Skip to content

Releases

v0.1.3

Choose a tag to compare

@github-actions github-actions released this 13 Nov 08:51

v0.1.3

v0.1.3: correctness and agentic release. Fixed: the v0.1.2 deep-mode known issue. The engine now sends an SSE keepalive while it reads a long append, and the controller and the OpenCode launch overlay wait up to 3600 s. Through OpenCode 2.2.7 in deep mode, a 28,420-token append at 111,025 tokens of context took 1,080.8 s to read, with 72 keepalives (longest gap 15.0 s) and no client retry. The controller and Studio are split into smaller modules (Studio context and workers, per-engine launch adapters for _begin, the proxy as prep, ledger and stream stages) with byte-for-byte contract tests on the old code. Proxy fixes: a client's Cookie and Authorization headers are never forwarded to a backend; an invalid Content-Length gets a 400 and closes the connection; client X-Oxe-* headers are stripped; a ledger error after admission no longer leaks the request slot; a query string on the legacy Flash route is refused (403). The engine refuses a non-loopback bind without an API key. New, off by default: LAN share (local-ai share; its own API key, inference routes only) and a tool-output cap (LOCAL_AI_TOOL_OUTPUT_CAP=on). New lab tool: the O-117 quality runner (teacher-forced KLD, top-1 agreement and an f32 reference for attention arms). Known issue: OpenCode 2.2.7 can fold a prompt submitted above the preset's compaction trigger into the compacted history as a "historical" checkpoint, and a local model may then continue the old objective instead of the new request. It was seen in deep mode at about 111K and also happens on v0.1.2. Fast mode's lower trigger does not prevent it. Workaround: after the compaction has finished, send the request again as a new message. Fixing it is the top v0.1.4 item.

This release makes no speed claim and changes no default. Optimisations measured in the lab and rejected by their pre-registered thresholds: fa2 and fa2x prefill attention reached 1.15-1.16x at 32K (the bar was 1.2x) and stay opt-in; fused decode reached 1.03x on Ornith and 0.94x on Qwen (the bar was 5%) and stays off. The tool-output cap gate was stopped as confounded (no output was actually capped), so the cap stays off. Tiled attention remains the default. CI binaries are rebuilds rather than the exact hardware-qualified executable.

Receipts: V013-INSTALL-G1, V013-BATCH, V013-AP1, V013-TOOL-OUTPUT-CAP, V013-SPLIT-SMOKE

Squashed snapshot of private source 8700bf12fd81c317f8ca5370ad27ff9894c6d516 (previous public release v0.1.2).

Binaries

Built by GitHub Actions from v0.1.3 (commit 96c5bf9) in run 43395197353; nobody uploads binaries by hand.
Linux aarch64, for ARMv8.6-A cores with dotprod, i8mm, fp16 and bf16 (measured on Snapdragon X Elite X1E80100 only); glibc requirement and toolchain in each tarball's README.txt.
CI builds pass the model-free tests and the Flash selftest; they have not been through the lab's model-based identity gates, and their outputs can differ from the measured laptop builds in the last floating-point bits (README, "Download").

File Contents
oxe-server-v0.1.3-linux-aarch64.tar.gz oxe-server: the oxe MoE engine (Ornith 1.5, Qwen3.6 35B-A3B)
oxe-server-flash-v0.1.3-linux-aarch64.tar.gz oxe-server with the Flash engine (--engine flash) + flash-m0 (expert sidecar) and flash-m3 (selftest); ggml built in the same run from the pinned public sources
oxi-v0.1.3-linux-aarch64.tar.gz oxi: FLUX.2 klein image engine (experimental)

Verify:

sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify oxe-server-v0.1.3-linux-aarch64.tar.gz --repo n0rpan/snapdragon-engine \
  --signer-workflow n0rpan/snapdragon-engine/.github/workflows/release.yml

No model weights are included.

v0.1.2

Choose a tag to compare

@github-actions github-actions released this 12 Nov 21:12

v0.1.2

v0.1.2: chat and agentic release. Ornith and Qwen3.6 run in a 128K window. OpenCode defaults to fast mode (compaction near 49.5K tokens); deep mode is selectable with LOCAL_AI_OPENCODE_MODE=deep. New durable state cache (64 GB default budget, 7-day TTL, local-ai cache): a session resumes after an engine restart or a model switch. History rendering keeps the prefix cache valid across turns (STABLE-HISTORY-B). The dense 27B model is retired. Bootstrap's default image set is FLUX.2 klein. Known issue: in deep mode a very large tool output above about 80K tokens of context can exceed OpenCode's 15-minute stream timeout; use fast mode.

