Skip to content

vllm.models.deepseek_v41.amd.mono.common

The mono kernels' shared mechanisms: ops (wave / math primitives, the TP group's peer buffers), sync (the tagged mailbox), mx (FP8 / MX arithmetic), plan (execution model, layouts, build keys) and peer_memory (symmetric peer buffers).

Modules:

  • mx –

    FP8 / MX arithmetic the mono kernels share: the E4M3 range, E8M0 scale

  • ops –

    Device-side primitives of the mono kernels: wave reductions and shuffles, the

  • peer_memory –

    Symmetric peer memory for the in-kernel TP all-reduces, and the agreement every

  • plan –

    Plain-Python planning of the mono kernels: the execution model (one CTA a CU,

  • sync –

    The tagged mailbox: hand-offs between the CTAs of one GPU (device scope) and