Repository navigation
[Weight Reload][Gemma4] Refresh MTP gathered LM head in stable storage - #59000
Draft
Ronald1995 wants to merge 1 commit into
Draft
Ronald1995 wants to merge 1 commit into
Ronald1995 wants to merge 1 commit into
Conversation
This was referenced Sep 28, 2026
Ronald1995
force-pushed
the
codex/gemma4-mtp-derived-reload
branch
from
September 28, 2026 09:03
cfd93e0 to
31f83f2
Compare
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Ronald1995
force-pushed
the
codex/gemma4-mtp-derived-reload
branch
from
September 28, 2026 09:05
31f83f2 to
728be8d
Compare
This was referenced Oct 5, 2026
Open
This branch has not been deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Purpose
Declare the TP gathered full LM head as a derived buffer and recompute it after source weights are stable. Return the owned stable tensor to logits consumers, and avoid clearing its allocation on load. Register suppression indices as a static non-persistent buffer rather than a weight-derived cache. Adds repeated refresh coverage that checks object identity, addresses, and values.
Merge order
Merge #53509 first. This branch starts directly from main and contains only this migration, with no prerequisite PR commits. It is not stacked. Rebase onto main after the core merges before running integration CI.
Tracks #57895; ports the relevant portion of aoshen02#56.
Validation
Local Ruff formatting and lint checks completed. Full Linux/backend integration and CUDA Graph replay remain pending; Windows pytest collection is blocked by unavailable uvloop. No remote GPU tests or H20 tests were run.
AI disclosure
Code and PR text prepared with AI assistance. Human review is pending; this PR is a draft.