vllm-cpu-avx512vnni 0.10.2.post2__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/transformers_utils/configs/step3_vl.py

@@ -0,0 +1,123 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any, Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class Step3VisionEncoderConfig(PretrainedConfig):
+    model_type = "step3_vision_encoder"
+
+    def __init__(
+        self,
+        hidden_size=1792,
+        intermediate_size=3072,
+        output_hidden_size=4096,
+        num_hidden_layers=63,
+        num_attention_heads=16,
+        num_channels=3,
+        image_size=728,
+        patch_size=14,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.output_hidden_size = output_hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        super().__init__(**kwargs)
+
+
+class Step3TextConfig(PretrainedConfig):
+    model_type = "step3_text"
+    architectures = ["Step3TextForCausalLM"]
+
+    def __init__(
+        self,
+        hidden_size: int = 7168,
+        intermediate_size: int = 18432,
+        num_attention_heads: int = 64,
+        num_attention_groups: int = 1,
+        num_hidden_layers: int = 61,
+        max_seq_len: int = 65536,
+        vocab_size: int = 128815,
+        rms_norm_eps: float = 1e-5,
+        moe_intermediate_size: int = 5120,
+        moe_num_experts: int = 48,
+        moe_top_k: int = 3,
+        rope_theta: float = 500000,
+        rope_scaling: Optional[dict[str, Any]] = None,
+        max_position_embedding: int = 65536,
+        share_expert_dim: int = 5120,
+        share_q_dim: int = 2048,
+        head_dim: int = 256,
+        norm_expert_weight: bool = False,
+        moe_layers_enum: tuple[int,
+                               ...] = (4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+                                       15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
+                                       25, 26, 27, 28, 29, 30, 31, 32, 33, 34,
+                                       35, 36, 37, 38, 39, 40, 41, 42, 43, 44,
+                                       45, 46, 47, 48, 49, 50, 51, 52, 53, 54,
+                                       55, 56, 57, 58, 59),
+        **kwargs,
+    ) -> None:
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_groups = num_attention_groups
+        self.num_hidden_layers = num_hidden_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.rms_norm_eps = rms_norm_eps
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.max_position_embedding = max_position_embedding
+        self.share_expert_dim = share_expert_dim
+        self.share_q_dim = share_q_dim
+        self.head_dim = head_dim
+        self.norm_expert_weight = norm_expert_weight
+        self.moe_layers_enum = moe_layers_enum
+
+        super().__init__(**kwargs)
+
+
+class Step3VLConfig(PretrainedConfig):
+    model_type = "step3_vl"
+
+    def __init__(
+        self,
+        vision_config: Optional[Union[dict, Step3VisionEncoderConfig]] = None,
+        text_config: Optional[Union[dict, Step3TextConfig]] = None,
+        understand_projector_stride: int = 1,
+        projector_bias: bool = True,
+        image_token_id: int = 128001,
+        **kwargs,
+    ) -> None:
+        if vision_config is None:
+            vision_config = Step3VisionEncoderConfig()
+        elif isinstance(vision_config, dict):
+            vision_config = Step3VisionEncoderConfig(**vision_config)
+        self.vision_config = vision_config
+
+        if text_config is None:
+            text_config = Step3TextConfig()
+        elif isinstance(text_config, dict):
+            text_config = Step3TextConfig(**text_config)
+        self.text_config = text_config
+
+        self.understand_projector_stride = understand_projector_stride
+        self.projector_bias = projector_bias
+        self.hidden_size = text_config.hidden_size
+        self.image_token_id = image_token_id
+
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/ultravox.py

