gpu_model_runner.py

# SPDX-License-Identifier: Apache-2.0

import copy
import gc
import time
import weakref
from typing import TYPE_CHECKING, Optional, Union

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

from vllm.attention import AttentionType, get_attn_backend
from vllm.attention.backends.abstract import (AttentionBackend,
                                              AttentionMetadataBuilder)
from vllm.attention.layer import Attention
from vllm.attention.utils.fa_utils import get_flash_attn_version
from vllm.config import (CompilationLevel, VllmConfig,
                         get_layers_from_vllm_config)
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group)
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
from vllm.distributed.parallel_state import (
    get_pp_group, get_tp_group, graph_capture,
    prepare_communication_buffer_for_model)
from vllm.forward_context import get_forward_context, set_forward_context
from vllm.logger import init_logger
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
from vllm.model_executor.model_loader import get_model
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
from vllm.multimodal.utils import group_mm_inputs_by_modality
from vllm.sampling_params import SamplingType
from vllm.sequence import IntermediateTensors
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                        GiB_bytes, LazyLoader, cdiv, check_use_alibi,
                        is_pin_memory_available)
from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
                                        KVCacheConfig, KVCacheSpec,
                                        SlidingWindowSpec)
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
                             ModelRunnerOutput)
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.spec_decode.utils import is_spec_decode_supported
from vllm.v1.utils import bind_kv_cache
from vllm.v1.worker.block_table import BlockTable
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin

from .utils import (gather_mm_placeholders, sanity_check_mm_encoder_outputs,
                    scatter_mm_placeholders)

if TYPE_CHECKING:
    import xgrammar as xgr

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

logger = init_logger(__name__)


class GPUModelRunner(LoRAModelRunnerMixin):

    def __init__(
        self,
        vllm_config: VllmConfig,
        device: torch.device,
    ):
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        self.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.prompt_adapter_config = vllm_config.prompt_adapter_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_multimodal_model = model_config.is_multimodal_model
        self.max_model_len = model_config.max_model_len
        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

        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

        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
            model_config=model_config,
            scheduler_config=scheduler_config,
            mm_registry=self.mm_registry,
        )
        self.max_num_encoder_input_tokens = encoder_compute_budget
        self.encoder_cache_size = encoder_cache_size

        # Sampler
        self.sampler = Sampler()

        # Lazy initializations
        # self.model: nn.Module  # Set after load_model
        # Initialize in initialize_kv_cache
        self.kv_caches: list[torch.Tensor] = []
        self.attn_metadata_builders: list[AttentionMetadataBuilder] = []
        self.attn_backends: list[type[AttentionBackend]] = []
        # self.kv_cache_config: KVCacheConfig
        # self.input_batch: InputBatch # Persistent batch.

        # req_id -> (input_id -> encoder_output)
        self.encoder_cache: dict[str, dict[int, torch.Tensor]] = {}

        # Set up speculative decoding.
        self.use_spec_decode = False
        self.use_aux_hidden_state_outputs = False
        if self.speculative_config:
            self.use_spec_decode = True
            if 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)  # 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] = {}

        self.use_cuda_graph = (self.vllm_config.compilation_config.level
                               == CompilationLevel.PIECEWISE
                               and not self.model_config.enforce_eager)
        # 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.
        self.cudagraph_batch_sizes = list(
            reversed(
                self.vllm_config.compilation_config.cudagraph_capture_sizes))

        # Cache the device properties.
        self.device_properties = torch.cuda.get_device_properties(self.device)
        self.num_sms = self.device_properties.multi_processor_count

        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
        self.query_start_loc = torch.zeros(self.max_num_reqs + 1,
                                           dtype=torch.int32,
                                           device=self.device)
        self.seq_lens = torch.zeros(self.max_num_reqs,
                                    dtype=torch.int32,
                                    device=self.device)
        self.slot_mapping = torch.zeros(self.max_num_tokens,
                                        dtype=torch.int64,
                                        device=self.device)

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

        # 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 = torch.zeros((3, self.max_num_tokens + 1),
                                               dtype=torch.int64,
                                               device=self.device)
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)

        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)

        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)

