vllm.models.mistral.mistral_large4.nvidia.dspark
¶
Mistral Large 4 DSpark draft model (dense MLA backbone + Markov head).
Draft layers are non-causal dense MLA blocks with the target's attention
geometry. Each draft layer caches latent K/V projected from the normalized
target taps (combine_hidden_states) and from its own query block. Draft
weights are BF16 and the draft RoPE is the target's YaRN without LeRoPE.
Checkpoint options select the trained variant:
dspark_head_type="rnn": the Markov bias comes from a gated recurrent state carried across the draft positions of one proposal instead of from the previous token alone.dspark_draft_input="position_specific":mask_embeddinghas one row per mask slot; slotkof a[anchor, mask_1, ...]block reads rowk - 1instead of sharing one vector.layer_types/sliding_window: windowed draft layers see the lastsliding_windowcontext tokens before the anchor plus their whole block.
Classes:
Functions:
-
draft_attention_windows–Per-layer decode window (keys counted back from the end of the block).
MistralLarge4DSparkModel
¶
Bases: Module
Methods:
-
embed_input_ids–Embed draft blocks, replacing the out-of-vocabulary mask slots.
-
markov_step–One recurrent Markov step: (output embedding, next state).
-
precompute_and_store_context_kv–Write each draft layer's latent K/V for the target context tokens.
Source code in vllm/models/mistral/mistral_large4/nvidia/dspark.py
83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | |
embed_input_ids(input_ids)
¶
Embed draft blocks, replacing the out-of-vocabulary mask slots.
Source code in vllm/models/mistral/mistral_large4/nvidia/dspark.py
markov_step(token_ids, draft_hidden, state)
¶
One recurrent Markov step: (output embedding, next state).
Source code in vllm/models/mistral/mistral_large4/nvidia/dspark.py
precompute_and_store_context_kv(context_states, context_positions, context_slot_mapping=None)
¶
Write each draft layer's latent K/V for the target context tokens.
The context is already normalized (target_norm), so it bypasses the
layers' input norms, as in training.
Source code in vllm/models/mistral/mistral_large4/nvidia/dspark.py
draft_attention_windows(config, num_speculative_tokens)
¶
Per-layer decode window (keys counted back from the end of the block).
The draft block is part of the attended sequence, so the window is widened
by its length: the queries then see the sliding_window context tokens up
to the anchor's source position, as in training.