IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
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 module

max.pipelines.speculative.cross_attention

The sequential proposers of a draft that keeps no cache of its own.

Gemma 4’s assistant has no K or V projections: each of its layers attends with its queries alone into the target’s own sliding-window or full-attention cache, so the driver hands it the target’s leaves. A ModuleV3 assistant runs through CrossAttentionModuleV3MTPProposer.

Drafts​

CrossAttentionDraftThe draft-side surface CrossAttentionMTPProposer drives.
CrossAttentionModuleV3DraftThe ModuleV3 draft CrossAttentionModuleV3MTPProposer drives.

Proposers​

CrossAttentionMTPProposerDrafts one token a step, attending into the target's caches.
CrossAttentionModuleV3MTPProposerDrafts one token a step on ModuleV3, attending into the target's caches.