merge main

vllm/py.typed

+# Marker file for PEP 561.
+# The vllm package uses inline types.

vllm/sampling_params.py

+"""Sampling parameters for text generation."""
+from enum import IntEnum
+from functools import cached_property
+from typing import Callable, List, Optional, Union
+
+import torch
+
+_SAMPLING_EPS = 1e-5
+
+
+class SamplingType(IntEnum):
+    GREEDY = 0
+    RANDOM = 1
+    BEAM = 2
+
+
+LogitsProcessor = Callable[[List[int], torch.Tensor], torch.Tensor]
+"""LogitsProcessor is a function that takes a list of previously generated
+tokens and a tensor of the logits for the next token, and returns a modified
+tensor of logits to sample from."""
+
+
+class SamplingParams:
+    """Sampling parameters for text generation.
+
+    Overall, we follow the sampling parameters from the OpenAI text completion
+    API (https://platform.openai.com/docs/api-reference/completions/create).
+    In addition, we support beam search, which is not supported by OpenAI.
+
+    Args:
+        n: Number of output sequences to return for the given prompt.
+        best_of: Number of output sequences that are generated from the prompt.
+            From these `best_of` sequences, the top `n` sequences are returned.
+            `best_of` must be greater than or equal to `n`. This is treated as
+            the beam width when `use_beam_search` is True. By default, `best_of`
+            is set to `n`.
+        presence_penalty: Float that penalizes new tokens based on whether they
+            appear in the generated text so far. Values > 0 encourage the model
+            to use new tokens, while values < 0 encourage the model to repeat
+            tokens.
+        frequency_penalty: Float that penalizes new tokens based on their
+            frequency in the generated text so far. Values > 0 encourage the
+            model to use new tokens, while values < 0 encourage the model to
+            repeat tokens.
+        repetition_penalty: Float that penalizes new tokens based on whether
+            they appear in the prompt and the generated text so far. Values > 1
+            encourage the model to use new tokens, while values < 1 encourage
+            the model to repeat tokens.
+        temperature: Float that controls the randomness of the sampling. Lower
+            values make the model more deterministic, while higher values make
+            the model more random. Zero means greedy sampling.
+        top_p: Float that controls the cumulative probability of the top tokens
+            to consider. Must be in (0, 1]. Set to 1 to consider all tokens.
+        top_k: Integer that controls the number of top tokens to consider. Set
+            to -1 to consider all tokens.
+        min_p: Float that represents the minimum probability for a token to be
+            considered, relative to the probability of the most likely token.
+            Must be in [0, 1]. Set to 0 to disable this.
+        use_beam_search: Whether to use beam search instead of sampling.
+        length_penalty: Float that penalizes sequences based on their length.
+            Used in beam search.
+        early_stopping: Controls the stopping condition for beam search. It
+            accepts the following values: `True`, where the generation stops as
+            soon as there are `best_of` complete candidates; `False`, where an
+            heuristic is applied and the generation stops when is it very
+            unlikely to find better candidates; `"never"`, where the beam search
+            procedure only stops when there cannot be better candidates
+            (canonical beam search algorithm).
+        stop: List of strings that stop the generation when they are generated.
+            The returned output will not contain the stop strings.
+        stop_token_ids: List of tokens that stop the generation when they are
+            generated. The returned output will contain the stop tokens unless
+            the stop tokens are special tokens.
+        include_stop_str_in_output: Whether to include the stop strings in output
+            text. Defaults to False.
+        ignore_eos: Whether to ignore the EOS token and continue generating
+            tokens after the EOS token is generated.
+        max_tokens: Maximum number of tokens to generate per output sequence.
+        logprobs: Number of log probabilities to return per output token.
+            Note that the implementation follows the OpenAI API: The return
+            result includes the log probabilities on the `logprobs` most likely
+            tokens, as well the chosen tokens. The API will always return the
+            log probability of the sampled token, so there  may be up to
+            `logprobs+1` elements in the response.
+        prompt_logprobs: Number of log probabilities to return per prompt token.
+        skip_special_tokens: Whether to skip special tokens in the output.
+        spaces_between_special_tokens: Whether to add spaces between special
+            tokens in the output.  Defaults to True.
+        logits_processors: List of functions that modify logits based on
+            previously generated tokens.
+    """
+
+    def __init__(
+        self,
+        n: int = 1,
+        best_of: Optional[int] = None,
+        presence_penalty: float = 0.0,
+        frequency_penalty: float = 0.0,
+        repetition_penalty: float = 1.0,
+        temperature: float = 1.0,
+        top_p: float = 1.0,
+        top_k: int = -1,
+        min_p: int = 0.0,
+        use_beam_search: bool = False,
+        length_penalty: float = 1.0,
+        early_stopping: Union[bool, str] = False,
+        stop: Optional[Union[str, List[str]]] = None,
+        stop_token_ids: Optional[List[int]] = None,
+        include_stop_str_in_output: bool = False,
+        ignore_eos: bool = False,
+        max_tokens: int = 16,
+        logprobs: Optional[int] = None,
+        prompt_logprobs: Optional[int] = None,
+        skip_special_tokens: bool = True,
+        spaces_between_special_tokens: bool = True,
+        logits_processors: Optional[List[LogitsProcessor]] = None,
+    ) -> None:
+        self.n = n
+        self.best_of = best_of if best_of is not None else n
+        self.presence_penalty = presence_penalty
+        self.frequency_penalty = frequency_penalty
+        self.repetition_penalty = repetition_penalty
+        self.temperature = temperature
+        self.top_p = top_p
+        self.top_k = top_k
+        self.min_p = min_p
+        self.use_beam_search = use_beam_search
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        if stop is None:
+            self.stop = []
+        elif isinstance(stop, str):
+            self.stop = [stop]
+        else:
+            self.stop = list(stop)
+        if stop_token_ids is None:
+            self.stop_token_ids = []
+        else:
+            self.stop_token_ids = list(stop_token_ids)
+        self.ignore_eos = ignore_eos
+        self.max_tokens = max_tokens
+        self.logprobs = logprobs
+        self.prompt_logprobs = prompt_logprobs
+        self.skip_special_tokens = skip_special_tokens
+        self.spaces_between_special_tokens = spaces_between_special_tokens
+        self.logits_processors = logits_processors
+        self.include_stop_str_in_output = include_stop_str_in_output
+        self._verify_args()
+        if self.use_beam_search:
+            self._verify_beam_search()
+        else:
+            self._verify_non_beam_search()
+            if self.temperature < _SAMPLING_EPS:
+                # Zero temperature means greedy sampling.
+                self.top_p = 1.0
+                self.top_k = -1
+                self.min_p = 0.0
+                self._verify_greedy_sampling()
+
+    def _verify_args(self) -> None:
+        if self.n < 1:
+            raise ValueError(f"n must be at least 1, got {self.n}.")
+        if self.best_of < self.n:
+            raise ValueError(f"best_of must be greater than or equal to n, "
+                             f"got n={self.n} and best_of={self.best_of}.")
+        if not -2.0 <= self.presence_penalty <= 2.0:
+            raise ValueError("presence_penalty must be in [-2, 2], got "
+                             f"{self.presence_penalty}.")
+        if not -2.0 <= self.frequency_penalty <= 2.0:
+            raise ValueError("frequency_penalty must be in [-2, 2], got "
+                             f"{self.frequency_penalty}.")
+        if not 0.0 < self.repetition_penalty <= 2.0:
+            raise ValueError("repetition_penalty must be in (0, 2], got "
+                             f"{self.repetition_penalty}.")
+        if self.temperature < 0.0:
+            raise ValueError(
+                f"temperature must be non-negative, got {self.temperature}.")
+        if not 0.0 < self.top_p <= 1.0:
+            raise ValueError(f"top_p must be in (0, 1], got {self.top_p}.")
+        if self.top_k < -1 or self.top_k == 0:
+            raise ValueError(f"top_k must be -1 (disable), or at least 1, "
+                             f"got {self.top_k}.")
+        if not 0.0 <= self.min_p <= 1.0:
+            raise ValueError("min_p must be in [0, 1], got "
+                             f"{self.min_p}.")
+        if self.max_tokens < 1:
+            raise ValueError(
+                f"max_tokens must be at least 1, got {self.max_tokens}.")
+        if self.logprobs is not None and self.logprobs < 0:
+            raise ValueError(
+                f"logprobs must be non-negative, got {self.logprobs}.")
+        if self.prompt_logprobs is not None and self.prompt_logprobs < 0:
+            raise ValueError(f"prompt_logprobs must be non-negative, got "
+                             f"{self.prompt_logprobs}.")
+
+    def _verify_beam_search(self) -> None:
+        if self.best_of == 1:
+            raise ValueError("best_of must be greater than 1 when using beam "
+                             f"search. Got {self.best_of}.")
+        if self.temperature > _SAMPLING_EPS:
+            raise ValueError("temperature must be 0 when using beam search.")
+        if self.top_p < 1.0 - _SAMPLING_EPS:
+            raise ValueError("top_p must be 1 when using beam search.")
+        if self.top_k != -1:
+            raise ValueError("top_k must be -1 when using beam search.")
+        if self.early_stopping not in [True, False, "never"]:
+            raise ValueError(
+                f"early_stopping must be True, False, or 'never', "
+                f"got {self.early_stopping}.")
+
+    def _verify_non_beam_search(self) -> None:
+        if self.early_stopping is not False:
+            raise ValueError("early_stopping is not effective and must be "
+                             "False when not using beam search.")
+        if (self.length_penalty < 1.0 - _SAMPLING_EPS
+                or self.length_penalty > 1.0 + _SAMPLING_EPS):
+            raise ValueError(
+                "length_penalty is not effective and must be the "
+                "default value of 1.0 when not using beam search.")
+
+    def _verify_greedy_sampling(self) -> None:
+        if self.best_of > 1:
+            raise ValueError("best_of must be 1 when using greedy sampling."
+                             f"Got {self.best_of}.")
+
+    @cached_property
+    def sampling_type(self) -> SamplingType:
+        if self.use_beam_search:
+            return SamplingType.BEAM
+        if self.temperature < _SAMPLING_EPS:
+            return SamplingType.GREEDY
+        return SamplingType.RANDOM
+
+    def __repr__(self) -> str:
+        return (
+            f"SamplingParams(n={self.n}, "
+            f"best_of={self.best_of}, "
+            f"presence_penalty={self.presence_penalty}, "
+            f"frequency_penalty={self.frequency_penalty}, "
+            f"repetition_penalty={self.repetition_penalty}, "
+            f"temperature={self.temperature}, "
+            f"top_p={self.top_p}, "
+            f"top_k={self.top_k}, "
+            f"min_p={self.min_p}, "
+            f"use_beam_search={self.use_beam_search}, "
+            f"length_penalty={self.length_penalty}, "
+            f"early_stopping={self.early_stopping}, "
+            f"stop={self.stop}, "
+            f"stop_token_ids={self.stop_token_ids}, "
+            f"include_stop_str_in_output={self.include_stop_str_in_output}, "
+            f"ignore_eos={self.ignore_eos}, "
+            f"max_tokens={self.max_tokens}, "
+            f"logprobs={self.logprobs}, "
+            f"prompt_logprobs={self.prompt_logprobs}, "
+            f"skip_special_tokens={self.skip_special_tokens}, "
+            "spaces_between_special_tokens="
+            f"{self.spaces_between_special_tokens})")