        # 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)
        # NOTE(woosuk): These tensors are "stateless", i.e., they are literally
        # a faster version of creating a new tensor every time. Thus, we should
        # not make any assumptions about the values in these tensors.
        self.input_ids_cpu = torch.zeros(self.max_num_tokens,
                                         dtype=torch.int32,
                                         device="cpu",
                                         pin_memory=self.pin_memory)
        self.positions_cpu = torch.zeros(self.max_num_tokens,
                                         dtype=torch.int64,
                                         device="cpu",
                                         pin_memory=self.pin_memory)
        self.positions_np = self.positions_cpu.numpy()
        self.query_start_loc_cpu = torch.zeros(self.max_num_reqs + 1,
                                               dtype=torch.int32,
                                               device="cpu",
                                               pin_memory=self.pin_memory)
        self.query_start_loc_np = self.query_start_loc_cpu.numpy()
        self.seq_lens_cpu = torch.zeros(self.max_num_reqs,
                                        dtype=torch.int32,
                                        device="cpu",
                                        pin_memory=self.pin_memory)
        self.seq_lens_np = self.seq_lens_cpu.numpy()

    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> bool:
        """
        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.

        Returns:
            True if the batch was reordered, False otherwise.
        """
        batch_reordered = self.attn_metadata_builders[0].reorder_batch(
            self.input_batch, scheduler_output)

        # For models with multiple KV cache groups, the groups should agree on
        # the same order of requests. We ensure this by only allowing the first
        # group to reorder the batch and asserting that all other groups do not
        # reorder the batch.
        for i in range(1, len(self.kv_cache_config.kv_cache_groups)):
            assert not self.attn_metadata_builders[i].reorder_batch(
                self.input_batch, scheduler_output)
        return batch_reordered

    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)
            self.encoder_cache.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.
        removed_req_indices: list[int] = []
        for req_id in scheduler_output.finished_req_ids:
            req_index = self.input_batch.remove_request(req_id)
            if req_index is not None:
                removed_req_indices.append(req_index)

        # Free the cached encoder outputs.
        for req_id, input_id in scheduler_output.free_encoder_input_ids:
            encoder_outputs = self.encoder_cache.get(req_id)
            if encoder_outputs is not None:
                encoder_outputs.pop(input_id, None)
                if not encoder_outputs:
                    self.encoder_cache.pop(req_id, 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:
            req_index = self.input_batch.remove_request(req_id)
            assert req_index is not None
            removed_req_indices.append(req_index)

        req_ids_to_add: list[str] = []
        # 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
            if sampling_params.sampling_type == SamplingType.RANDOM_SEED:
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
                prompt_token_ids=new_req_data.prompt_token_ids,
                mm_inputs=new_req_data.mm_inputs,
                mm_positions=new_req_data.mm_positions,
                sampling_params=sampling_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,
            )

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            if self.uses_mrope:
                image_grid_thw = []
                video_grid_thw = []
                second_per_grid_ts = []
                audio_feature_lengths = []
                use_audio_in_video = False
                for mm_input in self.requests[req_id].mm_inputs:
                    if mm_input.get("image_grid_thw") is not None:
                        image_grid_thw.extend(
                            mm_input["image_grid_thw"].tolist())
                    if mm_input.get("video_grid_thw") is not None:
                        video_grid_thw.extend(
                            mm_input["video_grid_thw"].tolist())
                    if mm_input.get("second_per_grid_ts") is not None:
                        second_per_grid_ts.extend(
                            mm_input["second_per_grid_ts"])
                    if mm_input.get("audio_feature_lengths") is not None:
                        audio_feature_lengths.extend(
                            mm_input["audio_feature_lengths"])
                    if mm_input.get("use_audio_in_video") is True:
                        use_audio_in_video = True

                hf_config = self.model_config.hf_config

                self.requests[req_id].mrope_positions, \
                    self.requests[req_id].mrope_position_delta = \
                    MRotaryEmbedding.get_input_positions_tensor(
                        self.requests[req_id].prompt_token_ids,
                        hf_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,
                    )

            req_ids_to_add.append(req_id)

        # Update the states of the running/resumed requests.
        for req_data in scheduler_output.scheduled_cached_reqs:
            req_id = req_data.req_id
            req_state = self.requests[req_id]

