vllm.model_executor.kernels.attention.dsa.dcp_topk_direct
¶
Direct DCP top-k candidate exchange over PyTorch symmetric memory.
Replaces the NCCL all-gather of the DCP indexer merge for small batches. Each rank packs its local (score, global token id) candidates, stores them into every peer's inbox slice, then publishes an epoch and waits for all peers. The stable top-k selector reads the completed local inbox directly. Two alternating inboxes let the next exchange start without a read acknowledgement: a rank can only overwrite a slot after every peer has published the exchange in between, i.e. after their selectors consumed it.
On SM90+, programmatic dependent launches overlap the dispatch of the pack, push, publish and select kernels; device-side waits keep the data dependencies.
Classes:
-
DirectDCPTopKWorkspace–Symmetric ping-pong inboxes for the DCP top-k candidate exchange.
Functions:
-
get_direct_dcp_topk_workspace–Collective on first call (symmetric-memory rendezvous): every DCP rank
DirectDCPTopKWorkspace
¶
Bases: DirectCPWorkspace
Symmetric ping-pong inboxes for the DCP top-k candidate exchange.
Source code in vllm/model_executor/kernels/attention/dsa/dcp_topk_direct.py
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 | |
get_direct_dcp_topk_workspace(group, device, max_rows, candidates_per_rank, use_direct)
cached
¶
Collective on first call (symmetric-memory rendezvous): every DCP rank must call it with the same arguments, outside CUDA graph capture.