vllm/sequence.py

+"""Sequence and its related classes."""
+import copy
+import enum
+from typing import Dict, List, Optional, Union
+
+from vllm.block import LogicalTokenBlock
+from vllm.sampling_params import SamplingParams
+
+PromptLogprobs = List[Optional[Dict[int, float]]]
+SampleLogprobs = List[Dict[int, float]]
+
+
+class SequenceStatus(enum.Enum):
+    """Status of a sequence."""
+    WAITING = enum.auto()
+    RUNNING = enum.auto()
+    SWAPPED = enum.auto()
+    FINISHED_STOPPED = enum.auto()
+    FINISHED_LENGTH_CAPPED = enum.auto()
+    FINISHED_ABORTED = enum.auto()
+    FINISHED_IGNORED = enum.auto()
+
+    @staticmethod
+    def is_finished(status: "SequenceStatus") -> bool:
+        return status in [
+            SequenceStatus.FINISHED_STOPPED,
+            SequenceStatus.FINISHED_LENGTH_CAPPED,
+            SequenceStatus.FINISHED_ABORTED,
+            SequenceStatus.FINISHED_IGNORED,
+        ]
+
+    @staticmethod
+    def get_finished_reason(status: "SequenceStatus") -> Union[str, None]:
+        if status == SequenceStatus.FINISHED_STOPPED:
+            finish_reason = "stop"
+        elif status == SequenceStatus.FINISHED_LENGTH_CAPPED:
+            finish_reason = "length"
+        elif status == SequenceStatus.FINISHED_ABORTED:
+            finish_reason = "abort"
+        elif status == SequenceStatus.FINISHED_IGNORED:
+            # The ignored sequences are the sequences whose prompt lengths
+            # are longer than the model's length cap. Therefore, the stop
+            # reason should also be "length" as in OpenAI API.
+            finish_reason = "length"
+        else:
+            finish_reason = None
+        return finish_reason
+
+
+class SequenceData:
+    """Data associated with a sequence.
+
+
+    Args:
+        prompt_token_ids: The token IDs of the prompt.
+
+    Attributes:
+        prompt_token_ids: The token IDs of the prompt.
+        output_token_ids: The token IDs of the output.
+        cumulative_logprob: The cumulative log probability of the output.
+    """
+
+    def __init__(
+        self,
+        prompt_token_ids: List[int],
+    ) -> None:
+        self.prompt_token_ids = prompt_token_ids
+        self.output_token_ids: List[int] = []
+        self.cumulative_logprob = 0.0
+
+    def append_token_id(self, token_id: int, logprob: float) -> None:
+        self.output_token_ids.append(token_id)
+        self.cumulative_logprob += logprob
+
+    def get_len(self) -> int:
+        return len(self.output_token_ids) + len(self.prompt_token_ids)
+
+    def get_prompt_len(self) -> int:
+        return len(self.prompt_token_ids)
+
+    def get_output_len(self) -> int:
+        return len(self.output_token_ids)
+
+    def get_token_ids(self) -> List[int]:
+        return self.prompt_token_ids + self.output_token_ids
+
+    def get_last_token_id(self) -> int:
+        if not self.output_token_ids:
+            return self.prompt_token_ids[-1]
+        return self.output_token_ids[-1]
+
+    def __repr__(self) -> str:
+        return (f"SequenceData("
+                f"prompt_token_ids={self.prompt_token_ids}, "
+                f"output_token_ids={self.output_token_ids}, "
+                f"cumulative_logprob={self.cumulative_logprob})")
+
+
+class Sequence:
+    """Stores the data, status, and block information of a sequence.
+
+    Args:
+        seq_id: The ID of the sequence.
+        prompt: The prompt of the sequence.
+        prompt_token_ids: The token IDs of the prompt.
+        block_size: The block size of the sequence. Should be the same as the
+            block size used by the block manager and cache engine.
+    """
+
+    def __init__(
+        self,
+        seq_id: int,
+        prompt: str,
+        prompt_token_ids: List[int],
+        block_size: int,
+    ) -> None:
+        self.seq_id = seq_id
+        self.prompt = prompt
+        self.block_size = block_size
+
+        self.data = SequenceData(prompt_token_ids)
+        self.output_logprobs: SampleLogprobs = []
+        self.output_text = ""
+
+        self.logical_token_blocks: List[LogicalTokenBlock] = []
+        # Initialize the logical token blocks with the prompt token ids.
+        self._append_tokens_to_blocks(prompt_token_ids)
+        self.status = SequenceStatus.WAITING
+
+        # Used for incremental detokenization
+        self.prefix_offset = 0
+        self.read_offset = 0
+        # Input + output tokens
+        self.tokens: Optional[List[str]] = None
+
+    def _append_logical_block(self) -> None:
+        block = LogicalTokenBlock(
+            block_number=len(self.logical_token_blocks),
+            block_size=self.block_size,
+        )
+        self.logical_token_blocks.append(block)
+
+    def _append_tokens_to_blocks(self, token_ids: List[int]) -> None:
+        cursor = 0
+        while cursor < len(token_ids):
+            if not self.logical_token_blocks:
+                self._append_logical_block()
+
+            last_block = self.logical_token_blocks[-1]
+            if last_block.is_full():
+                self._append_logical_block()
+                last_block = self.logical_token_blocks[-1]
+
+            num_empty_slots = last_block.get_num_empty_slots()
+            last_block.append_tokens(token_ids[cursor:cursor +
+                                               num_empty_slots])
+            cursor += num_empty_slots
+
+    def append_token_id(
+        self,
+        token_id: int,
+        logprobs: Dict[int, float],
+    ) -> None:
+        assert token_id in logprobs
+        self._append_tokens_to_blocks([token_id])
+        self.output_logprobs.append(logprobs)
+        self.data.append_token_id(token_id, logprobs[token_id])
+
+    def get_len(self) -> int:
+        return self.data.get_len()
+
+    def get_prompt_len(self) -> int:
+        return self.data.get_prompt_len()
+
+    def get_output_len(self) -> int:
+        return self.data.get_output_len()
+
+    def get_token_ids(self) -> List[int]:
+        return self.data.get_token_ids()
+
+    def get_last_token_id(self) -> int:
+        return self.data.get_last_token_id()
+
+    def get_output_token_ids(self) -> List[int]:
+        return self.data.output_token_ids
+
+    def get_cumulative_logprob(self) -> float:
+        return self.data.cumulative_logprob
+
+    def get_beam_search_score(self,
+                              length_penalty: float = 0.0,
+                              seq_len: Optional[int] = None,
+                              eos_token_id: Optional[int] = None) -> float:
+        """Calculate the beam search score with length penalty.
+
+        Adapted from
+
+        https://github.com/huggingface/transformers/blob/ccb92be23def445f2afdea94c31286f84b89eb5b/src/transformers/generation/beam_search.py#L938
+        """
+        if seq_len is None:
+            seq_len = self.get_len()
+            # NOTE: HF implementation does not count the EOS token
+            # towards the length, we align with that here for testing.
+            if (eos_token_id is not None
+                    and self.get_last_token_id() == eos_token_id):
+                seq_len -= 1
+        return self.get_cumulative_logprob() / (seq_len**length_penalty)
+
+    def is_finished(self) -> bool:
+        return SequenceStatus.is_finished(self.status)
+
+    def fork(self, new_seq_id: int) -> "Sequence":
+        new_seq = copy.deepcopy(self)
+        new_seq.seq_id = new_seq_id
+        return new_seq
+
+    def __repr__(self) -> str:
+        return (f"Sequence(seq_id={self.seq_id}, "
+                f"status={self.status.name}, "
+                f"num_blocks={len(self.logical_token_blocks)})")
+
+
+class SequenceGroup:
+    """A group of sequences that are generated from the same prompt.
+
+    Args:
+        request_id: The ID of the request.
+        seqs: The list of sequences.
+        sampling_params: The sampling parameters used to generate the outputs.
+        arrival_time: The arrival time of the request.
+    """
+
+    def __init__(
+        self,
+        request_id: str,
+        seqs: List[Sequence],
+        sampling_params: SamplingParams,
+        arrival_time: float,
+    ) -> None:
+        self.request_id = request_id
+        self.seqs_dict = {seq.seq_id: seq for seq in seqs}
+        self.sampling_params = sampling_params
+        self.arrival_time = arrival_time
+        self.prompt_logprobs: Optional[PromptLogprobs] = None
+
+    @property
+    def prompt(self) -> str:
+        # All sequences in the group should have the same prompt.
+        # We use the prompt of an arbitrary sequence.
+        return next(iter(self.seqs_dict.values())).prompt
+
+    @property
+    def prompt_token_ids(self) -> List[int]:
+        # All sequences in the group should have the same prompt.
+        # We use the prompt of an arbitrary sequence.
+        return next(iter(self.seqs_dict.values())).data.prompt_token_ids
+
+    def get_max_num_running_seqs(self) -> int:
+        """The maximum number of sequences running in parallel in the remaining
+        lifetime of the request."""
+        if self.sampling_params.use_beam_search:
+            # For beam search, maximally there will always be `best_of` beam
+            # candidates running in the future.
+            return self.sampling_params.best_of
+        else:
+            if self.sampling_params.best_of > self.num_seqs():
+                # At prompt stage, the sequence group is not yet filled up
+                # and only have one sequence running. However, in the
+                # generation stage, we will have `best_of` sequences running.
+                return self.sampling_params.best_of
+            # At sampling stages, return the number of actual sequences
+            # that are not finished yet.
+            return self.num_unfinished_seqs()
+
+    def get_seqs(
+        self,
+        status: Optional[SequenceStatus] = None,
+    ) -> List[Sequence]:
+        if status is None:
+            return list(self.seqs_dict.values())
+        else:
+            return [
+                seq for seq in self.seqs_dict.values() if seq.status == status
+            ]
+
+    def get_unfinished_seqs(self) -> List[Sequence]:
+        return [
+            seq for seq in self.seqs_dict.values() if not seq.is_finished()
+        ]
+
+    def get_finished_seqs(self) -> List[Sequence]:
+        return [seq for seq in self.seqs_dict.values() if seq.is_finished()]
+
+    def num_seqs(self, status: Optional[SequenceStatus] = None) -> int:
+        return len(self.get_seqs(status))
+
+    def num_unfinished_seqs(self) -> int:
+        return len(self.get_unfinished_seqs())
+
+    def num_finished_seqs(self) -> int:
+        return len(self.get_finished_seqs())
+
+    def find(self, seq_id: int) -> Sequence:
+        if seq_id not in self.seqs_dict:
+            raise ValueError(f"Sequence {seq_id} not found.")
+        return self.seqs_dict[seq_id]
+
+    def add(self, seq: Sequence) -> None:
+        if seq.seq_id in self.seqs_dict:
+            raise ValueError(f"Sequence {seq.seq_id} already exists.")
+        self.seqs_dict[seq.seq_id] = seq
+
+    def remove(self, seq_id: int) -> None:
+        if seq_id not in self.seqs_dict:
+            raise ValueError(f"Sequence {seq_id} not found.")
+        del self.seqs_dict[seq_id]
+
+    def is_finished(self) -> bool:
+        return all(seq.is_finished() for seq in self.get_seqs())
+
+    def __repr__(self) -> str:
+        return (f"SequenceGroup(request_id={self.request_id}, "
+                f"sampling_params={self.sampling_params}, "
+                f"num_seqs={len(self.seqs_dict)})")
+
+
+class SequenceGroupMetadata:
+    """Metadata for a sequence group. Used to create `InputMetadata`.
+
+
+    Args:
+        request_id: The ID of the request.
+        is_prompt: Whether the request is at prompt stage.
+        seq_data: The sequence data. (Seq id -> sequence data)
+        sampling_params: The sampling parameters used to generate the outputs.
+        block_tables: The block tables. (Seq id -> list of physical block
+            numbers)
+    """
+
+    def __init__(
+        self,
+        request_id: str,
+        is_prompt: bool,
+        seq_data: Dict[int, SequenceData],
+        sampling_params: SamplingParams,
+        block_tables: Dict[int, List[int]],
+    ) -> None:
+        self.request_id = request_id
+        self.is_prompt = is_prompt
+        self.seq_data = seq_data
+        self.sampling_params = sampling_params
+        self.block_tables = block_tables
+
+
+class SequenceOutput:
+    """The model output associated with a sequence.
+
+    Args:
+        parent_seq_id: The ID of the parent sequence (for forking in beam
+            search).
+        output_token: The output token ID.
+        logprobs: The logprobs of the output token.
+            (Token id -> logP(x_i+1 | x_0, ..., x_i))
+    """
+
+    def __init__(
+        self,
+        parent_seq_id: int,
+        output_token: int,
+        logprobs: Dict[int, float],
+    ) -> None:
+        self.parent_seq_id = parent_seq_id
+        self.output_token = output_token
+        self.logprobs = logprobs
+
+    def __repr__(self) -> str:
+        return (f"SequenceOutput(parent_seq_id={self.parent_seq_id}, "
+                f"output_token={self.output_token}, "
+                f"logprobs={self.logprobs})")
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, SequenceOutput):
+            raise NotImplementedError()
+        return (self.parent_seq_id == other.parent_seq_id
+                and self.output_token == other.output_token
+                and self.logprobs == other.logprobs)
+
+
+class SequenceGroupOutput:
+    """The model output associated with a sequence group."""
+
+    def __init__(
+        self,
+        samples: List[SequenceOutput],
+        prompt_logprobs: Optional[PromptLogprobs],
+    ) -> None:
+        self.samples = samples
+        self.prompt_logprobs = prompt_logprobs
+
+    def __repr__(self) -> str:
+        return (f"SequenceGroupOutput(samples={self.samples}, "
+                f"prompt_logprobs={self.prompt_logprobs})")
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, SequenceGroupOutput):
+            raise NotImplementedError()
+        return (self.samples == other.samples
+                and self.prompt_logprobs == other.prompt_logprobs)
+
+
+# For each sequence group, we generate a list of SequenceOutput object,
+# each of which contains one possible candidate for the next token.
+SamplerOutput = List[SequenceGroupOutput]

