vllm-npu 0.4.2__py3-none-any.whl

vllm/transformers_utils/tokenizer_group/ray_tokenizer_group.py

@@ -0,0 +1,169 @@
+import asyncio
+import os
+from typing import List, Optional
+
+from ray.util.scheduling_strategies import NodeAffinitySchedulingStrategy
+from transformers import PreTrainedTokenizer
+
+from vllm.config import TokenizerPoolConfig
+from vllm.executor.ray_utils import ray
+from vllm.lora.request import LoRARequest
+from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
+    BaseTokenizerGroup)
+from vllm.transformers_utils.tokenizer_group.tokenizer_group import (
+    TokenizerGroup)
+
+
+class RayTokenizerGroupPool(BaseTokenizerGroup):
+    """A Ray-based pool of TokenizerGroups for async tokenization."""
+
+    # Class to use for workers making up the pool.
+    _worker_cls = TokenizerGroup
+
+    @classmethod
+    def from_config(cls, tokenizer_pool_config: TokenizerPoolConfig,
+                    **init_kwargs) -> "RayTokenizerGroupPool":
+        ray_actor_options = (tokenizer_pool_config.extra_config or {
+            "num_cpus": 0
+        })
+        ray_actor_options.setdefault(
+            "scheduling_strategy",
+            NodeAffinitySchedulingStrategy(
+                node_id=ray.get_runtime_context().get_node_id(), soft=True))
+
+        # Carry over the env vars to the actors.
+        # This is necessary for API keys and such.
+        ray_actor_options.setdefault("runtime_env", {})
+        _carry_over_env_vars_to_runtime_env(ray_actor_options["runtime_env"])
+
+        init_kwargs["num_actors"] = tokenizer_pool_config.pool_size
+        init_kwargs["ray_actor_options"] = ray_actor_options
+
+        return cls(**init_kwargs)
+
+    def __init__(self, tokenizer_id: str, enable_lora: bool, max_num_seqs: int,
+                 max_input_length: Optional[int], num_actors: int,
+                 ray_actor_options: dict, **tokenizer_config):
+        # Store a local copy of the TokenizerGroup for quick access
+        # to underlying HF tokenizers.
+        self._local_tokenizer_group = self._worker_cls(
+            tokenizer_id=tokenizer_id,
+            enable_lora=enable_lora,
+            max_num_seqs=max_num_seqs,
+            max_input_length=max_input_length,
+            **tokenizer_config,
+        )
+
+        ray_tokenizer_group_cls = ray.remote(
+            self._worker_cls).options(**ray_actor_options)
+        self.tokenizer_actors = [
+            ray_tokenizer_group_cls.remote(tokenizer_id, enable_lora,
+                                           max_num_seqs, max_input_length,
+                                           **tokenizer_config)
+            for _ in range(num_actors)
+        ]
+        self._idle_actors: Optional[asyncio.Queue] = None
+
+    @property
+    def pool_size(self) -> int:
+        return len(self.tokenizer_actors)
+
+    def ping(self):
+        return ray.get(
+            [actor.ping.remote() for actor in self.tokenizer_actors])
+
+    def _ensure_queue_initialized(self):
+        if self._idle_actors is None:
+            self._idle_actors = asyncio.Queue()
+            for actor in self.tokenizer_actors:
+                self._idle_actors.put_nowait(actor)
+
+    def encode(self,
+               prompt: str,
+               request_id: Optional[str] = None,
+               lora_request: Optional[LoRARequest] = None) -> List[int]:
+        """Encode a prompt using the tokenizer group.
+
+        We pick an idle actor and use it to encode the prompt.
+        The actor is then put back in the queue for future use.
+        This is blocking.
+        """
+        self._ensure_queue_initialized()
+        assert self._idle_actors is not None
+
+        if self._idle_actors.empty():
+            raise RuntimeError("No idle actors available.")
+        actor = self._idle_actors.get_nowait()
+        try:
+            ret = ray.get(
+                actor.encode.remote(request_id=request_id,
+                                    prompt=prompt,
+                                    lora_request=lora_request))
+        finally:
+            # Put the actor back in the queue.
+            # This is done in a finally block to ensure that the actor is
+            # always put back in the queue, even if an exception/cancellation
+            # is raised.
+            self._idle_actors.put_nowait(actor)
+        return ret
+
+    async def encode_async(
+            self,
+            prompt: str,
+            request_id: Optional[str] = None,
+            lora_request: Optional[LoRARequest] = None) -> List[int]:
+        """Encode a prompt using the tokenizer group.
+
+        We pick an idle actor and use it to encode the prompt.
+        If there are no idle actors, we wait until one becomes
+        available.
+        The actor is then put back in the queue for future use.
+        This is non-blocking.
+        """
+        self._ensure_queue_initialized()
+        assert self._idle_actors is not None
+
+        actor = await self._idle_actors.get()
+        try:
+            ret = await actor.encode.remote(request_id=request_id,
+                                            prompt=prompt,
+                                            lora_request=lora_request)
+        finally:
+            # Put the actor back in the queue.
+            # This is done in a finally block to ensure that the actor is
+            # always put back in the queue, even if an exception/cancellation
+            # is raised.
+            self._idle_actors.put_nowait(actor)
+        return ret
+
+    def get_max_input_len(self,
+                          lora_request: Optional[LoRARequest] = None
+                          ) -> Optional[int]:
+        """Get the maximum input length for the LoRA request."""
+        return self._local_tokenizer_group.get_max_input_len(lora_request)
+
+    def get_lora_tokenizer(
+            self,
+            lora_request: Optional[LoRARequest] = None
+    ) -> "PreTrainedTokenizer":
+        return self._local_tokenizer_group.get_lora_tokenizer(lora_request)
+
+    async def get_lora_tokenizer_async(
+            self,
+            lora_request: Optional[LoRARequest] = None
+    ) -> "PreTrainedTokenizer":
+        return await self._local_tokenizer_group.get_lora_tokenizer_async(
+            lora_request)
+
+
+def _carry_over_env_vars_to_runtime_env(runtime_env: dict) -> None:
+    """Copy over all current process environment variables to the runtime_env.
+
+    The variables in runtime_env will take precedence over the current process
+    environment variables.
+
+    runtime_env will be modified in place."""
+    env_vars = os.environ.copy()
+    runtime_env.setdefault("env_vars", {})
+    env_vars.update(runtime_env["env_vars"])
+    runtime_env["env_vars"] = env_vars

