| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """HyperConnection (Gated Residual) utilities — NVIDIA model variant. |
| 4 | |
| 5 | Implements the HyperConnection residual scheme proposed in |
| 6 | "HyperConnections" (https://arxiv.org/abs/2409.19606). This NVIDIA variant |
| 7 | delays each HC combine to the following HC mix boundary. HC glue kernels, |
| 8 | including fused combine+RMSNorm, live in ``ops/hc.py``; projections remain |
| 9 | standard vLLM Linear modules. |
| 10 | |
| 11 | Hidden states between layers have shape ``[..., HC*HS]`` with HS inner |
| 12 | (HC outer, HS inner — checkpoint-native layout). |
| 13 | |
| 14 | Typical usage inside a transformer decoder layer:: |
| 15 | |
| 16 | self.attn_hc = GatedResidual(hc_config) |
| 17 | |
| 18 | hidden_states, block_input, injection = self.attn_hc.mix(hidden_states) |
| 19 | attention_output = attention(block_input) |
| 20 | hidden_states, block_input, injection = self.mlp_hc.combine_and_mix( |
| 21 | hidden_states, attention_output, injection |
| 22 | ) |
| 23 | """ |
| 24 | |
| 25 | import torch |
| 26 | from torch import nn |
| 27 | |
| 28 | import vllm.envs as envs |
| 29 | from vllm.model_executor.layers.linear import ( |
| 30 | MergedColumnParallelLinear, |
| 31 | ReplicatedLinear, |
| 32 | ) |
| 33 | from vllm.model_executor.models.utils import maybe_prefix |
| 34 | from vllm.platforms import current_platform |
| 35 | |
| 36 | from ..common.hyperconnection import ( |
| 37 | GroupedGemmaRMSNorm, |
| 38 | HyperConnectionConfig, |
| 39 | ) |
| 40 | from .ops.cute_dsl.hc_down_silu import MAX_FUSED_M, hc_down_silu |
| 41 | from .ops.hc import ( |
| 42 | grouped_gemma_rmsnorm, |
| 43 | hc_combine, |
| 44 | hc_combine_norm, |
| 45 | hc_gate_mix, |
| 46 | hc_silu, |
| 47 | ) |
| 48 | |
| 49 | |
| 50 | # --------------------------------------------------------------------------- |
| 51 | # Gated-residual variant |
| 52 | # --------------------------------------------------------------------------- |
| 53 | class GatedResidual(nn.Module): |
| 54 | """Gated HyperConnection with learnable low-rank mixing and injection. |
| 55 | |
| 56 | ``combine_and_mix()`` runs the pre pipeline (grouped GemmaRMSNorm -> merged |
| 57 | low-rank down+inject GEMM -> silu -> up GEMM -> sigmoid -> gated mean |
| 58 | over the HC streams). When passed a pending block output, it fuses its |
| 59 | residual combine with the RMSNorm. A missing injection selects unit-weight |
| 60 | combine. Final mixers use ``use_combine=False`` and do not produce a new |
| 61 | injection. |
| 62 | |
| 63 | Weights: the norm owns the grouped GemmaRMSNorm affine; the projections |
| 64 | are vLLM Linear modules (merged replicated linear for down+inject). |
| 65 | Eligible down+inject projections use the fused SiLU GEMM; other projections |
| 66 | use their Linear module's GEMM dispatch. |
| 67 | """ |
| 68 | |
| 69 | def __init__( |
| 70 | self, |
| 71 | config: HyperConnectionConfig, |
| 72 | use_combine: bool = True, |
| 73 | prefix: str = "", |
| 74 | ) -> None: |
| 75 | super().__init__() |
| 76 | self.config = config |
| 77 | self.lora_rank = config.hc_lowrank |
| 78 | self.hc_count = config.hc_count |
| 79 | self.hidden_size = config.hidden_size |
| 80 | self.use_combine = use_combine |
| 81 | |
| 82 | norm_size = ( |
| 83 | self.hyper_hidden_size if config.hc_per_branch_norm else config.hidden_size |
| 84 | ) |
| 85 | group_size = config.hidden_size if config.hc_per_branch_norm else None |
| 86 | # Normalize each H-sized HC stream independently while retaining a |
| 87 | # separate affine weight for every element of the HC*H layout. |
| 88 | self.hc_norm = GroupedGemmaRMSNorm( |
| 89 | norm_size, |
| 90 | eps=config.rms_norm_eps, |
| 91 | group_size=group_size, |
| 92 | dtype=config.params_dtype, |
| 93 | ) |
| 94 | |
| 95 | # -- vLLM Linear weights -------------------------------------------- |
| 96 | # The merged skinny-GEMM shape is physically padded to 16 rows to ensure |
| 97 | # good alignment and performant implementation chosen by CuBLAS heuristics. |
| 98 | self.pad_size = (-(self.lora_rank + self.hc_count)) % 16 if use_combine else 0 |
| 99 | if use_combine: |
| 100 | self.input_mix_weight_down_block_inject = MergedColumnParallelLinear( |
| 101 | self.hyper_hidden_size, |
| 102 | [self.lora_rank, self.hc_count] |
