Ember 2026.8.24

Added

• perf: show decode and prefill history together, not behind a toggle (9c654963)
• perf: show the number when one release is selected (33f01641)
• perf: group workload decode into columns, one per release (e093985c)
• perf: plot release history as a line, not a row of bars (8ae85095)
• bench: automate the per-release performance bundle (ac74130d)

Fixed

• ci: take the GPU lock through a wrapper, and repair the restore guard (11954c59)
• ci: block production from restarting during certification (02a520fc)
• ci: fail loudly when production restarts during certification (bb0f0824)
• ci: hold the documented GPU lock through certification (f6815452)
• ci: wait for memory before each certification model load (a20b8104)
• perf: repair the page, and test that it actually renders (6460b35b)
• ci: check out the harness in the job that runs it (885200f0)
• perf: stop peak callouts landing on each other and on the lines (b6eb39d2)
• perf: crop the release-history axis to the range the data occupies (994c6b8e)
• perf: make the performance page usable on a phone (789d4eb5)
• ci: stop TARGET_SHA meaning two things in the benchmark step (27ec72d8)
• bench: benchmark an old release as it shipped, and keep failure reasons (212bf2a7)
• bench: plot the depth that was asked for, not the one that drifted (dba6c388)
• bench: repair four defects the first dry run exposed (6c97629b)
• deepseek4: abandon speculation after one unqualified profitability pause (d616074c)
• deepseek4: bound the warmup exclusion so the profit scheduler can still see (d8993edd)
• deepseek4: correct the warmup bail-out's false monotonicity claim (b3720a50)

Performance

• bench: re-measure 2026.8.10, .22 and .23 against the current model (f8805583)
• deepseek4: qualify the wide verify path on a rate, not a lucky streak (b66595d6)
• ggml: size the rope block to the row instead of a fixed 256 (57f11cdf)
• ggml: collapse the short-row reduction to its exact minimal form (f3da20c2)
• ggml: stop launching one block per row in reduce_rows_f32 (00556dd1)
• deepseek4: hand back to AR when batch-verify warmup is unreachable (b2428029)

Build and CI

• benchmark the released image in its own job after promotion (87c2e099)
• do not certify a push that only changed documentation (140bbc24)
• gfx1151: benchmark the certified image as part of certification (adf64ec2)

Documentation

• operations: clarify storage and release validation (fa83e82f)
• refresh contributor and performance guidance (a9d51367)
• refresh release and performance guidance (07ef572c)
• refresh runtime and performance references (b4d6348c)
• link current performance bundle (40b6e065)
• refresh release performance summary (00cb4148)
• readme: name the performance page instead of printing its URL (865a21cf)
• readme: standard depths in the performance table, and a link to the page (e060f4e3)
• readme: drop the context-scaling chart from the performance section (170c772b)
• perf: a browsable performance page, built from benchmark bundles (f9dfa520)
• refresh performance numbers from a sweep on the current build (d24b6d5d)
• assets: give the context chart a legend and peak callouts (601ce8cd)
• readme: headline plus a measured performance section (f02650ed)
• readme: put measured performance above the fold (357c84c8)
• perf: correct the decode and roofline claims, record the prefill work (2d90f648)

Maintenance

• release: v2026.8.24 (721e6389)

Other

• hand-written release metadata the automation owns (0be0764f)
• ci: take benchmarking back out of the certification job (89bc6dc2)
• perf: dress the performance page in the otheru palette (487df24b)

Ember 2026.8.23

Fixed

• test: assert the compose environment syntax the file actually uses (5f6241cc)
• deepseek4: stop charging warmup cost to steady-state speculation (af16fe4b)
• docker: drop compose declarations that only restated engine defaults (7810971a)
• docker: stop the entrypoint overriding compose, make compose authoritative (8c999fee)
• engine: correct two profiling-harness bugs found on hardware (3da625c0)

Performance

• deepseek4: 33.60 -> 37.49 tok/s decode via width-6 verify (027ed106)
• engine: extend DPP wave reductions, verified against the RDNA 3.5 ISA (3a54084c)
• engine: DPP wave reduction in hand-written assembly for gfx1151 (5f460a0c)
• engine: select WGP vs CU mode per translation unit on gfx1151 (410d61d1)
• engine: branch-free UE4M3 scale decode for gfx1151 (b6291ce6)

Build and CI

• docker: add Hyperloom to the dev image (88898ac8)

