Repository navigation
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24949
Note: Links to docs will display an error until the docs builds have been completed. ❌ 136 New Failures, 7 Unrelated Failures, 10 Unclassified FailuresAs of commit ebd18ef with merge base c7f7931 ( 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. |
This was referenced Nov 12, 2026
meta-cla
Bot
added
the
CLA Signed
This was referenced Nov 12, 2026
gguf_hf takes tensor names from gguf-py's tensor map, which has no entry for Muse Glimmer's muse-glimmer architecture, so add a table from its GGUF names to the Hugging Face MuseGlimmerForConditionalGeneration parameters under model.language_model. The all-ones attn_q_norm/attn_k_norm tensors are skipped: the HF QK norm has no weight. The conversion undoes llama.cpp's Q/K row permutation, as for Llama, and the 1 it adds to the four centered decoder norms; the final norm is a plain gain and stays as stored. Checked on the first four layers of Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf against Muse Glimmer's own model and GGUF loader: FP32 logits agree to 5e-5. The tests need a transformers with Muse Glimmer and skip otherwise.
metascroy
force-pushed
the
offgraph-kv-mg-gguf
branch
from
November 14, 2026 00:18
2153eec to
ebd18ef
Compare
metascroy
force-pushed
the
offgraph-kv-ggml-q5k
branch
from
November 14, 2026 00:18
fd2123f to
acd0758
Compare
This branch had an error being 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
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.
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
Maps Muse Glimmer GGUF tensors onto the Hugging Face
MuseGlimmerForConditionalGenerationparameters undermodel.language_model.muse-glimmerarchitecture, sogguf_hfgets an explicit table, for architectures without one. Theattn_q_norm/attn_k_normtensors are skipped: the HF QK norm has no weight. llama.cpp's converter stores the query scale HF reads from the config (qk_scale_factor) asattn_q_normand ones asattn_k_norm, soload_ggufchecks them against the config (check_config_tensors) instead of dropping them silently.Test plan
test_gguf_hf.py: names, conversions, and the QK norm check on a tiny Muse Glimmer. Needs a transformers with Muse Glimmer; skips otherwise.attn_q_norm= 3.87); a config with another scale is rejected.Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf, HF with this mapping matches Muse Glimmer's own model and GGUF loader to 4.9e-5 at FP32, with the same top token at every position. Dropping the Q/K un-permute or the norm offset breaks it (max errors 7.5 and 13.8).Authored with Claude Code.