| 103 | + ([self.pad_size] if self.pad_size else []), |
| 104 | bias=False, |
| 105 | params_dtype=config.params_dtype, |
| 106 | quant_config=None, |
| 107 | prefix=maybe_prefix(prefix, "input_mix_weight_down_block_inject"), |
| 108 | return_bias=False, |
| 109 | disable_tp=True, |
| 110 | ) |
| 111 | weight = self.input_mix_weight_down_block_inject.weight |
| 112 | self._use_hc_down_silu = ( |
| 113 | weight.shape[1] % 8 == 0 |
| 114 | and weight.dtype == torch.bfloat16 |
| 115 | and current_platform.has_device_capability(90) |
| 116 | ) |
| 117 | else: |
| 118 | self.input_mix_weight_down = ReplicatedLinear( |
| 119 | self.hyper_hidden_size, |
| 120 | self.lora_rank, |
| 121 | bias=False, |
| 122 | params_dtype=config.params_dtype, |
| 123 | quant_config=None, |
| 124 | prefix=maybe_prefix(prefix, "input_mix_weight_down"), |
| 125 | return_bias=False, |
| 126 | ) |
| 127 | self.input_mix_weight_up = ReplicatedLinear( |
| 128 | self.lora_rank, |
| 129 | self.hyper_hidden_size, |
| 130 | bias=False, |
| 131 | params_dtype=config.params_dtype, |
| 132 | quant_config=None, |
| 133 | prefix=maybe_prefix(prefix, "input_mix_weight_up"), |
| 134 | return_bias=False, |
| 135 | ) |
| 136 | |
| 137 | def _down_and_inject( |
| 138 | self, xn: torch.Tensor |
| 139 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 140 | """Down projection + SiLU; also returns injection logits if combined.""" |
| 141 | if not self.use_combine: |
| 142 | return hc_silu(self.input_mix_weight_down(xn), self.hc_count), None |
| 143 | |
| 144 | use_fused = ( |
| 145 | self._use_hc_down_silu |
| 146 | and not envs.VLLM_BATCH_INVARIANT |
| 147 | and 1 <= xn.shape[0] <= MAX_FUSED_M |
| 148 | ) |
| 149 | if use_fused: |
| 150 | return hc_down_silu( |
| 151 | xn, |
| 152 | self.input_mix_weight_down_block_inject.weight, |
| 153 | self.lora_rank, |
| 154 | self.hc_count, |
| 155 | ) |
| 156 | down_and_injection = self.input_mix_weight_down_block_inject(xn) |
| 157 | split_sizes = [self.lora_rank, self.hc_count, self.pad_size] |
| 158 | lora, injection, _ = down_and_injection.split(split_sizes, dim=-1) |
| 159 | return hc_silu(lora, self.hc_count), injection |
| 160 | |
| 161 | def mix( |
| 162 | self, hidden_states: torch.Tensor |
| 163 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]: |
| 164 | xn = grouped_gemma_rmsnorm( |
| 165 | hidden_states, |
| 166 | self.hc_norm.weight, |
| 167 | self.config.rms_norm_eps, |
| 168 | self.hc_count, |
| 169 | ) |
| 170 | |
| 171 | lora, injection = self._down_and_inject(xn) |
| 172 | gate = self.input_mix_weight_up(lora) # [M, D] |
| 173 | block_input = hc_gate_mix(xn, gate, self.hc_count) |
| 174 | |
| 175 | return hidden_states, block_input, injection |
| 176 | |
| 177 | def combine_and_mix( |
| 178 | self, |
| 179 | hidden_states: torch.Tensor, |
| 180 | prev_block_output: torch.Tensor, |
| 181 | prev_injection: torch.Tensor | None, |
| 182 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]: |
| 183 | """Consume a pending combine, then prepare the next block input. |
| 184 | |
| 185 | ``hidden_states`` is the multi-stream state from before the pending |
| 186 | block's mix. Its combine with ``block_output`` is fused with this |
| 187 | module's input RMSNorm. A missing injection applies the block output |
| 188 | to every stream with unit weight. |
| 189 | """ |
| 190 | hidden_states, xn = hc_combine_norm( |
| 191 | hidden_states, |
| 192 | prev_block_output, |
| 193 | prev_injection, |
| 194 | self.hc_norm.weight, |
| 195 | self.config.rms_norm_eps, |
| 196 | self.hc_count, |
| 197 | ) |
| 198 | |
| 199 | lora, injection = self._down_and_inject(xn) |
| 200 | gate = self.input_mix_weight_up(lora) # [M, D] |
| 201 | block_input = hc_gate_mix(xn, gate, self.hc_count) |
| 202 | |
| 203 | return hidden_states, block_input, injection |
| 204 | |
| 205 | def combine( |
| 206 | self, |
| 207 | hidden_states: torch.Tensor, |
| 208 | block_output: torch.Tensor, |
| 209 | injection: torch.Tensor | None, |
| 210 | ) -> torch.Tensor: |
| 211 | return hc_combine(hidden_states, block_output, injection, self.hc_count) |
| 212 | |
| 213 | @property |
| 214 | def hyper_hidden_size(self) -> int: |
| 215 | return self.hc_count * self.hidden_size |
| 216 | |
| 217 | |
| 218 | __all__ = [ |
| 219 | "GatedResidual", |
| 220 | "GroupedGemmaRMSNorm", |
| 221 | "HyperConnectionConfig", |
| 222 | ] |