128K is the configured cap. Qualification covers an accepted approximately 107K Deep envelope on Ornith and approximately 83K restore on Qwen, not a full 128K fill. The installed runtime's source and accepted timing-only retry scope are recorded in V012-CHAT. Tiled attention remains the default; the optional fused arm failed its registered gate and stays off. No NPU or new image/text-encoder claim is made in this release. CI binaries are rebuilds rather than the exact hardware-qualified executable.

Receipts: V012-CHAT, STABLE-HISTORY-B, G1-prime, FLASH-PREFIX-2

Squashed snapshot of private source 93418149371a6716f87427849721412f3fdf33de (previous public release v0.1.1).

Binaries

Built by GitHub Actions from v0.1.2 (commit 2174e6d) in run 43306202394; nobody uploads binaries by hand.
Linux aarch64, for ARMv8.6-A cores with dotprod, i8mm, fp16 and bf16 (measured on Snapdragon X Elite X1E80100 only); glibc requirement and toolchain in each tarball's README.txt.
CI builds pass the model-free tests and the Flash selftest; they have not been through the lab's model-based identity gates, and their outputs can differ from the measured laptop builds in the last floating-point bits (README, "Download").

File Contents
oxe-server-v0.1.2-linux-aarch64.tar.gz oxe-server: the oxe MoE engine (Ornith 1.5, Qwen3.6 35B-A3B)
oxe-server-flash-v0.1.2-linux-aarch64.tar.gz oxe-server with the Flash engine (--engine flash) + flash-m0 (expert sidecar) and flash-m3 (selftest); ggml built in the same run from the pinned public sources
oxi-v0.1.2-linux-aarch64.tar.gz oxi: FLUX.2 klein image engine (experimental)

Verify:

sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify oxe-server-v0.1.2-linux-aarch64.tar.gz --repo n0rpan/snapdragon-engine \
  --signer-workflow n0rpan/snapdragon-engine/.github/workflows/release.yml

No model weights are included.

v0.1.1

Choose a tag to compare

@github-actions github-actions released this 11 Nov 18:37

v0.1.1

Security and robustness hotfix, a faster image VAE, an opt-in Flash disk prefix store, and ready-made binaries built by CI.

Fixed (from an independent code review of v0.1.0; every fix has a model-free test that fails on
v0.1.0 and passes here):

  • oxe-server access control: no wildcard CORS headers; 403 for a non-loopback Host, a
    non-loopback Origin (including null) and Sec-Fetch-Site: cross-site. Optional key
    (--api-key-file / OXE_API_KEY_FILE, llama-server format, or --api-key): with a key every
    route needs Authorization: Bearer or X-Api-Key, compared in constant time.
  • Sampler and request checks: a request with min_p > 1 could abort the server. sample() is
    now total; min_p / top_p / typical_p outside 0..1, non-finite temperature and out-of-range
    penalties get 400; n_probs capped; token ids above u32 no longer wrap.
  • GGUF reader: checked element/byte sizes, offsets and counts; a crafted model file can no
    longer cause out-of-bounds reads or writes (fuzz-style test: 20k mutated headers).
  • A streaming client that stops reading no longer holds the worker (stall timeout, default 30 s,
    cancels like a disconnect).
  • Flash cache preference 20/18 takes the swap-tolerant rule only when free swap covers what
    must leave RAM.
  • Exact routes (no trailing-slash aliases); Pool::run refuses concurrent use and aborts on a
    job panic in every profile.
  • kernels-sys library test linked only by luck of archive order.
    Installed and qualified on the test laptop with the lab's controller: all three models
    13/13 independent protected-task checks, security checks, five-case regressions PASS, no waiver.
    Receipts: v011-hotfix

Faster images: oxi's VAE no longer runs on one core. Pool::new pinned the calling thread to
cpu 0, and oxi kept that pool alive through the VAE, so the VAE's threads inherited a one-CPU mask.
The creator's CPU mask is now restored before the VAE runs.
Studio 512 px after the fix: 30.9-31.1 s per cold request in the VAE-1 qualification screen (was 51.53 s, not a matched comparison); installed live smoke 24.89 s.
Receipts: VAE-1

Opt-in Flash disk prefix store: reuse a long chat prefix across unload/reload, with
identical greedy output, private cache files and bounded on-disk entries. State
lengths are checked before allocation. Before allocating a save buffer, Flash
checks MemAvailable against a 1024 MiB save reserve plus twice the entry size;
when memory is short or unknown, or an earlier save is still running, it skips the
save without changing output. The store remains off unless explicitly enabled.
Installed qualification includes explicit O43/O43b timing exceptions, with all
five output identities, security and memory checks passing and cancel recovery
below five seconds. Raw failed timing reports remain part of the receipt.
Receipts: FLASH-PREFIX-2

Known limitation, FP-SHORTLAT: two-token chat/context decode added about 120-180 ms
versus the fresh control. The v0.1.2 target is <=50 ms of added short-request latency.
768/1024 px image qualification and NPU inference are outside this release.

New: ready-made binaries. Every v* tag builds, tests and publishes Linux aarch64 tarballs from
the tagged source on GitHub's ARM64 runner (Fedora 44 container, GCC 16, pinned Rust 1.99.1,
actions pinned by commit): oxe-server (oxe MoE), oxe-server-flash (Flash + flash-m0/flash-m3,
production ggml built in the same run from BigMoeOnEdge 930b2c9 + llama.cpp 7ac98f4 + the
demand-mapping patch) and oxi. Each tarball carries README.txt (commit, run, toolchain,
glibc >= 2.39), the licences of every vendored component and linked Rust crate, SHA256SUMS and
Sigstore build provenance (gh attestation verify FILE --repo n0rpan/snapdragon-engine
--signer-workflow n0rpan/snapdragon-engine/.github/workflows/release.yml). CI builds pass the
model-free tests and the Flash selftest; they are not gate-qualified and can differ from the
measured laptop builds in the last floating-point bits (README "Download").
Flash archives include the exact Apache-2.0 BigMoeOnEdge and MIT llama.cpp fork licence texts, pinned source attribution and licence hashes. Model-free archive tests also run when checks do not publish packages.
CI also runs on every push to main and every pull request.

Squashed snapshot of private source 9c76ef5ba9aa1fdd03331732e5149af60d564159 (previous public release v0.1.0).

Binaries

Built by GitHub Actions from v0.1.1 (commit 510d250) in run 43078406408; nobody uploads binaries by hand.
Linux aarch64, for ARMv8.6-A cores with dotprod, i8mm, fp16 and bf16 (Snapdragon X Elite / X Plus); glibc requirement and toolchain in each tarball's README.txt.
CI builds pass the model-free tests and the Flash selftest; they have not been through the lab's model-based identity gates, and their outputs can differ from the measured laptop builds in the last floating-point bits (README, "Download").

File Contents
oxe-server-v0.1.1-linux-aarch64.tar.gz oxe-server: the oxe MoE engine (Ornith 1.5, Qwen3.6 35B-A3B)
oxe-server-flash-v0.1.1-linux-aarch64.tar.gz oxe-server with the Flash engine (--engine flash) + flash-m0 (expert sidecar) and flash-m3 (selftest); ggml built in the same run from the pinned public sources
oxi-v0.1.1-linux-aarch64.tar.gz oxi: FLUX.2 klein image engine (experimental)

Verify:

sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify oxe-server-v0.1.1-linux-aarch64.tar.gz --repo n0rpan/snapdragon-engine \
  --signer-workflow n0rpan/snapdragon-engine/.github/workflows/release.yml

No model weights are included.