vllm/transformers_utils/config.py

+from typing import Optional
+
+from transformers import AutoConfig, PretrainedConfig
+
+from vllm.transformers_utils.configs import *
+
+_CONFIG_REGISTRY = {
+    "aquila": AquilaConfig,
+    "baichuan": BaiChuanConfig,
+    "chatglm": ChatGLMConfig,
+    "mpt": MPTConfig,
+    "qwen": QWenConfig,
+    "RefinedWeb": RWConfig,  # For tiiuae/falcon-40b(-instruct)
+    "RefinedWebModel": RWConfig,  # For tiiuae/falcon-7b(-instruct)
+    "yi": YiConfig,
+}
+
+
+def get_config(model: str,
+               trust_remote_code: bool,
+               revision: Optional[str] = None) -> PretrainedConfig:
+    try:
+        config = AutoConfig.from_pretrained(
+            model, trust_remote_code=trust_remote_code, revision=revision)
+    except ValueError as e:
+        if (not trust_remote_code and
+                "requires you to execute the configuration file" in str(e)):
+            err_msg = (
+                "Failed to load the model config. If the model is a custom "
+                "model not yet available in the HuggingFace transformers "
+                "library, consider setting `trust_remote_code=True` in LLM "
+                "or using the `--trust-remote-code` flag in the CLI.")
+            raise RuntimeError(err_msg) from e
+        else:
+            raise e
+    if config.model_type in _CONFIG_REGISTRY:
+        config_class = _CONFIG_REGISTRY[config.model_type]
+        config = config_class.from_pretrained(model, revision=revision)
+    return config

vllm/transformers_utils/configs/__init__.py

+from vllm.transformers_utils.configs.aquila import AquilaConfig
+from vllm.transformers_utils.configs.baichuan import BaiChuanConfig
+from vllm.transformers_utils.configs.chatglm import ChatGLMConfig
+from vllm.transformers_utils.configs.mpt import MPTConfig
+from vllm.transformers_utils.configs.qwen import QWenConfig
+# RWConfig is for the original tiiuae/falcon-40b(-instruct) and
+# tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
+# `FalconConfig` class from the official HuggingFace transformers library.
+from vllm.transformers_utils.configs.falcon import RWConfig
+from vllm.transformers_utils.configs.yi import YiConfig
+
+__all__ = [
+    "AquilaConfig",
+    "BaiChuanConfig",
+    "ChatGLMConfig",
+    "MPTConfig",
+    "QWenConfig",
+    "RWConfig",
+    "YiConfig",
+]

vllm/transformers_utils/configs/aquila.py

+# coding=utf-8
+# Copyright 2023 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Aquila model configuration"""
+
+from transformers import PretrainedConfig
+
+
+class AquilaConfig(PretrainedConfig):
+    model_type = "aquila"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=100008,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.006,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

vllm/transformers_utils/configs/baichuan.py

+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class BaiChuanConfig(PretrainedConfig):
+    model_type = "baichuan"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=64000,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

vllm/transformers_utils/configs/chatglm.py

+# coding=utf-8
+# Adapted from
+# https://github.com/THUDM/ChatGLM2-6B
+from transformers import PretrainedConfig
+
+
+class ChatGLMConfig(PretrainedConfig):
+    model_type = "chatglm"
+    attribute_map = {
+        "num_hidden_layers": "num_layers",
+        "n_head_kv": "multi_query_group_num",
+    }
+
+    def __init__(self,
+                 num_layers=28,
+                 padded_vocab_size=65024,
+                 hidden_size=4096,
+                 ffn_hidden_size=13696,
+                 kv_channels=128,
+                 num_attention_heads=32,
+                 seq_length=2048,
+                 hidden_dropout=0.0,
+                 attention_dropout=0.0,
+                 layernorm_epsilon=1e-5,
+                 rmsnorm=True,
+                 apply_residual_connection_post_layernorm=False,
+                 post_layer_norm=True,
+                 add_bias_linear=False,
+                 add_qkv_bias=False,
+                 interleaved_qkv=False,
+                 bias_dropout_fusion=True,
+                 multi_query_attention=False,
+                 multi_query_group_num=1,
+                 apply_query_key_layer_scaling=True,
+                 attention_softmax_in_fp32=True,
+                 fp32_residual_connection=False,
+                 quantization_bit=0,
+                 pre_seq_len=None,
+                 prefix_projection=False,
+                 **kwargs):
+        self.num_layers = num_layers
+        self.vocab_size = padded_vocab_size
+        self.padded_vocab_size = padded_vocab_size
+        self.hidden_size = hidden_size
+        self.ffn_hidden_size = ffn_hidden_size
+        self.kv_channels = kv_channels
+        self.num_attention_heads = num_attention_heads
+        self.seq_length = seq_length
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.layernorm_epsilon = layernorm_epsilon
+        self.rmsnorm = rmsnorm
+        self.apply_residual_connection_post_layernorm = (
+            apply_residual_connection_post_layernorm)
+        self.post_layer_norm = post_layer_norm
+        self.add_bias_linear = add_bias_linear
+        self.add_qkv_bias = add_qkv_bias
+        self.bias_dropout_fusion = bias_dropout_fusion
+        self.multi_query_attention = multi_query_attention
+        self.multi_query_group_num = multi_query_group_num
+        self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
+        self.attention_softmax_in_fp32 = attention_softmax_in_fp32
+        self.fp32_residual_connection = fp32_residual_connection
+        self.quantization_bit = quantization_bit
+        self.pre_seq_len = pre_seq_len
+        self.prefix_projection = prefix_projection
+        self.interleaved_qkv = interleaved_qkv
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/falcon.py