vllm/transformers_utils/tokenizer_group/tokenizer_group.py

@@ -0,0 +1,78 @@
+from typing import List, Optional
+
+from transformers import PreTrainedTokenizer
+
+from vllm.lora.request import LoRARequest
+from vllm.transformers_utils.tokenizer import (get_lora_tokenizer,
+                                               get_lora_tokenizer_async,
+                                               get_tokenizer)
+from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
+    BaseTokenizerGroup)
+from vllm.utils import LRUCache
+
+
+class TokenizerGroup(BaseTokenizerGroup):
+    """A group of tokenizers that can be used for LoRA adapters."""
+
+    def __init__(self, tokenizer_id: str, enable_lora: bool, max_num_seqs: int,
+                 max_input_length: Optional[int], **tokenizer_config):
+        self.tokenizer_id = tokenizer_id
+        self.tokenizer_config = tokenizer_config
+        self.enable_lora = enable_lora
+        self.max_input_length = max_input_length
+        self.tokenizer = get_tokenizer(self.tokenizer_id, **tokenizer_config)
+        self.lora_tokenizers = LRUCache[PreTrainedTokenizer](
+            capacity=max_num_seqs) if enable_lora else None
+
+    def ping(self) -> bool:
+        """Check if the tokenizer group is alive."""
+        return True
+
+    def get_max_input_len(self,
+                          lora_request: Optional[LoRARequest] = None
+                          ) -> Optional[int]:
+        """Get the maximum input length for the LoRA request."""
+        return self.max_input_length
+
+    def encode(self,
+               prompt: str,
+               request_id: Optional[str] = None,
+               lora_request: Optional[LoRARequest] = None) -> List[int]:
+        tokenizer = self.get_lora_tokenizer(lora_request)
+        return tokenizer.encode(prompt)
+
+    async def encode_async(
+            self,
+            prompt: str,
+            request_id: Optional[str] = None,
+            lora_request: Optional[LoRARequest] = None) -> List[int]:
+        tokenizer = await self.get_lora_tokenizer_async(lora_request)
+        return tokenizer.encode(prompt)
+
+    def get_lora_tokenizer(
+            self,
+            lora_request: Optional[LoRARequest] = None
+    ) -> "PreTrainedTokenizer":
+        if not lora_request or not self.enable_lora:
+            return self.tokenizer
+        if lora_request.lora_int_id not in self.lora_tokenizers:
+            tokenizer = (get_lora_tokenizer(
+                lora_request, **self.tokenizer_config) or self.tokenizer)
+            self.lora_tokenizers.put(lora_request.lora_int_id, tokenizer)
+            return tokenizer
+        else:
+            return self.lora_tokenizers.get(lora_request.lora_int_id)
+
+    async def get_lora_tokenizer_async(
+            self,
+            lora_request: Optional[LoRARequest] = None
+    ) -> "PreTrainedTokenizer":
+        if not lora_request or not self.enable_lora:
+            return self.tokenizer
+        if lora_request.lora_int_id not in self.lora_tokenizers:
+            tokenizer = (await get_lora_tokenizer_async(
+                lora_request, **self.tokenizer_config) or self.tokenizer)
+            self.lora_tokenizers.put(lora_request.lora_int_id, tokenizer)
+            return tokenizer
+        else:
+            return self.lora_tokenizers.get(lora_request.lora_int_id)

vllm/transformers_utils/tokenizers/__init__.py

@@ -0,0 +1,5 @@
+from vllm.transformers_utils.tokenizers.baichuan import BaichuanTokenizer
+
+__all__ = [
+    "BaichuanTokenizer",
+]

vllm/transformers_utils/tokenizers/baichuan.py

@@ -0,0 +1,255 @@
+# 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 = { # type: ignore
+    "vocab_file": {},
+    "tokenizer_file": {},
+}
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}  # type: ignore
+
+
+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: List[str]):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens: List[str] = []
+        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):
+            raise ValueError(f"Vocabulary path ({save_directory}) "
+                             "should be a directory")
+
+        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