common_attn_metadata

            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
                builder = attn_group.metadata_builder
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
                        num_common_prefix_blocks,
                        kv_cache_group_spec.kv_cache_spec,
                        builder,
                    )

                attn_metadata_i = builder.build(
                    common_prefix_len=common_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                )

                for layer_name in attn_group.layer_names:
                    attn_metadata[layer_name] = attn_metadata_i

        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

        return (attn_metadata, logits_indices, spec_decode_metadata,
                num_scheduled_tokens, spec_decode_common_attn_metadata,
                max_num_scheduled_tokens)

    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
    ) -> int:
        """Compute the length of the common prefix for cascade attention.

        NOTE(woosuk): The common prefix length returned by this function
        represents the length used specifically for cascade attention, not the
        actual number of tokens shared between requests. When cascade attention
        is disabled (use_cascade=False), this function returns 0 even if
        requests share common tokens. Additionally, the common prefix length is
        truncated to a multiple of the block size and may be further truncated
        due to implementation details explained below.

        Args:
            num_scheduled_tokens: Number of tokens scheduled per request.
            num_common_prefix_blocks: Number of shared KV cache blocks.

        Returns:
            int: Length of common prefix in tokens.
        """
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
        if common_prefix_len == 0:
            # Common case.
            return 0

        # NOTE(woosuk): Cascade attention uses two attention kernels: one
        # for the common prefix and the other for the rest. For the first
        # kernel, we concatenate all the query tokens (possibly from
        # different requests) and treat them as if they are from the same
        # request. Then, we use bi-directional attention to process the
        # common prefix in the KV cache. Importantly, this means that the
        # first kernel does not do any masking.

        # Consider the following example:
        # Request 1's input query: [D, E, X]
        # Request 1's kv cache: [A, B, C, D, E, X]
        # Request 1's num_computed_tokens: 3 (i.e., [A, B, C])
        # Request 2's input query: [E, Y]
        # Request 2's kv cache: [A, B, C, D, E, Y]
        # Request 2's num_computed_tokens: 4 (i.e., [A, B, C, D])

        # If we use [A, B, C, D, E] as the common prefix, then the
        # first kernel will compute the bi-directional attention between
        # input query [D, E, X, E, Y] and common prefix [A, B, C, D, E].
        # However, this is wrong because D in Request 1 should not attend to
        # E in the common prefix (i.e., we need masking).
        # To avoid this, [A, B, C, D] should be the common prefix.
        # That is, the common prefix should be capped by the minimum
        # num_computed_tokens among the requests, and plus one to include
        # the first token of the query.

        # In practice, we use [A, B, C] as the common prefix, instead of
        # [A, B, C, D] (i.e., the common prefix is capped by the minimum
        # num_computed_tokens, without plus one).
        # This is because of an implementation detail: We want to always
        # use two kernels for cascade attention. Let's imagine:
        # Request 3's input query: [D]
        # Request 3's kv cache: [A, B, C, D]
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
        # If we use [A, B, C, D] as the common prefix for Request 1-3,
        # then Request 3 will be processed only by the first kernel,
        # and the second kernel will get an empty input. While this is not
        # a fundamental problem, our current implementation does not support
        # this case.
        num_reqs = len(num_scheduled_tokens)
        common_prefix_len = min(
            common_prefix_len,
            self.input_batch.num_computed_tokens_cpu[:num_reqs].min())
        # common_prefix_len should be a multiple of the block size.
        common_prefix_len = (common_prefix_len // kv_cache_spec.block_size *
                             kv_cache_spec.block_size)
        use_sliding_window = (isinstance(kv_cache_spec, SlidingWindowSpec) or
                              (isinstance(kv_cache_spec, FullAttentionSpec)
                               and kv_cache_spec.sliding_window is not None))
        use_local_attention = (
            isinstance(kv_cache_spec, ChunkedLocalAttentionSpec)
            or (isinstance(kv_cache_spec, FullAttentionSpec)
                and kv_cache_spec.attention_chunk_size is not None))
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
            num_kv_heads=kv_cache_spec.num_kv_heads,
            use_alibi=self.use_alibi,
            use_sliding_window=use_sliding_window,
            use_local_attention=use_local_attention,
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
        for index, req_id in enumerate(self.input_batch.req_ids):
            req = self.requests[req_id]
            assert req.mrope_positions is not None

            num_computed_tokens = \
                self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = \
                scheduler_output.num_scheduled_tokens[req_id]
            num_prompt_tokens = len(req.prompt_token_ids)

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
                prompt_part_len = max(0,
                                      num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(
                    0, num_scheduled_tokens - prompt_part_len)
            else:
                prompt_part_len = num_scheduled_tokens
                completion_part_len = 0

            assert num_scheduled_tokens == prompt_part_len + completion_part_len

            if prompt_part_len > 0:
                # prompt's mrope_positions are pre-computed
                dst_start = mrope_pos_ptr
                dst_end = mrope_pos_ptr + prompt_part_len
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

                self.mrope_positions.cpu[:, dst_start:dst_end] = (
                    req.mrope_positions[:, src_start:src_end])
                mrope_pos_ptr += prompt_part_len

            if completion_part_len > 0:
                # compute completion's mrope_positions on-the-fly
                dst_start = mrope_pos_ptr
                dst_end = mrope_pos_ptr + completion_part_len

                MRotaryEmbedding.get_next_input_positions_tensor(
                    out=self.mrope_positions.np,
                    out_offset=dst_start,
                    mrope_position_delta=req.mrope_position_delta,
                    context_len=num_computed_tokens + prompt_part_len,
                    num_new_tokens=completion_part_len,
                )

                mrope_pos_ptr += completion_part_len

    def _calc_spec_decode_metadata(
        self,
        num_draft_tokens: np.ndarray,
        cu_num_scheduled_tokens: np.ndarray,
    ) -> SpecDecodeMetadata:
        # Inputs:
        # cu_num_scheduled_tokens:  [  4, 104, 107, 207, 209]
        # num_draft_tokens:         [  3,   0,   2,   0,   1]
        # Outputs:
        # cu_num_draft_tokens:      [  3,   3,   5,   5,   6]
        # logits_indices:           [  0,   1,   2,   3, 103, 104, 105, 106,
        #                            206, 207, 208]
        # target_logits_indices:    [  0,   1,   2,   5,   6,   9]
        # bonus_logits_indices:     [  3,   4,   7,   8,  10]

        # Compute the logits indices.
        # [4, 1, 3, 1, 2]
        num_sampled_tokens = num_draft_tokens + 1

        # Step 1. cu_num_sampled_tokens: [4, 5, 8, 9, 11]
        # arange: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        cu_num_sampled_tokens, arange = self._get_cumsum_and_arange(
            num_sampled_tokens, cumsum_dtype=np.int32)
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
        logits_indices = np.repeat(
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
        logits_indices += arange

        # Compute the bonus logits indices.
        bonus_logits_indices = cu_num_sampled_tokens - 1

        # Compute the draft logits indices.
        # cu_num_draft_tokens: [3, 3, 5, 5, 6]
        # arange: [0, 1, 2, 0, 1, 0]
        cu_num_draft_tokens, arange = self._get_cumsum_and_arange(
            num_draft_tokens, cumsum_dtype=np.int32)
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens)
        # [0, 1, 2, 5, 6, 9]
        target_logits_indices += arange

        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
            self.device, non_blocking=True)
        logits_indices = torch.from_numpy(logits_indices).to(self.device,
                                                             non_blocking=True)
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
            self.device, non_blocking=True)
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
            self.device, non_blocking=True)

        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
        draft_token_ids = self.input_ids.gpu[logits_indices]
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

        metadata = SpecDecodeMetadata(
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )
        return metadata

    def _prepare_kv_sharing_fast_prefill(
        self,
        logits_indices: torch.Tensor,
    ) -> torch.Tensor:
        assert self.kv_sharing_fast_prefill_logits_indices is not None
        num_logits = logits_indices.shape[0]
        assert num_logits > 0
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(
            logits_indices)
        # There might have leftover indices in logits_indices[num_logits:]
        # from previous iterations, whose values may be greater than the
        # batch size in the current iteration. To ensure indices are always
        # valid, we fill the padded indices with the last index.
        self.kv_sharing_fast_prefill_logits_indices[num_logits:].fill_(
            logits_indices[-1].item())
        if (self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
                and num_logits <= self.cudagraph_batch_sizes[-1]):
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
        else:
            num_logits_padded = num_logits
        logits_indices_padded = (
            self.kv_sharing_fast_prefill_logits_indices[:num_logits_padded])
        return logits_indices_padded

    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return
        # Batch the multi-modal inputs.
        mm_kwargs = list[MultiModalKwargsItem]()
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]

            for mm_input_id in encoder_input_ids:
                mm_hash = req_state.mm_hashes[mm_input_id]
                mm_kwargs.append(req_state.mm_kwargs[mm_input_id])
                mm_hashes_pos.append(
                    (mm_hash, req_state.mm_positions[mm_input_id]))

        # Batch mm inputs as much as we can: if a request in the batch has
        # multiple modalities or a different modality than the previous one,
        # we process it separately to preserve item order.
        # FIXME(ywang96): This is a hacky way to deal with multiple modalities
        # in the same batch while still being able to benefit from batching
        # multimodal inputs. The proper solution should be reordering the
        # encoder outputs.
        encoder_outputs = []
        for _, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
                mm_kwargs,
                device=self.device,
                pin_memory=self.pin_memory,
        ):
            # Run the encoder.
            # `curr_group_outputs` is either of the following:
            # 1. A tensor of shape (num_items, feature_size, hidden_size)
            # in case feature_size is fixed across all multimodal items.
            # 2. A list or tuple (length: num_items) of tensors, each of shape
            # (feature_size, hidden_size) in case the feature size is dynamic
            # depending on the input multimodal items.
            curr_group_outputs = self.model.get_multimodal_embeddings(
                **mm_kwargs_group)