+# Adapted from
+# https://huggingface.co/tiiuae/falcon-7b/blob/main/configuration_RW.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 the Big Science Workshop and HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Falcon configuration"""
+from transformers.configuration_utils import PretrainedConfig
+
+
+class RWConfig(PretrainedConfig):
+    model_type = "falcon"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_hidden_layers": "n_layer",
+        "num_attention_heads": "n_head",
+        "num_kv_heads": "n_head_kv",
+    }
+
+    def __init__(
+        self,
+        vocab_size=250880,
+        hidden_size=64,
+        n_layer=2,
+        n_head=8,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        hidden_dropout=0.0,
+        attention_dropout=0.0,
+        multi_query=True,
+        n_head_kv=None,
+        alibi=False,
+        bias=False,
+        parallel_attn=False,
+        new_decoder_architecture=False,
+        **kwargs,
+    ) -> None:
+        self.vocab_size = vocab_size
+        # Backward compatibility with n_embed kwarg
+        n_embed = kwargs.pop("n_embed", None)
+        self.hidden_size = hidden_size if n_embed is None else n_embed
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.multi_query = multi_query
+        self.n_head_kv = 1 if n_head_kv is None else n_head_kv
+        self.alibi = alibi
+        self.bias = bias
+        self.parallel_attn = parallel_attn
+        self.new_decoder_architecture = new_decoder_architecture
+
+        if self.hidden_size == 8192:
+            # Hack for falcon-40b
+            self.new_decoder_architecture = True
+
+        super().__init__(bos_token_id=bos_token_id,
+                         eos_token_id=eos_token_id,
+                         **kwargs)
+
+    @property
+    def head_dim(self):
+        return self.hidden_size // self.n_head
+
+    @property
+    def rotary(self):
+        return not self.alibi

vllm/transformers_utils/configs/mpt.py

+# coding=utf-8
+# Copied from
+# https://huggingface.co/mosaicml/mpt-7b/blob/main/configuration_mpt.py
+"""A HuggingFace-style model configuration."""
+import warnings
+from typing import Any, Dict, Optional, Union
+from transformers import PretrainedConfig
+
+attn_config_defaults: Dict = {
+    'attn_type': 'multihead_attention',
+    'attn_pdrop': 0.0,
+    'attn_impl': 'triton',
+    'qk_ln': False,
+    'clip_qkv': None,
+    'softmax_scale': None,
+    'prefix_lm': False,
+    'attn_uses_sequence_id': False,
+    'alibi': False,
+    'alibi_bias_max': 8
+}
+ffn_config_defaults: Dict = {'ffn_type': 'mptmlp'}
+init_config_defaults: Dict = {
+    'name': 'kaiming_normal_',
+    'fan_mode': 'fan_in',
+    'init_nonlinearity': 'relu',
+    'init_div_is_residual': True,
+    'emb_init_std': None,
+    'emb_init_uniform_lim': None,
+    'init_std': None,
+    'init_gain': 0.0
+}
+
+
+class MPTConfig(PretrainedConfig):
+    model_type = 'mpt'
+    attribute_map = {
+        'num_attention_heads': 'n_heads',
+        'hidden_size': 'd_model',
+        'num_hidden_layers': 'n_layers',
+    }
+
+    # pylint: disable=dangerous-default-value
+    def __init__(self,
+                 d_model: int = 2048,
+                 n_heads: int = 16,
+                 n_layers: int = 24,
+                 expansion_ratio: int = 4,
+                 max_seq_len: int = 2048,
+                 vocab_size: int = 50368,
+                 resid_pdrop: float = 0.0,
+                 emb_pdrop: float = 0.0,
+                 learned_pos_emb: bool = True,
+                 attn_config: Dict = attn_config_defaults,
+                 ffn_config: Dict = ffn_config_defaults,
+                 init_device: str = 'cpu',
+                 logit_scale: Optional[Union[float, str]] = None,
+                 no_bias: bool = False,
+                 embedding_fraction: float = 1.0,
+                 norm_type: str = 'low_precision_layernorm',
+                 use_cache: bool = False,
+                 init_config: Dict = init_config_defaults,
+                 fc_type: str = 'torch',
+                 verbose: Optional[int] = None,
+                 **kwargs: Any):
+        """The MPT configuration class.
+        Args:
+            d_model (int): The size of the embedding dimension of the model.
+            n_heads (int): The number of attention heads.
+            n_layers (int): The number of layers in the model.
+            expansion_ratio (int): The ratio of the up/down scale in the ffn.
+            max_seq_len (int): The maximum sequence length of the model.
+            vocab_size (int): The size of the vocabulary.
+            resid_pdrop (float): The dropout probability applied to the attention output before combining with residual.
+            emb_pdrop (float): The dropout probability for the embedding layer.
+            learned_pos_emb (bool): Whether to use learned positional embeddings
+            attn_config (Dict): A dictionary used to configure the model's attention module:
+                attn_type (str): type of attention to use. Options: multihead_attention, multiquery_attention, grouped_query_attention
+                attn_pdrop (float): The dropout probability for the attention layers.
+                attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
+                qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
+                clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
+                    this value.
+                softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,
+                    use the default scale of ``1/sqrt(d_keys)``.
+                prefix_lm (Optional[bool]): Whether the model should operate as a Prefix LM. This requires passing an
+                    extra `prefix_mask` argument which indicates which tokens belong to the prefix. Tokens in the prefix
+                    can attend to one another bi-directionally. Tokens outside the prefix use causal attention.
+                attn_uses_sequence_id (Optional[bool]): Whether to restrict attention to tokens that have the same sequence_id.
+                    When the model is in `train` mode, this requires passing an extra `sequence_id` argument which indicates
+                    which sub-sequence each token belongs to.
+                    Defaults to ``False`` meaning any provided `sequence_id` will be ignored.
+                alibi (bool): Whether to use the alibi bias instead of position embeddings.
+                alibi_bias_max (int): The maximum value of the alibi bias.
+                kv_n_heads (Optional[int]): For grouped_query_attention only, allow user to specify number of kv heads.
+            ffn_config (Dict): A dictionary used to configure the model's ffn module:
+                ffn_type (str): type of ffn to use. Options: mptmlp, te_ln_mlp
+            init_device (str): The device to use for parameter initialization.
+            logit_scale (Optional[Union[float, str]]): If not None, scale the logits by this value.
+            no_bias (bool): Whether to use bias in all layers.
+            verbose (int): The verbosity level. 0 is silent.
+            embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
+            norm_type (str): choose type of norm to use
+            use_cache (bool): Whether or not the model should return the last key/values attentions
+            init_config (Dict): A dictionary used to configure the model initialization:
+                init_config.name: The parameter initialization scheme to use. Options: 'default_', 'baseline_',
+                    'kaiming_uniform_', 'kaiming_normal_', 'neox_init_', 'small_init_', 'xavier_uniform_', or
+                    'xavier_normal_'. These mimic the parameter initialization methods in PyTorch.
+                init_div_is_residual (Union[int, float, str, bool]): Value to divide initial weights by if ``module._is_residual`` is True.
+                emb_init_std (Optional[float]): The standard deviation of the normal distribution used to initialize the embedding layer.
+                emb_init_uniform_lim (Optional[Union[Tuple[float, float], float]]): The lower and upper limits of the uniform distribution
+                    used to initialize the embedding layer. Mutually exclusive with ``emb_init_std``.
+                init_std (float): The standard deviation of the normal distribution used to initialize the model,
+                    if using the baseline_ parameter initialization scheme.
+                init_gain (float): The gain to use for parameter initialization with kaiming or xavier initialization schemes.
+                fan_mode (str): The fan mode to use for parameter initialization with kaiming initialization schemes.
+                init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
+                ---
+                See llmfoundry.models.utils.param_init_fns.py for info on other param init config options
+            fc_type (str): choose fc layer implementation. Options: torch and te. te layers support fp8 when using H100 GPUs.
+        """
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.expansion_ratio = expansion_ratio
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.resid_pdrop = resid_pdrop
+        self.emb_pdrop = emb_pdrop
+        self.learned_pos_emb = learned_pos_emb
+        self.attn_config = attn_config
+        self.ffn_config = ffn_config
+        self.init_device = init_device
+        self.logit_scale = logit_scale
+        self.no_bias = no_bias
+        self.embedding_fraction = embedding_fraction
+        self.norm_type = norm_type
+        self.use_cache = use_cache
+        self.init_config = init_config
+        self.fc_type = fc_type
+        if verbose is not None:
+            warnings.warn(DeprecationWarning(
+                'verbose argument for MPTConfig is now ignored and will be removed. Use python_log_level instead.'
+            ),
+                          stacklevel=2)
+        if 'name' in kwargs:
+            del kwargs['name']
+        if 'loss_fn' in kwargs:
+            del kwargs['loss_fn']
+        if self.attn_config.get('alibi', False):
+            self.learned_pos_emb = False
+            warnings.warn(
+                f'alibi is turned on, setting `learned_pos_emb` to {self.learned_pos_emb}`',
+                stacklevel=2)
+        super().__init__(**kwargs)
+        self._validate_config()
+
+    def _set_config_defaults(
+            self, config: Dict[str, Any],
+            config_defaults: Dict[str, Any]) -> Dict[str, Any]:
+        for (k, v) in config_defaults.items():
+            if k not in config:
+                config[k] = v
+        return config
+
+    def _validate_config(self) -> None:
+        self.attn_config = self._set_config_defaults(self.attn_config,
+                                                     attn_config_defaults)
+        self.ffn_config = self._set_config_defaults(self.ffn_config,
+                                                    ffn_config_defaults)
+        self.init_config = self._set_config_defaults(self.init_config,
+                                                     init_config_defaults)
+        if self.d_model % self.n_heads != 0:
+            raise ValueError('d_model must be divisible by n_heads')
+        if any((
+                prob < 0 or prob > 1 for prob in
+            [self.attn_config['attn_pdrop'], self.resid_pdrop, self.emb_pdrop]
+        )):
+            raise ValueError(
+                "self.attn_config['attn_pdrop'], resid_pdrop, emb_pdrop are probabilities and must be between 0 and 1"  # pylint: disable=line-too-long
+            )
+        if self.attn_config['attn_impl'] not in ['torch', 'flash', 'triton']:
+            raise ValueError(
+                f"Unknown attn_impl={self.attn_config['attn_impl']}")
+        if self.attn_config['prefix_lm'] and self.attn_config[
+                'attn_impl'] not in ['torch', 'triton']:
+            raise NotImplementedError(
+                'prefix_lm only implemented with torch and triton attention.')
+        if self.attn_config['alibi'] and self.attn_config['attn_impl'] not in [
+                'torch', 'triton'
+        ]:
+            raise NotImplementedError(
+                'alibi only implemented with torch and triton attention.')
+        if self.attn_config['attn_uses_sequence_id'] and self.attn_config[
+                'attn_impl'] not in ['torch', 'triton']:
+            raise NotImplementedError(
+                'attn_uses_sequence_id only implemented with torch and triton attention.'  # pylint: disable=line-too-long
+            )
+        if self.embedding_fraction > 1 or self.embedding_fraction <= 0:
+            raise ValueError(
+                'model.embedding_fraction must be between 0 (exclusive) and 1 (inclusive)!'  # pylint: disable=line-too-long
+            )
+        if isinstance(self.logit_scale,
+                      str) and self.logit_scale != 'inv_sqrt_d_model':
+            raise ValueError(
+                f"self.logit_scale={self.logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."  # pylint: disable=line-too-long
+            )
+        if self.init_config.get('name', None) is None:
+            raise ValueError(
+                f"self.init_config={self.init_config!r} 'name' needs to be set."
+            )
+        if not self.learned_pos_emb and (not self.attn_config['alibi']):
+            warnings.warn(
+                'Positional information not being provided to the model.',
+                stacklevel=2)
+        if self.fc_type == 'te' or self.ffn_config['ffn_type'] == 'te_ln_mlp':
+            try:
+                # pylint: disable=import-outside-toplevel
+                import transformer_engine.pytorch as te
+                del te
+            except Exception as exc:
+                raise ImportError(
+                    # pylint: disable=line-too-long
+                    'TransformerEngine import fail. `fc_type: te` requires TransformerEngine be installed. '
+                    +
+                    'The required version of transformer_engine also requires FlashAttention v1.0.6 is installed:\n'
+                    + 'pip install flash-attn==1.0.6 --no-build-isolation \n' +
+                    'pip install git+https://github.com/NVIDIA/TransformerEngine.git@144e4888b2cdd60bd52e706d5b7a79cb9c1a7156'
+                ) from exc
+        if self.ffn_config['ffn_type'] == 'mptmlp':
+            self.ffn_config['fc_type'] = self.fc_type
+        elif self.ffn_config['ffn_type'] == 'te_ln_mlp':
+            self.ffn_config['bias'] = not self.no_bias

