gpu_model_runner.py

# SPDX-License-Identifier: Apache-2.0

import gc
import time
from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union

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

from vllm.attention.backends.abstract import AttentionType
from vllm.attention.layer import Attention
from vllm.config import CompilationLevel, VllmConfig
from vllm.distributed.parallel_state import get_pp_group, graph_capture
from vllm.forward_context import set_forward_context
from vllm.inputs import INPUT_REGISTRY
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, MultiModalKwargs
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,
                        LayerBlockType, cdiv, is_pin_memory_available)
from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
                                                   FlashAttentionMetadata)
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
from vllm.v1.engine.mm_input_cache import MMInputCacheClient
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                        KVCacheSpec)
from vllm.v1.outputs import LogprobsTensors, ModelRunnerOutput
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.rejection_sampler import INVALID_TOKEN_ID
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.utils import bind_kv_cache
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin

if TYPE_CHECKING:
    from vllm.v1.core.scheduler_output import SchedulerOutput

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

        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.sliding_window = model_config.get_sliding_window()
        self.block_size = cache_config.block_size
        self.max_model_len = model_config.max_model_len
        self.max_num_blocks_per_req = cdiv(self.max_model_len, self.block_size)
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
        self.max_num_reqs = scheduler_config.max_num_seqs

        # Model-related.
        self.num_attn_layers = model_config.get_num_layers_by_block_type(
            parallel_config, LayerBlockType.attention)
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
        self.num_kv_heads = model_config.get_num_kv_heads(parallel_config)
        self.head_size = model_config.get_head_size()
        self.hidden_size = model_config.get_hidden_size()

        # Multi-modal data support
        self.input_registry = INPUT_REGISTRY
        self.mm_registry = MULTIMODAL_REGISTRY
        self.uses_mrope = model_config.uses_mrope

        if self.is_multimodal_model:
            # NOTE: Initialized client is only used for processing dummy
            # multimodal data into multimodal kwargs for GPU memory profiling.
            # Only applicable to multimodal models with legacy input mapper.
            self.mm_input_mapper_profiling = MMInputCacheClient(
                self.model_config)
            self.mm_input_mapper_profiling.use_cache = False

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

        # Lazy initialization
        # self.model: nn.Module  # Set after load_model
        self.kv_caches: List[torch.Tensor] = []
        # req_id -> (input_id -> encoder_output)
        self.encoder_cache: Dict[str, Dict[int, torch.Tensor]] = {}

        # Set up speculative decoding.
        self.use_spec_decode = False
        if self.speculative_config:
            self.use_spec_decode = True

            # TODO: find a better way to check if we are using ngram.
            assert self.speculative_config.ngram_prompt_lookup_min, \
                    "Currently, only ngram spec decode is supported in V1."
            if get_pp_group().is_last_rank:
                self.drafter = NgramProposer()
                # Trigger Numba JIT compilation for N-gram proposer.
                # This usually takes less than 1 second.
                self.drafter.propose(
                    np.zeros(1024, dtype=np.int32),
                    self.speculative_config.ngram_prompt_lookup_min,
                    self.speculative_config.num_speculative_tokens,
                )

        # Request states.
        self.requests: Dict[str, CachedRequestState] = {}
        # Persistent batch.
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
            max_model_len=self.max_model_len,
            max_num_blocks_per_req=self.max_num_blocks_per_req,
            device=self.device,
            pin_memory=self.pin_memory,
            vocab_size=model_config.get_vocab_size(),
        )

        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)
        # 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)

        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.
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
                                       self.max_model_len,
                                       self.max_num_tokens),
                                   dtype=np.int32)
        # 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.input_ids_np = self.input_ids_cpu.numpy()
        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.slot_mapping_cpu = torch.zeros(self.max_num_tokens,
                                            dtype=torch.int32,
                                            device="cpu",
                                            pin_memory=self.pin_memory)
        self.slot_mapping_np = self.slot_mapping_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 _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,
                prompt=new_req_data.prompt,
                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 = []
                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"])