            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
                expected_num_items=num_items,
            )

            for output in curr_group_outputs:
                encoder_outputs.append(output)

        # Cache the encoder outputs by mm_hash
        for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
            self.encoder_cache[mm_hash] = scatter_mm_placeholders(
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
        self,
        scheduler_output: "SchedulerOutput",
        shift_computed_tokens: int = 0,
    ) -> list[torch.Tensor]:
        mm_embeds: list[torch.Tensor] = []
        for req_id in self.input_batch.req_ids:
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                req_id]
            req_state = self.requests[req_id]
            num_computed_tokens = \
                req_state.num_computed_tokens + shift_computed_tokens
            mm_positions = req_state.mm_positions
            mm_hashes = req_state.mm_hashes
            for i, pos_info in enumerate(mm_positions):
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length

                # The encoder output is needed if the two ranges overlap:
                # [num_computed_tokens,
                #  num_computed_tokens + num_scheduled_tokens) and
                # [start_pos, start_pos + num_encoder_tokens)
                if start_pos >= num_computed_tokens + num_scheduled_tokens:
                    # The encoder output is not needed in this step.
                    break
                if start_pos + num_encoder_tokens <= num_computed_tokens:
                    # The encoder output is already processed and stored
                    # in the decoder's KV cache.
                    continue

                start_idx = max(num_computed_tokens - start_pos, 0)
                end_idx = min(
                    num_computed_tokens - start_pos + num_scheduled_tokens,
                    num_encoder_tokens,
                )
                assert start_idx < end_idx

                mm_hash = mm_hashes[i]
                encoder_output = self.encoder_cache.get(mm_hash, None)
                assert encoder_output is not None,\
                    f"Encoder cache miss for {mm_hash}."

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]

                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
                mm_embeds.append(mm_embeds_item)
        return mm_embeds

    def get_model(self) -> nn.Module:
        # get raw model out of the cudagraph wrapper.
        if isinstance(self.model, CUDAGraphWrapper):
            return self.model.unwrap()
        return self.model

    def get_supported_generation_tasks(self) -> list[GenerationTask]:
        model = self.get_model()
        supported_tasks = list[GenerationTask]()

        if is_text_generation_model(model):
            supported_tasks.append("generate")

        if supports_transcription(model):
            if model.supports_transcription_only:
                return ["transcription"]

            supported_tasks.append("transcription")

        return supported_tasks

    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

        supported_tasks = list(model.pooler.get_supported_tasks())

        if (self.scheduler_config.chunked_prefill_enabled
                and "encode" in supported_tasks):
            supported_tasks.remove("encode")

            logger.debug_once("Chunked prefill is not supported with "
                              "encode task which using ALL pooling. "
                              "Please turn off chunked prefill by "
                              "`--no-enable-chunked-prefill` before using it.")

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
                logger.debug_once(
                    "Score API is only enabled for num_labels == 1.")

        return supported_tasks

    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        tasks = list[SupportedTask]()

        if self.model_config.runner_type == "generate":
            tasks.extend(self.get_supported_generation_tasks())
        if self.model_config.runner_type == "pooling":
            tasks.extend(self.get_supported_pooling_tasks())

        return tuple(tasks)

    def apply_grammar_bitmask(
        self,
        scheduler_output: "SchedulerOutput",
        logits: torch.Tensor,
    ):
        grammar_bitmask = scheduler_output.grammar_bitmask
        if grammar_bitmask is None:
            return

        # We receive the structured output bitmask from the scheduler,
        # compacted to contain bitmasks only for structured output requests.
        # The order of the requests in the bitmask is not guaranteed to be the
        # same as the order of the requests in the gpu runner's batch. We need
        # to sort the bitmask to match the order of the requests used here.

        # Get the batch indices of the structured output requests.
        # Keep track of the number of speculative tokens scheduled for every
        # request in the batch, as the logit indices are offset by this amount.
        struct_out_req_batch_indices: dict[str, int] = {}
        cumulative_offset = 0
        seq = sorted(self.input_batch.req_id_to_index.items(),
                     key=lambda x: x[1])
        for req_id, batch_index in seq:
            logit_index = batch_index + cumulative_offset
            cumulative_offset += len(
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, []))
            if req_id in scheduler_output.structured_output_request_ids:
                struct_out_req_batch_indices[req_id] = logit_index

        out_indices = []

        # Reorder the bitmask to match the order of the requests in the batch.
        sorted_bitmask = np.full(shape=(logits.shape[0],
                                        grammar_bitmask.shape[1]),
                                 fill_value=-1,
                                 dtype=grammar_bitmask.dtype)
        cumulative_index = 0
        seq = sorted(scheduler_output.structured_output_request_ids.items(),
                     key=lambda x: x[1])
        for req_id, _ in seq:
            logit_index = struct_out_req_batch_indices[req_id]
            num_spec_tokens = len(
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, []))
            for i in range(1 + num_spec_tokens):
                sorted_bitmask[logit_index + i] = \
                    grammar_bitmask[cumulative_index + i]
                out_indices.append(logit_index + i)
            cumulative_index += 1 + num_spec_tokens
        grammar_bitmask = sorted_bitmask

        # If the length of out indices and the logits have the same shape
        # we don't need to pass indices to the kernel,
        # since the bitmask is already aligned with the logits.
        skip_out_indices = len(out_indices) == logits.shape[0]

        # Serialization of np.ndarray is much more efficient than a tensor,
        # so we receive it in that format.
        grammar_bitmask = torch.from_numpy(grammar_bitmask).contiguous()

        xgr.apply_token_bitmask_inplace(
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
            indices=out_indices if not skip_out_indices else None,
        )

    def sync_and_slice_intermediate_tensors(
            self, num_tokens: int, intermediate_tensors: IntermediateTensors,
            sync_self: bool) -> IntermediateTensors:

        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
        enabled_sp = self.compilation_config.pass_config. \
            enable_sequence_parallelism
        if enabled_sp:
            # When sequence parallelism is enabled, we always pad num_tokens
            # to be a multiple of tensor_parallel_size (tp) earlier
            assert num_tokens % tp == 0
        is_residual_scattered = tp > 1 and enabled_sp \
            and num_tokens % tp == 0

        # When sequence parallelism is enabled, the "residual" tensor is sharded
        # across tensor parallel ranks, so each rank only needs its own slice.
        if sync_self:
            assert intermediate_tensors is not None
            for k, v in intermediate_tensors.items():
                is_scattered = k == "residual" and is_residual_scattered
                copy_len = num_tokens // tp if is_scattered else \
                    num_tokens
                self.intermediate_tensors[k][:copy_len].copy_(
                    v[:copy_len], non_blocking=True)

        return IntermediateTensors({
            k:
            v[:num_tokens // tp]
            if k == "residual" and is_residual_scattered else v[:num_tokens]
            for k, v in self.intermediate_tensors.items()
        })

    def eplb_step(self,
                  is_dummy: bool = False,
                  is_profile: bool = False) -> None:
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
        model = self.get_model()
        assert is_mixture_of_experts(model)
        self.eplb_state.step(
            model,
            is_dummy,
            is_profile,
            log_stats=self.parallel_config.eplb_config.log_balancedness,
        )

    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank

        # For DP: Don't pad when setting enforce_eager.
        # This lets us set enforce_eager on the prefiller in a P/D setup and
        # still use CUDA graphs (enabled by this padding) on the decoder.
        #
        # TODO(tms) : There are many cases where padding is enabled for
        # prefills, causing unnecessary and excessive padding of activations.

        if dp_size == 1 or self.vllm_config.model_config.enforce_eager:
            # Early exit.
            return 0, None

        num_tokens_across_dp = DPMetadata.num_tokens_across_dp(
            num_tokens, dp_size, dp_rank)
        max_tokens_across_dp_cpu = torch.max(num_tokens_across_dp).item()
        num_tokens_after_padding = torch.tensor([max_tokens_across_dp_cpu] *
                                                dp_size,
                                                device="cpu",
                                                dtype=torch.int32)
        return max_tokens_across_dp_cpu - num_tokens, num_tokens_after_padding

    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
        kv_connector_output: Optional[KVConnectorOutput],
    ) -> ModelRunnerOutput:
        assert self.input_batch.num_reqs ==\
            len(self.input_batch.pooling_params), \
        "Either all or none of the requests in" \
        " a batch must be pooling request"

        hidden_states = hidden_states[:num_scheduled_tokens]
        pooling_metadata = self.input_batch.get_pooling_metadata()
        pooling_metadata.build_pooling_cursor(num_scheduled_tokens_np.tolist(),
                                              device=hidden_states.device)
        seq_lens_cpu = self.seq_lens.cpu[:self.input_batch.num_reqs]

        # Pooling models D2H & synchronize occurs in pooler.py:build_output
        raw_pooler_output = self.model.pooler(
            hidden_states=hidden_states, pooling_metadata=pooling_metadata)

        pooler_output: list[Optional[torch.Tensor]] = []
        for raw_output, seq_len, prompt_len in zip(
                raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens):

            output = raw_output.data if seq_len == prompt_len else None
            pooler_output.append(output)

        return ModelRunnerOutput(
            req_ids=self.input_batch.req_ids,
            req_id_to_index=self.input_batch.req_id_to_index,
            sampled_token_ids=[],
            logprobs=None,
            prompt_logprobs_dict={},
            pooler_output=pooler_output,
            kv_connector_output=kv_connector_output,
        )

    def _preprocess(
        self,
        scheduler_output: "SchedulerOutput",
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> tuple[int, int, Optional[torch.Tensor], Optional[torch.Tensor],
               Optional[torch.Tensor], torch.Tensor,
               Optional[IntermediateTensors], dict[str, Any]]:

        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if (self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
                and not envs.VLLM_DISABLE_PAD_FOR_CUDAGRAPH
                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
            # Use CUDA graphs.
            # Add padding to the batch size.
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
                num_scheduled_tokens)
        else:
            # Eager mode.
            # Pad tokens to multiple of tensor_parallel_size when
            # enabled collective fusion for SP
            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
            if self.compilation_config.pass_config. \
                enable_sequence_parallelism and tp_size > 1:
                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
            else:
                num_input_tokens = num_scheduled_tokens

        # Padding for DP
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
        num_input_tokens += num_pad

        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
        if self.supports_mm_inputs and get_pp_group().is_first_rank:
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
            mm_embeds = self._gather_mm_embeddings(scheduler_output)

            # NOTE(woosuk): To unify token ids and soft tokens (vision
            # embeddings), we always use embeddings (rather than token ids)
            # as input to the multimodal model, even when the input is text.
            inputs_embeds_scheduled = self.model.get_input_embeddings(
                input_ids=self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds or None,
            )

            # TODO(woosuk): Avoid the copy. Optimize.
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(
                inputs_embeds_scheduled)

            input_ids = None
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
        else:
            # For text-only models, we use token ids as input.
            # While it is possible to use embeddings as input just like the
            # multimodal models, it is not desirable for performance since
            # then the embedding layer is not included in the CUDA graph.
            input_ids = self.input_ids.gpu[:num_input_tokens]
            inputs_embeds = None
            model_kwargs = self._init_model_kwargs(num_input_tokens)
        if self.uses_mrope:
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
        else:
            positions = self.positions.gpu[:num_input_tokens]

        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                num_input_tokens, intermediate_tensors, True)

        return (
            num_scheduled_tokens,
            num_input_tokens,
            num_tokens_across_dp,
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )

    def _sample(
            self, logits: Optional[torch.Tensor],
            spec_decode_metadata: Optional[SpecDecodeMetadata]
    ) -> SamplerOutput:
        # Sample the next token and get logprobs if needed.
        sampling_metadata = self.input_batch.sampling_metadata
        if spec_decode_metadata is None:
            sampler_output = self.sampler(
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
            # When indexing with a tensor (bonus_logits_indices), PyTorch
            # creates a new tensor with separate storage from the original
            # logits tensor. This means any in-place operations on bonus_logits
            # won't affect the original logits tensor.
            assert logits is not None
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
            sampler_output = self.sampler(
                logits=bonus_logits,
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids

            # Just like `bonus_logits`, `target_logits` is a new tensor with
            # separate storage from the original `logits` tensor. Therefore,
            # it is safe to update `target_logits` in place.
            target_logits = logits[spec_decode_metadata.target_logits_indices]
            output_token_ids = self.rejection_sampler(
                spec_decode_metadata,
                None,  # draft_probs
                target_logits,
                bonus_token_ids,
                sampling_metadata,
            )
            sampler_output.sampled_token_ids = output_token_ids

        return sampler_output

    def _bookkeeping_sync(
        self, scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput, logits: Optional[torch.Tensor],
        hidden_states: torch.Tensor, num_scheduled_tokens: int
    ) -> tuple[
            dict[str, int],
            Optional[LogprobsLists],
            list[list[int]],
            dict[str, Optional[LogprobsTensors]],
            list[str],
            dict[str, int],
            list[int],
    ]:
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
        discard_sampled_tokens_req_indices = []
        for i, req_id in enumerate(self.input_batch.req_ids):
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
            if seq_len < req_state.num_tokens:
                # Ignore the sampled token for partial prefills.
                # Rewind the generator state as if the token was not sampled.
                # This relies on cuda-specific torch-internal impl details
                generator = self.input_batch.generators.get(i)
                if generator is not None:
                    generator.set_offset(generator.get_offset() - 4)
                # Record the index of the request that should not be sampled,
                # so that we could clear the sampled tokens before returning.
                discard_sampled_tokens_req_indices.append(i)

        # Copy some objects so they don't get modified after returning.
        # This is important when using async scheduling.
        req_ids_output_copy = self.input_batch.req_ids.copy()
        req_id_to_index_output_copy = \
            self.input_batch.req_id_to_index.copy()

        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
        logprobs_tensors = sampler_output.logprobs_tensors
        logprobs_lists = logprobs_tensors.tolists() \
            if logprobs_tensors is not None else None

        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
        sampled_token_ids = sampler_output.sampled_token_ids
        invalid_req_indices = []
        if not self.use_async_scheduling:
            # Get the valid generated tokens.
            max_gen_len = sampled_token_ids.shape[-1]
            if max_gen_len == 1:
                # No spec decode tokens.
                valid_sampled_token_ids = self._to_list(sampled_token_ids)
            else:
                # Includes spec decode tokens.
                valid_sampled_token_ids = self.rejection_sampler.parse_output(
                    sampled_token_ids,
                    self.input_batch.vocab_size,
                )
            # Mask out the sampled tokens that should not be sampled.
            for i in discard_sampled_tokens_req_indices:
                valid_sampled_token_ids[i].clear()
        else:
            valid_sampled_token_ids = []
            invalid_req_indices = list(discard_sampled_tokens_req_indices)
            invalid_req_indices_set = set(invalid_req_indices)
            assert sampled_token_ids.shape[-1] == 1

            # Cache the sampled tokens on the GPU and avoid CPU sync.
            # These will be copied into input_ids in the next step
            # when preparing inputs.
            self.input_batch.prev_sampled_token_ids = \
                sampled_token_ids
            self.input_batch.prev_sampled_token_ids_invalid_indices = \
                invalid_req_indices_set
            self.input_batch.prev_req_id_to_index = {
                req_id: i
                for i, req_id in enumerate(self.input_batch.req_ids)
                if i not in invalid_req_indices_set
            }

        # Cache the sampled tokens in the model runner, so that the scheduler
        # doesn't need to send them back.
        # NOTE(woosuk): As an exception, when using PP, the scheduler sends
        # the sampled tokens back, because there's no direct communication
        # between the first-stage worker and the last-stage worker.
        req_ids = self.input_batch.req_ids
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
                sampled_ids = [-1] if \
                    req_idx not in invalid_req_indices_set else None
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
            end_idx = start_idx + len(sampled_ids)
            assert end_idx <= self.max_model_len, (
                "Sampled token IDs exceed the max model length. "
                f"Total number of tokens: {end_idx} > max_model_len: "
                f"{self.max_model_len}")

            self.input_batch.token_ids_cpu[req_idx,
                                           start_idx:end_idx] = sampled_ids
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx

            req_id = req_ids[req_idx]
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

        return (
            num_nans_in_logits,
            logprobs_lists,
            valid_sampled_token_ids,
            prompt_logprobs_dict,
            req_ids_output_copy,
            req_id_to_index_output_copy,
            invalid_req_indices,
        )

    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> Union[ModelRunnerOutput, AsyncModelRunnerOutput, IntermediateTensors]:
        with record_function_or_nullcontext("Preprocess"):
            self._update_states(scheduler_output)
            if not scheduler_output.total_num_scheduled_tokens:
                if not has_kv_transfer_group():
                    # Return empty ModelRunnerOutput if there's no work to do.
                    return EMPTY_MODEL_RUNNER_OUTPUT
                return self.kv_connector_no_forward(scheduler_output,
                                                    self.vllm_config)
            if self.cache_config.kv_sharing_fast_prefill:
                assert not self.input_batch.num_prompt_logprobs, (
                    "--kv-sharing-fast-prefill produces incorrect logprobs for "
                    "prompt tokens, tokens, please disable it when the requests"
                    " need prompt logprobs")

            if self.prepare_inputs_event is not None:
                # Ensure prior step has finished with reused CPU tensors.
                self.prepare_inputs_event.synchronize()
            try:
                # Prepare the decoder inputs.
                (attn_metadata, logits_indices, spec_decode_metadata,
                 num_scheduled_tokens_np, spec_decode_common_attn_metadata,
                 max_query_len) = self._prepare_inputs(scheduler_output)

            finally:
                if self.prepare_inputs_event is not None:
                    self.prepare_inputs_event.record()

            (
                num_scheduled_tokens,
                num_input_tokens,
                num_tokens_across_dp,
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
            ) = self._preprocess(scheduler_output, intermediate_tensors)

            uniform_decode = (max_query_len
                              == self.uniform_decode_query_len) and (
                                  num_scheduled_tokens
                                  == self.input_batch.num_reqs * max_query_len)
            batch_descriptor = BatchDescriptor(num_tokens=num_input_tokens,
                                               uniform_decode=uniform_decode)
            cudagraph_runtime_mode, batch_descriptor = \
                self.cudagraph_dispatcher.dispatch(batch_descriptor)

        # Run the model.
        # Use persistent buffers for CUDA graphs.
        with (set_forward_context(
                attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                batch_descriptor=batch_descriptor,
        ), record_function_or_nullcontext("Forward"),
              self.maybe_get_kv_connector_output(scheduler_output) as
              kv_connector_output):
            model_output = self.model(
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

        with record_function_or_nullcontext("Postprocess"):
            if self.use_aux_hidden_state_outputs:
                hidden_states, aux_hidden_states = model_output
            else:
                hidden_states = model_output
                aux_hidden_states = None