            # Update the cached states.
            num_computed_tokens = req_data.num_computed_tokens
            req_state.num_computed_tokens = num_computed_tokens
            # Add the sampled token(s) from the previous step (if any).
            # This doesn't include "unverified" tokens like spec decode tokens.
            num_new_tokens = (num_computed_tokens +
                              len(req_data.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(req_data.new_token_ids[-1])
            elif num_new_tokens > 0:
                req_state.output_token_ids.extend(
                    req_data.new_token_ids[-num_new_tokens:])
            # Update the block IDs.
            if not req_data.resumed_from_preemption:
                # Append the new blocks to the existing block IDs.
                for i in range(len(self.kv_cache_config.kv_cache_groups)):
                    req_state.block_ids[i].extend(req_data.new_block_ids[i])
            else:
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
                req_state.block_ids = req_data.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.
                req_ids_to_add.append(req_id)
                continue

            # Update the persistent batch.
            self.input_batch.num_computed_tokens_cpu[req_index] = (
                num_computed_tokens)
            self.input_batch.block_table.append_row(req_data.new_block_ids,
                                                    req_index)
            # Add new_token_ids to token_ids_cpu.
            start_token_index = num_computed_tokens
            end_token_index = num_computed_tokens + len(req_data.new_token_ids)
            self.input_batch.token_ids_cpu[
                req_index,
                start_token_index:end_token_index] = req_data.new_token_ids
            self.input_batch.num_tokens_no_spec[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:
                start_index = end_token_index
                end_token_index += len(spec_token_ids)
                self.input_batch.token_ids_cpu[
                    req_index, start_index:end_token_index] = spec_token_ids
            # NOTE(woosuk): `num_tokens` here may include spec decode tokens.
            self.input_batch.num_tokens[req_index] = end_token_index

        # Check if the batch has changed. If not, we can skip copying the
        # sampling metadata from CPU to GPU.
        batch_changed = len(removed_req_indices) > 0 or len(req_ids_to_add) > 0

        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
        removed_req_indices.sort(reverse=True)
        for req_id in req_ids_to_add:
            req_state = self.requests[req_id]
            if removed_req_indices:
                # Fill the empty index.
                req_index = removed_req_indices.pop()
            else:
                # Append to the end.
                req_index = None
            self.input_batch.add_request(req_state, req_index)

        # Condense the batched states if there are empty indices.
        if removed_req_indices:
            self.input_batch.condense(removed_req_indices)

        batch_reordered = self._may_reorder_batch(scheduler_output)

        if batch_changed or batch_reordered:
            self.input_batch.refresh_sampling_metadata()

    def _prepare_inputs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> tuple[dict[str, FlashAttentionMetadata], torch.Tensor,
               Optional[SpecDecodeMetadata]]:
        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(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)

        # Get batched arange.
        # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        # Equivalent to but faster than:
        # np.concatenate([np.arange(n) for n in num_scheduled_tokens])
        # Step 1. [2, 5, 3] -> [2, 7, 10]
        cu_num_tokens = np.cumsum(num_scheduled_tokens)
        # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
        cumsums_offsets = np.repeat(cu_num_tokens - num_scheduled_tokens,
                                    num_scheduled_tokens)
        # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        arange = self.arange_np[:total_num_scheduled_tokens] - cumsums_offsets

        # 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])

        # Calculate the slot mapping for each KV cache group.
        for kv_cache_group_id, kv_cache_group_spec in enumerate(
                self.kv_cache_config.kv_cache_groups):
            block_size = kv_cache_group_spec.kv_cache_spec.block_size
            block_table: BlockTable = self.input_batch.block_table[
                kv_cache_group_id]
            # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
            # -> [0, 0, K, K, K + 1, K + 1, K + 2, 2 * K, 2 * K, 2 * K + 1]
            # where K is the max_num_blocks_per_req and the block size is 2.
            # NOTE(woosuk): We can't simply use `token_indices // block_size`
            # here because M (max_model_len) is not necessarily divisible by
            # block_size.
            block_table_indices = (
                req_indices * block_table.max_num_blocks_per_req +
                positions_np // block_size)
            block_table_cpu = block_table.get_cpu_tensor()
            block_numbers = block_table_cpu.flatten(
            )[block_table_indices].numpy()
            block_offsets = positions_np % block_size
            np.add(
                block_numbers * block_size,
                block_offsets,
                out=block_table.slot_mapping_np[: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

        self.seq_lens_np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)

