vllm-cpu-avx512bf16 0.14.0__cp313-cp313-manylinux_2_28_x86_64.whl

vllm/transformers_utils/configs/medusa.py

@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+
+from transformers import PretrainedConfig
+
+
+class MedusaConfig(PretrainedConfig):
+    model_type = "medusa"
+
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        vocab_size: int = 32001,
+        num_heads: int = 5,
+        num_hidden_layers: int = 1,
+        max_paths: int = 64,
+        topk: int = 10,
+        truncated_vocab_size: int | None = None,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.vocab_size = vocab_size
+        self.num_heads = num_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.max_paths = max_paths
+        self.topk = topk
+        self.max_seq_len = int(2**20)
+        self.truncated_vocab_size = (
+            vocab_size if truncated_vocab_size is None else truncated_vocab_size
+        )
+        if "architectures" not in kwargs:
+            kwargs["architectures"] = ["MedusaModel"]
+
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: str | os.PathLike,
+        **kwargs,
+    ) -> "MedusaConfig":
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+        for k in list(config_dict.keys()):
+            if "num" in k:
+                if "heads" in k:
+                    config_dict["num_heads"] = config_dict.pop(k)
+                elif "layers" in k:
+                    config_dict["num_hidden_layers"] = config_dict.pop(k)
+        return cls.from_dict(config_dict, **kwargs)
+
+    @property
+    def num_attention_heads(self):
+        return 0
+
+    @property
+    def num_lookahead_tokens(self):
+        return self.num_heads
+
+    @num_lookahead_tokens.setter
+    def num_lookahead_tokens(self, num_lookahead_tokens: int):
+        self.num_heads = num_lookahead_tokens

vllm/transformers_utils/configs/midashenglm.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Copyright 2025 Horizon team, Xiaomi MiLM Plus.
+# Copyright 2024 The Qwen team.
+# Copyright 2023 The vLLM team.
+# 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 import PretrainedConfig
+from transformers.models.qwen2_5_omni.configuration_qwen2_5_omni import (
+    Qwen2_5OmniTextConfig,
+)
+
+
+class DashengConfig(PretrainedConfig):
+    model_type = "midashenglm_dasheng_encoder"
+
+    def __init__(
+        self,
+        embed_dim: int = 768,
+        outputdim: int = 527,
+        patch_size: int | tuple[int, int] = 16,
+        patch_stride: int | tuple[int, int] = 16,
+        input_channels: int = 1,
+        target_length: int = 1012,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        init_values: float | None = None,
+        drop_rate: float = 0.0,
+        attn_drop_rate: float = 0.0,
+        f_min: float = 0.0,
+        f_max: float = 8000.0,
+        center: bool = True,
+        win_length: int = 512,
+        hop_length: int = 160,
+        sample_rate: int = 16000,
+        n_fft: int = 512,
+        n_mels: int = 64,
+        **kwargs,
+    ):
+        self.embed_dim = embed_dim
+        self.outputdim = outputdim
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.input_channels = input_channels
+        self.target_length = target_length
+        self.depth = depth
+        self.num_heads = num_heads
+        self.mlp_ratio = mlp_ratio
+        self.qkv_bias = qkv_bias
+        self.init_values = init_values
+        self.drop_rate = drop_rate
+        self.attn_drop_rate = attn_drop_rate
+        self.f_min = f_min
+        self.f_max = f_max
+        self.center = center
+        self.win_length = win_length
+        self.hop_length = hop_length
+        self.sample_rate = sample_rate
+        self.n_fft = n_fft
+        self.n_mels = n_mels
+        super().__init__(**kwargs)
+
+
+class MiDashengLMConfig(PretrainedConfig):
+    model_type = "midashenglm"
+
+    def __init__(
+        self,
+        audio_encoder_config: dict | None = None,
+        subsample_factor: int = 5,
+        text_config: dict | None = None,
+        audio_token_id: int | None = None,
+        **kwargs,
+    ):
+        self.audio_encoder_config = DashengConfig(**(audio_encoder_config or {}))
+        self.subsample_factor = subsample_factor
+        self.text_config = (
+            Qwen2_5OmniTextConfig(**text_config)
+            if text_config
+            else Qwen2_5OmniTextConfig()
+        )
+        self.text_config.rope_parameters = None  # uses_mrope is false
+        self.audio_token_id = audio_token_id
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/mistral.py

