vllm.models.deepseek_v41.amd.mono.keylist

K2's key lists without K1's stage_kt: the KT and KLEN that it writes.

On an index layer (2, 8, 14, 20, 24, 28, 32 and 36 of DeepSeek-V4.1-Flash) vLLM's compressor and indexer run between K1 and K2, and the indexer writes the layer's top-512, which the keys need. So K1 writes no keys there (FrontBuild.index), but K2 reads each token's keys from KT and KLEN, which stage_kt writes on the other layers. launch writes them with one Triton program a token, from the same inputs stage_kt reads, in the same format:

KT[t * KEYS + k] = the compressed slot | 1 << 31 of top-k entry k, for
                   k < ntopk (-1 where the entry is -1), then the window
                   slots, then -1 up to KEYS
KLEN[t] = ntopk + the window length, KLEN[S + t] = ntopk

ntopk is min((position + 1) // ratio, TOPK), as in stage_kt, and a token whose slot is -1 has no keys. tests/kernels/models/amd/test_dsv41_mono_gfx942.py compares the result with a reference built in torch.