        # Copy the tensors to the GPU.
        self.input_ids[:total_num_scheduled_tokens].copy_(
            self.input_ids_cpu[:total_num_scheduled_tokens], non_blocking=True)
        if self.uses_mrope:
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            self.mrope_positions[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions_cpu[:, :total_num_scheduled_tokens],
                non_blocking=True)
        else:
            # Common case (1D positions)
            self.positions[:total_num_scheduled_tokens].copy_(
                self.positions_cpu[:total_num_scheduled_tokens],
                non_blocking=True)

        self.query_start_loc[:num_reqs + 1].copy_(
            self.query_start_loc_cpu[:num_reqs + 1], non_blocking=True)
        self.seq_lens[:num_reqs].copy_(self.seq_lens_cpu[:num_reqs],
                                       non_blocking=True)

        # Fill unused with -1. Needed for reshape_and_cache
        self.seq_lens[num_reqs:].fill_(0)
        self.query_start_loc[num_reqs + 1:].fill_(-1)

        query_start_loc = self.query_start_loc[:num_reqs + 1]
        seq_lens = self.seq_lens[:num_reqs]

        common_attn_metadata = CommonAttentionMetadata(
            query_start_loc=query_start_loc, seq_lens=seq_lens)

        attn_metadata: dict[str, FlashAttentionMetadata] = {}
        # 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):

            # Prepare for cascade attention if enabled & beneficial.
            common_prefix_len = 0
            if self.cascade_attn_enabled:
                common_prefix_len = self._compute_cascade_attn_prefix_len(
                    num_scheduled_tokens,
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id],
                    kv_cache_group_spec.kv_cache_spec,
                    self.attn_metadata_builders[kv_cache_group_id],
                )

            attn_metadata_i = (
                self.attn_metadata_builders[kv_cache_group_id].build(
                    num_reqs=num_reqs,
                    num_actual_tokens=total_num_scheduled_tokens,
                    max_query_len=max_num_scheduled_tokens,
                    common_prefix_len=common_prefix_len,
                    common_attn_metadata=common_attn_metadata))
            for layer_name in kv_cache_group_spec.layer_names:
                attn_metadata[layer_name] = attn_metadata_i

        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

        # 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

    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))
        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,
            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

                self.mrope_positions_cpu[:, dst_start:dst_end] = \
                    MRotaryEmbedding.get_next_input_positions_tensor(
                        req.mrope_position_delta,
                        context_len=num_computed_tokens +
                        prompt_part_len,
                        seq_len=num_computed_tokens +
                        prompt_part_len +
                        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. [4, 5, 8, 9, 11]
        cu_num_sampled_tokens = np.cumsum(num_sampled_tokens, dtype=np.int32)
        total_num_sampled_tokens = cu_num_sampled_tokens[-1]
        # Step 2. [0, 0, 0, 0, 4, 5, 5, 5, 8, 9, 9]
        cumsums_offsets = np.repeat(cu_num_sampled_tokens - num_sampled_tokens,
                                    num_sampled_tokens)
        # Step 3. [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        arange = self.arange_np[:total_num_sampled_tokens] - cumsums_offsets
        # Step 4. [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 5. [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.
        # [3, 3, 5, 5, 6]
        cu_num_draft_tokens = np.cumsum(num_draft_tokens, dtype=np.int32)
        total_num_draft_tokens = cu_num_draft_tokens[-1]
        # [0, 0, 0, 3, 3, 5]
        cumsums_offsets = np.repeat(cu_num_draft_tokens - num_draft_tokens,
                                    num_draft_tokens)
        # [0, 1, 2, 0, 1, 0]
        arange = self.arange_np[:total_num_draft_tokens] - cumsums_offsets
        # [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[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 _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_inputs = list[MultiModalKwargs]()
        req_ids_pos = list[tuple[str, int, 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_inputs.append(req_state.mm_inputs[mm_input_id])
                req_ids_pos.append(
                    (req_id, mm_input_id, 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.
        grouped_mm_inputs_list = group_mm_inputs_by_modality(mm_inputs)

        encoder_outputs = []
        for grouped_mm_inputs in grouped_mm_inputs_list:
            batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs)
            batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs,
                                                           device=self.device)

            # 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(
                **batched_mm_inputs)

            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
                expected_num_items=len(grouped_mm_inputs),
            )

            for output in curr_group_outputs:
                encoder_outputs.append(output)

        # Cache the encoder outputs.
        for (req_id, input_id, pos_info), output in zip(
                req_ids_pos,
                encoder_outputs,
        ):
            if req_id not in self.encoder_cache:
                self.encoder_cache[req_id] = {}

            self.encoder_cache[req_id][input_id] = scatter_mm_placeholders(
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> 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
            mm_positions = req_state.mm_positions
            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
                assert req_id in self.encoder_cache
                assert i in self.encoder_cache[req_id]
                encoder_output = self.encoder_cache[req_id][i]