            # Broadcast PP output for external_launcher (torchrun)
            # to make sure we are synced across pp ranks
            # TODO: Support overlapping mirco-batches
            # https://github.com/vllm-project/vllm/issues/18019
            broadcast_pp_output = \
                self.parallel_config.distributed_executor_backend \
                == "external_launcher" and len(get_pp_group().ranks) > 0
            if not get_pp_group().is_last_rank:
                # For mid-pipeline stages, return the hidden states.
                assert isinstance(hidden_states, IntermediateTensors)
                if not broadcast_pp_output:
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
                get_pp_group().send_tensor_dict(
                    hidden_states.tensors, all_gather_group=get_tp_group())
                logits = None
            else:
                if self.is_pooling_model:
                    return self._pool(hidden_states, num_scheduled_tokens,
                                      num_scheduled_tokens_np,
                                      kv_connector_output)

                sample_hidden_states = hidden_states[logits_indices]
                logits = self.model.compute_logits(sample_hidden_states, None)
            if broadcast_pp_output:
                model_output_broadcast_data = {
                    "logits": logits.contiguous(),
                } if logits is not None else {}
                model_output_broadcast_data = get_pp_group(
                ).broadcast_tensor_dict(model_output_broadcast_data,
                                        src=len(get_pp_group().ranks) - 1)
                assert model_output_broadcast_data is not None
                logits = model_output_broadcast_data["logits"]

            # Apply structured output bitmasks if present
            if scheduler_output.grammar_bitmask is not None:
                self.apply_grammar_bitmask(scheduler_output, logits)

        with record_function_or_nullcontext("Sample"):
            sampler_output = self._sample(logits, spec_decode_metadata)

        with record_function_or_nullcontext("Bookkeep"):
            assert isinstance(hidden_states, torch.Tensor)
            (
                num_nans_in_logits,
                logprobs_lists,
                valid_sampled_token_ids,
                prompt_logprobs_dict,
                req_ids_output_copy,
                req_id_to_index_output_copy,
                invalid_req_indices,
            ) = self._bookkeeping_sync(scheduler_output, sampler_output,
                                       logits, hidden_states,
                                       num_scheduled_tokens)

        if self.speculative_config:
            assert spec_decode_common_attn_metadata is not None
            with record_function_or_nullcontext("Draft"):
                self._draft_token_ids = self.propose_draft_token_ids(
                    scheduler_output,
                    valid_sampled_token_ids,
                    self.input_batch.sampling_metadata,
                    hidden_states,
                    sample_hidden_states,
                    aux_hidden_states,
                    spec_decode_metadata,
                    spec_decode_common_attn_metadata,
                )

        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()

        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
            kv_connector_output=kv_connector_output,
            num_nans_in_logits=num_nans_in_logits,
        )

        if not self.use_async_scheduling:
            return output

        return AsyncGPUModelRunnerOutput(
            model_runner_output=output,
            sampled_token_ids=sampler_output.sampled_token_ids,
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

    def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
        if self._draft_token_ids is None:
            return None
        req_ids = self.input_batch.req_ids
        if isinstance(self._draft_token_ids, torch.Tensor):
            draft_token_ids = self._draft_token_ids.tolist()
        else:
            draft_token_ids = self._draft_token_ids
        self._draft_token_ids = None
        return DraftTokenIds(req_ids, draft_token_ids)

    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
        sampled_token_ids: list[list[int]],
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
        aux_hidden_states: Optional[torch.Tensor],
        spec_decode_metadata: Optional[SpecDecodeMetadata],
        common_attn_metadata: CommonAttentionMetadata,
    ) -> Union[list[list[int]], torch.Tensor]:
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
            assert isinstance(self.drafter, NgramProposer)
            draft_token_ids = self.propose_ngram_draft_token_ids(
                sampled_token_ids)
        elif self.speculative_config.method == "medusa":
            assert isinstance(self.drafter, MedusaProposer)
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
                        sampled_token_ids):
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
                indices = torch.tensor(indices, device=self.device)
                hidden_states = sample_hidden_states[indices]

            draft_token_ids = self.drafter.propose(
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
        elif self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            # TODO(woosuk): Refactor the loop.
            req_ids = self.input_batch.req_ids
            next_token_ids: list[int] = []
            for i, token_ids in enumerate(sampled_token_ids):
                if token_ids:
                    # Common case.
                    next_token_id = token_ids[-1]
                else:
                    # Partial prefill (rare case).
                    # Get the next token id from the request state.
                    req_id = req_ids[i]
                    req_state = self.requests[req_id]
                    seq_len = (req_state.num_computed_tokens +
                               scheduler_output.num_scheduled_tokens[req_id])
                    next_token_id = req_state.get_token_id(seq_len)
                next_token_ids.append(next_token_id)
            next_token_ids = torch.tensor(next_token_ids,
                                          dtype=torch.int32,
                                          device=self.device)

            if spec_decode_metadata is None:
                # input_ids can be None for multimodal models.
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
                # TODO(woosuk): Support M-RoPE.
                target_positions = self.positions.gpu[:num_scheduled_tokens]
                if self.use_aux_hidden_state_outputs:
                    target_hidden_states = torch.cat(
                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
                        dim=-1)
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
            else:
                # TODO(woosuk): Refactor this.
                num_draft_tokens = spec_decode_metadata.num_draft_tokens
                num_rejected_tokens = [
                    n + 1 - len(sampled_token_ids[i]) if n > 0 else 0
                    for i, n in enumerate(num_draft_tokens)
                ]
                num_rejected_tokens_cpu = torch.tensor(num_rejected_tokens,
                                                       dtype=torch.int32)
                common_attn_metadata, token_indices =\
                    self.drafter.prepare_inputs(
                    common_attn_metadata, num_rejected_tokens_cpu)

                target_token_ids = self.input_ids.gpu[token_indices]
                # TODO(woosuk): Support M-RoPE.
                target_positions = self.positions.gpu[token_indices]
                if self.use_aux_hidden_state_outputs:
                    target_hidden_states = torch.cat(
                        [h[token_indices] for h in aux_hidden_states], dim=-1)
                else:
                    target_hidden_states = hidden_states[token_indices]
            mm_embeds = None
            if self.supports_mm_inputs:
                mm_embeds = self._gather_mm_embeddings(scheduler_output,
                                                       shift_computed_tokens=1)

            draft_token_ids = self.drafter.propose(
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
                sampling_metadata=sampling_metadata,
                common_attn_metadata=common_attn_metadata,
                mm_embeds=mm_embeds,
            )
        return draft_token_ids

    def propose_ngram_draft_token_ids(
        self,
        sampled_token_ids: list[list[int]],
    ) -> list[list[int]]:
        # TODO(woosuk): Optimize.
        req_ids = self.input_batch.req_ids
        draft_token_ids: list[list[int]] = []
        for i, sampled_ids in enumerate(sampled_token_ids):
            num_sampled_ids = len(sampled_ids)
            if not num_sampled_ids:
                # Skip speculative decoding.
                draft_token_ids.append([])
                continue

            # Skip requests that require sampling parameters that are not
            # supported with speculative decoding.
            req_id = req_ids[i]
            if req_id in self.input_batch.spec_decode_unsupported_reqs:
                draft_token_ids.append([])
                continue

            num_tokens = self.input_batch.num_tokens_no_spec[i]
            if num_tokens >= self.max_model_len:
                # Skip requests that have already reached the max model length.
                draft_token_ids.append([])
                continue

            drafter_output = self.drafter.propose(
                self.input_batch.token_ids_cpu[i, :num_tokens])
            if drafter_output is None or len(drafter_output) == 0:
                draft_token_ids.append([])
            else:
                draft_token_ids.append(drafter_output.tolist())
        return draft_token_ids

    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
            assert config_name in allowed_config_names, \
                f"Config `{config_name}` not supported. " \
                f"Allowed configs: {allowed_config_names}"
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
        logger.info("Starting to load model %s...", self.model_config.model)
        if eep_scale_up:
            from vllm.distributed.parallel_state import get_ep_group
            num_local_physical_experts = torch.empty(1,
                                                     dtype=torch.int32,
                                                     device="cpu")
            torch.distributed.broadcast(num_local_physical_experts,
                                        group=get_ep_group().cpu_group,
                                        group_src=0)
            num_local_physical_experts = int(num_local_physical_experts.item())
            new_ep_size = get_ep_group().world_size
            global_expert_load, old_global_expert_indices = (
                EplbState.recv_state())
            num_logical_experts = global_expert_load.shape[1]
            self.parallel_config.eplb_config.num_redundant_experts = (
                num_local_physical_experts * new_ep_size - num_logical_experts)
            assert old_global_expert_indices.shape[
                1] % num_local_physical_experts == 0
            old_ep_size = old_global_expert_indices.shape[
                1] // num_local_physical_experts
            rank_mapping = {
                old_ep_rank: old_ep_rank
                for old_ep_rank in range(old_ep_size)
            }
        else:
            global_expert_load = None
            old_global_expert_indices = None
            rank_mapping = None

        with DeviceMemoryProfiler() as m:
            time_before_load = time.perf_counter()
            model_loader = get_model_loader(self.load_config)
            logger.info("Loading model from scratch...")
            self.model = model_loader.load_model(
                vllm_config=self.vllm_config, model_config=self.model_config)
            if self.lora_config:
                self.model = self.load_lora_model(self.model,
                                                  self.model_config,
                                                  self.scheduler_config,
                                                  self.lora_config,
                                                  self.device)
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
            if self.use_aux_hidden_state_outputs:
                if supports_eagle3(self.model):
                    self.model.set_aux_hidden_state_layers(
                        self.model.get_eagle3_aux_hidden_state_layers())
                else:
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
                        "aux_hidden_state_outputs was requested")
            time_after_load = time.perf_counter()
        self.model_memory_usage = m.consumed_memory
        logger.info("Model loading took %.4f GiB and %.6f seconds",
                    self.model_memory_usage / GiB_bytes,
                    time_after_load - time_before_load)
        prepare_communication_buffer_for_model(self.model)

