Add the Eagle3Speculator module for speculative decoding#20153
Draft
digantdesai wants to merge 1 commit into
Draft
Add the Eagle3Speculator module for speculative decoding#20153digantdesai wants to merge 1 commit into
digantdesai wants to merge 1 commit into
Conversation
Contributor
Author
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20153
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 2 PendingAs of commit 66da947 with merge base dc55469 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Jun 9, 2026
This was referenced Jun 9, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Eagle3Speculator holds a shared target and an EAGLE-3 draft head and exposes
prefill, target_verify, draft_decode, and decode. It is written against the
TapTarget protocol (target.py) rather than a concrete model, so any target that
exposes EAGLE-3 hidden-state taps works; gemma4-31B is the reference, registered
in target.py's TARGETS (loader + forward-length bounds). The export lowers only
the first three methods; under the shifted (vLLM-EAGLE) runner scheme decode is
unnecessary.
The target methods return the fused (hidden-size) draft feature, so draft_decode
takes a single uniform feature whether it comes from the target (confirmed
positions) or the draft's own recurrent output (proposed positions).
target_verify returns the per-position greedy argmax target id; verify_ids[i] is
the greedy token after token i. The draft pairs target hidden_state_t with
token_{t+1}, so after verification the runner reseeds the draft cache from the
feature target_verify already produced for the accepted positions, paired with
the corrected token; the corrected/bonus token never needs its own target
forward, so the three exported methods suffice at one target forward per round.
Draft heads without their own embedding table are rejected.
Tests cover each method, target_verify reusing a prefilled cache, the
verification alignment, a cached draft chain matching the stateless recompute,
the full shifted speculative loop reproducing greedy decoding, and TapTarget
protocol conformance.
Authored with assistance from Claude Code.