Repository navigation
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24944
Note: Links to docs will display an error until the docs builds have been completed. ❌ 137 New Failures, 7 Unrelated Failures, 10 Unclassified FailuresAs of commit 3f908ee 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. |
meta-cla
Bot
added
the
CLA Signed
This was referenced Nov 12, 2026
--offgraph_cache now exports models with sliding-window layers instead of rejecting them. The window is a property of the cache, not the graph, so the export writes the cache geometry (get_n_caches, get_kv_heads, get_head_dims, get_windows) to the package's metadata, and the golden samples come from a reference cache whose sliding layers are rings. A layer's window follows HF: layer_types when the config has them, else sliding_window on every layer. The runner reads the windows from the metadata, checks the shapes against the graph, and builds ring layers sized for the method's token input. GgmlSequenceCache holds a ring layer at window + max_write - 1 rows from the start. The executor fills one set of row and mask uploads per window from the cache's plan: each row is masked by the logical position it holds, so a ring layer attends all its rows in any physical order, and a query sees only its own window, including within a prefill wider than the window.
metascroy
force-pushed
the
offgraph-kv-ptn-metadata
branch
from
November 14, 2026 00:18
f5f4802 to
f387dc6
Compare
metascroy
force-pushed
the
offgraph-kv-swa
branch
from
November 14, 2026 00:18
fc7c865 to
3f908ee
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
--offgraph_cachenow exports models with sliding-window layers instead of rejecting them. The window belongs to the cache, not the graph, so it travels in the package's metadata.get_n_caches,get_kv_heads,get_head_dims,get_windows). A layer's window follows HF:layer_typeswhen the config has them, elsesliding_windowon every layer (Mistral). Golden samples come from a reference cache whose sliding layers are rings.max_write). A package withoutget_windowsis refused rather than run as full attention.--warmup_prefillclears the cache after its warmup prefill, since a ring layer can't rewind to the start of a prompt longer than its window.GgmlSequenceCacheholds a ring layer atwindow + max_write - 1rows from the start; flat layers still grow on demand.Test plan
--warmup_prefillwith a chunked 37-token prompt on Mistral works; without the clear it fails with "cannot rewind to position 0". A 121-token chunked prefill matches HF eager to 2.4e-7 on Mistral (0.175 without the windows).Authored with Claude Code.