        if is_mixture_of_experts(
                self.model) and self.parallel_config.enable_eplb:
            logger.info("EPLB is enabled for model %s.",
                        self.model_config.model)
            self.eplb_state = EplbState.build(
                self.model,
                self.device,
                self.parallel_config,
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
            )

        if (
            self.vllm_config.compilation_config.level == \
                CompilationLevel.DYNAMO_AS_IS and supports_dynamo()
        ):
            backend = self.vllm_config.compilation_config.init_backend(
                self.vllm_config)
            compilation_counter.dynamo_as_is_count += 1
            self.model.compile(
                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
                backend=backend)
            return
        # for other compilation levels, cudagraph behavior is controlled by
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
            self.model = CUDAGraphWrapper(self.model,
                                          self.vllm_config,
                                          runtime_mode=CUDAGraphMode.FULL)

    def reload_weights(self) -> None:
        assert getattr(self, "model", None) is not None, \
            "Cannot reload weights before model is loaded."
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
        model = self.get_model()
        model_loader.load_weights(model, model_config=self.model_config)

    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        model = self.get_model()
        TensorizerLoader.save_model(
            model,
            tensorizer_config=tensorizer_config,
            model_config=self.model_config,
        )

    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: dict[str, int],
    ) -> dict[str, Optional[LogprobsTensors]]:
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}

        # Since prompt logprobs are a rare feature, prioritize simple,
        # maintainable loop over optimal performance.
        completed_prefill_reqs = []
        for req_id, num_prompt_logprobs in num_prompt_logprobs_dict.items():
            num_tokens = num_scheduled_tokens[req_id]

            # Get metadata for this request.
            request = self.requests[req_id]
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
                self.device, non_blocking=True)

            # Set up target LogprobsTensors object.
            logprobs_tensors = in_progress_dict.get(req_id)
            if not logprobs_tensors:
                # Create empty logprobs CPU tensors for the entire prompt.
                # If chunked, we'll copy in slice by slice.
                logprobs_tensors = LogprobsTensors.empty_cpu(
                    num_prompt_tokens - 1, num_prompt_logprobs + 1)
                in_progress_dict[req_id] = logprobs_tensors

            # Determine number of logits to retrieve.
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
            num_remaining_tokens = num_prompt_tokens - start_tok
            if num_tokens <= num_remaining_tokens:
                # This is a chunk, more tokens remain.
                # In the == case, there are no more prompt logprobs to produce
                # but we want to defer returning them to the next step where we
                # have new generated tokens to return.
                num_logits = num_tokens
            else:
                # This is the last chunk of prompt tokens to return.
                num_logits = num_remaining_tokens
                completed_prefill_reqs.append(req_id)
                prompt_logprobs_dict[req_id] = logprobs_tensors

            if num_logits <= 0:
                # This can happen for the final chunk if we prefilled exactly
                # (num_prompt_tokens - 1) tokens for this request in the prior
                # step. There are no more prompt logprobs to produce.
                continue

            # Get the logits corresponding to this req's prompt tokens.
            # If this is a partial request (i.e. chunked prefill),
            # then there is prompt logprob generated for each index.
            req_idx = self.input_batch.req_id_to_index[req_id]
            offset = self.query_start_loc.np[req_idx].item()
            prompt_hidden_states = hidden_states[offset:offset + num_logits]
            logits = self.model.compute_logits(prompt_hidden_states, None)

            # Get the "target" tokens for each index. For prompt at index i,
            # the token at prompt index i+1 is the "sampled" token we want
            # to gather the logprob for.
            tgt_token_ids = prompt_token_ids[start_tok:start_tok + num_logits]

            # Compute prompt logprobs.
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
                token_ids, non_blocking=True)
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs,
                                                         non_blocking=True)
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
                ranks, non_blocking=True)

        # Remove requests that have completed prefill from the batch
        # num_prompt_logprobs_dict.
        for req_id in completed_prefill_reqs:
            del num_prompt_logprobs_dict[req_id]
            del in_progress_dict[req_id]

        # Must synchronize the non-blocking GPU->CPU transfers.
        if prompt_logprobs_dict:
            self._sync_device()

        return prompt_logprobs_dict

    def _get_nans_in_logits(
        self,
        logits: Optional[torch.Tensor],
    ) -> dict[str, int]:
        try:
            if logits is None:
                return {req_id: 0 for req_id in self.input_batch.req_ids}

            num_nans_in_logits = {}
            num_nans_for_index = logits.isnan().sum(dim=-1).cpu().numpy()
            for req_id in self.input_batch.req_ids:
                req_index = self.input_batch.req_id_to_index[req_id]
                num_nans_in_logits[req_id] = (
                    int(num_nans_for_index[req_index])
                    if num_nans_for_index is not None
                    and req_index < logits.shape[0] else 0)
            return num_nans_in_logits
        except IndexError:
            return {}

    @contextmanager
    def maybe_randomize_inputs(self, input_ids: torch.Tensor):
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
         - during DP rank dummy run
        """
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        randomize_inputs = envs.VLLM_RANDOMIZE_DP_DUMMY_INPUTS and dp_size > 1
        if not randomize_inputs:
            yield
        else:
            import functools

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
                    self.input_ids.gpu,
                    low=0,
                    high=self.model_config.get_vocab_size(),
                    dtype=input_ids.dtype)

            logger.debug_once("Randomizing dummy data for DP Rank")
            input_ids.copy_(rand_input_ids()[:input_ids.size(0)],
                            non_blocking=True)
            yield
            input_ids.fill_(0)

    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
        assert self.mm_budget is not None

        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
            seq_len=self.max_num_tokens,
            mm_counts={modality: 1},
            cache=self.mm_budget.cache,
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch

        return next(mm_kwargs_group
                    for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                        dummy_mm_items,
                        device=self.device,
                        pin_memory=self.pin_memory,
                    ))

    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
        cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
        force_attention: bool = False,
        uniform_decode: bool = False,
        skip_eplb: bool = False,
        is_profile: bool = False,
        remove_lora: bool = True,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Run a dummy forward pass to warm up/profile run or capture the
        CUDA graph for the model.

        Args:
            num_tokens: Number of tokens to run the dummy forward pass.
            cudagraph_runtime_mode: used to control the behavior.
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
            force_attention: If True, always create attention metadata. Used to
                warm up attention backend when mode is NONE.
            uniform_decode: If True, the batch is a uniform decode batch.
            skip_eplb: If True, skip EPLB state update.
            is_profile: If True, this is a profile run.
            remove_lora: If False, dummy LoRAs are not destroyed after the run
        """
        assert cudagraph_runtime_mode in {
            CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL
        }

        # Padding for DP
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
        num_tokens += num_pad

        # If cudagraph_mode.decode_mode() == FULL and
        # cudagraph_mode.seperate_routine(). This means that we are using
        # different graphs and/or modes for mixed prefill-decode batches vs.
        # uniform decode batches. A uniform decode batch means that all
        # requests have identical query length, except a potential virtual
        # request (shorter) in the batch account for padding.
        # Uniform decode batch could either be common pure decode, where
        # max_query_len == 1, or speculative decode, where
        # max_query_len == 1 + num_spec_decode_tokens.

        # When setting max_query_len = 1, we switch to and capture the optimized
        # routine of FA2 for pure decode, i.e., Flashdecode + an optimization
        # for GQA/MQA.
        max_query_len = self.uniform_decode_query_len if uniform_decode else \
                                                                num_tokens

        # Set num_scheduled_tokens based on num_tokens and max_num_seqs
        # for dummy run with LoRA so that the num_reqs collectively
        # has num_tokens in total.
        assert num_tokens <= self.scheduler_config.max_num_batched_tokens
        max_num_reqs = self.scheduler_config.max_num_seqs
        if uniform_decode:
            num_reqs = cdiv(num_tokens, max_query_len)
            assert num_reqs <= max_num_reqs, \
                "Do not capture num_reqs > max_num_reqs for uniform batch"
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
        else:
            num_reqs = min(num_tokens, max_num_reqs)
            min_tokens_per_req = num_tokens // num_reqs
            num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
            num_scheduled_tokens_list[-1] += num_tokens % num_reqs

        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
        num_scheduled_tokens = np.array(num_scheduled_tokens_list,
                                        dtype=np.int32)

        attn_metadata: Optional[dict[str, Any]] = None

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
            attn_metadata = {}

            # Make sure max_model_len is used at the graph capture time.
            self.seq_lens.np[:num_reqs] = self.max_model_len
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()