vllm/transformers_utils/configs/qwen.py

+# Copyright (c) Alibaba Cloud.
+# LICENSE: https://huggingface.co/Qwen/Qwen-7B/blob/main/LICENSE
+
+from transformers import PretrainedConfig
+
+
+class QWenConfig(PretrainedConfig):
+    model_type = "qwen"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        emb_dropout_prob=0.0,
+        attn_dropout_prob=0.0,
+        layer_norm_epsilon=1e-6,
+        initializer_range=0.02,
+        max_position_embeddings=8192,
+        scale_attn_weights=True,
+        use_cache=True,
+        bf16=False,
+        fp16=False,
+        fp32=False,
+        kv_channels=128,
+        rotary_pct=1.0,
+        rotary_emb_base=10000,
+        use_dynamic_ntk=True,
+        use_logn_attn=True,
+        use_flash_attn="auto",
+        intermediate_size=22016,
+        no_bias=True,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.emb_dropout_prob = emb_dropout_prob
+        self.attn_dropout_prob = attn_dropout_prob
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.bf16 = bf16
+        self.fp16 = fp16
+        self.fp32 = fp32
+        self.kv_channels = kv_channels
+        self.rotary_pct = rotary_pct
+        self.rotary_emb_base = rotary_emb_base
+        self.use_dynamic_ntk = use_dynamic_ntk
+        self.use_logn_attn = use_logn_attn
+        self.use_flash_attn = use_flash_attn
+        self.no_bias = no_bias
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

vllm/transformers_utils/configs/yi.py

+""" Yi model configuration"""
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+Yi_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class YiConfig(PretrainedConfig):
+    r"""
+        Reference:
+        https://huggingface.co/01-ai/Yi-6B/blob/main/configuration_yi.py
+    """
+    model_type = "Yi"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=64000,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=4,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        output_attentions=False,
+        rope_theta=5000000.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.output_attentions = output_attentions
+        self.rope_theta = rope_theta
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

vllm/transformers_utils/tokenizer.py

+from typing import List, Optional, Tuple, Union
+
+from transformers import (AutoTokenizer, PreTrainedTokenizer,
+                          PreTrainedTokenizerFast)
+
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizers import *
+
+logger = init_logger(__name__)
+
+
+def get_tokenizer(
+    tokenizer_name: str,
+    *args,
+    tokenizer_mode: str = "auto",
+    trust_remote_code: bool = False,
+    tokenizer_revision: Optional[str] = None,
+    **kwargs,
+) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
+    """Gets a tokenizer for the given model name via Huggingface."""
+    if tokenizer_mode == "slow":
+        if kwargs.get("use_fast", False):
+            raise ValueError(
+                "Cannot use the fast tokenizer in slow tokenizer mode.")
+        kwargs["use_fast"] = False
+
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(
+            tokenizer_name,
+            *args,
+            trust_remote_code=trust_remote_code,
+            tokenizer_revision=tokenizer_revision,
+            **kwargs)
+    except ValueError as e:
+        # If the error pertains to the tokenizer class not existing or not
+        # currently being imported, suggest using the --trust-remote-code flag.
+        if (not trust_remote_code and
+            ("does not exist or is not currently imported." in str(e)
+             or "requires you to execute the tokenizer file" in str(e))):
+            err_msg = (
+                "Failed to load the tokenizer. If the tokenizer is a custom "
+                "tokenizer not yet available in the HuggingFace transformers "
+                "library, consider setting `trust_remote_code=True` in LLM "
+                "or using the `--trust-remote-code` flag in the CLI.")
+            raise RuntimeError(err_msg) from e
+        else:
+            raise e
+    except AttributeError as e:
+        if "BaichuanTokenizer" in str(e):
+            # This is for the error "'BaichuanTokenizer' object has no
+            # attribute 'sp_model'".
+            tokenizer = BaichuanTokenizer.from_pretrained(
+                tokenizer_name,
+                *args,
+                trust_remote_code=trust_remote_code,
+                tokenizer_revision=tokenizer_revision,
+                **kwargs)
+        else:
+            raise e
+
+    if not isinstance(tokenizer, PreTrainedTokenizerFast):
+        logger.warning(
+            "Using a slow tokenizer. This might cause a significant "
+            "slowdown. Consider using a fast tokenizer instead.")
+    return tokenizer
+
+
+def _convert_tokens_to_string_with_added_encoders(
+    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
+    output_tokens: List[str],
+    skip_special_tokens: bool,
+    spaces_between_special_tokens: bool,
+) -> str:
+    # Adapted from
+    # https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/tokenization_utils.py#L921
+    # NOTE(woosuk): The following code is slow because it runs a for loop over
+    # the output_tokens. In Python, running a for loop over a list can be slow
+    # even when the loop body is very simple.
+    sub_texts = []
+    current_sub_text = []
+    all_special_tokens = set(tokenizer.all_special_tokens)
+    for token in output_tokens:
+        if skip_special_tokens and token in all_special_tokens:
+            continue
+        if token in tokenizer.get_added_vocab():
+            if current_sub_text:
+                sub_text = tokenizer.convert_tokens_to_string(current_sub_text)
+                sub_texts.append(sub_text)
+                current_sub_text = []
+            sub_texts.append(token)
+        else:
+            current_sub_text.append(token)
+    if current_sub_text:
+        sub_text = tokenizer.convert_tokens_to_string(current_sub_text)
+        sub_texts.append(sub_text)
+    if spaces_between_special_tokens:
+        return " ".join(sub_texts)
+    else:
+        return "".join(sub_texts)
+
+
+# Based on
+# https://github.com/huggingface/text-generation-inference/blob/v0.9.4/server/text_generation_server/models/model.py#L62C9-L62C15
+# under Apache 2.0 license
+def detokenize_incrementally(
+    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
+    all_input_ids: List[int],
+    prev_tokens: Optional[List[str]],
+    prefix_offset: int = 0,
+    read_offset: int = 0,
+    skip_special_tokens: bool = False,
+    spaces_between_special_tokens: bool = True,
+) -> Tuple[List[str], str, int, int]:
+    new_token_id = all_input_ids[-1]
+    # This is the first iteration for this sequence
+    if prev_tokens is None:
+        new_tokens = tokenizer.convert_ids_to_tokens(
+            all_input_ids, skip_special_tokens=skip_special_tokens)
+        output_tokens = new_tokens
+        # 5 is an arbitrary value that should work for all
+        # tokenizers (bigger = more conservative).
+        # Subtract 1 extra to account for the generated token.
+        prefix_offset = max(len(output_tokens) - 6, 0)
+        # If the first new token is a special token, we can't skip 1 extra token
+        if skip_special_tokens and new_token_id in tokenizer.all_special_ids:
+            read_offset = max(len(output_tokens), 0)
+        else:
+            read_offset = max(len(output_tokens) - 1, 0)
+    else:
+        # Put new_token_id in a list so skip_special_tokens is respected
+        new_tokens = tokenizer.convert_ids_to_tokens(
+            [new_token_id], skip_special_tokens=skip_special_tokens)
+        output_tokens = prev_tokens + new_tokens
+
+    # The prefix text is necessary only to defeat cleanup algorithms in
+    # the decode which decide to add a space or not depending on the
+    # surrounding ids.
+    if tokenizer.is_fast or not tokenizer.get_added_vocab():
+        prefix_text = tokenizer.convert_tokens_to_string(
+            output_tokens[prefix_offset:read_offset])
+        new_text = tokenizer.convert_tokens_to_string(
+            output_tokens[prefix_offset:])
+    else:
+        prefix_text = _convert_tokens_to_string_with_added_encoders(
+            tokenizer,
+            output_tokens[prefix_offset:read_offset],
+            skip_special_tokens=skip_special_tokens,
+            spaces_between_special_tokens=spaces_between_special_tokens,
+        )
+        new_text = _convert_tokens_to_string_with_added_encoders(
+            tokenizer,
+            output_tokens[prefix_offset:],
+            skip_special_tokens=skip_special_tokens,
+            spaces_between_special_tokens=spaces_between_special_tokens,
+        )
+
+    if len(new_text) > len(prefix_text) and not new_text.endswith("�"):
+        # utf-8 char at the end means it's a potential unfinished byte sequence
+        # from byte fallback tokenization.
+        # If it's in the middle, it's probably a real invalid id generated
+        # by the model
+        new_text = new_text[len(prefix_text):]
+        return new_tokens, new_text, read_offset, len(output_tokens)
+    else:
+        return new_tokens, "", prefix_offset, read_offset

