vllm.models.deepseek_v41.nvidia.specialized_decode.logits.logits

Logits of the specialized decode: the low-batch LM head GEMM (lm_head) and vocab-parallel logits gathered with the executor's all-gather.

For decode batches a one-shot all-gather over NVLink multicast is several times faster than the NCCL ring. The executor's LamportAllGather takes the logits when they fit, by size alone, so every TP rank takes the same path; larger ones go to NCCL. vLLM's custom all-gather is not used: its MNNVL Lamport kernel never waits for a rank's own multicast copy (see ops.cute_dsl.all_gather). The all-gather's mailbox also serves the row-sharded indexers, so the logits must be gathered on the forward stream after any side stream has rejoined it.