            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
                common_attn_metadata = CommonAttentionMetadata(
                    query_start_loc=self.query_start_loc.gpu[:num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs +
                                                                 1],
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
                    num_computed_tokens_cpu=self.input_batch.
                    num_computed_tokens_cpu_tensor[:num_reqs],
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
                    max_query_len=max_query_len,
                    max_seq_len=self.max_model_len,
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id].get_device_tensor()[:num_reqs],
                    slot_mapping=self.input_batch.
                    block_table[kv_cache_group_id].slot_mapping[:num_tokens],
                    causal=True)

                for attn_group in self.attn_groups[kv_cache_group_id]:
                    attn_metadata_i = attn_group.metadata_builder\
                        .build_for_cudagraph_capture(common_attn_metadata)
                    for layer_name in kv_cache_group_spec.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i

        with self.maybe_dummy_run_with_lora(self.lora_config,
                                            num_scheduled_tokens, remove_lora):
            if self.supports_mm_inputs:
                input_ids = None
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens]
                model_kwargs = {
                    **self._init_model_kwargs(num_tokens),
                    **self._dummy_mm_kwargs(num_reqs),
                }
            else:
                input_ids = self.input_ids.gpu[:num_tokens]
                inputs_embeds = None
                model_kwargs = self._init_model_kwargs(num_tokens)

            if self.uses_mrope:
                positions = self.mrope_positions.gpu[:, :num_tokens]
            else:
                positions = self.positions.gpu[:num_tokens]

            if get_pp_group().is_first_rank:
                intermediate_tensors = None
            else:
                if self.intermediate_tensors is None:
                    self.intermediate_tensors = (
                        self.model.make_empty_intermediate_tensors(
                            batch_size=self.max_num_tokens,
                            dtype=self.model_config.dtype,
                            device=self.device))

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                    num_tokens, None, False)
            if cudagraph_runtime_mode == CUDAGraphMode.NONE:
                batch_descriptor = None
            else:
                # filter out the valid batch descriptor
                _cg_mode, batch_descriptor = \
                    self.cudagraph_dispatcher.dispatch(
                        BatchDescriptor(num_tokens=num_tokens,
                                        uniform_decode=uniform_decode))
                # sanity check
                assert cudagraph_runtime_mode == _cg_mode, (
                    f"Cudagraph runtime mode mismatch at dummy_run. "
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}.")

            with self.maybe_randomize_inputs(input_ids), set_forward_context(
                    attn_metadata,
                    self.vllm_config,
                    num_tokens=num_tokens,
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
                    batch_descriptor=batch_descriptor):
                outputs = self.model(
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
                    **model_kwargs,
                )

            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs

            if self.speculative_config and self.speculative_config.use_eagle():
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

        # This is necessary to avoid blocking DP.
        # For dummy runs, we typically skip EPLB since we don't have any real
        # requests to process.
        # However, in DP settings, there may be cases when some DP ranks do
        # not have any requests to process, so they're executing dummy batches.
        # In such cases, we still have to trigger EPLB to make sure
        # ranks execute the rearrangement in synchronization.
        if not skip_eplb:
            self.eplb_step(is_dummy=True, is_profile=is_profile)

        logit_indices = np.cumsum(num_scheduled_tokens) - 1
        return hidden_states, hidden_states[logit_indices]

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
        hidden_states = torch.rand_like(hidden_states)

        logits = self.model.compute_logits(hidden_states, None)
        num_reqs = logits.size(0)

        dummy_tensors = lambda v: torch.full(
            (num_reqs, ), v, device=self.device)

        dummy_metadata = SamplingMetadata(
            temperature=dummy_tensors(0.5),
            all_greedy=False,
            all_random=False,
            top_p=dummy_tensors(0.9),
            top_k=dummy_tensors(logits.size(1) - 1),
            generators={},
            max_num_logprobs=None,
            no_penalties=True,
            prompt_token_ids=None,
            frequency_penalties=dummy_tensors(0.1),
            presence_penalties=dummy_tensors(0.1),
            repetition_penalties=dummy_tensors(0.1),
            output_token_ids=[[] for _ in range(num_reqs)],
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
            logitsprocs=LogitsProcessors(),
        )
        try:
            sampler_output = self.sampler(logits=logits,
                                          sampling_metadata=dummy_metadata)
        except RuntimeError as e:
            if 'out of memory' in str(e):
                raise RuntimeError(
                    "CUDA out of memory occurred when warming up sampler with "
                    f"{num_reqs} dummy requests. Please try lowering "
                    "`max_num_seqs` or `gpu_memory_utilization` when "
                    "initializing the engine.") from e
            else:
                raise e
        if self.speculative_config:
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
                draft_token_ids, self.device)

            num_tokens = sum(len(ids) for ids in draft_token_ids)
            # draft_probs = torch.randn(
            #     num_tokens, logits.shape[-1], device=self.device,
            #     dtype=logits.dtype)
            draft_probs = None
            target_logits = torch.randn(num_tokens,
                                        logits.shape[-1],
                                        device=self.device,
                                        dtype=logits.dtype)
            # NOTE(woosuk): Here, we should use int32 because the sampler uses
            # int32 for bonus_token_ids. If the dtype mismatches, re-compilation
            # will occur at runtime.
            bonus_token_ids = torch.zeros(num_reqs,
                                          device=self.device,
                                          dtype=torch.int32)
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
                target_logits,
                bonus_token_ids,
                dummy_metadata,
            )
        return sampler_output

    def _dummy_pooler_run_task(
        self,
        hidden_states: torch.Tensor,
        task: PoolingTask,
    ) -> PoolerOutput:
        num_tokens = hidden_states.shape[0]
        max_num_reqs = self.scheduler_config.max_num_seqs
        num_reqs = min(num_tokens, max_num_reqs)
        min_tokens_per_req = num_tokens // num_reqs
        num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
        num_scheduled_tokens_list[-1] += num_tokens % num_reqs
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs

        req_num_tokens = num_tokens // num_reqs

        dummy_prompt_lens = torch.tensor(
            num_scheduled_tokens_list,
            device="cpu",
        )
        dummy_token_ids = torch.zeros((num_reqs, req_num_tokens),
                                      dtype=torch.int32,
                                      device=self.device)

        model = cast(VllmModelForPooling, self.get_model())
        dummy_pooling_params = PoolingParams(task=task)
        to_update = model.pooler.get_pooling_updates(task)
        to_update.apply(dummy_pooling_params)

        dummy_metadata = PoolingMetadata(
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )

        dummy_metadata.build_pooling_cursor(num_scheduled_tokens_list,
                                            device=hidden_states.device)

        try:
            return model.pooler(hidden_states=hidden_states,
                                pooling_metadata=dummy_metadata)
        except RuntimeError as e:
            if 'out of memory' in str(e):
                raise RuntimeError(
                    "CUDA out of memory occurred when warming up pooler "
                    f"({task=}) with {num_reqs} dummy requests. Please try "
                    "lowering `max_num_seqs` or `gpu_memory_utilization` when "
                    "initializing the engine.") from e
            else:
                raise e

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
        output_size = dict[PoolingTask, float]()
        for task in self.get_supported_pooling_tasks():
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
            output_size[task] = output.get_data_nbytes()
            del output  # Allow GC

        max_task = max(output_size.items(), key=lambda x: x[1])[0]
        return self._dummy_pooler_run_task(hidden_states, max_task)

    def profile_run(self) -> None:
        # Profile with multimodal encoder & encoder cache.
        if self.supports_mm_inputs:
            if self.model_config.multimodal_config.skip_mm_profiling:
                logger.info(
                    "Skipping memory profiling for multimodal encoder and "
                    "encoder cache.")
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                # TODO: handle encoder-decoder models once we support them.
                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
                    # NOTE: Currently model is profiled with a single non-text
                    # modality with the max possible input tokens even when
                    # it supports multiple.
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
                    max_mm_items_per_batch = mm_budget \
                        .max_items_per_batch_by_modality[dummy_modality]

                    logger.info(
                        "Encoder cache will be initialized with a budget of "
                        "%s tokens, and profiled with %s %s items of the "
                        "maximum feature size.",
                        encoder_budget,
                        max_mm_items_per_batch,
                        dummy_modality,
                    )

                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )

                    # Run multimodal encoder.
                    dummy_encoder_outputs = \
                        self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs)

                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )

                    # Cache the dummy encoder outputs.
                    self.encoder_cache["tmp"] = dict(
                        enumerate(dummy_encoder_outputs))

        # Add `is_profile` here to pre-allocate communication buffers
        hidden_states, last_hidden_states \
            = self._dummy_run(self.max_num_tokens, is_profile=True)
        if get_pp_group().is_last_rank:
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
        else:
            output = None
        self._sync_device()
        del hidden_states, output
        self.encoder_cache.clear()
        gc.collect()

    def capture_model(self) -> None:
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
            logger.warning(
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
                "ensure `cudagraph_mode` was not manually set to `NONE`")
            return
        else:
            self.initialize_cudagraph_capture()

        compilation_counter.num_gpu_runner_capture_triggers += 1

        start_time = time.perf_counter()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

        @contextmanager
        def freeze_gc():
            # Optimize garbage collection during CUDA graph capture.
            # Clean up, then freeze all remaining objects from being included
            # in future collections.
            gc.collect()
            should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
            if should_freeze:
                gc.freeze()
            try:
                yield
            finally:
                if should_freeze:
                    gc.unfreeze()
                    gc.collect()

        # Trigger CUDA graph capture for specific shapes.
        # Capture the large shapes first so that the smaller shapes
        # can reuse the memory pool allocated for the large shapes.
        set_cudagraph_capturing_enabled(True)
        with freeze_gc(), graph_capture(device=self.device):
            cudagraph_mode = self.compilation_config.cudagraph_mode
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()

                compilation_cases = list(reversed(self.cudagraph_batch_sizes))
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
                    uniform_decode=False)

            # Capture full cudagraph for uniform decode batches if we have
            # dont already have full mixed prefill-decode cudagraphs
            if cudagraph_mode.decode_mode() == CUDAGraphMode.FULL and \
                cudagraph_mode.separate_routine():
                max_num_tokens = self.scheduler_config.max_num_seqs * \
                        self.uniform_decode_query_len
                decode_cudagraph_batch_sizes = [
                    x for x in self.cudagraph_batch_sizes if
                    x <= max_num_tokens and x >= self.uniform_decode_query_len
                ]
                compilation_cases_decode = list(
                    reversed(decode_cudagraph_batch_sizes))
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
                    uniform_decode=True)

