Gemma4-31B-IT with MTP Speculative Decoding on Trillium TPU v6e-4#261
Draft
jawadaminGOOG wants to merge 5 commits into
Draft
Gemma4-31B-IT with MTP Speculative Decoding on Trillium TPU v6e-4#261jawadaminGOOG wants to merge 5 commits into
jawadaminGOOG wants to merge 5 commits into
Conversation
Add server startup script, benchmarking scripts, hotpatches, configurations, benchmark performance summaries, and technical hotpatch integration guide to enable speculative decoding stably on TPU v6e. TAG=agy CONV=0408a73a-f24b-4400-97f9-706785f355f1
…n scripts - Patched API server startup to use block size 32 and propose length 5. - Fixed multimodal benchmarks command to use updated random-mm dataset flags. - Staged all assistant MTP model attention layer overrides and loaders. TAG=agy CONV=0408a73a-f24b-4400-97f9-706785f355f1
TAG=agy CONV=53a929f7-98b1-4674-bb2b-632a7ce5aa01
jawadaminGOOG
marked this pull request as draft
June 18, 2026 15:50
Replaces the Bare Metal TPU VM serving instructions with GKE configurations. Creates a unified YAML manifest gemma4-mtp-gke.yaml containing the Namespace, StorageClass, PersistentVolumeClaim, Service, and Deployment specs. Modifies README.md to guide nodepool setup, configmap patches installation, and documents verified GKE serving performance metrics. TAG=agy CONV=c4055879-301f-4770-a69a-87b9593ce043
…port for Gemma 4 MTP TAG=agy CONV=c4055879-301f-4770-a69a-87b9593ce043
jawadaminGOOG
force-pushed
the
spec-decode-gemma
branch
from
July 8, 2026 19:00
5b43efe to
a46f24d
Compare
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This guide shows how to serve the Gemma 4 IT model (google/gemma-4-31B-it) with vLLM using speculative decoding on Trillium (TPU v6e) VMs. We use the official google/gemma-4-31B-it-assistant companion model as the draft model.
Note: Speculative decoding for Gemma 4 on TPU currently requires specific python package hotpatches and source overrides to run stably without Out-of-Memory (OOM) or shape mismatch crashes. All necessary python files and hotpatch scripts are included in this directory.