Repository navigation
Add optional tree speculation to the serving runtime - #1
Merged
Merged
Conversation
Share candidate ancestors with BF16 parent-state semantics and explicit commit/discard ownership. Add an eight-node best-first DFlash policy and bound PLE hashing history. Keep tree generation experimental and separate from serving. Validation: 688 CPU tests (13 tokenizer skips), 36 exact-weight GPU node/state/continuation checks, and 30 virtual Agent workflows. No production promotion or release approval.
Select tree decoding at startup while retaining MTP by default. Keep per-request target/DFlash state, paired prefill-cache reuse, streaming/tools, reasoning limits and cancellation on the existing owner-thread scheduler. Reuse linear verification graphs for chains and detach committed graph views. Validation: 719 CPU tests passed (13 tokenizer skips); separately started MTP/tree containers passed 14/16 real HTTP checks with the current checkpoint. Tree lanes cooperate without GPU batching; full quality/capacity gates remain open.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add startup-selectable tree speculation to the normal serving path. The Docker launcher uses
SPECULATION=tree; direct entrypoints acceptSPARK_SPECULATION=treeor--speculation tree. MTP remains the default, and health/scheduler metadata identifies the selected mode.Tree sessions preserve private target/DFlash histories, the existing paired prefill cache, SSE/tool framing, reasoning controls and cancellation. Requests run cooperatively; unbranched trees reuse linear verification graphs. Committed graph views are detached regardless of lazy-clone configuration.
Validation:
This is an experimental opt-in feature. Tree GPU batching, broad quality evaluation and four full 262K contexts remain unverified. Historical in-process TPS figures are not presented as measured speedups of the serving switch.