| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import math |
| 5 | from collections.abc import Iterable |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | |
| 10 | from vllm.config import VllmConfig |
| 11 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 12 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 13 | ParallelLMHead, |
| 14 | VocabParallelEmbedding, |
| 15 | ) |
| 16 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 17 | |
| 18 | from .utils import maybe_prefix |
| 19 | |
| 20 | |
| 21 | class MLPSpeculatorLayerNorm(nn.Module): |
| 22 | """A L2 normalization implementation |
| 23 | ... |
| 24 | |
| 25 | Args: |
| 26 | ---- |
| 27 | normalized_shape : int |
| 28 | Dimensionality of input data (size of final tensor axis) |
| 29 | eps : float |
| 30 | Safety term to prevent division by zero. Make sure the chosen value |
| 31 | fits in the range of your encoding scheme |
| 32 | (i.e. fp16 requires eps >= 6e-8). |
| 33 | elementwise_scale_and_shift : bool |
| 34 | Include a learned scaling and shift term after normalization. |
| 35 | |
| 36 | """ |
| 37 | |
| 38 | def __init__( |
| 39 | self, |
| 40 | normalized_shape, |
| 41 | eps=1e-06, |
| 42 | elementwise_scale_and_shift=True, |
| 43 | ): |
| 44 | super().__init__() |
| 45 | self.elementwise_scale_and_shift = elementwise_scale_and_shift |
| 46 | if self.elementwise_scale_and_shift: |
| 47 | self.weight = nn.Parameter(torch.empty(normalized_shape)) |
| 48 | self.bias = nn.Parameter(torch.empty(normalized_shape)) |
| 49 | self.eps = eps |
| 50 | |
| 51 | def forward(self, x): |
| 52 | xf = x |
| 53 | xf = xf * torch.rsqrt(xf.pow(2).mean(-1, keepdim=True) + self.eps) |
| 54 | x = xf.type_as(x) |
| 55 | if self.elementwise_scale_and_shift: |
| 56 | x = self.weight * x |
| 57 | x = x + self.bias |
| 58 | return x |
| 59 | |
| 60 | |
| 61 | class MLPSpeculator(nn.Module): |
| 62 | """An implementation of the speculative models introduced in |
| 63 | "Accelerating Production LLMs with Combined Token/Embedding |
| 64 | Speculators" |
| 65 | https://arxiv.org/pdf/2404.19124 |
| 66 | |
| 67 | Trained speculators of this type are available on HF hub at: |
| 68 | https://huggingface.co/ibm-ai-platform and https://huggingface.co/ibm-granite |
| 69 | """ |
| 70 | |
| 71 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 72 | super().__init__() |
| 73 | config = vllm_config.model_config.hf_config |
| 74 | self.n_predict = config.n_predict |
| 75 | self.vocab_size = config.vocab_size |
| 76 | self.emb_dim = config.emb_dim |
| 77 | self.inner_dim = config.inner_dim if config.inner_dim != 0 else config.emb_dim |
| 78 | |
| 79 | self.max_speculative_tokens = config.num_lookahead_tokens |
| 80 | |
| 81 | self.tie_weights = config.tie_weights |
| 82 | self.scale_input = config.scale_input |
| 83 | |
| 84 | if self.tie_weights: |
| 85 | assert self.n_predict > 1, ( |
| 86 | "You cannot tie weights between stages when only 1 exists" |
| 87 | ) |
| 88 | embedding = VocabParallelEmbedding( |
| 89 | config.vocab_size, self.inner_dim, org_num_embeddings=config.vocab_size |
| 90 | ) |
| 91 | self.emb = nn.ModuleList([embedding] * self.max_speculative_tokens) |
| 92 | |
| 93 | # the initial projection from the base model may |
| 94 | # have a different size, so that stays separate. |
| 95 | proj_first = nn.Linear(self.emb_dim, self.inner_dim, bias=False) |
| 96 | proj_tied = nn.Linear(self.inner_dim, self.inner_dim, bias=False) |
| 97 | self.proj = nn.ModuleList( |
| 98 | [proj_first] + [proj_tied] * (self.max_speculative_tokens - 1) |
| 99 | ) |
| 100 | |
| 101 | self.head = nn.ModuleList( |
| 102 | [ |
| 103 | ParallelLMHead( |
| 104 | self.vocab_size, |
| 105 | self.inner_dim, |
| 106 | bias=False, |
| 107 | prefix=maybe_prefix(prefix, f"head.{i}"), |
| 108 | ) |
| 109 | for i in range(self.max_speculative_tokens) |
| 110 | ] |
| 111 | ) |
| 112 | |
| 113 | ln = MLPSpeculatorLayerNorm( |
| 114 | self.inner_dim, elementwise_scale_and_shift=True |
| 115 | ) |
| 116 | self.ln = nn.ModuleList([ln] * self.max_speculative_tokens) |
| 117 | |
| 118 | else: |
| 119 | self.emb = nn.ModuleList( |
| 120 | [ |
| 121 | VocabParallelEmbedding( |
| 122 | config.vocab_size, |
| 123 | self.inner_dim, |
| 124 | ) |
| 125 | for _ in range(self.max_speculative_tokens) |
| 126 | ] |
| 127 | ) |
| 128 | |
| 129 | self.proj = nn.ModuleList( |
| 130 | [ |
| 131 | nn.Linear( |
| 132 | (self.emb_dim if i == 0 else self.inner_dim), |
| 133 | self.inner_dim, |
| 134 | bias=False, |
| 135 | ) |
| 136 | for i in range(self.max_speculative_tokens) |
| 137 | ] |
| 138 | ) |
| 139 | |
| 140 | self.head = nn.ModuleList( |
| 141 | [ |
| 142 | ParallelLMHead( |
| 143 | self.vocab_size, |
| 144 | self.inner_dim, |
| 145 | bias=False, |
| 146 | prefix=maybe_prefix(prefix, f"head.{i}"), |
| 147 | ) |
| 148 | for i in range(self.max_speculative_tokens) |
| 149 | ] |
| 150 | ) |
| 151 | self.ln = nn.ModuleList( |
| 152 | [ |
| 153 | MLPSpeculatorLayerNorm( |
| 154 | self.inner_dim, elementwise_scale_and_shift=True |
| 155 | ) |
| 156 | for _ in range(self.max_speculative_tokens) |
| 157 | ] |
| 158 | ) |
| 159 | if self.scale_input: |
| 160 | self.ln0 = MLPSpeculatorLayerNorm( |
| 161 | self.emb_dim, elementwise_scale_and_shift=False |
| 162 | ) |
| 163 | |
| 164 | self.state_weight = 0.5 ** (0.5 / config.n_predict) |
| 165 | self.emb_weight = math.sqrt((1 - self.state_weight**2) * (self.inner_dim / 2)) |
| 166 | self.activation = nn.GELU() |
| 167 | self.config = config |
| 168 | self.logits_processor = LogitsProcessor( |
| 169 | config.vocab_size, config.vocab_size, 1.0 |
| 170 | ) |
| 171 | |
| 172 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 173 | params_dict = dict(self.named_parameters()) |
| 174 | loaded_params: set[str] = set() |
| 175 | for name, loaded_weight in weights: |
| 176 | name = name.replace("speculator.", "") |
| 177 | param = params_dict.get(name) |
| 178 | if param is not None: |
| 179 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 180 | weight_loader(param, loaded_weight) |
| 181 | loaded_params.add(name) |
| 182 | return loaded_params |