vllm-cpu-avx512bf16 0.9.0.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/transformers_utils/detokenizer.py

@@ -0,0 +1,167 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional
+
+from vllm.sequence import (VLLM_INVALID_TOKEN_ID, Logprob, SamplingParams,
+                           Sequence, SequenceGroup)
+
+from .detokenizer_utils import (convert_prompt_ids_to_tokens,
+                                detokenize_incrementally)
+from .tokenizer import AnyTokenizer
+from .tokenizer_group import TokenizerGroup
+
+
+class Detokenizer:
+    """Provides methods to decode the output of a model into text."""
+
+    def __init__(self, tokenizer_group: TokenizerGroup):
+        self.tokenizer_group = tokenizer_group
+
+    def get_tokenizer_for_seq(self, sequence: Sequence) -> AnyTokenizer:
+        """Returns the HF tokenizer to use for a given sequence."""
+        return self.tokenizer_group.get_lora_tokenizer(sequence.lora_request)
+
+    def decode_prompt_logprobs_inplace(self, seq_group: SequenceGroup,
+                                       prompt_logprobs: list[Optional[dict[
+                                           int, Logprob]]],
+                                       position_offset: int) -> None:
+        """Decodes the logprobs for the prompt of a sequence group.
+
+        Args:
+            seq_group: The sequence group to decode.
+            prompt_logprobs: The logprobs to decode.
+            position_offset: Offset of the first index of the logprobs 
+                relative to the start of the sequence (for chunked prefill).
+        
+        Returns:
+            The prompt logprobs with the decoded tokens.
+        """
+        prms = seq_group.sampling_params
+        assert prms is not None
+
+        # We can pick any sequence for the prompt.
+        seq = seq_group.get_seqs()[0]
+        # Only prompt, without the generated token.
+        all_token_ids = seq.get_token_ids()
+        prompt_token_ids = all_token_ids[:-1]
+        tokenizer = self.get_tokenizer_for_seq(seq)
+        prefix_offset = 0
+        read_offset = 0
+        next_iter_prefix_offset = 0
+        next_iter_read_offset = 0
+        next_iter_tokens: list[str] = []
+        prev_tokens = None
+
+        for token_position_in_logprob, prompt_logprobs_for_token in enumerate(
+                prompt_logprobs):
+
+            # Absolute token position equals the index in the logprobs
+            # list plus the offset of the entire logprobs list relative
+            # to the start of the sequence.
+            token_position = token_position_in_logprob + position_offset
+            if not prompt_logprobs_for_token:
+                continue
+            for token_id, sample_logprob in prompt_logprobs_for_token.items():
+                if (sample_logprob.decoded_token is None
+                        and token_id != VLLM_INVALID_TOKEN_ID):
+                    prompt_token_ids_with_token = (
+                        prompt_token_ids[:token_position] + [token_id])
+                    (new_tokens, new_text, new_prefix_offset,
+                     new_read_offset) = detokenize_incrementally(
+                         tokenizer=tokenizer,
+                         all_input_ids=prompt_token_ids_with_token,
+                         prev_tokens=prev_tokens,
+                         prefix_offset=prefix_offset,
+                         read_offset=read_offset,
+                         skip_special_tokens=prms.skip_special_tokens,
+                         spaces_between_special_tokens=prms.
+                         spaces_between_special_tokens,
+                     )
+
+                    sample_logprob.decoded_token = new_text
+
+                    # Use the offsets & prev tokens corresponding to
+                    # real tokens to ensure detokenization is consistent
+                    # actual with prompt.
+                    if token_id == all_token_ids[token_position]:
+                        next_iter_prefix_offset = new_prefix_offset
+                        next_iter_read_offset = new_read_offset
+                        next_iter_tokens = new_tokens
+
+            # Advance to the next token position.
+            prefix_offset = next_iter_prefix_offset
+            read_offset = next_iter_read_offset
+            if prev_tokens is None:
+                prev_tokens = next_iter_tokens.copy()
+            else:
+                prev_tokens.extend(next_iter_tokens)
+
+    def decode_sequence_inplace(self, seq: Sequence,
+                                prms: SamplingParams) -> int:
+        """Decodes the new token for a sequence. In-place operation.
+
+        Args:
+            seq: The sequence to decode.
+            prms: The sampling parameters used to generate the sequence.
+
+        Returns:
+            The number of characters added to the output text.
+        """
+        all_input_ids = seq.get_token_ids()
+        token_id_generated_this_iteration = all_input_ids[-1]
+        tokenizer = self.get_tokenizer_for_seq(seq)
+
+        # Convert prompt token IDs to tokens if necessary.
+        # Do it here so that we don't have to repeat this
+        # computation for each logprob.
+        if seq.tokens is None:
+            (seq.tokens, seq.prefix_offset,
+             seq.read_offset) = convert_prompt_ids_to_tokens(
+                 tokenizer=tokenizer,
+                 prompt_ids=all_input_ids[:-1],
+                 skip_special_tokens=prms.skip_special_tokens,
+             )
+
+        (new_tokens, new_decoded_token_text, prefix_offset,
+         read_offset) = detokenize_incrementally(
+             tokenizer=tokenizer,
+             all_input_ids=all_input_ids,
+             prev_tokens=seq.tokens,
+             prefix_offset=seq.prefix_offset,
+             read_offset=seq.read_offset,
+             skip_special_tokens=prms.skip_special_tokens,
+             spaces_between_special_tokens=prms.spaces_between_special_tokens,
+         )
+
+        # Decode logprobs
+        logprobs = seq.output_logprobs[-1]
+        if logprobs:
+            previous_tokens = all_input_ids[:-1]
+            for token_id, sample_logprob in logprobs.items():
+                # If the token was generated this iteration,
+                # use the provided text.
+                if token_id == token_id_generated_this_iteration:
+                    sample_logprob.decoded_token = new_decoded_token_text
+                    continue
+
+                if (sample_logprob.decoded_token is None
+                        and token_id != VLLM_INVALID_TOKEN_ID):
+                    all_input_ids_with_logprob = previous_tokens + [token_id]
+                    (_, new_text, _, _) = detokenize_incrementally(
+                        tokenizer=tokenizer,
+                        all_input_ids=all_input_ids_with_logprob,
+                        prev_tokens=seq.tokens,
+                        prefix_offset=seq.prefix_offset,
+                        read_offset=seq.read_offset,
+                        skip_special_tokens=prms.skip_special_tokens,
+                        spaces_between_special_tokens=prms.
+                        spaces_between_special_tokens,
+                    )
+                    sample_logprob.decoded_token = new_text
+
+        seq.tokens.extend(new_tokens)
+        seq.prefix_offset = prefix_offset
+        seq.read_offset = read_offset
+        seq.output_text += new_decoded_token_text
+
+        return len(new_decoded_token_text)

vllm/transformers_utils/detokenizer_utils.py

@@ -0,0 +1,188 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional
+
+from .tokenizer import AnyTokenizer
+
+
+def _replace_none_with_empty(tokens: list[Optional[str]]):
+    for i, token in enumerate(tokens):
+        if token is None:
+            tokens[i] = ""
+
+
+def _convert_tokens_to_string_with_added_encoders(
+    tokenizer: AnyTokenizer,
+    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: list[str] = []
+    current_sub_text: list[str] = []
+    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)
+
+
+# 5 is an arbitrary value that should work for all
+# tokenizers (bigger = more conservative).
+INITIAL_INCREMENTAL_DETOKENIZATION_OFFSET = 5
+
+
+def convert_prompt_ids_to_tokens(
+    tokenizer: AnyTokenizer,
+    prompt_ids: list[int],
+    skip_special_tokens: bool = False,
+) -> tuple[list[str], int, int]:
+    """Converts the prompt ids to tokens and returns the tokens and offsets
+    for incremental detokenization.
+
+    Note that not all tokens are converted to strings. Only the tokens that
+    are necessary for incremental detokenization are converted to strings.
+    """
+    # We do not need to convert the whole prompt to tokens.
+    # Offset a little more in case we have special tokens.
+    new_tokens = tokenizer.convert_ids_to_tokens(
+        prompt_ids[-INITIAL_INCREMENTAL_DETOKENIZATION_OFFSET - 2:],
+        skip_special_tokens=skip_special_tokens)
+    read_offset = len(new_tokens)
+    prefix_offset = max(
+        read_offset - INITIAL_INCREMENTAL_DETOKENIZATION_OFFSET, 0)
+    # This is required to guard against out-of-vocab prompt token ids
+    _replace_none_with_empty(new_tokens)  # type: ignore[arg-type]
+    return new_tokens, prefix_offset, read_offset
+
+
+def convert_ids_list_to_tokens(
+    tokenizer: AnyTokenizer,
+    token_ids: list[int],
+) -> list[str]:
+    """Detokenize the input ids individually.
+
+    Args:
+      tokenizer: tokenizer used by model under test
+      token_ids: convert these tokens (Python list form)
+
+    Returns:
+      Python list of token string representations
+    
+    """
+    token_str_lst = tokenizer.convert_ids_to_tokens(token_ids)
+    _replace_none_with_empty(token_str_lst)  # type: ignore
+    return token_str_lst
+
+
+# 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: AnyTokenizer,
+    all_input_ids: list[int],
+    prev_tokens: Optional[list[str]],
+    prefix_offset: int,
+    read_offset: int,
+    skip_special_tokens: bool = False,
+    spaces_between_special_tokens: bool = True,
+) -> tuple[list[str], str, int, int]:
+    """Detokenizes the input ids incrementally and returns the new tokens
+    and the new text.
+
+    If `prev_tokens` is None, this function will convert the input ids to
+    tokens and return the tokens and the new text. Otherwise, it will return the
+    new tokens and the new text.
+
+    This function will also return the new prefix offset and the new read
+    offset to be used in the next iteration.
+
+    The offsets are necessary to defeat cleanup algorithms in the decode which
+    decide to add a space or not depending on the surrounding ids.
+
+    Args:
+        tokenizer: The tokenizer to use.
+        all_input_ids: The input ids. The last id is the new token id.
+        prev_tokens: The previous tokens. If None, this function will convert
+            the input ids to tokens and return the tokens and the new text.
+        prefix_offset: The prefix offset.
+        read_offset: The read offset.
+        skip_special_tokens: Whether to skip special tokens.
+        spaces_between_special_tokens: Whether to add spaces between special
+            tokens.
+    """
+    new_token_id = all_input_ids[-1]
+    # This is the first iteration for this sequence
+    is_first_iter = prev_tokens is None
+    if is_first_iter:
+        (prev_tokens, prefix_offset,
+         read_offset) = convert_prompt_ids_to_tokens(
+             tokenizer,
+             all_input_ids[:-1],
+             skip_special_tokens=skip_special_tokens)
+    assert prev_tokens is not None
+
+    # If the new token id is out of bounds, return an empty string.
+    if 0 <= new_token_id < len(tokenizer):
+        # 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)
+        if isinstance(new_tokens, str):
+            new_tokens = [new_tokens]
+    else:
+        new_tokens = [""]
+    output_tokens = prev_tokens + new_tokens
+
+    # If this is the first iteration, return all tokens.
+    if is_first_iter:
+        new_tokens = output_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) or 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
+        return new_tokens, "", prefix_offset, read_offset
+
+    new_text = new_text[len(prefix_text):]
+    return new_tokens, new_text, read_offset, len(output_tokens)