vllm/transformers_utils/tokenizers/__init__.py

+from vllm.transformers_utils.tokenizers.baichuan import BaichuanTokenizer
+
+__all__ = [
+    "BaichuanTokenizer",
+]

vllm/transformers_utils/tokenizers/baichuan.py

+# yapf: disable
+# Adapted from
+# https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/blob/8f6e343d545c503b91429582231d1d354dac2740/tokenization_baichuan.py
+# This includes a fix suggested in
+# https://github.com/vllm-project/vllm/issues/1403#issuecomment-1767503058
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {},
+    "tokenizer_file": {},
+}
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
+
+
+class BaichuanTokenizer(PreTrainedTokenizer):
+    """
+    Construct a Baichuan tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        pad_token=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = (
+            AddedToken(bos_token, lstrip=False, rstrip=False)
+            if isinstance(bos_token, str)
+            else bos_token
+        )
+        eos_token = (
+            AddedToken(eos_token, lstrip=False, rstrip=False)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+        unk_token = (
+            AddedToken(unk_token, lstrip=False, rstrip=False)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        pad_token = (
+            AddedToken(pad_token, lstrip=False, rstrip=False)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Returns a tokenized string."""
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for i, token in enumerate(tokens):
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special and i != 0:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    def save_vocabulary(
+        self, save_directory, filename_prefix: Optional[str] = None
+    ) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "")
+            + VOCAB_FILES_NAMES["vocab_file"],
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(
+            out_vocab_file
+        ) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    def get_special_tokens_mask(
+        self,
+        token_ids_0: List[int],
+        token_ids_1: Optional[List[int]] = None,
+        already_has_special_tokens: bool = False,
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0,
+                token_ids_1=token_ids_1,
+                already_has_special_tokens=True,
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        if token_ids_1 is None, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output

vllm/utils.py

+import enum
+import socket
+import uuid
+from platform import uname
+
+import psutil
+import torch
+
+from vllm._C import cuda_utils
+
+
+class Device(enum.Enum):
+    GPU = enum.auto()
+    CPU = enum.auto()
+
+
+class Counter:
+
+    def __init__(self, start: int = 0) -> None:
+        self.counter = start
+
+    def __next__(self) -> int:
+        i = self.counter
+        self.counter += 1
+        return i
+
+    def reset(self) -> None:
+        self.counter = 0
+
+
+def is_hip() -> bool:
+    return torch.version.hip is not None
+
+
+def get_max_shared_memory_bytes(gpu: int = 0) -> int:
+    """Returns the maximum shared memory per thread block in bytes."""
+    # https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
+    cudaDevAttrMaxSharedMemoryPerBlockOptin = 97 if not is_hip() else 74
+    max_shared_mem = cuda_utils.get_device_attribute(
+        cudaDevAttrMaxSharedMemoryPerBlockOptin, gpu)
+    return int(max_shared_mem)
+
+
+def get_cpu_memory() -> int:
+    """Returns the total CPU memory of the node in bytes."""
+    return psutil.virtual_memory().total
+
+
+def random_uuid() -> str:
+    return str(uuid.uuid4().hex)
+
+
+def in_wsl() -> bool:
+    # Reference: https://github.com/microsoft/WSL/issues/4071
+    return "microsoft" in " ".join(uname()).lower()
+
+
+def get_open_port():
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("", 0))
+        return s.getsockname()[1]

vllm/worker/cache_engine.py

+"""CacheEngine class for managing the KV cache."""
+from typing import Dict, List, Tuple
+
+import torch
+
+from vllm._C import cache_ops
+from vllm.config import CacheConfig, ModelConfig, ParallelConfig
+from vllm.logger import init_logger
+from vllm.utils import in_wsl
+
+logger = init_logger(__name__)
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class CacheEngine:
+    """Manages the KV cache.
+
+    This class is responsible for initializing and managing the GPU and CPU KV
+    caches. It also provides methods for performing KV cache operations, such
+    as swapping and copying.
+    """
+
+    def __init__(
+        self,
+        cache_config: CacheConfig,
+        model_config: ModelConfig,
+        parallel_config: ParallelConfig,
+    ) -> None:
+        self.cache_config = cache_config
+        self.model_config = model_config
+        self.parallel_config = parallel_config
+
+        self.head_size = model_config.get_head_size()
+        self.num_layers = model_config.get_num_layers(parallel_config)
+        self.num_heads = model_config.get_num_kv_heads(parallel_config)
+        self.dtype = model_config.dtype
+
+        self.block_size = cache_config.block_size
+        self.num_gpu_blocks = cache_config.num_gpu_blocks
+        self.num_cpu_blocks = cache_config.num_cpu_blocks
+
+        # Initialize the cache.
+        self.gpu_cache = self.allocate_gpu_cache()
+        self.cpu_cache = self.allocate_cpu_cache()
+
+        # Initialize the stream for caching operations.
+        self.cache_stream = torch.cuda.Stream()
+        assert self.cache_stream != torch.cuda.current_stream()
+        # Initialize the events for stream synchronization.
+        self.events = [torch.cuda.Event() for _ in range(self.num_layers)]
+
+    def get_key_block_shape(self) -> Tuple[int, int, int, int]:
+        element_size = torch.tensor([], dtype=self.dtype).element_size()
+        x = 16 // element_size
+        return (
+            self.num_heads,
+            self.head_size // x,
+            self.block_size,
+            x,
+        )
+
+    def get_value_block_shape(self) -> Tuple[int, int, int]:
+        return (
+            self.num_heads,
+            self.head_size,
+            self.block_size,
+        )
+
+    def allocate_gpu_cache(self) -> List[KVCache]:
+        gpu_cache: List[KVCache] = []
+        key_block_shape = self.get_key_block_shape()
+        value_block_shape = self.get_value_block_shape()
+        for _ in range(self.num_layers):
+            key_blocks = torch.empty(
+                size=(self.num_gpu_blocks, *key_block_shape),
+                dtype=self.dtype,
+                device="cuda",
+            )
+            value_blocks = torch.empty(
+                size=(self.num_gpu_blocks, *value_block_shape),
+                dtype=self.dtype,
+                device="cuda",
+            )
+            gpu_cache.append((key_blocks, value_blocks))
+        return gpu_cache
+
+    def allocate_cpu_cache(self) -> List[KVCache]:
+        cpu_cache: List[KVCache] = []
+        key_block_shape = self.get_key_block_shape()
+        value_block_shape = self.get_value_block_shape()
+        pin_memory = not in_wsl()
+        if not pin_memory:
+            # Pinning memory in WSL is not supported.
+            # https://docs.nvidia.com/cuda/wsl-user-guide/index.html#known-limitations-for-linux-cuda-applications
+            logger.warning("Using 'pin_memory=False' as WSL is detected. "
+                           "This may slow down the performance.")
+        for _ in range(self.num_layers):
+            key_blocks = torch.empty(
+                size=(self.num_cpu_blocks, *key_block_shape),
+                dtype=self.dtype,
+                pin_memory=pin_memory,
+            )
+            value_blocks = torch.empty(
+                size=(self.num_cpu_blocks, *value_block_shape),
+                dtype=self.dtype,
+                pin_memory=pin_memory,
+            )
+            cpu_cache.append((key_blocks, value_blocks))
+        return cpu_cache
+
+    def _swap(
+        self,
+        src: List[KVCache],
+        dst: List[KVCache],
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        with torch.cuda.stream(self.cache_stream):
+            for i in range(self.num_layers):
+                src_key_cache, src_value_cache = src[i]
+                dst_key_cache, dst_value_cache = dst[i]
+                # Copy the key blocks.
+                cache_ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
+                # Copy the value blocks.
+                cache_ops.swap_blocks(src_value_cache, dst_value_cache,
+                                      src_to_dst)
+                event = self.events[i]
+                event.record(stream=self.cache_stream)
+
+    def swap_in(self, src_to_dst: Dict[int, int]) -> None:
+        self._swap(self.cpu_cache, self.gpu_cache, src_to_dst)
+
+    def swap_out(self, src_to_dst: Dict[int, int]) -> None:
+        self._swap(self.gpu_cache, self.cpu_cache, src_to_dst)
+
+    def copy(self, src_to_dsts: Dict[int, List[int]]) -> None:
+        key_caches = [key_cache for key_cache, _ in self.gpu_cache]
+        value_caches = [value_cache for _, value_cache in self.gpu_cache]
+        # NOTE(woosuk): This operation implicitly synchronizes the CPU and GPU.
+        cache_ops.copy_blocks(key_caches, value_caches, src_to_dsts)
+
+    @staticmethod
+    def get_cache_block_size(
+        block_size: int,
+        model_config: ModelConfig,
+        parallel_config: ParallelConfig,
+    ) -> int:
+        head_size = model_config.get_head_size()
+        num_heads = model_config.get_num_kv_heads(parallel_config)
+        num_layers = model_config.get_num_layers(parallel_config)
+
+        key_cache_block = block_size * num_heads * head_size
+        value_cache_block = key_cache_block
+        total = num_layers * (key_cache_block + value_cache_block)
+        dtype_size = _get_dtype_size(model_config.dtype)
+        return dtype_size * total
+
+
+def _get_dtype_size(dtype: torch.dtype) -> int:
+    return torch.tensor([], dtype=dtype).element_size()

