| 130 | |
| 131 | |
| 132 | class DFlashDraftModel(nn.Module): |
| 133 | def __init__(self, config: DFlashConfig): |
| 134 | super().__init__() |
| 135 | self.config = config |
| 136 | if not self.config.layer_types: |
| 137 | self.config.layer_types = ("full_attention",) * self.config.num_hidden_layers |
| 138 | concat_dim = len(config.target_layer_ids) * config.hidden_size |
| 139 | self.fc = nn.Linear(concat_dim, config.hidden_size, bias=False) |
| 140 | self.hidden_norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 141 | self.layers = [DFlashDecoderLayer(config, i) for i in range(config.num_hidden_layers)] |
| 142 | self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 143 | self.rope = _build_rope( |
| 144 | config.head_dim, |
| 145 | config.rope_theta, |
| 146 | config.max_position_embeddings, |
| 147 | config.rope_scaling, |
| 148 | ) |
| 149 | self.embed_tokens = None |
| 150 | self.lm_head = None |
| 151 | self.embed_scale = 1.0 |
| 152 | |
| 153 | def bind(self, target_model): |
| 154 | if hasattr(target_model, "embed_tokens"): |
| 155 | inner = target_model |
| 156 | elif hasattr(target_model, "model") and hasattr(target_model.model, "embed_tokens"): |
| 157 | inner = target_model.model |
| 158 | elif (hasattr(target_model, "language_model") and |
| 159 | hasattr(target_model.language_model, "model") and |
| 160 | hasattr(target_model.language_model.model, "embed_tokens")): |
| 161 | inner = target_model.language_model.model |
| 162 | else: |
| 163 | raise AttributeError(f"Cannot find embed_tokens in {type(target_model).__name__}") |
| 164 | self.embed_tokens = inner.embed_tokens |
| 165 | self.embed_scale = getattr(self.embed_tokens, "embed_scale", getattr(inner, "embed_scale", 1.0)) |
| 166 | lm = getattr(target_model, "language_model", target_model) |
| 167 | self.lm_head = getattr(target_model, "lm_head", None) or getattr(lm, "lm_head", None) or self.embed_tokens.as_linear |
| 168 | return self |
| 169 | |
| 170 | def make_cache(self): |
| 171 | caches = [] |
| 172 | for layer_type in self.config.layer_types: |
| 173 | if layer_type == "sliding_attention": |
| 174 | if self.config.sliding_window is None: |
| 175 | raise ValueError("Draft config must define sliding_window for sliding_attention layers.") |
| 176 | caches.append(RotatingKVCache(max_size=self.config.sliding_window - 1, keep=0)) |
| 177 | else: |
| 178 | caches.append(KVCache()) |
| 179 | return caches |
| 180 | |
| 181 | def __call__( |
| 182 | self, |
| 183 | inputs, |
| 184 | target_hidden, |
| 185 | cache, |
| 186 | logits_start: int = 0, |
| 187 | ): |
| 188 | h = self.embed_tokens(inputs) * self.embed_scale |
| 189 | h_ctx = self.hidden_norm(self.fc(target_hidden)) |