        # Disable cudagraph capturing globally, so any unexpected cudagraph
        # capturing will be detected and raise an error after here.
        # Note: We don't put it into graph_capture context manager because
        # we may do lazy capturing in future that still allows capturing
        # after here.
        set_cudagraph_capturing_enabled(False)

        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))

    def _capture_cudagraphs(self, compilation_cases: list[int],
                            cudagraph_runtime_mode: CUDAGraphMode,
                            uniform_decode: bool):
        assert cudagraph_runtime_mode != CUDAGraphMode.NONE and \
            cudagraph_runtime_mode in [CUDAGraphMode.FULL,
                                        CUDAGraphMode.PIECEWISE]

        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
            compilation_cases = tqdm(
                compilation_cases,
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
                    cudagraph_runtime_mode.name))
        # We skip EPLB here since we don't want to record dummy metrics
        for num_tokens in compilation_cases:
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
                # But be careful, warm up with `NONE`is orthogonal to
                # if we want to warm up attention or not. This is
                # different from the case where `FULL` implies capture
                # attention while `PIECEWISE` implies no attention.
                force_attention = (
                    cudagraph_runtime_mode == CUDAGraphMode.FULL)
                self._dummy_run(num_tokens,
                                cudagraph_runtime_mode=CUDAGraphMode.NONE,
                                force_attention=force_attention,
                                uniform_decode=uniform_decode,
                                skip_eplb=True,
                                remove_lora=False)
            self._dummy_run(num_tokens,
                            cudagraph_runtime_mode=cudagraph_runtime_mode,
                            uniform_decode=uniform_decode,
                            skip_eplb=True,
                            remove_lora=False)
        self.maybe_remove_all_loras(self.lora_config)

    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
        assert len(self.attn_groups) == 0, \
            "Attention backends are already initialized"

        def get_attn_backends_for_layers(
                layer_names: list[str]
        ) -> dict[type[AttentionBackend], list[str]]:
            layers = get_layers_from_vllm_config(self.vllm_config,
                                                 AttentionLayerBase,
                                                 layer_names)
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
            # Dedupe based on full class name; this is a bit safer than
            # using the class itself as the key because when we create dynamic
            # attention backend subclasses (e.g. ChunkedLocalAttention) unless
            # they are cached correctly, there will be different objects per
            # layer.
            for layer_name in layer_names:
                attn_backend = layers[layer_name].get_attn_backend()

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
                        attn_backend,
                    )

                key = attn_backend.full_cls_name()
                attn_backends[key] = attn_backend
                attn_backend_layers[key].append(layer_name)
            return {
                attn_backends[k]: v
                for k, v in attn_backend_layers.items()
            }

        def create_attn_groups(
            attn_backends_map: dict[AttentionBackend, list[str]],
            kv_cache_spec: KVCacheSpec,
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
            for attn_backend, layer_names in attn_backends_map.items():
                attn_metadata_builder_i = attn_backend.get_builder_cls()(
                    kv_cache_spec,
                    layer_names,
                    self.vllm_config,
                    self.device,
                )
                attn_group = AttentionGroup(attn_backend,
                                            attn_metadata_builder_i,
                                            layer_names)
                attn_groups.append(attn_group)
            return attn_groups

        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
            kv_cache_spec = kv_cache_group_spec.kv_cache_spec
            attn_backends = get_attn_backends_for_layers(
                kv_cache_group_spec.layer_names)
            self.attn_groups.append(
                create_attn_groups(attn_backends, kv_cache_spec))

        # Calculate reorder batch threshold (if needed)
        self.calculate_reorder_batch_threshold()

    def initialize_cudagraph_capture(self) -> None:
        min_cg_support = AttentionCGSupport.ALWAYS
        min_cg_builder_name = None

        for attn_group in self._attn_group_iterator():
            builder = attn_group.metadata_builder
            if builder.cudagraph_support.value < min_cg_support.value:
                min_cg_support = builder.cudagraph_support
                min_cg_builder_name = builder.__class__.__name__

        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
        # check cudagraph for mixed batch is supported
        if cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL \
            and min_cg_support != AttentionCGSupport.ALWAYS:
            msg = (f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                   f"with {min_cg_builder_name} backend (support: "
                   f"{min_cg_support})")
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
                msg += "; please try cudagraph_mode=PIECEWISE, and "\
                    "make sure compilation level is piecewise"
                raise ValueError(msg)

            # attempt to resolve the full cudagraph related mode
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=FULL_AND_PIECEWISE"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.FULL_AND_PIECEWISE
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.FULL_DECODE_ONLY
            logger.warning(msg)

        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
        if (cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and self.uniform_decode_query_len > 1 and min_cg_support.value
                < AttentionCGSupport.UNIFORM_BATCH.value):
            msg = (f"CUDAGraphMode.{cudagraph_mode.name} is not supported"
                   f" with spec-decode for attention backend "
                   f"{min_cg_builder_name} (support: {min_cg_support})")
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.PIECEWISE
            else:
                msg += "; setting cudagraph_mode=NONE"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.NONE
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
        if cudagraph_mode.has_full_cudagraphs() \
            and min_cg_support == AttentionCGSupport.NEVER:
            raise ValueError(f"CUDAGraphMode.{cudagraph_mode.name} is not "
                             f"supported with {min_cg_builder_name} backend ("
                             f"support:{min_cg_support}) "
                             "; please try cudagraph_mode=PIECEWISE, "
                             "and make sure compilation level is piecewise")

        # Trigger cudagraph dispatching keys initialization here (after
        # initializing attn backends).
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
            self.compilation_config.cudagraph_mode,
            self.uniform_decode_query_len)

    def calculate_reorder_batch_threshold(self) -> None:
        """
        Check that if any backends reorder batches; that the reordering
        is compatible (e.g., decode threshold is the same)
        """
        for group in self._attn_group_iterator():
            attn_metadata_builder_i = group.metadata_builder

            # check that if any backends reorder batches; that the reordering
            # is compatible (e.g., decode threshold is the same)
            reorder_batch_threshold_i = (
                attn_metadata_builder_i.reorder_batch_threshold)
            if reorder_batch_threshold_i is not None:
                if self.reorder_batch_threshold is not None:
                    if reorder_batch_threshold_i != \
                        self.reorder_batch_threshold:
                        raise ValueError(
                            f"Attention backend reorders decodes with "
                            f"threshold {reorder_batch_threshold_i} but other "
                            f"backend uses threshold "
                            f"{self.reorder_batch_threshold}")
                else:
                    self.reorder_batch_threshold = reorder_batch_threshold_i

    def may_reinitialize_input_batch(self,
                                     kv_cache_config: KVCacheConfig) -> None:
        """
        Re-initialize the input batch if the block sizes are different from
        `[self.cache_config.block_size]`. This usually happens when there
        are multiple KV cache groups.

        Args:
            kv_cache_config: The KV cache configuration.
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
        ]
        if block_sizes != [self.cache_config.block_size]:
            assert self.cache_config.cpu_offload_gb == 0, (
                "Cannot re-initialize the input batch when CPU weight "
                "offloading is enabled. See https://github.com/vllm-project/vllm/pull/18298 "  # noqa: E501
                "for more details.")
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
                max_model_len=self.max_model_len,
                max_num_batched_tokens=self.max_num_tokens,
                device=self.device,
                pin_memory=self.pin_memory,
                vocab_size=self.model_config.get_vocab_size(),
                block_sizes=block_sizes,
                is_spec_decode=bool(self.vllm_config.speculative_config),
                logitsprocs=self.input_batch.logitsprocs,
                is_pooling_model=self.is_pooling_model,
            )

    def _allocate_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
        """
        Initializes the KV cache buffer with the correct size. The buffer needs
        to be reshaped to the desired shape before being used by the models.

        Args:
            kv_cache_config: The KV cache config
        Returns:
            dict[str, torch.Tensor]: A map between layer names to their
            corresponding memory buffer for KV cache.
         """
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
            tensor = torch.zeros(kv_cache_tensor.size,
                                 dtype=torch.int8,
                                 device=self.device)
            for layer_name in kv_cache_tensor.shared_by:
                kv_cache_raw_tensors[layer_name] = tensor

        layer_names = set()
        for group in kv_cache_config.kv_cache_groups:
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
        assert layer_names == set(kv_cache_raw_tensors.keys(
        )), "Some layers are not correctly initialized"
        return kv_cache_raw_tensors

    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

    def _kv_cache_spec_attn_group_iterator(
            self) -> Iterator[tuple[KVCacheSpec, AttentionGroup]]:
        if not self.kv_cache_config.kv_cache_groups:
            return
        for kv_cache_spec_id, attn_groups in enumerate(self.attn_groups):
            for attn_group in attn_groups:
                yield self.kv_cache_config.kv_cache_groups[
                    kv_cache_spec_id].kv_cache_spec, attn_group

    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
        """
        Reshape the KV cache tensors to the desired shape and dtype.

