Skip to content

[Native] Add load-time options to engine contexts - #24636

Draft
metascroy wants to merge 1 commit into
offgraph-kv-native-exportfrom
offgraph-kv-engine-options
Draft

metascroy wants to merge 1 commit into
offgraph-kv-native-exportfrom
offgraph-kv-engine-options

Conversation

@metascroy

@metascroy metascroy commented Nov 5, 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

Adds EngineHost::create_context_with_options(program, package, options), where EngineOptions is a std::map<std::string, std::string> of named load-time settings for one context. The next PR uses it to pass the registry key of a KV cache the caller installed (llm_cache_registry_key).

The default implementation throws std::invalid_argument on any option and otherwise calls create_context, so each engine opts in to the options it understands. create_context and its callers (Session, Module) are unchanged.

An option can bind caller-owned state, like a KV cache, into several contexts. The doc comment says those contexts aren't isolated from each other: the caller must serialize their executables as if they shared one context.

This PR doesn't depend on the two below it; it's stacked to keep the series linear.

Test plan

  • Built the ggml engine and run_hf_llm_ggml, and ran native_ggml_engine_test: 36 passed. GpuUnavailableOnCpuVariant fails as it does on main, because this build has Metal enabled.
  • The next PR tests the ggml engine's handling of options, including rejecting an unknown one.

@pytorch-bot

pytorch-bot Bot commented Nov 5, 2026 •

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24636

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

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

As of commit b2aebff 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 5, 2026
@metascroy
metascroy force-pushed the offgraph-kv-engine-options branch from 4b5de65 to 0c97fdf Compare November 5, 2026 22:29
@metascroy
metascroy force-pushed the offgraph-kv-native-export branch from 01fcf32 to 05200ca Compare November 5, 2026 22:29
@metascroy
metascroy force-pushed the offgraph-kv-engine-options branch from 0c97fdf to 4ec9257 Compare November 11, 2026 18:05
@metascroy
metascroy force-pushed the offgraph-kv-native-export branch from 05200ca to 655cb68 Compare November 11, 2026 18:05
virtual std::unique_ptr<EngineContext> create_context_with_options(
std::shared_ptr<const Program> program,
std::shared_ptr<const Package> package,
const EngineOptions& options);

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

EngineOptions reads like the generic form of GgmlEngineOptions. Maybe EngineContextOptions, to match EngineContext?

EngineHost::create_context_with_options takes named string options for one
context, such as the registry key of a KV cache the caller installed. The
default implementation rejects any option, so engines opt in to the options
they understand; create_context and its callers are unchanged.

This branch had an error being deployed

1 failed deployment
cadence — b2aebff3 Deployed Nov 14, 2026 by metascroy via hifi-build / hifi4 #38328
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.

2 participants