For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Python class
DraftCache
DraftCache
class max.pipelines.speculative.driver.DraftCache(value, names=<not given>, *values, module=None, qualname=None, type=None, start=1, boundary=None)
Bases: Enum
Which cache the draft writes, and so what the driver advances.
OWN
OWN = 'own'
The draft has a cache of its own.
The driver advances its lengths by the accepted count before the loop and by one per step, and substitutes them into the collections.
PER_DEPTH
PER_DEPTH = 'per_depth'
Each depth owns a cache and writes a single token into it.
A step’s window therefore starts empty rather than extending the one before it, so the driver zeroes the lengths instead of advancing them. Declaring the ownership is what selects that; a draft cannot ask for the zeroing and the advance at once.
TARGET
TARGET = 'target'
The draft only reads the target’s cache.
It writes no KV slot, so its position stays fixed across the steps and the
per-step advance is skipped. The accepted-count advance still runs,
reaching the draft as SequentialBatch.draft_cache_lengths rather
than being substituted into the collections, which would lengthen the
window it attends over.