feat: integrate model guard v2.3 and multi-platform v1.1 release - #101

Merged

athenawxd merged 123 commits into

mainfrom

feat/model-guard-v2.3

Aug 18, 2026

Commits on Jul 31, 2026

refactor(nn): make bmrt ownership explicit
Fernandez81
committed
eedc9a3
build(model-guard): admit the v2.3.3 runtime SDK
Fernandez81
committed
b6b5aa1
feat(model-guard): load preset CEMC through CMG_2.0
Fernandez81
committed
4ef5570
feat(models): publish protected BM1688 presets
Fernandez81
committed
6a136fa
feat(release): add signed atomic release transactions
Fernandez81
committed
a95203a
feat(upgrade): accept signed release archives
Fernandez81
committed
b66d7ea
build(sophon): produce verified protected packages
Fernandez81
committed
e5a3a49
feat(source): install SOURCE runtime without backups
Fernandez81
committed
aefb227
docs(security): document protected model release flows
Fernandez81
committed
964fa43
docs(source): align Chinese installer behavior
Fernandez81
committed
d0f9efe
refactor(model-guard): remove redundant admission gates
Fernandez81
committed
4e618bc
docs(model-guard): update deployment operations
Fernandez81
committed
472dae4

Commits on Aug 4, 2026

Commits on Aug 5, 2026

build(guard): select staged production sdk
athenawxd
committed
9415c63
fix(guard): stream device authorization request
athenawxd
committed
1ef9557
build: add RK3576 RKNN toolchain
samuel--hu
committed
60c6b24
feat: integrate RK3576 RKNN backend
samuel--hu
committed
0064b59
perf: harden RK3576 runtime admission
samuel--hu
committed
ae2e800
feat: add Rockchip native media pipeline
samuel--hu
committed
670d0ec
test: harden RK3576 long-run qualification
samuel--hu
committed
5a5169e
fix: harden websocket HTTP probe handling
samuel--hu
committed
5856058
test: add RK3576 customer journey fixture
samuel--hu
committed
9478cfc
chore: curate RK3576 integration assets
samuel--hu
committed
bba6c3a
fix: harden runtime logs and startup contracts
samuel--hu
committed
8a7d77a
fix: mark unreliable RK3576 NPU telemetry unavailable
samuel--hu
committed
81f3f24
fix: clean up scenario bench on shutdown signals
samuel--hu
committed
ec4ac42
fix: report RK3576 NPU busy-time load
samuel--hu
committed
d3b52fb
fix: model RK3576 accelerator memory as shared
samuel--hu
committed
515a149
Merge remote-tracking branch 'origin/main' into codex/rk3576-integration
samuel--hu
committed
dec90b4
perf: add RKNN preprocessing stage metrics
samuel--hu
committed
933ca6a
perf: add RK3576 RGA detector preprocessing
samuel--hu
committed
44575a5
perf: feed RKNN detector with native int8 NHWC input
samuel--hu
committed
29a155c
fix: preserve RKNN classifier preprocessing compatibility
samuel--hu
committed
63107ee
test: qualify RK3576 detector fast path
samuel--hu
committed
dbc5b21
perf: add quantized RKNN output reconstruction
samuel--hu
committed
14ed613
perf: use native RKNN detector outputs
samuel--hu
committed
05583dc
perf: optimize RKNN YOLOv8 output consumption
samuel--hu
committed
6025cd2
fix: make algorithm orchestration paths portable
samuel--hu
committed
686d68b
test: disable default OSD pixel heuristic
samuel--hu
committed
fd6e575

