Repository navigation
Playground chat: Gemma 3 1B Instruct replaces Llama 3.2 1B (#24014) - #24014
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24014
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@psiddh has exported this pull request. If you are a Meta employee, you can view the originating Diff in D126233262. |
This PR needs a
|
…4014) Summary: On GPUs with `shader-f16` (Apple GPUs, most others), Llama 3.2 1B in the playground's On-device Chat fails on its first reply with "linear_q4gsw_bk64_qkv(resize): separate projection dispatch was not restored before the QKV resize hook ran". That is a WebGPU runtime bug in the fused Q/K/V path, which only Llama's shapes take; D126153972 fixes it upstream. Until the fixed runtime ships, the chat offers Qwen 3 0.6B and Gemma 3 1B Instruct. - Gemma 3 1B Instruct is loaded from `executorch-community/Gemma-3-1B-IT-ExecuTorch-WebGPU` on Hugging Face, pinned to commit `91864e51`, with a sha256 check per file. Export: 4-bit groupwise weights, exact sliding-window and global attention, 8,960-token context. Its repo carries the Gemma Terms of Use, the Prohibited Use Policy and the required NOTICE. Legal approved publishing it. - The chat shows Gemma's required notice ("Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms") when Gemma is selected. Llama leaves the chat and the Models page; its Hugging Face repo stays. - `llm.js`: when a model's chat template can't render the conversation, the chat reports it instead of silently falling back to ChatML, which only Qwen understands. - Model cards lay out in one row. `fbcode/executorch/website` and `xplat/executorch/website` stay byte-identical. Differential Revision: D126233262
meta-codesync
Bot
changed the title
psiddh
force-pushed
the
export-D126233262
branch
from
October 20, 2026 07:01
4075b7c to
f8374ed
Compare
Gasoonjia
approved these changes
Oct 20, 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.
Summary:
On GPUs with
shader-f16(Apple GPUs, most others), Llama 3.2 1B in the playground's On-device Chat fails on its first reply with "linear_q4gsw_bk64_qkv(resize): separate projection dispatch was not restored before the QKV resize hook ran". That is a WebGPU runtime bug in the fused Q/K/V path, which only Llama's shapes take; D126153972 fixes it upstream. Until the fixed runtime ships, the chat offers Qwen 3 0.6B and Gemma 3 1B Instruct.executorch-community/Gemma-3-1B-IT-ExecuTorch-WebGPUon Hugging Face, pinned to commit91864e51, with a sha256 check per file. Export: 4-bit groupwise weights, exact sliding-window and global attention, 8,960-token context. Its repo carries the Gemma Terms of Use, the Prohibited Use Policy and the required NOTICE. Legal approved publishing it.llm.js: when a model's chat template can't render the conversation, the chat reports it instead of silently falling back to ChatML, which only Qwen understands.fbcode/executorch/websiteandxplat/executorch/websitestay byte-identical.Differential Revision: D126233262