| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """NVIDIA multi-token predictor for Dots3Note.""" |
| 4 | |
| 5 | from collections.abc import Iterable, Iterator |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | |
| 10 | from vllm.config import VllmConfig |
| 11 | from vllm.distributed import ( |
| 12 | tensor_model_parallel_all_gather, |
| 13 | tensor_model_parallel_all_reduce, |
| 14 | ) |
| 15 | from vllm.model_executor.layers.fused_embed_norm import ( |
| 16 | fused_embed_eh_norm, |
| 17 | has_full_vocab_on_rank, |
| 18 | ) |
| 19 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 20 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 21 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 22 | from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding |
| 23 | from vllm.model_executor.models.deepseek_mtp import SharedHead |
| 24 | from vllm.model_executor.models.utils import maybe_prefix |
| 25 | from vllm.models.deepseek_v32.common.kernels import fused_eh_norm |
| 26 | from vllm.models.deepseek_v32.nvidia.mtp import ( |
| 27 | DeepseekV32MTP, |
| 28 | DeepseekV32MultiTokenPredictor, |
| 29 | ) |
| 30 | |
| 31 | from .model import Dots3NoteDecoderLayer, _pad_dense_mlp_weight |
| 32 | |
| 33 | |
| 34 | class Dots3NoteMultiTokenPredictorLayer(nn.Module): |
| 35 | def __init__(self, vllm_config: VllmConfig, prefix: str) -> None: |
| 36 | super().__init__() |
| 37 | assert vllm_config.speculative_config is not None |
| 38 | config = vllm_config.speculative_config.draft_model_config.hf_config |
| 39 | quant_config = vllm_config.quant_config |
| 40 | |
| 41 | self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 42 | self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 43 | self.eh_proj = ReplicatedLinear( |
| 44 | config.hidden_size * 2, |
| 45 | config.hidden_size, |
| 46 | bias=False, |
| 47 | quant_config=quant_config, |
| 48 | prefix=f"{prefix}.eh_proj", |
| 49 | ) |
| 50 | self.shared_head = SharedHead( |
| 51 | config=config, prefix=prefix, quant_config=quant_config |
| 52 | ) |
| 53 | self.mtp_block = Dots3NoteDecoderLayer( |
| 54 | vllm_config=vllm_config, |
| 55 | config=config, |
| 56 | prefix=prefix, |
| 57 | ) |
| 58 | |
| 59 | def forward( |
| 60 | self, |
| 61 | input_ids: torch.Tensor, |
| 62 | positions: torch.Tensor, |
| 63 | previous_hidden_states: torch.Tensor, |
| 64 | inputs_embeds: torch.Tensor | None = None, |
| 65 | embed_table: torch.Tensor | None = None, |
| 66 | spec_step_index: int = 0, |
| 67 | ) -> torch.Tensor: |
| 68 | del spec_step_index |
| 69 | if embed_table is not None: |
| 70 | eh_input = fused_embed_eh_norm( |
| 71 | positions, |
| 72 | input_ids, |
| 73 | embed_table, |
| 74 | previous_hidden_states, |
| 75 | self.enorm.weight, |
| 76 | self.hnorm.weight, |
| 77 | self.enorm.variance_epsilon, |
| 78 | ) |
| 79 | else: |
| 80 | assert inputs_embeds is not None |
| 81 | eh_input = fused_eh_norm( |
| 82 | positions, |
| 83 | inputs_embeds, |
| 84 | previous_hidden_states, |
| 85 | self.enorm.weight, |
| 86 | self.hnorm.weight, |
| 87 | self.enorm.variance_epsilon, |
| 88 | ) |
| 89 | hidden_states = self.eh_proj(eh_input)[0] |
| 90 | hidden_states, residual = self.mtp_block( |
| 91 | positions=positions, hidden_states=hidden_states, residual=None |
| 92 | ) |
| 93 | is_sequence_parallel = self.mtp_block.use_sequence_parallel_moe |
| 94 | if not is_sequence_parallel: |
| 95 | hidden_states = tensor_model_parallel_all_reduce(hidden_states) |
| 96 | hidden_states, _ = self.shared_head.norm(hidden_states, residual) |
| 97 | if is_sequence_parallel: |
| 98 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 99 | hidden_states = hidden_states[: positions.shape[0]] |
| 100 | return hidden_states |
| 101 | |
| 102 | |
| 103 | class Dots3NoteMultiTokenPredictor(DeepseekV32MultiTokenPredictor): |
| 104 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 105 | nn.Module.__init__(self) |
| 106 | assert vllm_config.speculative_config is not None |
| 107 | config = vllm_config.speculative_config.draft_model_config.hf_config |
| 108 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 109 | self.num_mtp_layers = config.num_nextn_predict_layers |
| 110 | self.layers = nn.ModuleDict( |
| 111 | { |
| 112 | str(idx): Dots3NoteMultiTokenPredictorLayer( |
| 113 | vllm_config, f"{prefix}.layers.{idx}" |
| 114 | ) |
| 115 | for idx in range( |
| 116 | self.mtp_start_layer_idx, |
| 117 | self.mtp_start_layer_idx + self.num_mtp_layers, |
| 118 | ) |
| 119 | } |
| 120 | ) |
| 121 | self.embed_tokens = VocabParallelEmbedding( |
| 122 | config.vocab_size, |
| 123 | config.hidden_size, |
| 124 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 125 | ) |
| 126 | self.replicated_embed = has_full_vocab_on_rank(self.embed_tokens) |
| 127 | self.register_buffer( |
| 128 | "max_token_id", |
| 129 | torch.tensor(config.vocab_size - 1, dtype=torch.int64), |
| 130 | persistent=False, |
| 131 | ) |
| 132 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 133 | |
| 134 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 135 | input_ids = torch.minimum(input_ids, self.max_token_id) |
| 136 | return self.embed_tokens(input_ids) |
| 137 | |
| 138 | def forward( |
| 139 | self, |
| 140 | input_ids: torch.Tensor, |
| 141 | positions: torch.Tensor, |
| 142 | previous_hidden_states: torch.Tensor, |
| 143 | inputs_embeds: torch.Tensor | None = None, |
| 144 | spec_step_idx: int = 0, |
| 145 | ) -> torch.Tensor: |
| 146 | if inputs_embeds is None: |
| 147 | input_ids = torch.minimum(input_ids, self.max_token_id) |
| 148 | return super().forward( |
| 149 | input_ids, |
| 150 | positions, |
| 151 | previous_hidden_states, |
| 152 | inputs_embeds, |
| 153 | spec_step_idx, |
| 154 | ) |
| 155 | |
| 156 | |
| 157 | class Dots3NoteMTP(DeepseekV32MTP): |
| 158 | has_own_embed_tokens = True |
| 159 | has_own_lm_head = False |
| 160 | |
| 161 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 162 | nn.Module.__init__(self) |
| 163 | self.config = vllm_config.model_config.hf_config |
| 164 | self.quant_config = vllm_config.quant_config |
| 165 | self.model = Dots3NoteMultiTokenPredictor( |
| 166 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 167 | ) |
| 168 | self.set_moe_parameters() |
| 169 | |
| 170 | def _adapt_weights( |
| 171 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 172 | ) -> Iterator[tuple[str, torch.Tensor]]: |
| 173 | mtp_layer = self.config.num_hidden_layers |
| 174 | for name, weight in weights: |
| 175 | if name.startswith("model.mtp.embed_tokens."): |
| 176 | name = name.replace( |
| 177 | "model.mtp.embed_tokens.", |
| 178 | f"model.layers.{mtp_layer}.embed_tokens.", |
| 179 | 1, |
| 180 | ) |
| 181 | yield ( |
| 182 | name, |
| 183 | _pad_dense_mlp_weight( |
| 184 | name, |
| 185 | weight, |
| 186 | getattr(self.quant_config, "weight_block_size", None), |
| 187 | ), |
| 188 | ) |
| 189 | |
| 190 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 191 | return super().load_weights(self._adapt_weights(weights)) |