Commits on Aug 6, 2026

perf: decode RKNN YOLOv8 candidates directly
samuel--hu
committed
3615b50
perf: register RKNN output tensor contracts
samuel--hu
committed
1825b00
perf: bind RKNN detector input memory
samuel--hu
committed
11ebdff
perf: bind RKNN detector input to persistent RGA target
samuel--hu
committed
7be57b2
fix(upgrade): restore legacy installer lifecycle
athenawxd
committed
5f61160
fix(package): allow plaintext VLLM presets
athenawxd
committed
68139c3
perf: feed RKNN detector from MPP DMA-BUF
samuel--hu
committed
f512d7e
fix(package): require protected VLLM presets
athenawxd
committed
d696ad3
perf: qualify RKNN detector core scheduling
samuel--hu
committed
cbd74a1
fix: measure shared detector throughput per channel
samuel--hu
committed
25e2657
fix: exclude event reports from CV throughput
samuel--hu
committed
deb63eb
perf: expose RKNN fused input telemetry
samuel--hu
committed
956417f
perf: gate RKNN uint8 binding by model contract
samuel--hu
committed
e00ff7c
fix: preserve RKNN FP16 input contract
samuel--hu
committed
3b9d8a0
test: separate pool warmup from long-run growth
samuel--hu
committed
69e9c72
fix: reuse RK3576 decoder across stream loops
samuel--hu
committed
8f8b4b8

Commits on Aug 7, 2026

merge: integrate latest main into model guard v2.3
athenawxd
committed
9124d1b
Merge origin/main into rk3576-integration
samuel--hu
committed
fd434fc
merge: integrate rk3576 support into model guard v2.3
athenawxd
committed
34e315b
fix(build): link cpu allocator in cross-backend tests
athenawxd
committed
67e15f6
docs(readme): draft next-release landing pages
samuel--hu
committed
8d5bd15
docs(readme): address draft review feedback
samuel--hu
committed
bd4e103
docs(readme): refresh RK3576 evidence and links
samuel--hu
committed
2564d35
docs(changelog): document changes since v1.0.0
samuel--hu
committed
a8b283b

Commits on Aug 9, 2026

fix(stream): reduce preview and OSD startup latency
samuel--hu
committed
6f77799
feat(macos): add Docker Desktop preview workflow
samuel--hu
committed
dffd87b

Commits on Aug 10, 2026

Commits on Aug 11, 2026

chore(build): ignore rknn build output
Fernandez81
committed
93a0c9d
build(rk3576): add reproducible docker package entry
Fernandez81
committed
f318c00
ci(rk3576): add nightly cross-build workflow
Fernandez81
committed
aa9d2c1
docs(rk3576): document stable release workflow
Fernandez81
committed
c690e78
fix(system): preserve model authorization on factory reset
Fernandez81
committed
b7b65ea

Commits on Aug 12, 2026

feat(web): restore network speaker menus
athenawxd
committed
ea4414d
fix(web): clear channel id when adding camera
athenawxd
committed
b76ba84
fix(audio): initialize device operations with valid values
athenawxd
committed
a2c0733
feat: enable Qwen3.5 RKLLM vision inference on RK3576

samuel--hu
authored and
nquyencubas523-cloud
committed
36afe4b
fix(rknn): restore model uploads and release version

samuel--hu
authored and
nquyencubas523-cloud
committed
22d4a75
perf(rkllm): reduce visual prefill latency

samuel--hu
authored and
nquyencubas523-cloud
committed
43772c7
fix(rk3576): expose only validated model types

samuel--hu
authored and
nquyencubas523-cloud
committed
46d8f87
fix(rkllm): keep baseline RK3576 builds compatible

samuel--hu
authored and
nquyencubas523-cloud
committed
791f035
fix(linkage): allow empty workflow drafts
athenawxd
committed
e547c8b
fix(web): rebuild empty linkage canvas
athenawxd
committed
6cd8419
fix(web): restore linkage editor interactions
athenawxd
committed
1f10148
Merge pull request #96 from cosmo-wander-ai/perf/rk3576-qwen35-validated
Fernandez81
authored
e2d6170

Commits on Aug 13, 2026

Commits on Aug 14, 2026

fix: harden multi-platform release validation
nquyencubas523-cloud
committed
a0f8b04
fix: embed target chip in release archives
nquyencubas523-cloud
committed
22f41f9
fix: cache and further bound npm fetches
nquyencubas523-cloud
committed
87e6b58
fix: serialize npm fetches on validated mirror
nquyencubas523-cloud
committed
3f7c7de
fix: install web dependencies from a serialized cache
nquyencubas523-cloud
committed
72aab13
docs: qualify CV186X and RK3576 model evidence
nquyencubas523-cloud
committed
7ac23b1
build: pin published RK3576 RKLLM image
nquyencubas523-cloud
committed
b3c7ce9

Commits on Aug 17, 2026

Commits on Aug 18, 2026