| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from copy import copy |
| 4 | from dataclasses import fields as dataclass_fields |
| 5 | from typing import Any |
| 6 | |
| 7 | from pydantic import ConfigDict |
| 8 | from pydantic.dataclasses import dataclass |
| 9 | |
| 10 | from vllm.logger import init_logger |
| 11 | |
| 12 | logger = init_logger(__name__) |
| 13 | |
| 14 | |
| 15 | @dataclass(config=ConfigDict(arbitrary_types_allowed=True)) |
| 16 | class ModelArchitectureConfig: |
| 17 | """Configuration for model architecture that required by vLLM runtime.""" |
| 18 | |
| 19 | architectures: list[str] |
| 20 | """List of model architecture class names (e.g., ['LlamaForCausalLM']). |
| 21 | It can be None upon calling `vllm_config.with_hf_config(config.text_config)`""" |
| 22 | |
| 23 | model_type: str |
| 24 | """Model type identifier (e.g., 'llama', 'gpt_oss').""" |
| 25 | |
| 26 | text_model_type: str | None |
| 27 | """Text model type identifier (e.g., 'llama4_text').""" |
| 28 | |
| 29 | hidden_size: int |
| 30 | """Hidden size of the model.""" |
| 31 | |
| 32 | total_num_hidden_layers: int |
| 33 | """Number of hidden layers in the model.""" |
| 34 | |
| 35 | total_num_attention_heads: int |
| 36 | """Number of attention heads in the model.""" |
| 37 | |
| 38 | head_size: int |
| 39 | """Head dimension of the model.""" |
| 40 | |
| 41 | vocab_size: int |
| 42 | """Vocabulary size of the model.""" |
| 43 | |
| 44 | total_num_kv_heads: int |
| 45 | """Number of key value heads in the model.""" |
| 46 | |
| 47 | num_experts: int |
| 48 | """Number of experts in the model.""" |
| 49 | |
| 50 | num_experts_per_token: int |
| 51 | """Number of routed experts selected per token.""" |
| 52 | |
| 53 | quantization_config: dict[str, Any] | None |
| 54 | """Quantization configuration dictionary containing quantization parameters.""" |
| 55 | |
| 56 | is_deepseek_mla: bool |
| 57 | """Whether the model is a DeepSeek MLA model.""" |
| 58 | |
| 59 | is_mm_prefix_lm: bool |
| 60 | """Whether the model uses image bidirectional attention.""" |
| 61 | |
| 62 | rswa_window: int | None |
| 63 | """Reference Sliding Window Attention window size (None disables R-SWA).""" |
| 64 | |
| 65 | derived_max_model_len_and_key: tuple[float, str | None] |
| 66 | """Derived maximum model length and key from the hf config.""" |
| 67 | |
| 68 | per_layer_overrides: "list[dict[str, Any]] | None" = None |
| 69 | """Per-layer values for the fields that vary, `None` unless some field does. |
| 70 | |
| 71 | One dict per layer, holding only the fields whose value differs from the |
| 72 | whole-model value above. Everything else is read from the whole-model config, |
| 73 | so later edits to it are visible through `self[layer_idx]`.""" |
| 74 | |
| 75 | def __getitem__(self, layer_idx: int) -> "ModelArchitectureConfig": |
| 76 | """ModelArchitectureConfig for a specific layer. |
| 77 | |
| 78 | Returns `self` when no field varies by layer, so callers never need to |
| 79 | branch on heterogeneity. Mirrors `PreTrainedConfig.per_layer_config[i]`. |
| 80 | """ |
| 81 | if layer_idx < 0: |
| 82 | # Homogeneous configs would ignore a negative index and heterogeneous |
| 83 | # ones would wrap, so an off-by-one would only show up on the models |
| 84 | # this exists for. Reject it for both. |
| 85 | raise IndexError(f"layer index must not be negative, got {layer_idx}") |
| 86 | if self.per_layer_overrides is None: |
| 87 | return self |
| 88 | layer = copy(self) |
| 89 | object.__setattr__(layer, "per_layer_overrides", None) |
| 90 | for name, value in self.per_layer_overrides[layer_idx].items(): |
| 91 | object.__setattr__(layer, name, value) |
| 92 | return layer |
| 93 | |
| 94 | @classmethod |
| 95 | def from_layers( |
| 96 | cls, layers: "list[ModelArchitectureConfig]" |
| 97 | ) -> "ModelArchitectureConfig": |
| 98 | """Whole-model config for a checkpoint whose layers differ. |
| 99 | |
| 100 | Fields that agree across layers are taken as they are. Fields that differ |
| 101 | are collapsed with `max`, so buffers are sized for the largest layer, and |
| 102 | the differing values are kept per layer. No field is named here: which |
| 103 | ones vary is whatever the checkpoint says. |
| 104 | """ |
| 105 | if not layers: |
| 106 | raise ValueError("a model must have at least one layer") |
| 107 | |
| 108 | merged: dict[str, Any] = {} |
| 109 | overrides: list[dict[str, Any]] = [{} for _ in layers] |
| 110 | for f in dataclass_fields(cls): |
| 111 | if f.name == "per_layer_overrides": |
| 112 | continue |
| 113 | values = [getattr(layer, f.name) for layer in layers] |
| 114 | if all(value == values[0] for value in values): |
| 115 | merged[f.name] = values[0] |
| 116 | continue |
| 117 | # `bool` is an `int`, so an exact type check is what keeps a varying |
| 118 | # flag from collapsing to `any`. `is_deepseek_mla` doing that would |
| 119 | # make `use_mla` true model wide, and `get_num_kv_heads` then returns |
| 120 | # 1 for every layer, discarding the overrides built here. |
| 121 | if not all(type(value) in (int, float) for value in values): |
| 122 | raise ValueError( |
| 123 | f"{f.name!r} varies across layers and has no whole-model " |
| 124 | f"value: {sorted(set(map(repr, values)))}. Only numeric " |
| 125 | f"fields collapse (with `max`, to size buffers for the " |
| 126 | f"largest layer); give this one an explicit rule in " |
| 127 | f"ModelArchitectureConfig.from_layers." |
| 128 | ) |
| 129 | merged[f.name] = max(values) |
| 130 | for override, value in zip(overrides, values): |
| 131 | if value != merged[f.name]: |
| 132 | override[f.name] = value |
| 133 | |
| 134 | if len(layers) != merged["total_num_hidden_layers"]: |
| 135 | raise ValueError( |
| 136 | f"got {len(layers)} per-layer configs for a model with " |
| 137 | f"{merged['total_num_hidden_layers']} layers" |
| 138 | ) |
| 139 | # A checkpoint can be heterogeneous over attributes vLLM never reads, in |
| 140 | # which case there is nothing to keep the layers apart for. |
| 141 | return cls(**merged, per_layer_overrides=overrides if any(overrides) else None) |