@@ -0,0 +1,263 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any
+
+from transformers import PretrainedConfig, WhisperConfig
+
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def adapt_config_dict(
+    config_dict: dict[str, Any],
+    defaults: dict[str, Any],
+) -> PretrainedConfig:
+    config_dict = _remap_general_mistral_args(config_dict)
+
+    if bool(config_dict.get("quantization")):
+        config_dict = _remap_mistral_quantization_args(config_dict)
+
+    is_moe = bool(config_dict.get("moe"))
+    is_mistral_large_3 = (
+        is_moe and (config_dict["moe"].get("num_shared_experts") or 0) > 0
+    )
+    if config_dict.get("model_type") == "mamba":
+        config_dict["architectures"] = ["Mamba2ForCausalLM"]
+    elif is_moe and is_mistral_large_3:
+        config_dict = _remap_moe_args(config_dict)
+        config_dict["model_type"] = "deepseek_v3"
+        config_dict["architectures"] = ["MistralLarge3ForCausalLM"]
+
+        assert "llama_4_scaling" in config_dict, (
+            "MistralLarge3 expect llama4 scaling config."
+        )
+        llama_4_scaling_config_keys = ["original_max_position_embeddings", "beta"]
+        assert all(
+            [
+                key in config_dict["llama_4_scaling"]
+                for key in llama_4_scaling_config_keys
+            ]
+        ), (
+            "llama_4_scaling config should define the keys: "
+            f"{','.join(llama_4_scaling_config_keys)}"
+        )
+    elif is_moe:
+        config_dict["architectures"] = ["MixtralForCausalLM"]
+    else:
+        config_dict["architectures"] = ["MistralForCausalLM"]
+
+    if bool(config_dict.get("yarn")):
+        config_dict = _remap_mistral_yarn_args(config_dict)
+
+    if bool(config_dict.get("llama_4_scaling")):
+        llama_4_scaling_config_keys = ["original_max_position_embeddings", "beta"]
+        assert all(
+            [
+                key in config_dict["llama_4_scaling"]
+                for key in llama_4_scaling_config_keys
+            ]
+        ), (
+            "llama_4_scaling config should define the keys: "
+            f"{','.join(llama_4_scaling_config_keys)}"
+        )
+
+    is_vision = (config_dict.get("multimodal") or {}).get(
+        "vision_encoder_args"
+    ) or config_dict.get("vision_encoder")
+    is_audio = bool(
+        ((config_dict.get("multimodal") or {}).get("whisper_model_args") or {}).get(
+            "encoder_args"
+        )
+    )
+
+    assert not (is_vision and is_audio), "Vision and audio are mutually exclusive"
+
+    if is_vision:
+        config_dict = _remap_mistral_vision_args(config_dict)
+    if is_audio:
+        config_dict = _remap_mistral_audio_args(config_dict)
+
+    for k, v in defaults.items():
+        config_dict.setdefault(k, v)
+
+    config = PretrainedConfig.from_dict(config_dict)
+
+    logger.debug("Initialized config %s", config)
+
+    return config
+
+
+def _remap_mistral_vision_args(config: dict) -> dict:
+    if config.get("multimodal"):
+        vision_config = config.pop("multimodal")
+    else:
+        vision_config = config.pop("vision_encoder")
+
+    quant_config = config.get("quantization_config")
+    config = {
+        "model_type": "pixtral",
+        "architectures": ["PixtralForConditionalGeneration"],
+        "text_config": PretrainedConfig.from_dict(config),
+        "vision_config": PretrainedConfig.from_dict(vision_config),
+    }
+    if quant_config:
+        config["quantization_config"] = quant_config
+    return config
+
+
+def _remap_mistral_yarn_args(config: dict) -> dict:
+    yarn_config_map = {
+        "factor": "factor",
+        "original_max_position_embeddings": "original_max_position_embeddings",
+        "beta": "beta_fast",
+        "alpha": "beta_slow",
+        "apply_scale": "apply_yarn_scaling",
+    }
+    yarn_config = config.get("yarn") or {}
+    config["rope_parameters"] = {
+        "rope_type": "yarn",
+        "mscale_all_dim": 1,
+    }
+
+    if rope_theta := config.pop("rope_theta", None):
+        config["rope_parameters"]["rope_theta"] = rope_theta
+
+    for old_name, new_name in yarn_config_map.items():
+        if old_name in yarn_config:
+            config["rope_parameters"][new_name] = yarn_config.pop(old_name)
+
+    assert len(yarn_config) == 0, f"Unparsed yarn config: {yarn_config}"
+
+    return config
+
+
+def _remap_general_mistral_args(config: dict) -> dict:
+    # Mistral key -> HF key
+    config_mapping = {
+        "dim": "hidden_size",
+        "norm_eps": "rms_norm_eps",
+        "n_kv_heads": "num_key_value_heads",
+        "n_layers": "num_hidden_layers",
+        "n_heads": "num_attention_heads",
+        "hidden_dim": "intermediate_size",
+    }
+    # HF key -> (Mistral key, default value)
+    top_level_mapping_with_default = {
