| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Gemma4 MTP (Multi-Token Prediction) speculator for speculative decoding. |
| 4 | |
| 5 | The Gemma4 assistant model runs all decoder layers per draft step |
| 6 | (producing one token), and all its attention layers share KV cache |
| 7 | with the target model via cross-model KV sharing. |
| 8 | """ |
| 9 | |
| 10 | from collections import defaultdict |
| 11 | |
| 12 | import torch.nn as nn |
| 13 | |
| 14 | from vllm.compilation.backends import set_model_tag |
| 15 | from vllm.config import VllmConfig, replace |
| 16 | from vllm.distributed.parallel_state import get_pp_group |
| 17 | from vllm.logger import init_logger |
| 18 | from vllm.model_executor.model_loader import get_model |
| 19 | from vllm.model_executor.model_loader.utils import get_draft_load_config |
| 20 | from vllm.v1.worker.gpu.spec_decode.autoregressive.speculator import ( |
| 21 | AutoRegressiveSpeculator, |
| 22 | ) |
| 23 | |
| 24 | logger = init_logger(__name__) |
| 25 | |
| 26 | |
| 27 | def _copy_target_kv_scales(attn: nn.Module, target_attn: nn.Module) -> None: |
| 28 | """Copy target KV scales while preserving their tensor representation. |
| 29 | |
| 30 | Default attention scales are scalar buffers, while some quantization |
| 31 | methods replace them with length-one or per-head parameters. Preserve the |
| 32 | draft layer's registration type so the shared KV cache is interpreted with |
| 33 | the target's values and shapes without aliasing target parameters. |
| 34 | """ |
| 35 | for scale_name in ("_k_scale", "_v_scale"): |
| 36 | draft_scale = getattr(attn, scale_name) |
| 37 | target_scale = getattr(target_attn, scale_name) |
| 38 | scale = target_scale.detach().clone() |
| 39 | if isinstance(draft_scale, nn.Parameter): |
| 40 | scale = nn.Parameter(scale, requires_grad=False) |
| 41 | setattr(attn, scale_name, scale) |
| 42 | for scale_name in ("_k_scale_float", "_v_scale_float"): |
| 43 | setattr(attn, scale_name, getattr(target_attn, scale_name)) |
| 44 | for scale_name in ("_k_scale_cpu", "_v_scale_cpu"): |
| 45 | getattr(attn, scale_name).copy_(getattr(target_attn, scale_name)) |
| 46 | |
| 47 | |
| 48 | class Gemma4Speculator(AutoRegressiveSpeculator): |
| 49 | @property |
| 50 | def advance_draft_positions(self) -> bool: |
| 51 | # Gemma4 MTP is Q-only and reads K/V from the target's existing cache. |
| 52 | # No new KV slots are written, so positions and seq_lens stay fixed. |
| 53 | return False |
| 54 | |
| 55 | def load_draft_model( |
| 56 | self, |
| 57 | target_model: nn.Module, |
| 58 | target_attn_layer_names: set[str], |
| 59 | ) -> nn.Module: |
| 60 | draft_vllm_config = self._create_draft_vllm_config() |
| 61 | with set_model_tag("eagle_head"): |
| 62 | draft_model = get_model( |
| 63 | vllm_config=draft_vllm_config, |
| 64 | model_config=self.speculative_config.draft_model_config, |
| 65 | load_config=get_draft_load_config(draft_vllm_config), |
| 66 | ) |
| 67 | self._setup_gemma4_kv_sharing(draft_model, target_attn_layer_names) |
| 68 | self._share_embeddings(draft_model, target_model) |
| 69 | return draft_model |
| 70 | |
| 71 | def _create_draft_vllm_config(self) -> VllmConfig: |
| 72 | """Preserve the target's forced TRITON_ATTN backend for draft layers. |
| 73 | |
| 74 | Gemma4 forces TRITON_ATTN due to heterogeneous head dimensions |
| 75 | (head_dim=256 sliding, global_head_dim=512 full). The base class |
| 76 | resets attention_config.backend to None for draft models, causing |
| 77 | sliding layers to fall back to FLASH_ATTN which cannot handle |
| 78 | KV-shared cache. Override to carry the target's backend through. |
| 79 | """ |
| 80 | draft_model_config = self.speculative_config.draft_model_config |
| 81 | draft_vllm_config = replace( |
| 82 | self.vllm_config, |
| 83 | model_config=draft_model_config, |
| 84 | ) |
| 85 | target_backend = self.vllm_config.attention_config.backend |
| 86 | if target_backend is not None: |
| 87 | draft_vllm_config = replace( |
| 88 | draft_vllm_config, |
| 89 | attention_config=replace( |
| 90 | draft_vllm_config.attention_config, |
| 91 | backend=target_backend, |
| 92 | ), |
| 93 | ) |
| 94 | return draft_vllm_config |
| 95 | |
| 96 | def _setup_gemma4_kv_sharing( |
| 97 | self, |
| 98 | model: nn.Module, |
| 99 | target_attn_layer_names: set[str], |
| 100 | ) -> None: |
| 101 | """Wire draft layers to share KV and KV scales with the target model. |
| 102 | |
| 103 | Each draft decoder layer is mapped to the last non-KV-shared |
| 104 | target layer of the same attention type (sliding or full). |
| 105 | """ |
| 106 | draft_config = self.speculative_config.draft_model_config.hf_config |
| 107 | draft_text_config = draft_config.get_text_config() |
| 108 | target_config = self.vllm_config.model_config.hf_config |
| 109 | target_text_config = target_config.get_text_config() |
| 110 | target_layer_types = getattr(target_text_config, "layer_types", []) |
| 111 | |
| 112 | if not (hasattr(model, "model") and hasattr(model.model, "layers")): |
| 113 | return |
| 114 | |
| 115 | target_num_kv_shared = getattr(target_text_config, "num_kv_shared_layers", 0) |
| 116 | target_num_layers = getattr( |
| 117 | target_text_config, "num_hidden_layers", len(target_layer_types) |
| 118 | ) |
| 119 | num_non_shared = target_num_layers - target_num_kv_shared |
| 120 | type_to_target_indices: dict[str, list[int]] = defaultdict(list) |
| 121 | for idx, lt in enumerate(target_layer_types[:num_non_shared]): |
| 122 | type_to_target_indices[lt].append(idx) |
| 123 | |
| 124 | target_prefix = "model.layers" |
| 125 | for name in target_attn_layer_names: |
| 126 | if ".layers." in name: |
| 127 | target_prefix = name.split(".layers.")[0] + ".layers" |
| 128 | break |
| 129 | |
| 130 | draft_layer_types = getattr(draft_text_config, "layer_types", []) |
| 131 | for draft_idx, layer in enumerate(model.model.layers): |
| 132 | if not hasattr(layer, "self_attn"): |
| 133 | continue |
| 134 | attn = getattr(layer.self_attn, "attn", None) |
| 135 | if attn is None: |
| 136 | continue |
| 137 | |
| 138 | draft_layer_type = ( |
| 139 | draft_layer_types[draft_idx] |
| 140 | if draft_idx < len(draft_layer_types) |
| 141 | else "full_attention" |
| 142 | ) |
| 143 | candidates = type_to_target_indices.get(draft_layer_type, []) |
| 144 | if not candidates: |
| 145 | logger.warning( |
| 146 | "No target layer of type '%s' for draft layer %d", |
| 147 | draft_layer_type, |
| 148 | draft_idx, |
| 149 | ) |
| 150 | continue |
| 151 | |
| 152 | target_idx = candidates[-1] |
| 153 | target_layer_name = f"{target_prefix}.{target_idx}.self_attn.attn" |
| 154 | attn.kv_sharing_target_layer_name = target_layer_name |
| 155 | |
| 156 | # KV-cache sharing aliases the cache tensor during allocation, but |
| 157 | # the quantization scales live on the Attention modules themselves. |
| 158 | # The BF16 draft model has no quantization config, so its K/V scales |
| 159 | # otherwise remain at the default 1.0 while it reads the target's |
| 160 | # calibrated FP8 cache. |
| 161 | target_attn = self.vllm_config.compilation_config.static_forward_context[ |
| 162 | target_layer_name |
| 163 | ] |
| 164 | _copy_target_kv_scales(attn, target_attn) |
| 165 | |
| 166 | logger.info( |
| 167 | "Gemma4 MTP: draft layer %d (%s) -> %s", |
| 168 | draft_idx, |
| 169 | draft_layer_type, |
| 170 | target_layer_name, |
| 171 | ) |
| 172 | |
| 173 | def _share_embeddings( |
| 174 | self, |
| 175 | draft_model: nn.Module, |
| 176 | target_model: nn.Module, |
| 177 | ) -> None: |
| 178 | target_language_model = ( |
| 179 | target_model.get_language_model() |
| 180 | if hasattr(target_model, "get_language_model") |
| 181 | else target_model |
| 182 | ) |
| 183 | if get_pp_group().world_size == 1: |
| 184 | target_embed = getattr(target_language_model.model, "embed_tokens", None) |
| 185 | if target_embed is not None: |
| 186 | del draft_model.model.embed_tokens |
| 187 | draft_model.model.embed_tokens = target_embed |