                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:
        return self.model

    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.zeros_like(grammar_bitmask,
                                       shape=(logits.shape[0],
                                              grammar_bitmask.shape[1]))
        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

        # 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)

        xgr.apply_token_bitmask_inplace(
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
            indices=out_indices,
        )

    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.vllm_config.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 = "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()
        })

    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> Union[ModelRunnerOutput, IntermediateTensors]:

        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)

        # Prepare the decoder inputs.
        attn_metadata, logits_indices, spec_decode_metadata = (
            self._prepare_inputs(scheduler_output))
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if (self.use_cuda_graph
                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
            # Use piecewise 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.vllm_config.compilation_config.pass_config. \
                enable_sequence_parallelism and tp_size > 1:
                from vllm.utils import round_up
                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
            else:
                num_input_tokens = num_scheduled_tokens

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

        if self.is_multimodal_model and get_pp_group().is_first_rank:
            # 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.
            input_ids = self.input_ids[:num_scheduled_tokens]
            if mm_embeds:
                inputs_embeds = self.model.get_input_embeddings(
                    input_ids, mm_embeds)
            else:
                inputs_embeds = self.model.get_input_embeddings(input_ids)
            # TODO(woosuk): Avoid the copy. Optimize.
            self.inputs_embeds[:num_scheduled_tokens].copy_(inputs_embeds)
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
            input_ids = None
        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[:num_input_tokens]
            inputs_embeds = None
        if self.uses_mrope:
            positions = self.mrope_positions[:, :num_input_tokens]
        else:
            positions = self.positions[: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)

        # Run the decoder.
        # Use persistent buffers for CUDA graphs.
        with set_forward_context(attn_metadata,
                                 self.vllm_config,
                                 num_tokens=num_input_tokens):
            self.maybe_setup_kv_connector(scheduler_output)

            model_output = self.model(
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
            )

            self.maybe_wait_for_kv_save()
            finished_sending, finished_recving = (
                self.get_finished_kv_transfers(scheduler_output))

        if self.use_aux_hidden_state_outputs:
            hidden_states, aux_hidden_states = model_output
        else:
            hidden_states = model_output
        # 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.
            if not broadcast_pp_output:
                return hidden_states
            assert isinstance(hidden_states, IntermediateTensors)
            get_pp_group().send_tensor_dict(hidden_states.tensors,
                                            all_gather_group=get_tp_group())
            logits = None
        else:
            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)

        # 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

        # 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)

        # 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,
        )

        # Get the valid generated tokens.
        sampled_token_ids = sampler_output.sampled_token_ids
        max_gen_len = sampled_token_ids.shape[-1]
        if max_gen_len == 1:
            # No spec decode tokens.
            valid_sampled_token_ids = sampled_token_ids.tolist()
        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()

        if not self.use_spec_decode:
            # Speculative decoding is not enabled.
            spec_token_ids = None
        elif self.speculative_config.method == "ngram":
            assert isinstance(self.drafter, NgramProposer)
            spec_token_ids = self.generate_draft_token_ids(
                valid_sampled_token_ids, sampling_metadata)
        elif self.speculative_config.method == "medusa":
            assert isinstance(self.drafter, MedusaProposer)
            if max_gen_len == 1:
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
                        valid_sampled_token_ids):
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1

                indices = torch.tensor(indices,
                                       device=sample_hidden_states.device)
                hidden_states = sample_hidden_states[indices]

            spec_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.
            next_token_ids: list[int] = []
            for i, token_ids in enumerate(valid_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 = self.input_batch.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)
            eagle_attn_metadata = attn_metadata[self.drafter.attn_layer_name]

            if spec_decode_metadata is None:
                # input_ids can be None for multimodal models.
                target_token_ids = self.input_ids[:num_scheduled_tokens]
                target_positions = positions[: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]
                target_slot_mapping = eagle_attn_metadata.slot_mapping
                cu_num_tokens = eagle_attn_metadata.query_start_loc
            else:
                # TODO(woosuk): Refactor this.
                num_draft_tokens = spec_decode_metadata.num_draft_tokens
                num_rejected_tokens = [
                    n + 1 - len(valid_sampled_token_ids[i]) if n > 0 else 0
                    for i, n in enumerate(num_draft_tokens)
                ]
                num_rejected_tokens = torch.tensor(
                    num_rejected_tokens,
                    dtype=torch.int32,
                    device=self.device,
                )
                cu_num_tokens, token_indices = self.drafter.prepare_inputs(
                    eagle_attn_metadata.query_start_loc,
                    num_rejected_tokens,
                )
                target_token_ids = self.input_ids[token_indices]
                target_positions = positions[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]
                target_slot_mapping = eagle_attn_metadata.slot_mapping[
                    token_indices]

            draft_token_ids = self.drafter.propose(
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                target_slot_mapping=target_slot_mapping,
                next_token_ids=next_token_ids,
                cu_num_tokens=cu_num_tokens,
                block_table=eagle_attn_metadata.block_table,
                sampling_metadata=sampling_metadata,
            )
            spec_token_ids = draft_token_ids.tolist()

        # Clear KVConnector state after all KVs are generated.
        if has_kv_transfer_group():
            get_kv_transfer_group().clear_connector_metadata()

        return ModelRunnerOutput(
            req_ids=self.input_batch.req_ids,
            req_id_to_index=self.input_batch.req_id_to_index,
            sampled_token_ids=valid_sampled_token_ids,
            spec_token_ids=spec_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            finished_sending=finished_sending,
            finished_recving=finished_recving,
        )

    def kv_connector_no_forward(
            self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput:
        # KV send/recv even if no work to do.
        with set_forward_context(None, self.vllm_config):
            self.maybe_setup_kv_connector(scheduler_output)
            finished_sending, finished_recving = (
                self.get_finished_kv_transfers(scheduler_output))

        if not finished_sending and not finished_recving:
            return EMPTY_MODEL_RUNNER_OUTPUT

        output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
        output.finished_sending = finished_sending
        output.finished_recving = finished_recving
        return output

    @staticmethod
    def maybe_setup_kv_connector(scheduler_output: "SchedulerOutput"):
        # Update KVConnector with the KVConnector metadata forward().
        if has_kv_transfer_group():
            kv_connector = get_kv_transfer_group()
            assert isinstance(kv_connector, KVConnectorBase_V1)
            assert scheduler_output.kv_connector_metadata is not None
            kv_connector.bind_connector_metadata(
                scheduler_output.kv_connector_metadata)

            # Background KV cache transfers happen here.
            # These transfers are designed to be async and the requests
            # involved may be disjoint from the running requests.
            # Do this here to save a collective_rpc.
            kv_connector.start_load_kv(get_forward_context())

    @staticmethod
    def maybe_wait_for_kv_save() -> None:
        if has_kv_transfer_group():
            get_kv_transfer_group().wait_for_save()

    @staticmethod
    def get_finished_kv_transfers(
        scheduler_output: "SchedulerOutput",
    ) -> tuple[Optional[set[str]], Optional[set[str]]]:
        if has_kv_transfer_group():
            return get_kv_transfer_group().get_finished(
                scheduler_output.finished_req_ids)
        return None, None

    def generate_draft_token_ids(
        self,
        sampled_token_ids: list[list[int]],
        sampling_metadata: SamplingMetadata,
    ) -> list[list[int]]:
        # TODO(woosuk): Optimize.
        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 = self.input_batch.req_ids[i]
            if not is_spec_decode_supported(req_id, self.input_batch):
                draft_token_ids.append([])
                continue

            # Add sampled_token_ids to token_ids_cpu.
            start_idx = self.input_batch.num_tokens_no_spec[i]
            end_idx = start_idx + num_sampled_ids
            if end_idx >= self.max_model_len:
                # Skip requests that have already reached the max model length.
                draft_token_ids.append([])
                continue

            self.input_batch.token_ids_cpu[i, start_idx:end_idx] = sampled_ids
            drafter_output = self.drafter.propose(
                self.input_batch.token_ids_cpu[i, :end_idx])
            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 load_model(self) -> None:
        logger.info("Starting to load model %s...", self.model_config.model)
        with DeviceMemoryProfiler() as m:  # noqa: SIM117
            time_before_load = time.perf_counter()
            self.model = get_model(vllm_config=self.vllm_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:
                self.model.set_aux_hidden_state_layers(
                    self.model.get_eagle3_aux_hidden_state_layers())
            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)

