Skip to content

SGL: Add Gemma4 31b support#542

Draft
xiangyuT wants to merge 5 commits into
intel:mainfrom
xiangyuT:gemma4-bmg-sync
Draft

SGL: Add Gemma4 31b support#542
xiangyuT wants to merge 5 commits into
intel:mainfrom
xiangyuT:gemma4-bmg-sync

Conversation

@xiangyuT

Copy link
Copy Markdown
Contributor

No description provided.

gc-fu and others added 5 commits July 15, 2026 09:07
Sync the updated SGLang and XPU kernel patches, port the validated Gemma4 ESIMD decode paths into the unified kernel package, and add a TP=2 eager launch recipe without local development artifacts.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Restore the original README scope, migrate all XPU fast-path gates to the SGLANG prefix, and remove model-specific labels from general ESIMD operators.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Keep the README and launch recipe aligned with the llm-scaler upstream naming contract and update every corresponding fast-path reader in the SGLang patch.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Include new files that were lost when regenerating the upstream-base patch from an unstaged worktree.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Port the validated oneDNN W8A16 path into the unified SGLang custom kernel wheel and route XPU FP8 prefill through it while preserving the ESIMD decode path.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants