gpu_model_runner.py

import gc
import time
from typing import TYPE_CHECKING, Dict, List, Tuple, cast

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

from vllm.config import CompilationLevel, VllmConfig
from vllm.distributed.parallel_state import 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.model_loader import get_model
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
from vllm.sampling_params import SamplingType
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.engine.mm_input_mapper import MMHasher, MMInputMapperClient
from vllm.v1.outputs import ModelRunnerOutput
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch

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

logger = init_logger(__name__)


class GPUModelRunner:

    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

        # NOTE: mm_input_mapper_client and mm_hasher are only used for memory
        # profiling.
        self.mm_input_mapper_client = MMInputMapperClient(self.model_config)
        self.mm_hasher = MMHasher()
        self.use_hash = (not model_config.disable_mm_preprocessor_cache) or \
            cache_config.enable_prefix_caching

        self.max_num_encoder_input_tokens = self.scheduler_config.max_num_encoder_input_tokens  # noqa: E501
        self.encoder_cache_size = self.scheduler_config.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]] = {}

        # 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.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.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),
                                   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_start_loc_cpu = torch.zeros(self.max_num_reqs + 1,
                                             dtype=torch.int32,
                                             device="cpu",
                                             pin_memory=self.pin_memory)
        self.seq_start_loc_np = self.seq_start_loc_cpu.numpy()

    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
        # Remove stopped requests from the cached states.
        # Keep the states of the pre-empted requests.
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
            self.encoder_cache.pop(req_id, None)

        # 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 requests from the persistent batch.
        stopped_req_ids = set().union(
            scheduler_output.preempted_req_ids,
            scheduler_output.finished_req_ids,
        )
        removed_req_indices: List[int] = []
        for req_id in stopped_req_ids:
            req_index = self.input_batch.remove_request(req_id)
            if req_index is not None:
                removed_req_indices.append(req_index)

        # Update the states of the running requests.
        for req_data in scheduler_output.scheduled_running_reqs:
            req_id = req_data.req_id
            req_state = self.requests[req_id]
            req_index = self.input_batch.req_id_to_index[req_id]

            # Update the num_computed_tokens.
            req_state.num_computed_tokens = req_data.num_computed_tokens
            self.input_batch.num_computed_tokens_cpu[req_index] = (
                req_data.num_computed_tokens)

            # Update the block table.
            num_new_blocks = len(req_data.new_block_ids)
            if num_new_blocks == 0:
                continue
            start_index = len(req_state.block_ids)
            req_state.block_ids.extend(req_data.new_block_ids)
            self.input_batch.block_table.append_row(req_index, start_index,
                                                    req_data.new_block_ids)

        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=[],
            )
            req_ids_to_add.append(req_id)

        # Update the cached states of the resumed requests.
        for res_req_data in scheduler_output.scheduled_resumed_reqs:
            req_id = res_req_data.req_id
            req_state = self.requests[req_id]

            req_state.block_ids = res_req_data.block_ids
            req_state.num_computed_tokens = res_req_data.num_computed_tokens
            req_ids_to_add.append(req_id)

        # 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)

    def _prepare_inputs(self, scheduler_output: "SchedulerOutput"):
        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 = []
        max_num_scheduled_tokens = 0
        for req_id in self.input_batch.req_ids[:num_reqs]:
            assert req_id is not None
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
            num_scheduled_tokens.append(num_tokens)
            max_num_scheduled_tokens = max(max_num_scheduled_tokens,
                                           num_tokens)
        num_scheduled_tokens = np.array(num_scheduled_tokens, dtype=np.int32)
        assert max_num_scheduled_tokens > 0

        # 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]
        arange = np.concatenate(
            [self.arange_np[:n] for n in num_scheduled_tokens])

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

        # 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
        np.cumsum(num_scheduled_tokens,
                  out=self.query_start_loc_np[1:num_reqs + 1])

        seq_lens = (self.input_batch.num_computed_tokens_cpu[:num_reqs] +
                    num_scheduled_tokens)
        max_seq_len = seq_lens.max()
        self.seq_start_loc_np[0] = 0
        np.cumsum(seq_lens, out=self.seq_start_loc_np[1:num_reqs + 1])

        # 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)
        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_start_loc = self.seq_start_loc_cpu[:num_reqs + 1].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 = (scheduler_output.num_common_prefix_blocks *
                             self.block_size)
        if common_prefix_len == 0:
            # Common case.
            use_cascade = False
        else:
            # 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.
            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,
            )

        if use_cascade:
            # TODO: Optimize.
            cu_prefix_query_lens = torch.tensor(
                [0, total_num_scheduled_tokens],
                dtype=torch.int32,
                device=self.device)
            cu_prefix_kv_lens = torch.tensor([0, common_prefix_len],
                                             dtype=torch.int32,
                                             device=self.device)
            cu_suffix_kv_lens = (
                self.seq_start_loc_np[:num_reqs + 1] -
                self.arange_np[:num_reqs + 1] * common_prefix_len)
            cu_suffix_kv_lens = torch.from_numpy(cu_suffix_kv_lens).to(
                self.device)
        else:
            cu_prefix_query_lens = None
            cu_prefix_kv_lens = None
            cu_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_start_loc=seq_start_loc,
            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,
            cu_prefix_kv_lens=cu_prefix_kv_lens,
            cu_suffix_kv_lens=cu_suffix_kv_lens,
        )
        # NOTE(woosuk): Due to chunked prefills, there can be at most 1 partial
        # request in the batch. While we should not sample any token from this
        # partial request, we do so for simplicity. We will ignore the sampled
        # token from the partial request.
        # TODO: Support prompt logprobs.
        logits_indices = query_start_loc[1:] - 1
        return attn_metadata, logits_indices

    def _prepare_sampling(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> SamplingMetadata:
        skip_copy = True
        if (scheduler_output.finished_req_ids
                or scheduler_output.preempted_req_ids):
            skip_copy = False
        if (scheduler_output.scheduled_new_reqs
                or scheduler_output.scheduled_resumed_reqs):
            skip_copy = False
        # Create the sampling metadata.
        req_id_output_token_ids: Dict[str, List[int]] = \
            {req_id: req.output_token_ids \
                for req_id, req in self.requests.items()}

        sampling_metadata = self.input_batch.make_sampling_metadata(
            req_id_output_token_ids, skip_copy)
        return sampling_metadata

    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))
        batched_mm_inputs = MultiModalKwargs.batch(mm_inputs)
        batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs,
                                                       device=self.device)

        # Run the encoder.
        # `encoder_outputs` is either of the following:
        # 1. A tensor of shape [num_images, feature_size, hidden_size]
        # in case when feature_size is fixed across all images.
        # 2. A list (length: num_images) of tensors, each of shape
        # [feature_size, hidden_size] in case when the feature size is
        # dynamic depending on input images.
        encoder_outputs = self.model.get_multimodal_embeddings(
            **batched_mm_inputs)

        # 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] = []
        num_reqs = self.input_batch.num_reqs
        for req_id in self.input_batch.req_ids[:num_reqs]:
            assert req_id is not None
            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

    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> ModelRunnerOutput:
        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

        # 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=self.positions[:num_input_tokens],
                kv_caches=self.kv_caches,
                attn_metadata=None,
                inputs_embeds=inputs_embeds,
            )
        hidden_states = hidden_states[:num_scheduled_tokens]
        hidden_states = hidden_states[logits_indices]
        logits = self.model.compute_logits(hidden_states, None)

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

        sampled_token_ids = sampler_output.sampled_token_ids
        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
        num_reqs = self.input_batch.num_reqs
        for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
            assert req_id is not None
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
            assert seq_len <= req_state.num_tokens
            if seq_len == req_state.num_tokens:
                # Append the sampled token to the output token ids.
                token_id = sampled_token_ids[i]
                self.input_batch.token_ids_cpu[i, seq_len] = token_id
                self.input_batch.num_tokens[i] += 1
                req_state.output_token_ids.append(token_id)
            else:
                # Ignore the sampled token from the partial request.
                # 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)

        if sampler_output.logprob_token_ids is None:
            logprob_token_ids = None
        else:
            logprob_token_ids = sampler_output.logprob_token_ids.cpu()
        if sampler_output.logprobs is None:
            logprobs = None
        else:
            logprobs = sampler_output.logprobs.cpu()

        # num_reqs entries should be non-None
        assert all(
            req_id is not None for req_id in
            self.input_batch.req_ids[:num_reqs]), "req_ids contains None"
        req_ids = cast(List[str], self.input_batch.req_ids[:num_reqs])

        model_runner_output = ModelRunnerOutput(
            req_ids=req_ids,
            req_id_to_index=self.input_batch.req_id_to_index,
            sampled_token_ids=sampled_token_ids,
            logprob_token_ids_cpu=logprob_token_ids,
            logprobs_cpu=logprobs,
        )
        return model_runner_output

    def load_model(self) -> None:
        logger.info("Starting to load model %s...", self.model_config.model)
        with DeviceMemoryProfiler() as m:  # noqa: SIM117
            self.model = get_model(vllm_config=self.vllm_config)

        self.model_memory_usage = m.consumed_memory
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))

    @torch.inference_mode()
    def _dummy_run(
        self,
        model: nn.Module,
        num_tokens: int,
        kv_caches: List[torch.Tensor],
    ) -> torch.Tensor:
        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
        with set_forward_context(None, self.vllm_config):
            hidden_states = model(
                input_ids=input_ids,
                positions=self.positions[:num_tokens],
                kv_caches=kv_caches,
                attn_metadata=None,
                inputs_embeds=inputs_embeds,
            )
        return hidden_states

    def profile_run(self) -> None:
        # use an empty tensor instead of `None`` to force Dynamo to pass
        # it by reference, rather by specializing on the value `None`.
        # the `dtype` argument does not matter, and we use `float32` as
        # a placeholder (it has wide hardware support).
        # it is important to create tensors inside the loop, rather than
        # multiplying the list, to avoid Dynamo from treating them as
        # tensor aliasing.
        dummy_kv_caches = [
            torch.tensor([], dtype=torch.float32, device=self.device)
            for _ in range(self.num_attn_layers)
        ]

        # Profile with multimodal encoder & encoder cache.
        # TODO (ywang96): generalize this beyond image modality since
        # mm_input_mapper only supports image inputs.
        if self.is_multimodal_model:

            # 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

            # NOTE: Currently model is profiled with a single non-text
            # modality even when it supports multiple.
            max_tokens_per_mm_item = max(
                self.mm_registry.get_max_tokens_per_item_by_modality(
                    self.model_config).values())

            max_num_mm_items_encoder_budget = min(
                self.max_num_encoder_input_tokens,
                self.encoder_cache_size) // max_tokens_per_mm_item

            max_mm_items_per_req = max(
                self.mm_registry.get_mm_limits_per_prompt(
                    self.model_config).values())

            # 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)

            # 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 need to "unbatch"
            # and take the first item in each batched tensor.
            # TODO (ywang96): This is somewhat hacky. Refactor this to be
            # consistent with the other case.
            if isinstance(dummy_mm_data, MultiModalKwargs):
                dummy_mm_kwargs = {
                    k: v[0].unsqueeze(0)
                    for k, v in dummy_mm_data.items()
                }

            # Case when models have dummy data explicitly defined as
            # `MultiModalDataDict`, so they need to be processed through input
            # mapper.
            else:
                # Compute MM hashes (if enabled)
                mm_hashes = None
                if self.use_hash:
                    mm_hashes = self.mm_hasher.hash_dummy_mm_data(
                        dummy_mm_data)

                mm_kwargs_list = self.mm_input_mapper_client.process_inputs(
                    mm_data=dummy_mm_data,
                    mm_hashes=mm_hashes,
                    mm_processor_kwargs=None,
                    precomputed_mm_inputs=None)

                # Take the first `MultiModalKwargs`
                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))

        # Trigger compilation for general shape.
        hidden_states = self._dummy_run(self.model, self.max_num_tokens,
                                        dummy_kv_caches)
        logits = self.model.compute_logits(hidden_states, None)
        logits = logits[:self.max_num_tokens]
        # TODO(woosuk): Consider the memory usage of the sampler.
        torch.cuda.synchronize()
        del hidden_states, logits
        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(self.model, num_tokens, self.kv_caches)
                self._dummy_run(self.model, num_tokens, self.kv_caches)

        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, num_blocks: int) -> None:
        assert len(self.kv_caches) == 0
        kv_cache_shape = FlashAttentionBackend.get_kv_cache_shape(
            num_blocks, self.block_size, self.num_kv_heads, self.head_size)
        for _ in range(self.num_attn_layers):
            self.kv_caches.append(
                torch.zeros(kv_cache_shape,
                            dtype=self.kv_cache_dtype,
                            device=self.device))