                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,
                    )

            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.
                req_state.block_ids.extend(req_data.new_block_ids)
            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)
            start_index = (len(req_state.block_ids) -
                           len(req_data.new_block_ids))
            self.input_batch.block_table.append_row(req_index, start_index,
                                                    req_data.new_block_ids)
            # 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 = sorted(removed_req_indices, 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)

        if batch_changed:
            self.input_batch.refresh_sampling_metadata()

    def _prepare_inputs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> Tuple[FlashAttentionMetadata, torch.Tensor]:
        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.
        # TODO: The Python loop can be slow. Optimize.
        num_scheduled_tokens = np.empty(num_reqs, dtype=np.int32)
        max_num_scheduled_tokens = 0
        for i, req_id in enumerate(self.input_batch.req_ids):
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
            num_scheduled_tokens[i] = num_tokens
            max_num_scheduled_tokens = max(max_num_scheduled_tokens,
                                           num_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.
        # 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 * self.max_num_blocks_per_req +
                               positions_np // self.block_size)
        # 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.
        block_table_cpu = self.input_batch.block_table.get_cpu_tensor()
        block_numbers = block_table_cpu.flatten()[block_table_indices].numpy()
        block_offsets = positions_np % self.block_size
        np.add(block_numbers * self.block_size,
               block_offsets,
               out=self.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)
        max_seq_len = self.seq_lens_np[:num_reqs].max()

        # 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)
        query_start_loc = self.query_start_loc_cpu[:num_reqs + 1].to(
            self.device, non_blocking=True)
        seq_lens = self.seq_lens_cpu[:num_reqs].to(self.device,
                                                   non_blocking=True)
        slot_mapping = self.slot_mapping_cpu[:total_num_scheduled_tokens].to(
            self.device, non_blocking=True).long()

        # Prepare for cascade attention if needed.
        common_prefix_len = self._compute_cascade_attn_prefix_len(
            num_scheduled_tokens,
            scheduler_output.num_common_prefix_blocks,
        )
        use_cascade = common_prefix_len > 0
        if use_cascade:
            # TODO: Optimize.
            cu_prefix_query_lens = torch.tensor(
                [0, total_num_scheduled_tokens],
                dtype=torch.int32,
                device=self.device)
            prefix_kv_lens = torch.tensor([common_prefix_len],
                                          dtype=torch.int32,
                                          device=self.device)
            suffix_kv_lens = (self.seq_lens_np[:num_reqs] - common_prefix_len)
            suffix_kv_lens = torch.from_numpy(suffix_kv_lens).to(self.device)
        else:
            cu_prefix_query_lens = None
            prefix_kv_lens = None
            suffix_kv_lens = None

        attn_metadata = FlashAttentionMetadata(
            num_actual_tokens=total_num_scheduled_tokens,
            max_query_len=max_num_scheduled_tokens,
            query_start_loc=query_start_loc,
            max_seq_len=max_seq_len,
            seq_lens=seq_lens,
            block_table=(
                self.input_batch.block_table.get_device_tensor()[:num_reqs]),
            slot_mapping=slot_mapping,
            use_cascade=use_cascade,
            common_prefix_len=common_prefix_len,
            cu_prefix_query_lens=cu_prefix_query_lens,
            prefix_kv_lens=prefix_kv_lens,
            suffix_kv_lens=suffix_kv_lens,
        )

        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
        if use_spec_decode:
            logits_indices = self._calc_spec_decode_metadata(
                scheduler_output, cu_num_tokens)
        else:
            # 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

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

        return attn_metadata, logits_indices

    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
    ) -> 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 * self.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: 4 (i.e., [A, B, C, D])
        # 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 // self.block_size *
                             self.block_size)
        use_cascade = FlashAttentionBackend.use_cascade_attention(
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
            num_kv_heads=self.num_kv_heads,
            use_alibi=False,  # FIXME
            use_sliding_window=self.sliding_window is not None,
            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,
        scheduler_output: "SchedulerOutput",
        cu_num_tokens: np.ndarray,
    ) -> torch.Tensor:
        # Get the number of spec decode tokens for each request.
        num_reqs = self.input_batch.num_reqs
        num_spec_decode_tokens = np.empty(num_reqs, dtype=np.int32)
        for i, req_id in enumerate(self.input_batch.req_ids):
            num_spec_decode_tokens[i] = len(
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, ()))

        # Get spec decode logits indices.
        # E.g.,   num_scheduled_tokens: [4, 100, 3,   100, 2]
        #         cu_num_tokens:        [4, 104, 107, 207, 209]
        #         num_spec_tokens_list: [3, 0,   2,   0,   1]
        #         num_sampled_tokens:   [4, 1,   3,   1,   2]
        #         spec_decode_logits_indices:
        #                 [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
        num_sampled_tokens = num_spec_decode_tokens + 1
        # logits_start_loc: [0, 103, 104, 206, 207]
        logits_start_loc = cu_num_tokens - num_sampled_tokens
        # [0, 103, 104, 206, 207] ->
        #               [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
        logits_start_loc = np.repeat(logits_start_loc, num_sampled_tokens)
        # The following three lines:
        # [4, 1,   3,   1,   2] -> [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        # Step 1. [4, 1, 3, 1, 2] -> [4, 5, 8, 9, 11]
        cu_num_sampled_tokens = np.cumsum(num_sampled_tokens)
        # Step 2. [4, 5, 8, 9, 11] -> [0, 4, 5, 8, 9]
        #         -> [0, 0, 0, 0, 4, 5, 5, 5, 8, 9, 9]
        cumsums_sampled_offsets = np.repeat(
            cu_num_sampled_tokens - num_sampled_tokens, num_sampled_tokens)
        # Step 3.  [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
        #       -  [0, 0, 0, 0, 4, 5, 5, 5, 8, 9, 9]
        #      -> [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        total_num_sampled_tokens = num_sampled_tokens.sum()
        sampled_arange = (self.arange_np[:total_num_sampled_tokens] -
                          cumsums_sampled_offsets)

        # [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207] ->
        # [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
        spec_decode_logits_indices = logits_start_loc + sampled_arange
        return torch.from_numpy(spec_decode_logits_indices).to(
            self.device, non_blocking=True)

    def _execute_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_input_ids: List[Tuple[str, int]] = []
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
            for input_id in encoder_input_ids:
                mm_inputs.append(req_state.mm_inputs[input_id])
                req_input_ids.append((req_id, 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)

            for output in curr_group_outputs:
                encoder_outputs.append(output)

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

    def _gather_encoder_outputs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> List[torch.Tensor]:
        encoder_outputs: 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]
                encoder_outputs.append(encoder_output[start_idx:end_idx])
        return encoder_outputs

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

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

        if self.is_multimodal_model:
            # Run the multimodal encoder if any.
            self._execute_encoder(scheduler_output)
            encoder_outputs = self._gather_encoder_outputs(scheduler_output)
        else:
            encoder_outputs = []

        # Prepare the decoder inputs.
        attn_metadata, logits_indices = 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.
            num_input_tokens = num_scheduled_tokens
        attn_metadata.num_input_tokens = num_input_tokens

        if self.is_multimodal_model:
            # 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 encoder_outputs:
                inputs_embeds = self.model.get_input_embeddings(
                    input_ids, encoder_outputs)
            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:
            assert intermediate_tensors is not None
            assert self.intermediate_tensors is not None
            for k, v in intermediate_tensors.items():
                self.intermediate_tensors[k][:num_input_tokens].copy_(
                    v[:num_input_tokens], non_blocking=True)
            intermediate_tensors = IntermediateTensors({
                k: v[:num_input_tokens]
                for k, v in self.intermediate_tensors.items()
            })