+        "model_type": ("model_type", "transformer"),
+        "hidden_act": ("activation", "silu"),
+        "tie_word_embeddings": ("tied_embeddings", False),
+        "max_seq_len": ("max_seq_len", config.get("max_position_embeddings", 128_000)),
+        "max_position_embeddings": ("max_position_embeddings", 128_000),
+    }
+
+    for key, new_key in config_mapping.items():
+        if key in config:
+            config[new_key] = config.pop(key)
+
+    for new_key, (key, default_value) in top_level_mapping_with_default.items():
+        config[new_key] = config.pop(key, default_value)
+
+    return config
+
+
+def _remap_mistral_quantization_args(config: dict) -> dict:
+    if config.get("quantization"):
+        quantization = config.pop("quantization", {})
+        if quantization.get("qformat_weight") == "fp8_e4m3":
+            qscheme_act = quantization.get("qscheme_act")
+            assert qscheme_act in ("NO_SCALES", "TENSOR", None), (
+                "Only NO_SCALES and TENSOR (default) are supported for qscheme_act"
+            )
+            is_dynamic = qscheme_act == "NO_SCALES"
+            config["quantization_config"] = {
+                "quant_method": "fp8",
+                "activation_scheme": "dynamic" if is_dynamic else "static",
+            }
+        else:
+            raise ValueError(f"Found unknown quantization='{quantization}' in config")
+
+    return config
+
+
+def _remap_mistral_audio_args(config: dict) -> dict:
+    whisper_args = config["multimodal"].pop("whisper_model_args")
+    encoder_args = whisper_args["encoder_args"]
+    downsample_args = whisper_args["downsample_args"]
+    downsample_factor = downsample_args["downsample_factor"]
+
+    # make sure that k/v blocks can be allocated with
+    # unified k/v cache class and pool whisper k/v cache blocks
+    # with downsample_factor:1 ratio
+    if encoder_args.get("causal"):
+        block_pool_size = downsample_factor
+        config["projection_size"] = downsample_factor * encoder_args["dim"]
+    else:
+        block_pool_size = 1
+
+    _maybe_sliding_window = encoder_args.get("ragged_attention", None)
+    if _maybe_sliding_window is None:
+        sliding_window = None
+    elif _maybe_sliding_window.isdigit():
+        sliding_window = int(_maybe_sliding_window)
+    else:
+        raise NotImplementedError(f"Unsupported: {_maybe_sliding_window=}")
+
+    architecture = (
+        "VoxtralStreamingGeneration"
+        if encoder_args.get("causal")
+        else "VoxtralForConditionalGeneration"
+    )
+
+    quant_config = config.get("quantization_config")
+    config = {
+        "model_type": "voxtral",
+        "architectures": [architecture],
+        "text_config": PretrainedConfig.from_dict(config),
+        "audio_config": WhisperConfig(
+            num_mel_bins=encoder_args["audio_encoding_args"]["num_mel_bins"],
+            window_size=encoder_args["audio_encoding_args"]["window_size"],
+            sampling_rate=encoder_args["audio_encoding_args"]["sampling_rate"],
+            hop_length=encoder_args["audio_encoding_args"]["hop_length"],
+            downsample_factor=downsample_factor,
+            d_model=encoder_args["dim"],
+            encoder_layers=encoder_args["n_layers"],
+            encoder_ffn_dim=encoder_args["hidden_dim"],
+            encoder_attention_heads=encoder_args["n_heads"],
+            vocab_size=encoder_args["vocab_size"],
+            max_source_positions=encoder_args["max_source_positions"],
+            is_encoder_decoder=False,  # Override WhisperConfig default
+            is_causal=encoder_args.get("causal", False),
+            sliding_window=sliding_window,
+            block_pool_size=block_pool_size,
+            pos_embed=encoder_args.get("pos_embed", "sinusoidal"),
+        ),
+    }
+    if quant_config:
+        config["quantization_config"] = quant_config
+    return config
+
+
+def _remap_moe_args(config: dict) -> dict:
+    moe_config_map = {
+        "route_every_n": "moe_layer_freq",
+        "first_k_dense_replace": "first_k_dense_replace",
+        "num_experts_per_tok": "num_experts_per_tok",
+        "num_experts": "n_routed_experts",
+        "expert_hidden_dim": "moe_intermediate_size",
+        "routed_scale": "routed_scaling_factor",
+        "num_shared_experts": "n_shared_experts",
+        "num_expert_groups": "n_group",
+        "num_expert_groups_per_tok": "topk_group",
+    }
+    moe_config = config.get("moe", {})
+    for old_name, new_name in moe_config_map.items():
+        if old_name in moe_config:
+            value = moe_config.pop(old_name)
+            config[new_name] = value
+
+    config["topk_method"] = None
+    config["norm_topk_prob"] = True
+    config["scoring_func"] = "softmax"
+
+    return config

vllm/transformers_utils/configs/mlp_speculator.py

@@ -0,0 +1,69 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+
+from transformers import PretrainedConfig
+
+
+class MLPSpeculatorConfig(PretrainedConfig):
+    model_type = "mlp_speculator"
+
+    attribute_map = {
+        "hidden_size": "emb_dim",
+    }
+
+    def __init__(
+        self,
+        vocab_size: int = 32000,
+        emb_dim: int = 4096,
+        inner_dim: int = 0,
+        n_predict: int = 3,
+        top_k_tokens_per_head: list[int] | None = None,
+        n_candidates: int = 5,
+        tie_weights: bool = False,
+        scale_input: bool = False,
+        **kwargs,
+    ):
+        """
+        Initialize an MLPSpeculatorConfig
+
+        Args:
+            vocab_size: int
+                the model vocab size
+            emb_dim: int
+                the model embedding dimension
+            inner_dim: int
+                the inner dimension of the model. If 0, will be the emb_dim.
+            n_predict: int
+                the number of lookaheads for the speculator
+            top_k_tokens_per_head: list[int]
+                Number of tokens to consider from each head when forming the
+                candidate tree.
+                For each candidate branch in the tree, head n produces topk[n]
+                additional sub-branches.
+                NOTE: This parameter is currently unused.
+            n_candidates: int
+                number of child candidates to create per sequence
+            tie_weights: bool
+                If true, use a single set of weights for every model
+                head/stage after the first. The initial projection
+                from the base model may have a different size, so that
+                stays separate.
+            scale_input: bool
+                if True, will scale the initial hidden states from
+                the base model.
+        """
+        if top_k_tokens_per_head is None:
+            top_k_tokens_per_head = [5, 4, 3]
+        assert len(top_k_tokens_per_head) == n_predict
+        self.vocab_size = vocab_size
+        self.emb_dim = emb_dim
+        self.inner_dim = inner_dim
+        self.n_predict = n_predict
+        self.top_k_tokens_per_head = top_k_tokens_per_head
+        self.n_candidates = n_candidates
+        self.num_lookahead_tokens = n_predict
+        self.tie_weights = tie_weights
+        self.scale_input = scale_input
+
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/moonvit.py

@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
+from transformers.configuration_utils import PretrainedConfig
+
+
+class MoonViTConfig(PretrainedConfig):
+    model_type = "moonvit"
+
+    def __init__(
+        self,
+        patch_size: int = 14,
+        init_pos_emb_height: int = 64,
+        init_pos_emb_width: int = 64,
+        num_attention_heads: int = 16,
+        num_hidden_layers: int = 27,
+        hidden_size: int = 1152,
+        intermediate_size: int = 4304,
+        merge_kernel_size: tuple[int, int] = (2, 2),
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.patch_size = patch_size
+        # Positional embedding config
+        self.init_pos_emb_height = init_pos_emb_height
+        self.init_pos_emb_width = init_pos_emb_width
+        # Transformer config
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        # Patch merger config
+        self.merge_kernel_size = merge_kernel_size