vllm/worker/model_runner.py

+import time
+from typing import Dict, List, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from vllm.config import ModelConfig, ParallelConfig, SchedulerConfig
+from vllm.logger import init_logger
+from vllm.model_executor import get_model, InputMetadata, SamplingMetadata
+from vllm.sampling_params import SamplingParams, SamplingType
+from vllm.sequence import SamplerOutput, SequenceData, SequenceGroupMetadata
+from vllm.utils import in_wsl
+
+logger = init_logger(__name__)
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+_PAD_SLOT_ID = -1
+# Capture graphs for batch size 1, 2, 4, 8, 16, 24, 32, 40, ..., 256.
+# NOTE: _get_graph_batch_size needs to be updated if this list is changed.
+_BATCH_SIZES_TO_CAPTURE = [1, 2, 4] + [8 * i for i in range(1, 33)]
+
+
+class ModelRunner:
+
+    def __init__(
+        self,
+        model_config: ModelConfig,
+        parallel_config: ParallelConfig,
+        scheduler_config: SchedulerConfig,
+    ):
+        self.model_config = model_config
+        self.parallel_config = parallel_config
+        self.scheduler_config = scheduler_config
+
+        # model_config can be None in tests/samplers/test_sampler.py.
+        # FIXME(woosuk): This is a hack to make the tests work. Refactor this.
+        self.sliding_window = (model_config.get_sliding_window()
+                               if model_config is not None else None)
+        self.model = None
+        self.block_size = None  # Set after initial profiling.
+
+        self.graph_runners: Dict[int, CUDAGraphRunner] = {}
+        self.graph_memory_pool = None  # Set during graph capture.
+
+        self.max_context_len_to_capture = (
+            self.model_config.max_context_len_to_capture
+            if self.model_config is not None else 0)
+        # When using CUDA graph, the input block tables must be padded to
+        # max_context_len_to_capture. However, creating the block table in
+        # Python can be expensive. To optimize this, we cache the block table
+        # in numpy and only copy the actual input content at every iteration.
+        # The shape of the cached block table will be
+        # (max batch size to capture, max context len to capture / block size).
+        self.graph_block_tables = None  # Set after initial profiling.
+        # cache in_wsl result
+        self.in_wsl = in_wsl()
+
+    def load_model(self) -> None:
+        self.model = get_model(self.model_config)
+
+    def set_block_size(self, block_size: int) -> None:
+        self.block_size = block_size
+
+        max_num_blocks = (self.max_context_len_to_capture + block_size -
+                          1) // block_size
+        self.graph_block_tables = np.zeros(
+            (max(_BATCH_SIZES_TO_CAPTURE), max_num_blocks), dtype=np.int32)
+
+    def _prepare_prompt(
+        self,
+        seq_group_metadata_list: List[SequenceGroupMetadata],
+    ) -> Tuple[torch.Tensor, torch.Tensor, InputMetadata]:
+        assert len(seq_group_metadata_list) > 0
+        input_tokens: List[List[int]] = []
+        input_positions: List[List[int]] = []
+        slot_mapping: List[List[int]] = []
+
+        prompt_lens: List[int] = []
+        for seq_group_metadata in seq_group_metadata_list:
+            assert seq_group_metadata.is_prompt
+            seq_ids = list(seq_group_metadata.seq_data.keys())
+            assert len(seq_ids) == 1
+            seq_id = seq_ids[0]
+
+            seq_data = seq_group_metadata.seq_data[seq_id]
+            prompt_tokens = seq_data.get_token_ids()
+            prompt_len = len(prompt_tokens)
+            prompt_lens.append(prompt_len)
+
+            input_tokens.append(prompt_tokens)
+            # NOTE(woosuk): Here we assume that the first token in the prompt
+            # is always the first token in the sequence.
+            input_positions.append(list(range(prompt_len)))
+
+            if seq_group_metadata.block_tables is None:
+                # During memory profiling, the block tables are not initialized
+                # yet. In this case, we just use a dummy slot mapping.
+                slot_mapping.append([_PAD_SLOT_ID] * prompt_len)
+                continue
+
+            # Compute the slot mapping.
+            slot_mapping.append([])
+            block_table = seq_group_metadata.block_tables[seq_id]
+            # Mask the [0, start_idx) tokens of the prompt with _PAD_SLOT_ID,
+            # where start_idx is max(0, prompt_len - sliding_window).
+            # For example, if the prompt len is 10, sliding window is 8, and
+            # block size is 4, the first two tokens are masked and the slot
+            # mapping will be [-1, -1, 2, 3, 4, 5, 6, 7, 0, 1].
+            start_idx = 0
+            if self.sliding_window is not None:
+                start_idx = max(0, prompt_len - self.sliding_window)
+            for i in range(prompt_len):
+                if i < start_idx:
+                    slot_mapping[-1].append(_PAD_SLOT_ID)
+                    continue
+
+                block_number = block_table[i // self.block_size]
+                block_offset = i % self.block_size
+                slot = block_number * self.block_size + block_offset
+                slot_mapping[-1].append(slot)
+
+        max_prompt_len = max(prompt_lens)
+        input_tokens = _make_tensor_with_pad(input_tokens,
+                                             max_prompt_len,
+                                             pad=0,
+                                             dtype=torch.long)
+        input_positions = _make_tensor_with_pad(input_positions,
+                                                max_prompt_len,
+                                                pad=0,
+                                                dtype=torch.long)
+        slot_mapping = _make_tensor_with_pad(slot_mapping,
+                                             max_prompt_len,
+                                             pad=_PAD_SLOT_ID,
+                                             dtype=torch.long)
+
+        input_metadata = InputMetadata(
+            prompt_lens=prompt_lens,
+            slot_mapping=slot_mapping,
+            max_context_len=None,
+            context_lens=None,
+            block_tables=None,
+            use_cuda_graph=False,
+        )
+        return input_tokens, input_positions, input_metadata
+
+    def _prepare_decode(
+        self,
+        seq_group_metadata_list: List[SequenceGroupMetadata],
+    ) -> Tuple[torch.Tensor, torch.Tensor, InputMetadata]:
+        assert len(seq_group_metadata_list) > 0
+        input_tokens: List[List[int]] = []
+        input_positions: List[List[int]] = []
+        slot_mapping: List[List[int]] = []
+        context_lens: List[int] = []
+        block_tables: List[List[int]] = []
+
+        for seq_group_metadata in seq_group_metadata_list:
+            assert not seq_group_metadata.is_prompt
+
+            seq_ids = list(seq_group_metadata.seq_data.keys())
+            for seq_id in seq_ids:
+                seq_data = seq_group_metadata.seq_data[seq_id]
+                generation_token = seq_data.get_last_token_id()
+                input_tokens.append([generation_token])
+
+                seq_len = seq_data.get_len()
+                position = seq_len - 1
+                input_positions.append([position])
+
+                context_len = seq_len if self.sliding_window is None else min(
+                    seq_len, self.sliding_window)
+                context_lens.append(context_len)
+
+                block_table = seq_group_metadata.block_tables[seq_id]
+                block_number = block_table[position // self.block_size]
+                block_offset = position % self.block_size
+                slot = block_number * self.block_size + block_offset
+                slot_mapping.append([slot])
+
+                if self.sliding_window is not None:
+                    sliding_window_blocks = (self.sliding_window //
+                                             self.block_size)
+                    block_table = block_table[-sliding_window_blocks:]
+                block_tables.append(block_table)
+
+        batch_size = len(input_tokens)
+        max_context_len = max(context_lens)
+        use_captured_graph = (
+            not self.model_config.enforce_eager
+            and batch_size <= _BATCH_SIZES_TO_CAPTURE[-1]
+            and max_context_len <= self.max_context_len_to_capture)
+        if use_captured_graph:
+            # Pad the input tokens, positions, and slot mapping to match the
+            # batch size of the captured graph.
+            graph_batch_size = _get_graph_batch_size(batch_size)
+            assert graph_batch_size >= batch_size
+            for _ in range(graph_batch_size - batch_size):
+                input_tokens.append([])
+                input_positions.append([])
+                slot_mapping.append([])
+                context_lens.append(1)
+                block_tables.append([])
+            batch_size = graph_batch_size
+
+        # When using CUDA graph, we don't need to make the tensors on the GPU
+        # because they will be eventually copied to the designated GPU buffer.
+        device = "cpu" if use_captured_graph else "cuda"
+        pin_memory = use_captured_graph and not self.in_wsl
+        input_tokens = _make_tensor_with_pad(input_tokens,
+                                             max_len=1,
+                                             pad=0,
+                                             dtype=torch.long,
+                                             device=device,
+                                             pin_memory=pin_memory)
+        input_positions = _make_tensor_with_pad(input_positions,
+                                                max_len=1,
+                                                pad=0,
+                                                dtype=torch.long,
+                                                device=device,
+                                                pin_memory=pin_memory)
+        slot_mapping = _make_tensor_with_pad(slot_mapping,
+                                             max_len=1,
+                                             pad=_PAD_SLOT_ID,
+                                             dtype=torch.long,
+                                             device=device,
+                                             pin_memory=pin_memory)
+        context_lens = torch.tensor(context_lens,
+                                    dtype=torch.int,
+                                    device=device,
+                                    pin_memory=pin_memory)
+
+        if use_captured_graph:
+            # The shape of graph_block_tables is
+            # [max batch size, max context len // block size].
+            input_block_tables = self.graph_block_tables[:batch_size]
+            for i, block_table in enumerate(block_tables):
+                if block_table:
+                    input_block_tables[i, :len(block_table)] = block_table
+            block_tables = torch.tensor(input_block_tables, device=device)
+        else:
+            block_tables = _make_tensor_with_pad(
+                block_tables,
+                max_len=max_context_len,
+                pad=0,
+                dtype=torch.int,
+            )
+
+        input_metadata = InputMetadata(
+            prompt_lens=[],
+            slot_mapping=slot_mapping,
+            max_context_len=max_context_len,
+            context_lens=context_lens,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+        return input_tokens, input_positions, input_metadata
+
+    def _prepare_sample(
+        self,
+        seq_group_metadata_list: List[SequenceGroupMetadata],
+        prompt_lens: List[int],
+    ) -> SamplingMetadata:
+        seq_groups: List[Tuple[List[int], SamplingParams]] = []
+        selected_token_indices: List[int] = []
+        selected_token_start_idx = 0
+        categorized_sample_indices = {t: [] for t in SamplingType}