vllm/transformers_utils/processor.py

@@ -0,0 +1,220 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from functools import lru_cache
+from typing import TYPE_CHECKING, Any, Optional, Union, cast
+
+from transformers.processing_utils import ProcessorMixin
+from typing_extensions import TypeVar
+
+if TYPE_CHECKING:
+    from vllm.config import ModelConfig
+
+_P = TypeVar("_P", bound=ProcessorMixin, default=ProcessorMixin)
+
+
+class HashableDict(dict):
+    """
+    A dictionary that can be hashed by lru_cache.
+    """
+
+    # NOTE: pythonic dict is not hashable,
+    # we override on it directly for simplicity
+    def __hash__(self) -> int:  # type: ignore[override]
+        return hash(frozenset(self.items()))
+
+
+class HashableList(list):
+    """
+    A list that can be hashed by lru_cache.
+    """
+
+    def __hash__(self) -> int:  # type: ignore[override]
+        return hash(tuple(self))
+
+
+def _merge_mm_kwargs(model_config: "ModelConfig", **kwargs):
+    mm_config = model_config.get_multimodal_config()
+    base_kwargs = mm_config.mm_processor_kwargs
+    if base_kwargs is None:
+        base_kwargs = {}
+
+    merged_kwargs = {**base_kwargs, **kwargs}
+
+    # NOTE: Pythonic dict is not hashable and will raise unhashable type
+    # error when calling `cached_get_processor`, therefore we need to
+    # wrap it to a hashable dict.
+    for key, value in merged_kwargs.items():
+        if isinstance(value, dict):
+            merged_kwargs[key] = HashableDict(value)
+        if isinstance(value, list):
+            merged_kwargs[key] = HashableList(value)
+    return merged_kwargs
+
+
+def get_processor(
+    processor_name: str,
+    *args: Any,
+    revision: Optional[str] = None,
+    trust_remote_code: bool = False,
+    processor_cls: Union[type[_P], tuple[type[_P], ...]] = ProcessorMixin,
+    **kwargs: Any,
+) -> _P:
+    """Load a processor for the given model name via HuggingFace."""
+    # don't put this import at the top level
+    # it will call torch.cuda.device_count()
+    from transformers import AutoProcessor
+
+    processor_factory = (AutoProcessor if processor_cls == ProcessorMixin or
+                         isinstance(processor_cls, tuple) else processor_cls)
+
+    try:
+        processor = processor_factory.from_pretrained(
+            processor_name,
+            *args,
+            revision=revision,
+            trust_remote_code=trust_remote_code,
+            **kwargs,
+        )
+    except ValueError as e:
+        # If the error pertains to the processor class not existing or not
+        # currently being imported, suggest using the --trust-remote-code flag.
+        # Unlike AutoTokenizer, AutoProcessor does not separate such errors
+        if not trust_remote_code:
+            err_msg = (
+                "Failed to load the processor. If the processor is "
+                "a custom processor 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 not isinstance(processor, processor_cls):
+        raise TypeError("Invalid type of HuggingFace processor. "
+                        f"Expected type: {processor_cls}, but "
+                        f"found type: {type(processor)}")
+
+    return processor
+
+
+cached_get_processor = lru_cache(get_processor)
+
+
+def cached_processor_from_config(
+    model_config: "ModelConfig",
+    processor_cls: Union[type[_P], tuple[type[_P], ...]] = ProcessorMixin,
+    **kwargs: Any,
+) -> _P:
+    return cached_get_processor(
+        model_config.model,
+        revision=model_config.revision,
+        trust_remote_code=model_config.trust_remote_code,
+        processor_cls=processor_cls,  # type: ignore[arg-type]
+        **_merge_mm_kwargs(model_config, **kwargs),
+    )
+
+
+def get_feature_extractor(
+    processor_name: str,
+    *args: Any,
+    revision: Optional[str] = None,
+    trust_remote_code: bool = False,
+    **kwargs: Any,
+):
+    """Load an audio feature extractor for the given model name 
+    via HuggingFace."""
+    # don't put this import at the top level
+    # it will call torch.cuda.device_count()
+    from transformers import AutoFeatureExtractor
+    from transformers.feature_extraction_utils import FeatureExtractionMixin
+    try:
+        feature_extractor = AutoFeatureExtractor.from_pretrained(
+            processor_name,
+            *args,
+            revision=revision,
+            trust_remote_code=trust_remote_code,
+            **kwargs)
+    except ValueError as e:
+        # If the error pertains to the processor class not existing or not
+        # currently being imported, suggest using the --trust-remote-code flag.
+        # Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
+        if not trust_remote_code:
+            err_msg = (
+                "Failed to load the feature extractor. If the feature "
+                "extractor is a custom extractor 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
+    return cast(FeatureExtractionMixin, feature_extractor)
+
+
+cached_get_feature_extractor = lru_cache(get_feature_extractor)
+
+
+def cached_feature_extractor_from_config(
+    model_config: "ModelConfig",
+    **kwargs: Any,
+):
+    return cached_get_feature_extractor(
+        model_config.model,
+        revision=model_config.revision,
+        trust_remote_code=model_config.trust_remote_code,
+        **_merge_mm_kwargs(model_config, **kwargs),
+    )
+
+
+def get_image_processor(
+    processor_name: str,
+    *args: Any,
+    revision: Optional[str] = None,
+    trust_remote_code: bool = False,
+    **kwargs: Any,
+):
+    """Load an image processor for the given model name via HuggingFace."""
+    # don't put this import at the top level
+    # it will call torch.cuda.device_count()
+    from transformers import AutoImageProcessor
+    from transformers.image_processing_utils import BaseImageProcessor
+
+    try:
+        processor = AutoImageProcessor.from_pretrained(
+            processor_name,
+            *args,
+            revision=revision,
+            trust_remote_code=trust_remote_code,
+            **kwargs)
+    except ValueError as e:
+        # If the error pertains to the processor class not existing or not
+        # currently being imported, suggest using the --trust-remote-code flag.
+        # Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
+        if not trust_remote_code:
+            err_msg = (
+                "Failed to load the image processor. If the image processor is "
+                "a custom processor 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
+
+    return cast(BaseImageProcessor, processor)
+
+
+cached_get_image_processor = lru_cache(get_image_processor)
+
+
+def cached_image_processor_from_config(
+    model_config: "ModelConfig",
+    **kwargs: Any,
+):
+    return cached_get_image_processor(
+        model_config.model,
+        revision=model_config.revision,
+        trust_remote_code=model_config.trust_remote_code,
+        **_merge_mm_kwargs(model_config, **kwargs),
+    )

vllm/transformers_utils/processors/__init__.py

@@ -0,0 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.transformers_utils.processors.deepseek_vl2 import (
+    DeepseekVLV2Processor)
+from vllm.transformers_utils.processors.ovis import OvisProcessor
+
+__all__ = ["DeepseekVLV2Processor", "OvisProcessor"]