vllm.models.deepseek_v41.nvidia.specialized_decode.ampere.indexer_metadata
¶
The DeepSeek-V4.1 indexer metadata on SM80.
Each step, the indexer metadata builder tells the indexer which tokens are decodes and which are prefills, which compressed keys each one may score, and where the key cache pages of each request are. The stock builder also plans DeepGEMM's paged-logits schedule for every decode step, and DeepGEMM has no SM80 kernels (a vLLM install bundles it, so it is present but fails). This builder keeps the stock one's state and its prefill chunk planning, and builds the decode part the SM80 logits kernel reads:
- one row per decode token, holding that token's causal compressed length and a copy of its request's block table (the stock flattening kernel);
- no DeepGEMM schedule.
Neither decode nor prefill context parallelism is supported, and the index
K cache is always MXFP4 (indexer_uses_fp4).
Classes:
-
AmpereIndexerBackend–The index K cache backend of the SM80 layer.
-
AmpereIndexerMetadataBuilder–The stock indexer metadata on the MXFP4 cache, with one decode row per
AmpereIndexerBackend
¶
Bases: DeepseekV41IndexerBackend
The index K cache backend of the SM80 layer.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/indexer_metadata.py
AmpereIndexerMetadataBuilder
¶
Bases: DeepseekV32IndexerMetadataBuilder
The stock indexer metadata on the MXFP4 cache, with one decode row per token and no DeepGEMM schedule.
Methods:
-
build–Decode rows per token, stock prefill chunks.
-
update_draft_decode_metadata–The stock draft-step update (one token per request), without the
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/indexer_metadata.py
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 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 | |
_build_decode(m, num_decodes, num_decode_tokens)
¶
One row per decode token: its causal compressed length (zero for a CUDA-graph padding row) and its request's block table.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/indexer_metadata.py
_build_prefill(m, num_decodes, num_prefills)
¶
The stock prefill chunks under the logits budget, and the TP row
shares once the step's scoring work reaches _ROW_SHARD_MIN_WORK.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/indexer_metadata.py
build(common_prefix_len, common_attn_metadata, fast_build=False)
¶
Decode rows per token, stock prefill chunks.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/indexer_metadata.py
update_draft_decode_metadata(metadata)
¶
The stock draft-step update (one token per request), without the DeepGEMM schedule.