fix: select one discovered HTP session for Hexagon inference - #907
Merged
Conversation
![pocketpal-dev-team[bot]](https://avatars.githubusercontent.com/in/4815089?s=60&v=4){kind=small}
Contributor
Author
There was a problem hiding this comment.
Independent review of 47a26875 is complete across architecture, QA, security, performance, mobile, data, UX and local invariants. No source-code findings were identified. Independent lint, typecheck, full tests and coverage pass; the paired device and actual-generation evidence was reviewed with its documented artifact and experimental limits.
Approval remains pending the required visual evidence: all eight captures exist and were inspected, but uploading them failed because the GitHub browser session expired (HTTP 302). Refreshing that authentication and posting the captures resolves this prerequisite; no code change or new device run is requested for it.
This draft targets the pinned base for #901. GitHub currently lists no status checks for this temporary base; retarget after #901 merges and run the applicable checks on the resulting branch.
Generated by PocketPal Dev Team
a-ghorbani
force-pushed
the
feature/TASK-20260909-1820
branch
from
September 10, 2026 09:37
47a2687 to
d50f964
Compare
Owner
PR #907 benchmark results
poco-myron · cpu · 29 cells
poco-myron · hexagon · 29 cells
poco-myron · gpu · 29 cells
samsung-s23 · cpu · 29 cells
samsung-s23 · hexagon · 29 cells
samsung-s23 · gpu · 29 cells
poco-x7-klee · cpu · 29 cells
PR #907 feature e2e
Failures were rerun alone on the #907 APK and on the #901 APK (
|
Owner
PR #907 compared with release v1.17.2
Myron GPU Q4_K_M regression
Smoke tier, prefill / decode tok/s. v1.16.0 uses its cool smoke rerun (its focused tier is contaminated).
Feature e2e compared with v1.17.2
|
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
1 participant
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.
Hexagon selection currently emits
['HTP*'], which requests every discovered HTP session. On devices exposing several sessions, this substantially increases native compute-buffer allocation. Select the first exact runtime HTP name instead, and resolve existing saved wildcard, stale, or multiple HTP selections before normal model initialization.Resolution preserves saved preferences and the GPU-layer setting. If discovery fails or no HTP device is available, that load uses CPU with zero GPU layers; a later load can recover Hexagon. Settings and benchmarks share the same device choice, while unavailable benchmark backends still fail their cell rather than measure CPU. CPU/OpenCL/iOS selection and flash policy are unchanged.
Fixes #904. Based on
mainafter #901 merged; the diff contains only the HTP selection fix and its tests.Validation
Paired Android measurements
Benchmarks use identical per-model configs: pp256/tg64, three repetitions, context2048, six threads, mmap off; Hexagon/CPU flash ON and OpenCL flash OFF. These are individual matrix trials, not statistical performance guarantees. All before/after thermal snapshots report status 0, but temperature and OS memory/cache state differ; Myron candidate runs occurred the following day.
All eight Myron candidate cells passed, including CPU/OpenCL controls whose native allocations remain unchanged. Native buffer totals are allocation-log measurements, not process PSS or total system memory. Qwen reported peak memory was 1,681.10 → 1,548.58 MiB.
S23 already uses one HTP session in the baseline: Qwen native buffers remain 1,764.11 MiB; prefill/decode measured 95.30/12.72 → 90.09/14.38 tok/s. Known Phi Q4_0 native abort, Phi Q8_0 allocation failure and Gemma Q6_K/Q8_0 memory kills remain. Phi Q6_K initially passed only on baseline; a reversed-order repeat passed only on candidate, leaving each build with one pass and one confirmed memory kill. This does not establish a repeatable candidate-specific regression or workload reliability.
With flash OFF, Myron Phi Q4_K_M and Gemma Q4_K_M both completed correct Paris answers on candidate; baseline attempts ended in confirmed LOW_MEMORY without a complete answer. S23 Phi Q4_0 still aborts after loading and Gemma Q6_K still dies during load on both builds. Qwen completed correct answers before Settings and after selection/restart on both phones.
Artifact provenance and limits
Paired measurements use preserved PR #901 CI build
6408291(runtime-equivalent to base986f402a) and a controlled comparison APK with only its Hermes JavaScript bundle replaced by the freshly compiled candidate bundle. Every other non-signing APK entry, including native libraries/assets, is byte-identical. Installed APK hashes and model hashes were verified. The independent complete Gradle build uses package native prebuilts and passes; it is a separate artifact from the controlled measurement APK. Candidate runtime source175a7016is unchanged by test-only commit47a26875. After rebasing onto merged #901, current headd50f9646has the identical complete Git tree as the reviewed and tested47a26875; all four patches are unchanged.Myron E2E data had to be recreated after an installation problem, so its legacy selection was reseeded; S23 retained its E2E datastore across replacement installs. UI text/captures establish actual generation; full native arguments and allocation evidence come from benchmark traces and complementary native-boundary tests. These results do not claim that all Hexagon model failures are fixed.
Visual captures are complete; uploading them is pending GitHub browser-session authentication.
Generated by PocketPal Dev Team