        Args:
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
                correct size but uninitialized shape.
        Returns:
            Dict[str, torch.Tensor]: A map between layer names to their
            corresponding memory buffer for KV cache.
        """
        kv_caches: dict[str, torch.Tensor] = {}
        has_attn, has_mamba = False, False
        for kv_cache_spec, group in self._kv_cache_spec_attn_group_iterator():
            attn_backend = group.backend
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
                num_blocks = (raw_tensor.numel() //
                              kv_cache_spec.page_size_bytes)
                if isinstance(kv_cache_spec, AttentionSpec):
                    has_attn = True
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    dtype = kv_cache_spec.dtype
                    try:
                        kv_cache_stride_order = \
                            attn_backend.get_kv_cache_stride_order()
                        assert len(kv_cache_stride_order) == len(
                            kv_cache_shape)
                    except (AttributeError, NotImplementedError):
                        kv_cache_stride_order = tuple(
                            range(len(kv_cache_shape)))
                    # The allocation respects the backend-defined stride order
                    # to ensure the semantic remains consistent for each
                    # backend. We first obtain the generic kv cache shape and
                    # then permute it according to the stride order which could
                    # result in a non-contiguous tensor.
                    kv_cache_shape = tuple(kv_cache_shape[i]
                                           for i in kv_cache_stride_order)
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
                    kv_caches[layer_name] = kv_cache_raw_tensors[
                        layer_name].view(dtype).view(kv_cache_shape).permute(
                            *inv_order)
                elif isinstance(kv_cache_spec, MambaSpec):
                    has_mamba = True
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
                    storage_offset_bytes = 0
                    for (shape, dtype) in zip(kv_cache_spec.shapes,
                                              kv_cache_spec.dtypes):
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
                            kv_cache_spec.page_size_bytes // dtype_size)
                        target_shape = (num_blocks, *shape)
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
                        assert storage_offset_bytes % dtype_size == 0
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
                            storage_offset=storage_offset_bytes // dtype_size,
                        )
                        state_tensors.append(tensor)
                        storage_offset_bytes += stride[0] * dtype_size

                    kv_caches[layer_name] = state_tensors
                else:
                    raise NotImplementedError

        if has_attn and has_mamba:
            self._update_hybrid_attention_mamba_layout(kv_caches)

        return kv_caches

    def _update_hybrid_attention_mamba_layout(
            self, kv_caches: dict[str, torch.Tensor]) -> None:
        """
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).

        Args:
            kv_caches: The KV cache buffer of each layer.
        """

        for kv_cache_spec, group in self._kv_cache_spec_attn_group_iterator():
            for layer_name in group.layer_names:
                kv_cache = kv_caches[layer_name]
                if (isinstance(kv_cache_spec, AttentionSpec)
                        and kv_cache.shape[0] == 2):
                    assert kv_cache.shape[1] != 2, \
                        "Fail to determine whether the layout is " \
                        "(2, num_blocks, ...) or (num_blocks, 2, ...) for " \
                        f"a tensor of shape {kv_cache.shape}"
                    hidden_size = kv_cache.shape[2:].numel()
                    kv_cache.as_strided_(size=kv_cache.shape,
                                         stride=(hidden_size, 2 * hidden_size,
                                                 *kv_cache.stride()[2:]))

    def initialize_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
        Returns:
            Dict[str, torch.Tensor]: A map between layer names to their
            corresponding memory buffer for KV cache.
        """
        # Initialize the memory buffer for KV cache
        kv_cache_raw_tensors = self._allocate_kv_cache_tensors(kv_cache_config)
        # Change the memory buffer to the desired shape
        kv_caches = self._reshape_kv_cache_tensors(kv_cache_config,
                                                   kv_cache_raw_tensors)

        # Set up cross-layer KV cache sharing
        for layer_name, target_layer_name in self.shared_kv_cache_layers.items(
        ):
            logger.debug("%s reuses KV cache of %s", layer_name,
                         target_layer_name)
            kv_caches[layer_name] = kv_caches[target_layer_name]

        bind_kv_cache(kv_caches,
                      self.compilation_config.static_forward_context,
                      self.kv_caches)
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
            self, kv_cache_config: KVCacheConfig) -> None:
        """
        Add layers that re-use KV cache to KV cache group of its target layer.
        Mapping of KV cache tensors happens in `initialize_kv_cache_tensors()`
        """
        if not self.shared_kv_cache_layers:
            # No cross-layer KV sharing, return
            return

        add_kv_sharing_layers_to_kv_cache_groups(
            self.shared_kv_cache_layers,
            kv_cache_config.kv_cache_groups,
            self.runner_only_attn_layers,
        )

        if self.cache_config.kv_sharing_fast_prefill:
            # In You Only Cache Once (https://arxiv.org/abs/2405.05254) or other
            # similar KV sharing setups, only the layers that generate KV caches
            # are involved in the prefill phase, enabling prefill to early exit.
            attn_layers = get_layers_from_vllm_config(self.vllm_config,
                                                      Attention)
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
                    self.kv_sharing_fast_prefill_eligible_layers.add(
                        layer_name)
                else:
                    break

    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize KV cache based on `kv_cache_config`.
        Args:
            kv_cache_config: Configuration for the KV cache, including the KV
            cache size of each layer
        """
        kv_cache_config = deepcopy(kv_cache_config)
        self.kv_cache_config = kv_cache_config
        self.may_reinitialize_input_batch(kv_cache_config)
        self.may_add_encoder_only_layers_to_kv_cache_config()
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
        self.initialize_attn_backend(kv_cache_config)
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)
            if self.device.type == 'xpu':
                get_kv_transfer_group().set_host_xfer_buffer_ops(
                    copy_kv_blocks)

        if self.dcp_world_size > 1:
            layer_names = self.attn_groups[0][0].layer_names
            layers = get_layers_from_vllm_config(self.vllm_config,
                                                 AttentionLayerBase,
                                                 layer_names)
            for layer in layers.values():
                assert layer.impl.need_to_return_lse_for_decode, (
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
                    f"{layer.impl.__class__.__name__} "
                    "does not return the softmax lse for decode.")

    def may_add_encoder_only_layers_to_kv_cache_config(self) -> None:
        """
        Add encoder-only layers to the KV cache config.
        """
        block_size = self.vllm_config.cache_config.block_size
        use_mla = self.vllm_config.model_config.use_mla
        encoder_only_attn_specs: dict[AttentionSpec,
                                      list[str]] = defaultdict(list)
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
            if attn_module.attn_type == AttentionType.ENCODER_ONLY:
                attn_spec = EncoderOnlyAttentionSpec(
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
                    dtype=self.kv_cache_dtype,
                    use_mla=use_mla)
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
            assert len(
                encoder_only_attn_specs
            ) == 1, "Only support one encoder-only attention spec now"
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec))

    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
        """
        Generates the KVCacheSpec by parsing the kv cache format from each
        Attention module in the static forward context.
        Returns:
            KVCacheSpec: A dictionary mapping layer names to their KV cache
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
        use_mla = self.vllm_config.model_config.use_mla
        kv_cache_spec: dict[str, KVCacheSpec] = {}
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
            if (kv_tgt_layer :=
                    attn_module.kv_sharing_target_layer_name) is not None:
                # The layer doesn't need its own KV cache and will use that of
                # the target layer. We skip creating a KVCacheSpec for it, so
                # that KV cache management logic will act as this layer does
                # not exist, and doesn't allocate KV cache for the layer. This
                # enables the memory saving of cross-layer kv sharing, allowing
                # a given amount of memory to accommodate longer context lengths
                # or enable more requests to be processed simultaneously.
                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
                continue

            # TODO: Support other attention modules, e.g., cross-attention
            # TODO(lucas): move the attention specs into the model layers like
            # the attention backends
            if attn_module.attn_type == AttentionType.DECODER:
                if attn_module.sliding_window is not None:
                    kv_cache_spec[layer_name] = SlidingWindowSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
                        sliding_window=attn_module.sliding_window,
                        use_mla=use_mla)
                elif self.attention_chunk_size is not None \
                        and isinstance(attn_module, ChunkedLocalAttention):
                    kv_cache_spec[layer_name] = ChunkedLocalAttentionSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
                        attention_chunk_size=self.attention_chunk_size,
                        use_mla=use_mla)
                else:
                    kv_cache_spec[layer_name] = FullAttentionSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
                        use_mla=use_mla)
            elif attn_module.attn_type in (AttentionType.ENCODER,
                                           AttentionType.ENCODER_ONLY):
                # encoder-only attention does not need KV cache.
                continue
            elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
                raise NotImplementedError
            else:
                raise ValueError(
                    f"Unknown attention type: {attn_module.attn_type}")

        mamba_layers = get_layers_from_vllm_config(self.vllm_config, MambaBase)
        if len(mamba_layers) > 0:
            if self.vllm_config.speculative_config is not None:
                raise NotImplementedError(
                    "Mamba with speculative decoding is not supported yet.")
            if self.vllm_config.cache_config.enable_prefix_caching:
                raise NotImplementedError(
                    "Prefix caching is not supported for Mamba yet.")
            max_model_len = self.vllm_config.model_config.max_model_len

            page_size_padded = (
                self.vllm_config.cache_config.mamba_page_size_padded)

            # Set block_size to max_model_len, so that mamba model will always
            # have only one block in the KV cache.
            for layer_name, mamba_module in mamba_layers.items():
                kv_cache_spec[layer_name] = MambaSpec(
                    shapes=mamba_module.get_state_shape(),
                    dtypes=mamba_module.get_state_dtype(),
                    block_size=max_model_len,
                    page_size_padded=page_size_padded,
                    mamba_type=mamba_module.mamba_type)

        return kv_cache_spec

    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
        # This is a short term mitigation for issue mentioned in
        # https://github.com/vllm-project/vllm/issues/22754.
        # `tolist` would trigger a cuda wise stream sync, which
        # would block other copy ops from other cuda streams.
        # A cuda event sync would avoid such a situation. Since
        # this is in the critical path of every single model
        # forward loop, this has caused perf issue for a disagg
        # setup.
        pinned = self.sampled_token_ids_pinned_cpu[:sampled_token_ids.shape[0]]
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()