Skip to content

[Native] Run GGUF Q5_K weights in the ggml engine - #24948

Draft
metascroy wants to merge 1 commit into
offgraph-kv-gguf-layersfrom
offgraph-kv-ggml-q5k
Draft

metascroy wants to merge 1 commit into
offgraph-kv-gguf-layersfrom
offgraph-kv-ggml-q5k

Conversation

@metascroy

@metascroy metascroy commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Stack

Off-graph KV cache for the native ggml engine, through serving Muse Glimmer on it. Review bottom-up; each PR builds and passes its tests on its own.

Summary

The ggml engine took packed GGUF weights only as Q4_K or Q6_K, but K-quant mixes such as Q4_K_M store some tensors as Q5_K; Muse Glimmer's LM head is one. Export already carries any GGUF type as a gguf:<type> PackedQuant, so only the engine needed it. A table of K-quant codecs (ggml type and block bytes) replaces the Q4_K/Q6_K conditionals, with Q5_K added.

Test plan

New engine test: a packed Q5_K linear with every fifth bit set computes the expected dot product, on CPU and on Metal.


Authored with Claude Code.

@pytorch-bot

pytorch-bot Bot commented Nov 12, 2026 •

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24948

Note: Links to docs will display an error until the docs builds have been completed.

❌ 138 New Failures, 7 Unrelated Failures, 10 Unclassified Failures

As of commit acd0758 with merge base c7f7931 (image):

NEW FAILURES - The following jobs have failed:

UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Nov 12, 2026
The engine took packed GGUF weights only as Q4_K or Q6_K, but K-quant mixes
such as Q4_K_M store some tensors as Q5_K; Muse Glimmer's LM head is one.
Export already carries any GGUF type as a gguf:<type> PackedQuant, so only
the engine needed it. A table of K-quant codecs (ggml type and block bytes)
now replaces the Q4_K/Q6_K conditionals, with Q5_K added.

This branch had an error being deployed

1 failed deployment
cadence — acd07589 Deployed Nov 14, 2026 by metascroy via vision-build / vision #38333
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant