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

Proposed

`Proposed`

class max.pipelines.speculative.driver.Proposed(logits, hidden, reuse=<factory>, carry=<factory>, token=None)

source

One draft invocation’s contribution.

Carries the step’s logits alongside the hidden state the next step consumes.

Parameters:

carry: list[TensorValue]

source

Step 0’s hidden state, already gathered at each accepted position.

Empty for a draft that hands back its whole verified window and lets the driver do the gather

hidden: list[TensorValue]

source

logits: TensorValue | None

source

Logits over this call’s query, None only alongside token.

reuse: list[TensorValue]

source

Per-device work step 0 did that every later step reuses unchanged.

Read from the prefill only. A later step returns nothing here, because reusing step 0’s result is the point.

token: TensorValue | None = None

source

Step 0’s proposed token, for a draft that produced carry.