+        categorized_sample_indices_start_idx = 0
+
+        max_prompt_len = max(prompt_lens) if prompt_lens else 1
+        for i, seq_group_metadata in enumerate(seq_group_metadata_list):
+            seq_ids = list(seq_group_metadata.seq_data.keys())
+            sampling_params = seq_group_metadata.sampling_params
+            seq_groups.append((seq_ids, sampling_params))
+
+            if seq_group_metadata.is_prompt:
+                assert len(seq_ids) == 1
+                prompt_len = prompt_lens[i]
+                if sampling_params.prompt_logprobs is not None:
+                    # NOTE: prompt token positions do not need sample, skip
+                    categorized_sample_indices_start_idx += prompt_len - 1
+
+                categorized_sample_indices[
+                    sampling_params.sampling_type].append(
+                        categorized_sample_indices_start_idx)
+                categorized_sample_indices_start_idx += 1
+
+                if sampling_params.prompt_logprobs is not None:
+                    selected_token_indices.extend(
+                        range(selected_token_start_idx,
+                              selected_token_start_idx + prompt_len - 1))
+                selected_token_indices.append(selected_token_start_idx +
+                                              prompt_len - 1)
+                selected_token_start_idx += max_prompt_len
+            else:
+                num_seqs = len(seq_ids)
+                selected_token_indices.extend(
+                    range(selected_token_start_idx,
+                          selected_token_start_idx + num_seqs))
+                selected_token_start_idx += num_seqs
+
+                categorized_sample_indices[
+                    sampling_params.sampling_type].extend(
+                        range(categorized_sample_indices_start_idx,
+                              categorized_sample_indices_start_idx + num_seqs))
+                categorized_sample_indices_start_idx += num_seqs
+
+        selected_token_indices = _async_h2d(selected_token_indices,
+                                            dtype=torch.long,
+                                            pin_memory=not self.in_wsl)
+        categorized_sample_indices = {
+            t: _async_h2d(seq_ids, dtype=torch.int, pin_memory=not self.in_wsl)
+            for t, seq_ids in categorized_sample_indices.items()
+        }
+
+        seq_data: Dict[int, SequenceData] = {}
+        for seq_group_metadata in seq_group_metadata_list:
+            seq_data.update(seq_group_metadata.seq_data)
+
+        sampling_metadata = SamplingMetadata(
+            seq_groups=seq_groups,
+            seq_data=seq_data,
+            prompt_lens=prompt_lens,
+            selected_token_indices=selected_token_indices,
+            categorized_sample_indices=categorized_sample_indices,
+        )
+        return sampling_metadata
+
+    @torch.inference_mode()
+    def execute_model(
+        self,
+        seq_group_metadata_list: List[SequenceGroupMetadata],
+        kv_caches: List[Tuple[torch.Tensor, torch.Tensor]],
+    ) -> SamplerOutput:
+        # NOTE: We assume that all sequences in the group are all prompts or
+        # all decodes.
+        is_prompt = seq_group_metadata_list[0].is_prompt
+        # Prepare input tensors.
+        if is_prompt:
+            inputs = self._prepare_prompt(seq_group_metadata_list)
+            input_tokens, input_positions, input_metadata = inputs
+        else:
+            inputs = self._prepare_decode(seq_group_metadata_list)
+            input_tokens, input_positions, input_metadata = inputs
+
+        # Execute the model.
+        if input_metadata.use_cuda_graph:
+            graph_batch_size = input_tokens.shape[0]
+            model_executable = self.graph_runners[graph_batch_size]
+        else:
+            model_executable = self.model
+        hidden_states = model_executable(
+            input_ids=input_tokens,
+            positions=input_positions,
+            kv_caches=kv_caches,
+            input_metadata=input_metadata,
+        )
+
+        sampling_metadata = self._prepare_sample(seq_group_metadata_list,
+                                                 input_metadata.prompt_lens)
+
+        # Sample the next token.
+        output = self.model.sample(
+            hidden_states=hidden_states,
+            sampling_metadata=sampling_metadata,
+        )
+        return output
+
+    @torch.inference_mode()
+    def profile_run(self) -> None:
+        # Enable top-k sampling to reflect the accurate memory usage.
+        vocab_size = self.model_config.get_vocab_size()
+        sampling_params = SamplingParams(top_p=0.99, top_k=vocab_size - 1)
+        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
+        max_num_seqs = self.scheduler_config.max_num_seqs
+
+        # Profile memory usage with max_num_sequences sequences and the total
+        # number of tokens equal to max_num_batched_tokens.
+        seqs: List[SequenceGroupMetadata] = []
+        for group_id in range(max_num_seqs):
+            seq_len = (max_num_batched_tokens // max_num_seqs +
+                       (group_id < max_num_batched_tokens % max_num_seqs))
+            seq_data = SequenceData([0] * seq_len)
+            seq = SequenceGroupMetadata(
+                request_id=str(group_id),
+                is_prompt=True,
+                seq_data={group_id: seq_data},
+                sampling_params=sampling_params,
+                block_tables=None,
+            )
+            seqs.append(seq)
+
+        # Run the model with the dummy inputs.
+        num_layers = self.model_config.get_num_layers(self.parallel_config)
+        kv_caches = [(None, None)] * num_layers
+        self.execute_model(seqs, kv_caches)
+        torch.cuda.synchronize()
+        return
+
+    @torch.inference_mode()
+    def capture_model(self, kv_caches: List[KVCache]) -> None:
+        assert not self.model_config.enforce_eager
+        logger.info("Capturing the model for CUDA graphs. This may lead to "
+                    "unexpected consequences if the model is not static. To "
+                    "run the model in eager mode, set 'enforce_eager=True' or "
+                    "use '--enforce-eager' in the CLI.")
+        logger.info("CUDA graphs can take additional 1~3 GiB memory per GPU. "
+                    "If you are running out of memory, consider decreasing "
+                    "`gpu_memory_utilization` or enforcing eager mode.")
+        start_time = time.perf_counter()
+
+        # Prepare dummy inputs. These will be reused for all batch sizes.
+        max_batch_size = max(_BATCH_SIZES_TO_CAPTURE)
+        input_tokens = torch.zeros(max_batch_size, 1, dtype=torch.long).cuda()
+        input_positions = torch.zeros(max_batch_size, 1,
+                                      dtype=torch.long).cuda()
+        slot_mapping = torch.empty(max_batch_size, 1, dtype=torch.long).cuda()
+        slot_mapping.fill_(_PAD_SLOT_ID)
+        context_lens = torch.ones(max_batch_size, dtype=torch.int32).cuda()
+        block_tables = torch.from_numpy(self.graph_block_tables).cuda()
+
+        # NOTE: Capturing the largest batch size first may help reduce the
+        # memory usage of CUDA graph.
+        for batch_size in reversed(_BATCH_SIZES_TO_CAPTURE):
+            # Create dummy input_metadata.
+            input_metadata = InputMetadata(
+                prompt_lens=[],
+                slot_mapping=slot_mapping[:batch_size],
+                max_context_len=self.max_context_len_to_capture,
+                context_lens=context_lens[:batch_size],
+                block_tables=block_tables[:batch_size],
+                use_cuda_graph=True,
+            )
+
+            graph_runner = CUDAGraphRunner(self.model)
+            graph_runner.capture(
+                input_tokens[:batch_size],
+                input_positions[:batch_size],
+                kv_caches,
+                input_metadata,
+                memory_pool=self.graph_memory_pool,
+            )
+            self.graph_memory_pool = graph_runner.graph.pool()
+            self.graph_runners[batch_size] = graph_runner
+
+        end_time = time.perf_counter()
+        elapsed_time = end_time - start_time
+        # This usually takes < 10 seconds.
+        logger.info(f"Graph capturing finished in {elapsed_time:.0f} secs.")
+
+
+class CUDAGraphRunner:
+
+    def __init__(self, model: nn.Module):
+        self.model = model
+        self.graph = None
+        self.input_buffers: Dict[str, torch.Tensor] = {}
+        self.output_buffers: Dict[str, torch.Tensor] = {}
+
+    def capture(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+        memory_pool,
+    ) -> None:
+        assert self.graph is None
+        # Run the model once without capturing the graph.
+        # This is to make sure that the captured graph does not include the
+        # kernel launches for initial benchmarking (e.g., Triton autotune).
+        self.model(
+            input_ids,
+            positions,
+            kv_caches,
+            input_metadata,
+        )
+        torch.cuda.synchronize()
+
+        # Capture the graph.
+        self.graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(self.graph, pool=memory_pool):
+            hidden_states = self.model(
+                input_ids,
+                positions,
+                kv_caches,
+                input_metadata,
+            )
+        torch.cuda.synchronize()
+
+        # Save the input and output buffers.
+        self.input_buffers = {
+            "input_ids": input_ids,
+            "positions": positions,
+            "kv_caches": kv_caches,
+            "slot_mapping": input_metadata.slot_mapping,
+            "context_lens": input_metadata.context_lens,
+            "block_tables": input_metadata.block_tables,
+        }
+        self.output_buffers = {"hidden_states": hidden_states}
+        return
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[Tuple[torch.Tensor, torch.Tensor]],
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        # KV caches are fixed tensors, so we don't need to copy them.
+        del kv_caches
+
+        # Copy the input tensors to the input buffers.
+        self.input_buffers["input_ids"].copy_(input_ids, non_blocking=True)
+        self.input_buffers["positions"].copy_(positions, non_blocking=True)
+        self.input_buffers["slot_mapping"].copy_(input_metadata.slot_mapping,
+                                                 non_blocking=True)
+        self.input_buffers["context_lens"].copy_(input_metadata.context_lens,
+                                                 non_blocking=True)
+        self.input_buffers["block_tables"].copy_(input_metadata.block_tables,
+                                                 non_blocking=True)
+
+        # Run the graph.
+        self.graph.replay()
+
+        # Return the output tensor.
+        return self.output_buffers["hidden_states"]
+
+    def __call__(self, *args, **kwargs):
+        return self.forward(*args, **kwargs)
+
+
+def _pad_to_max(x: List[int], max_len: int, pad: int) -> List[int]:
+    assert len(x) <= max_len
+    return x + [pad] * (max_len - len(x))
+
+
+def _make_tensor_with_pad(
+    x: List[List[int]],
+    max_len: int,
+    pad: int,
+    dtype: torch.dtype,
+    device: Union[str, torch.device] = "cuda",
+    pin_memory: bool = False,
+) -> torch.Tensor:
+    padded_x = [_pad_to_max(x_i, max_len, pad) for x_i in x]
+    return torch.tensor(padded_x,
+                        dtype=dtype,
+                        device=device,
+                        pin_memory=pin_memory and str(device) == "cpu")
+
+
+def _get_graph_batch_size(batch_size: int) -> int:
+    if batch_size <= 2:
+        return batch_size
+    elif batch_size <= 4:
+        return 4
+    else:
+        return (batch_size + 7) // 8 * 8
+
+
+def _async_h2d(data: list, dtype, pin_memory):
+    t = torch.tensor(data, dtype=dtype, pin_memory=pin_memory)
+    return t.to(device="cuda", non_blocking=True)