        # Run the decoder.
        # Use persistent buffers for CUDA graphs.
        with set_forward_context(attn_metadata, self.vllm_config):
            hidden_states = self.model(
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
            )
        if not get_pp_group().is_last_rank:
            # For mid-pipeline stages, return the hidden states.
            return hidden_states

        hidden_states = hidden_states[:num_scheduled_tokens]
        sample_hidden_states = hidden_states[logits_indices]
        logits = self.model.compute_logits(sample_hidden_states, None)

        # Sample the next token and get logprobs if needed.
        sampling_metadata = self.input_batch.get_sampling_metadata(
            scheduler_output.scheduled_spec_decode_tokens)
        sampler_output = self.model.sample(
            logits=logits,
            sampling_metadata=sampling_metadata,
        )

        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
        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.
                # Rewind the generator state as if the token was not sampled.
                generator = self.input_batch.generators.get(i)
                if generator is not None:
                    # This relies on cuda-specific torch-internal impl details
                    generator.set_offset(generator.get_offset() - 4)

        # 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,
            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_mask = sampled_token_ids != INVALID_TOKEN_ID
            gen_lens = valid_mask.sum(dim=1).tolist()
            # TODO(woosuk): Optimize this.
            valid_sampled_token_ids = [
                seq.tolist()
                for seq in sampled_token_ids[valid_mask].split(gen_lens)
            ]

        if not self.use_spec_decode:
            spec_token_ids = None
        else:
            spec_token_ids = self.generate_draft_token_ids(
                valid_sampled_token_ids)

        model_runner_output = 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,
        )
        return model_runner_output

    def generate_draft_token_ids(
        self,
        sampled_token_ids: List[List[int]],
    ) -> 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

            # 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
            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],
                self.speculative_config.ngram_prompt_lookup_min,
                self.speculative_config.num_speculative_tokens,
            )
            if drafter_output is None or len(drafter_output) == 0:
                draft_token_ids.append([])
            else:
                draft_token_ids.append(drafter_output.tolist())
        return draft_token_ids

    def 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)
            time_after_load = time.perf_counter()
        self.model_memory_usage = m.consumed_memory
        logger.info("Loading model weights took %.4f GB and %.6f seconds",
                    self.model_memory_usage / float(2**30),
                    time_after_load - time_before_load)

    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        scheduler_output: "SchedulerOutput",
    ) -> Dict[str, LogprobsTensors]:
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

        prompt_logprobs_dict: Dict[str, 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)

            # Determine number of logits to retrieve.
            start_tok = request.num_computed_tokens + 1
            num_remaining_tokens = num_prompt_tokens - start_tok
            if num_tokens < num_remaining_tokens:
                # This is a chunk, more tokens remain.
                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)

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

            # Transfer GPU->CPU async.
            prompt_logprobs_dict[req_id] = LogprobsTensors(
                token_ids.to("cpu", non_blocking=True),
                logprobs.to("cpu", non_blocking=True),
                ranks.to("cpu", 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]

        # Must synchronize the non-blocking GPU->CPU transfers.
        torch.cuda.synchronize()

        return prompt_logprobs_dict

    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
    ) -> torch.Tensor:
        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 = IntermediateTensors({
                k: v[:num_tokens]
                for k, v in self.intermediate_tensors.items()
            })

        with set_forward_context(None, self.vllm_config,
                                 num_tokens=num_tokens):
            hidden_states = model(
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
            )
        return hidden_states

    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 = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality(  # noqa: E501
                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_request_data = self.input_registry.dummy_data_for_profiling(
                model_config=self.model_config,
                seq_len=self.max_num_tokens,
                mm_registry=self.mm_registry,
            )
            dummy_mm_data = dummy_request_data.multi_modal_data

            # Dummy data definition in V0 may contain multiple multimodal items
            # (e.g, multiple images) for a single request, therefore here we
            # always replicate first item by max_num_mm_items times since in V1
            # they are scheduled to be processed separately.

            # Case when models have a merged processor, their dummy data is
            # already batched `MultiModalKwargs`, therefore we take the first
            # `MultiModalKwargsItem` from the desired modality to profile on.
            if isinstance(dummy_mm_data, MultiModalKwargs):
                dummy_mm_item = dummy_mm_data.get_item(
                    modality=dummy_data_modality, item_index=0)
                dummy_mm_kwargs = MultiModalKwargs.from_items([dummy_mm_item])

            # Case when models have dummy data explicitly defined as
            # `MultiModalDataDict`, so they need to be processed through input
            # mapper.
            # TODO (ywang96): deprecate this path once merged processor is
            # supported on all models.
            else:
                mm_kwargs_list = self.mm_input_mapper_profiling.process_inputs(
                    mm_data=dummy_mm_data,
                    mm_hashes=None,
                    mm_processor_kwargs=None,
                    precomputed_mm_inputs=None)
                dummy_mm_kwargs = mm_kwargs_list[0]

            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)
            assert len(dummy_encoder_outputs) == max_num_mm_items, (
                "Expected dimension 0 of encoder outputs to match the number "
                f"of multimodal data items: {max_num_mm_items}, got "
                f"{len(dummy_encoder_outputs)=} instead. This is most likely "
                "due to the 'get_multimodal_embeddings' method of the model "
                "not implemented correctly.")

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

        # For profile, have maximum num_reqs and that collectively have
        # maximum num_tokens.
        num_reqs = self.scheduler_config.max_num_seqs
        num_tokens = self.max_num_tokens
        min_tokens_per_req: int = num_tokens // num_reqs

        num_scheduled_tokens_list: List[int] = [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.ndarray = np.array(num_scheduled_tokens_list,
                                                    dtype=np.int32)
        logit_indices = np.cumsum(num_scheduled_tokens) - 1

        with self.maybe_profile_with_lora(self.lora_config,
                                          num_scheduled_tokens):
            # Trigger compilation for general shape.
            hidden_states = self._dummy_run(self.max_num_tokens)
            if get_pp_group().is_last_rank:
                hidden_states = hidden_states[logit_indices]
                logits = self.model.compute_logits(hidden_states, None)
                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,
                    spec_token_ids=None,
                    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=torch.ones_like(logits,
                                                     dtype=torch.int64),
                    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,
                )
                sampler_output = self.model.sample(
                    logits=logits, sampling_metadata=dummy_metadata)
            else:
                logits = None
                sampler_output = None
                dummy_metadata = None
            torch.cuda.synchronize()
            del hidden_states, logits, sampler_output, dummy_metadata
            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):
            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)
                self._dummy_run(num_tokens)

        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_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
        """
        if len(kv_cache_config.groups) > 1:
            raise NotImplementedError(
                "Hybrid models with more than one KV cache type are not "
                "supported yet.")

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

        for layer_name, layer_spec in kv_cache_config.kv_cache_spec.items():
            tensor_config = kv_cache_config.tensors[layer_name]
            assert tensor_config.size % layer_spec.page_size_bytes == 0
            num_blocks = tensor_config.size // layer_spec.page_size_bytes
            if isinstance(layer_spec, FullAttentionSpec):
                kv_cache_shape = FlashAttentionBackend.get_kv_cache_shape(
                    num_blocks, layer_spec.block_size, layer_spec.num_kv_heads,
                    layer_spec.head_size)
                dtype = layer_spec.dtype
                kv_caches[layer_name] = torch.zeros(kv_cache_shape,
                                                    dtype=dtype,
                                                    device=self.device)
            else:
                raise NotImplementedError

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

    def get_kv_cache_spec(self) -> 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.
        """

        forward_ctx = self.vllm_config.compilation_config.static_forward_context
        block_size = self.vllm_config.cache_config.block_size
        kv_cache_spec: KVCacheSpec = {}
        for layer_name, attn_module in forward_ctx.items():
            # TODO: Support other attention modules, e.g., sliding window,
            # cross-attention, MLA.
            assert isinstance(attn_module, Attention)
            if attn_module.attn_type == AttentionType.DECODER:
                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=attn_module.dtype,
                )
            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