Repository navigation
Conversation
Collaborator
|
hey thanks for this, seems like a good idea! I can review and merge when you are ready @joe-sama |
joe-sama
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Cinque Terre V3 can compute historical attention independently of the current image. This adds an opt-in path that runs HTP vision and GPU history concurrently, then HTP current-token decoding and CPU output/state assembly. Original weights, precision and same-frame feedback are preserved.
QNN discovery is generic rather than Xiaomi/SM8750-pinned. The route uses strict HTP placement, real shared allocations, dependency/error barriers and validated model/part hashes. Unsupported models select whole-NPU execution before preparation, and the UI reports the actual route. Compatibility still requires device accuracy and sustained-performance checks. The shared QNN foundation is also isolated in #19; merge that first, then rebase this PR to remove the overlap.
The diff now contains 41 source/test/doc files (+2,174/−55). Research telemetry and device-pinned adapter settings were removed; historical evidence remains linked at its original commit. No SDK binaries, model weights or APKs are committed.
Validation: 42 focused Swift tests, 59 Android unit tests in each default/opt-in build, both release APK builds, Android native compilation, and adapter forwarding/failure/lifetime tests passed. All four generated graphs and the manifest reproduce the phone-tested bundle byte-for-byte. All 41 published blobs and the complete tree match the tested source. An expanded WSL run had eight large-UDP frame-assembly issues in unchanged code; those remain outside this patch.
The earlier Xiaomi 15 prototype measured 27.2 ms mean / 29.5 ms P99 over 12,001 frames (10 minutes), with zero frames above 50 ms; unchanged original-reference gates passed 128 frames before and after. These are historical prototype results: this exact portable APK has not been phone-qualified, other Snapdragon devices are unqualified, and provider-call overlap does not measure physical NPU occupancy. Kept as a draft pending device qualification.
Build options and reproduction. Related: #17, #19.