@@ -0,0 +1,120 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from https://github.com/fixie-ai/ultravox/blob/ecd58c4041030bae2ad15aa6bcf04ab43199ea02/ultravox/model/ultravox_config.py
+from typing import Any, Optional
+
+import transformers
+
+
+class UltravoxConfig(transformers.PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a
+    [`UltravoxForConditionalGeneration`]. It is used to instantiate an
+    Ultravox model according to the specified arguments, defining the model
+    architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to
+    control the model outputs. Read the documentation from [`PretrainedConfig`]
+    for more information.
+
+    Args:
+        audio_config (`Union[AutoConfig, dict]`,  *optional*):
+            Custom audio config or dict.
+        text_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object of the text backbone.
+        audio_model_id (`str`, *optional*):
+            The model ID of the audio backbone.
+        text_model_id (`str`, *optional*):
+            The model ID of the text backbone.
+        ignore_index (`int`, *optional*, defaults to -100):
+            The ignore index for the loss function.
+        audio_token_index (`int`, *optional*, defaults to 32000):
+            The audio token index to encode the audio prompt.
+        stack_factor (`int`, *optional*, defaults to 8):
+            Audio downsampling factor for the multimodal projector.
+        norm_init (`float`, *optional*, defaults to 0.4):
+            The initialization value for the layer normalization.
+        projector_act (`str`, *optional*, defaults to `"swiglu"`):
+            The activation function used by the multimodal projector.
+        text_model_lora_config (`LoraConfigSimplified`, *optional*):
+            The LoRA configuration for finetuning the text model.
+        audio_model_lora_config (`LoraConfigSimplified`, *optional*):
+            The LoRA configuration for finetuning the audio model.
+        projector_ln_mid (`bool`, *optional*, defaults to `False`):
+            Whether to apply layer normalization at the middle of the
+            projector or at the end. Versions v0.4.1 and below
+            use `False`, but v0.5 and above use `True`.
+    """
+    wrapped_model_config: transformers.PretrainedConfig
+    model_type = "ultravox"
+    audio_token = "<|audio|>"
+    is_composition = False
+
+    def __init__(
+        self,
+        audio_config: Optional[dict[str, Any]] = None,
+        text_config: Optional[dict[str, Any]] = None,
+        audio_model_id: Optional[str] = None,
+        text_model_id: Optional[str] = None,
+        ignore_index: int = -100,
+        audio_token_index: int = 32000,
+        hidden_size: int = 4096,
+        stack_factor: int = 8,
+        norm_init: float = 0.4,
+        projector_act: str = "swiglu",
+        projector_ln_mid: bool = False,
+        **kwargs,
+    ):
+        self.ignore_index = ignore_index
+        self.audio_token_index = audio_token_index
+
+        self.hidden_size = hidden_size
+        self.stack_factor = stack_factor
+        self.norm_init = norm_init
+        self.projector_act = projector_act
+        self.projector_ln_mid = projector_ln_mid
+
+        # N.B. May set the wrapped_model_config below.
+        self.text_model_id = text_model_id
+        if text_model_id is None:
+            text_config = text_config or {}
+            self.wrapped_model_config = transformers.CONFIG_MAPPING[
+                text_config.get("model_type", "llama")](**text_config)
+
+        # N.B. May set the audio_config below.
+        self.audio_model_id = audio_model_id
+        if audio_model_id is None:
+            self.audio_model_id = None
+            audio_config = audio_config or {}
+            self.audio_config = transformers.CONFIG_MAPPING[audio_config.get(
+                "model_type", "whisper")](**audio_config)
+
+        super().__init__(**kwargs)
+
+    def __setattr__(self, key, value):
+        # Since --hf-overrides are applied _after_ the UltravoxConfig is
+        # instantiated, load the configs implicitly when assigning text_model_id
+        # or audio_model_id. This allows:
+        #
+        #   --hf-overrides.text_model_id=<quantized variant>
+        #
+        # to behave as intended.
+        if key == "text_model_id" and value is not None:
+            from vllm.transformers_utils.config import get_config
+
+            self.wrapped_model_config = get_config(value,
+                                                   trust_remote_code=False)
+        elif key == "audio_model_id" and value is not None:
+            from vllm.transformers_utils.config import get_config
+
+            self.audio_config = get_config(value, trust_remote_code=False)
+
+        return super().__setattr__(key, value)
+
+    @property
+    def text_config(self) -> transformers.PretrainedConfig:
+        # When Ultravox wraps a multi-modal model (e.g. Gemma), we instantiate
+        # the full model, but the text config is the text config of the inner
+        # model.
+        return self.wrapped_model_config.get_text_config()

vllm/transformers_utils/detokenizer.py

@@ -0,0 +1,169 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Optional
+
+from vllm.logprobs import Logprob
+from vllm.sequence import (VLLM_INVALID_TOKEN_ID, 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 +1,199 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+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.
+    # Performance improvements: avoid repeated attribute and function lookups;
+    # localize frequently used objects;
+
+    sub_texts: list[str] = []
+    current_sub_text: list[str] = []
+    convert_tokens_to_string = tokenizer.convert_tokens_to_string
+    added_vocab_set = set(tokenizer.get_added_vocab())
+    all_special_tokens = set(
+        tokenizer.all_special_tokens) if skip_special_tokens else ()
+
+    for token in output_tokens:
+        # Use precomputed set for skip-special check
+        if token in all_special_tokens:
+            continue
+        if token in added_vocab_set:
+            if current_sub_text:
+                sub_texts.append(convert_tokens_to_string(current_sub_text))
+                current_sub_text.clear()
+            sub_texts.append(token)
+        else:
+            current_sub_text.append(token)
+    if current_sub_text:
+        sub_texts.append(convert_tokens_to_string(current_sub_text))
+    if spaces_between_special_tokens:
+        return " ".join(sub_texts)
+    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 = []
+    for token_id in token_ids:
+        # use default skip_special_tokens.
+        token_str = tokenizer.decode([token_id])
+        if token_str is None:
+            token_str = ""
+        token_str_lst.append(token_str)
+    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/dynamic_module.py

@@ -0,0 +1,60 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import os
+from typing import Optional, Union
+
+from transformers.dynamic_module_utils import get_class_from_dynamic_module
+
+import vllm.envs as envs
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def try_get_class_from_dynamic_module(
+    class_reference: str,
+    pretrained_model_name_or_path: str,
+    cache_dir: Optional[Union[str, os.PathLike]] = None,
+    force_download: bool = False,
+    resume_download: Optional[bool] = None,
+    proxies: Optional[dict[str, str]] = None,
+    token: Optional[Union[bool, str]] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
+    repo_type: Optional[str] = None,
+    code_revision: Optional[str] = None,
+    warn_on_fail: bool = True,
+    **kwargs,
+) -> Optional[type]:
+    """
+    As [transformers.dynamic_module_utils.get_class_from_dynamic_module][],
+    but ignoring any errors.
+    """
+    try:
+        return get_class_from_dynamic_module(
+            class_reference,
+            pretrained_model_name_or_path,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            resume_download=resume_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            local_files_only=local_files_only,
+            repo_type=repo_type,
+            code_revision=code_revision,
+            **kwargs,
+        )
+    except Exception:
+        location = "ModelScope" if envs.VLLM_USE_MODELSCOPE else "HF Hub"
+
+        if warn_on_fail:
+            logger.warning(
+                "Unable to load %s from %s on %s.",
+                class_reference,
+                pretrained_model_name_or_path,
+                location,
+                exc_info=True,
+            )
+
+        return None