[#9241][feat] AutoDeploy: Support Eagle3 Speculative Decoding#9869

Merged

govind-ramnarayan merged 18 commits intoNVIDIA:mainfrom

nv-auto-deploy:gramnarayan/mtp-eagle

Dec 25, 2025

Commits on Dec 24, 2025

Init commit for EagleDecodingConfig support in AutoDeploy. Adds: file for running AD with Eagle1, modifies existing TRTLLM example to use Eagle1 instead of Eagle3 for comparison. Eagle1 is chosen (…
govind-ramnarayan
committed
removing print statements for num accepted tokens
govind-ramnarayan
committed
removed some prints and added license
govind-ramnarayan
committed
fixing comments from AI
govind-ramnarayan
committed
revert _compute_logits() args
govind-ramnarayan
committed
point integration test to existing eagle3 directory
govind-ramnarayan
committed
fix error with hidden state transform on rebase - implement abstract class method
govind-ramnarayan
committed
remove triton backend workaround from test
govind-ramnarayan
committed
slimming down cached attention op for hidden states
govind-ramnarayan
committed
more robust sharing of embedding and lm_head weights with PyTorchModelEngine drafter
govind-ramnarayan
committed
counting number of hidden layers inside transform
govind-ramnarayan
committed
set default layers to capture instead of classmethod
govind-ramnarayan
committed
adding review comments addressing again
govind-ramnarayan
committed
finished review comments
govind-ramnarayan
committed
restoring original spec dec tests for draft-target. Adding test for acceptance rates for Eagle3 + Llama
govind-ramnarayan
committed
some small cleanups to the integration test after reverting
govind-ramnarayan
committed
fix precommit
govind-ramnarayan
committed
reverting batch_size > 1 test from before; it seems to fail due to IFB nondeterminism. Need to figure out how to meaningfully test batch_size > 1 in a deterministic way
govind-ramnarayan
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[#9241][feat] AutoDeploy: Support Eagle3 Speculative Decoding#9869

[#9241][feat] AutoDeploy: Support Eagle3 Speculative Decoding#9869
govind-ramnarayan merged 18 commits intoNVIDIA:mainfrom
nv-auto-deploy:gramnarayan/mtp-eagle

Commits on Dec 24, 2025

Init commit for EagleDecodingConfig support in AutoDeploy. Adds: file for running AD with Eagle1, modifies existing TRTLLM example to use Eagle1 instead of Eagle3 for comparison. Eagle1 is chosen (…

removing print statements for num accepted tokens

removed some prints and added license

fixing comments from AI

revert _compute_logits() args

point integration test to existing eagle3 directory

fix error with hidden state transform on rebase - implement abstract class method

remove triton backend workaround from test

slimming down cached attention op for hidden states

more robust sharing of embedding and lm_head weights with PyTorchModelEngine drafter

counting number of hidden layers inside transform

set default layers to capture instead of classmethod

adding review comments addressing again

finished review comments

restoring original spec dec tests for draft-target. Adding test for acceptance rates for Eagle3 + Llama

some small cleanups to the integration test after reverting

fix precommit

reverting batch_size > 1 test from before; it seems to fail due to IFB nondeterminism. Need to figure out how to meaningfully test batch_size > 1 in a deterministic way