    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        scheduler_output: "SchedulerOutput",
    ) -> 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 = scheduler_output.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:
            torch.cuda.synchronize()

        return prompt_logprobs_dict

    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
        skip_attn: bool = True,
    ) -> torch.Tensor:

        # 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
        num_reqs = max_num_reqs if num_tokens >= max_num_reqs else num_tokens
        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)

        if skip_attn:
            attn_metadata: Optional[dict[str, FlashAttentionMetadata]] = None
        else:
            query_start_loc = self.query_start_loc[:num_reqs + 1]
            seq_lens = self.seq_lens[:num_reqs]

            common_attn_metadata = CommonAttentionMetadata(
                query_start_loc=query_start_loc, seq_lens=seq_lens)

            attn_metadata = {}
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
                attn_metadata_i = (
                    self.attn_metadata_builders[kv_cache_group_id].build(
                        num_reqs=num_tokens,
                        num_actual_tokens=num_tokens,
                        max_query_len=num_tokens,
                        common_prefix_len=0,
                        common_attn_metadata=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):
            model = self.model
            if self.is_multimodal_model:
                input_ids = None
                inputs_embeds = self.inputs_embeds[:num_tokens]
            else:
                input_ids = self.input_ids[:num_tokens]
                inputs_embeds = None
            if self.uses_mrope:
                positions = self.mrope_positions[:, :num_tokens]
            else:
                positions = self.positions[: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)

            with set_forward_context(attn_metadata,
                                     self.vllm_config,
                                     num_tokens=num_tokens):
                outputs = model(
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
                )
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs

            if self.use_spec_decode and \
                self.speculative_config.method in ('eagle', 'eagle3'):
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

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

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:

        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),
            min_p=None,
            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)],
            min_tokens={},
            logit_bias=[None for _ in range(num_reqs)],
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
        )
        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.use_spec_decode:
            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 profile_run(self) -> None:
        # Profile with multimodal encoder & encoder cache.
        # TODO: handle encoder-decoder models once we support them.
        if (self.is_multimodal_model and self.max_num_encoder_input_tokens > 0
                and self.encoder_cache_size > 0):

            # NOTE: Currently model is profiled with a single non-text
            # modality with the max possible input tokens even when
            # it supports multiple.
            max_tokens_by_modality_dict = self.mm_registry \
                .get_max_tokens_per_item_by_nonzero_modality(self.model_config)
            dummy_data_modality, max_tokens_per_mm_item = max(
                max_tokens_by_modality_dict.items(), key=lambda item: item[1])

            # Check how many items of this modality can be supported by
            # the encoder budget.
            encoder_budget = min(self.max_num_encoder_input_tokens,
                                 self.encoder_cache_size)

            max_num_mm_items_encoder_budget = cdiv(encoder_budget,
                                                   max_tokens_per_mm_item)

            # Check how many items of this modality can be supported by
            # the decoder budget.
            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
                self.model_config)[dummy_data_modality]

            # NOTE: We do not consider max_num_batched_tokens on purpose
            # because the multimodal embeddings can be generated in advance
            # and chunked prefilled.
            max_num_mm_items_decoder_budget = self.max_num_reqs * \
                max_mm_items_per_req

            max_num_mm_items = min(max_num_mm_items_encoder_budget,
                                   max_num_mm_items_decoder_budget)

            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_num_mm_items, dummy_data_modality)

            # Create dummy batch of multimodal inputs.
            dummy_mm_kwargs = self.mm_registry.get_decoder_dummy_data(
                model_config=self.model_config,
                seq_len=self.max_num_tokens,
                mm_counts={
                    dummy_data_modality: 1
                },
            ).multi_modal_data

            batched_dummy_mm_inputs = MultiModalKwargs.batch(
                [dummy_mm_kwargs] * max_num_mm_items)
            batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
                batched_dummy_mm_inputs, device=self.device)

            # 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_num_mm_items,
            )

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

        hidden_states = self._dummy_run(self.max_num_tokens)
        if get_pp_group().is_last_rank:
            sampler_output = self._dummy_sampler_run(hidden_states)
        else:
            sampler_output = None
        torch.cuda.synchronize()
        del hidden_states, sampler_output
        self.encoder_cache.clear()
        gc.collect()

    def capture_model(self) -> None:
        if not self.use_cuda_graph:
            logger.warning(
                "Skipping CUDA graph capture. Please add "
                "-O %s to use CUDA graphs.", CompilationLevel.PIECEWISE)
            return

