Skip to content

vllm.models.kimi_k3.amd.specialized_decode.mono.abi

The layout every MonoKernel's error words follow.

A rank's health words: a header whose word FAILED is set once any of its in-kernel waits gave up (sticky: a failed rank stays failed, and its later waits exit at once), then one record a CTA of the first wait it gave up on (flag, region id, pair index, tag, seen tag, CTA, clock, pad).

The output rank's flags: one int32 word a TP rank in the first PEER_HEAD bytes of the output rank's peer buffer. A rank whose wait gives up writes its word there, drains the store, then sets FAILED, so nothing it stores later can overtake the flag. The runner reads these words after every step.

Functions:

  • health_words –

    A rank's health words for a grid of ctas: the header, a record a CTA.

health_words(ctas)

A rank's health words for a grid of ctas: the header, a record a CTA.

Source code in vllm/models/kimi_k3/amd/specialized_decode/mono/abi.py 
24
25
26
def health_words(ctas: int) -> int:
    """A rank's health words for a grid of ``ctas``: the header, a record a CTA."""
    return (ctas + 1) * RECORD_WORDS