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

        # 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.
        with graph_capture(device=self.device):
            skip_attn = not self.vllm_config.compilation_config.full_cuda_graph
            for num_tokens in reversed(self.cudagraph_batch_sizes):
                for _ in range(self.vllm_config.compilation_config.
                               cudagraph_num_of_warmups):
                    self._dummy_run(num_tokens, skip_attn=skip_attn)
                self._dummy_run(num_tokens, skip_attn=skip_attn)

        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 initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
        assert len(self.attn_backends) == 0 and len(
            self.attn_metadata_builders
        ) == 0, "Attention backends are already initialized"
        for i, kv_cache_group_spec in enumerate(
                kv_cache_config.kv_cache_groups):
            kv_cache_spec = kv_cache_group_spec.kv_cache_spec
            if not isinstance(kv_cache_spec, AttentionSpec):
                raise NotImplementedError(
                    "Only AttentionSpec is supported for now.")
            attn_backend_i = get_attn_backend(
                kv_cache_spec.head_size,
                self.dtype,
                kv_cache_spec.dtype,
                kv_cache_spec.block_size,
                self.model_config.is_attention_free,
                use_mla=kv_cache_spec.use_mla,
            )
            if attn_backend_i is None:
                error_msg = (
                    f"Error with get_attn_backend: {kv_cache_spec.head_size=}, "
                    f"{self.dtype=}, {kv_cache_spec.dtype=}, "
                    f"{kv_cache_spec.block_size=}, "
                    f"{self.model_config.is_attention_free=}, "
                    f"{kv_cache_spec.use_mla=}")
                logger.error(error_msg)
                raise NotImplementedError(
                    "Non-Attention backend is not supported by V1 "
                    "GPUModelRunner.")

            if self.vllm_config.compilation_config.full_cuda_graph:
                attn_backend_name = attn_backend_i.__name__
                flash_attn_version = get_flash_attn_version()
                if attn_backend_name != "FlashAttentionBackend" or \
                    flash_attn_version != 3:
                    raise ValueError(
                        f"full_cuda_graph is only supported with "
                        f"FA3. Current attention backend is "
                        f"{attn_backend_name}, FlashAttention version is "
                        f"{flash_attn_version}.")

            block_table_i = self.input_batch.block_table[i]
            attn_metadata_builder_i = attn_backend_i.get_builder_cls()(
                weakref.proxy(self), kv_cache_spec, block_table_i)
            self.attn_backends.append(attn_backend_i)
            self.attn_metadata_builders.append(attn_metadata_builder_i)

    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
        """
        self.kv_cache_config = 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(),
            kv_cache_config=kv_cache_config,
        )
        self.initialize_attn_backend(kv_cache_config)

        kv_caches: dict[str, torch.Tensor] = {}

        for i, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            kv_cache_spec = kv_cache_group.kv_cache_spec
            for layer_name in kv_cache_group.layer_names:
                tensor_config = kv_cache_config.tensors[layer_name]
                assert tensor_config.size % kv_cache_spec.page_size_bytes == 0
                num_blocks = tensor_config.size // kv_cache_spec.page_size_bytes
                # `num_blocks` is the number of blocks the model runner can use.
                # `kv_cache_config.num_blocks` is the number of blocks that
                # KVCacheManager may allocate.
                # Since different GPUs may have different number of layers and
                # different memory capacities, `num_blocks` can be different on
                # different GPUs, and `kv_cache_config.num_blocks` is set to
                # the min of all `num_blocks`. Verify it here.
                assert num_blocks >= kv_cache_config.num_blocks
                if isinstance(kv_cache_spec, AttentionSpec):
                    kv_cache_shape = self.attn_backends[i].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
                    kv_caches[layer_name] = torch.zeros(kv_cache_shape,
                                                        dtype=dtype,
                                                        device=self.device)
                else:
                    # TODO: add new branches when introducing more types of
                    # KV cache specs.
                    raise ValueError("Unknown KV cache spec type.")

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

        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)

    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.
        """

        layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        block_size = self.vllm_config.cache_config.block_size
        use_mla = self.vllm_config.model_config.use_mla
        kv_cache_spec: dict[str, KVCacheSpec] = {}
        for layer_name, attn_module in layers.items():
            # TODO: Support other attention modules, e.g., cross-attention
            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)
                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}")

        return kv_cache_spec