Documentation

• perf: document TTFT, and record why narrowing the FA KV read failed (1cb30f76)
• engine: link the two competing LUCE_MMVQ_MAX_NCOLS defaults (4a1ccfef)
• engine: correct the HIP-graph diagnosis, which blamed the wrong tensor (dcc00326)

Maintenance

• release: add gfx1151 kernel profiling harness (e5a0d064)
• release: document CI gates and clear prune residue (f02f7d10)
• release: document Forgejo sanitizer split (b760adb8)
• release: stabilize Forgejo queue (02305c07)
• release: recover release-note publication (08f4f9f8)
• release: repair certified tag publication (45ac7d3c)

Ember 2026.8.22

Curated notes

• Resolved #2: Please add an option to specify the host to listen on: added --host and EMBER_HOST so trusted container/Kubernetes gateways can
  reach Ember while preserving the unauthenticated loopback default.
• Resolved #3: Add option to disable shasum check: added EMBER_VERIFY_EXISTING_SHA256=0 for trusted immutable model stores;
  downloaded artifacts remain pinned and checksum-verified before promotion.
• Integrated and documented the opt-in CPU/GPU/XDNA2 resident DSpark
  prototype. The XDNA Compose overlay now selects the measured two-session
  proposal pipeline instead of the slower target-expert placement experiment.
• Kept heterogeneous inference experimental: the fixed throughput fixture
  passed, while capture-graph output equivalence remains a promotion blocker.
• Linked README installation guidance to the current GitHub release instead of
  embedding a version that becomes stale after every release.
• Made candidate certification and release promotion fully unattended: trusted
  main commits proceed from immutable GHCR images through the dedicated
  gfx1151 runner to changelog generation, tagging, and package publication.
• Pinned the certification model pair to the published Hugging Face artifacts,
  added IOMMU/device preflight checks, and made the hardware job quiesce and
  restore production around exclusive GPU validation.
• Upgraded GitHub checkout steps to the pinned Node 24 release so routine CI
  and release logs no longer carry Node 20 deprecation warnings.
• Added persistent BuildKit, ccache, and Trivy database reuse across GitHub and
  Forgejo runners, including compiler cache mounts inside the ROCm Docker build.
• Kept periodic full model digest verification while caching unchanged results
  for seven days and using direct I/O on cache misses, preventing the 96 GiB
  checksum pass from consuming UMA and disabling monolithic DSpark validation.

Added

• release: automate gfx1151 certification (62205ed3)
• release: automate certified promotion (324956c7)
• ci: automate candidate release chain (a44a8205)
• ci: mirror GitHub issues to Forgejo (ad46d678)
• release: finish deployment controls and XDNA docs (4c37043a)
• xdna: validate asynchronous DSpark provider (1b439f8f)
• xdna: schedule weighted expert runlists (666c612d)
• xdna: add async DSpark draft provider seam (ac541642)
• xdna: package ROCMFP2 XRT prototype (e6f98a8d)
• engine: prototype XDNA2 expert offload (d41d2642)

Fixed

• release: compare sorted metadata paths (fff01830)
• ci: mount Forgejo CA in job containers (4f64a037)
• release: quiesce production supervisor (27a9effb)
• ci: make IOMMU preflight explicit (3803c075)
• release: avoid stale README version (9b918f43)
• ci: isolate release note permissions (39de69f2)
• ci: use native Forgejo push mirror (d0ddd613)
• engine: make XDNA corpus target-exact (7472ac92)
• engine: make resident XDNA acceptance target-exact (6d0cb138)
• engine: classify custom clamp types (83a7b839)
• engine: isolate resident graphs and XDNA context (115d3a79)
• xdna: validate mmap-backed model path (907d0064)
• ci: let the vulnerability scanner reach its database (0e57f256)
• docker: declare the image source so GHCR links the package (0c33af6b)
• ci: give the BuildKit container a working network (3df0939a)
• release: publish the container image where the docs promise it (3ab6fada)

Performance

• release: cache model verification and expose hits (a629fb25)
• release: cache builds and preserve UMA capacity (f2627063)
• engine: batch resident verification safely (06d15d7b)
• xdna: attribute resident pipeline phases (5a3171c1)
• xdna: gate resident promotion by confidence (d7d604b0)
• xdna: capture current design control (56c25ad2)
• xdna: pipeline resident draft proposals (6f8ba53c)
• xdna: vectorize draft routed experts on CPU (0c98d313)
• xdna: keep draft Q8 stages resident (cdec0696)
• engine: fuse DSpark context KV on GPU (4cbba655)
• xdna: unify draft projection overlay (358e1794)
• xdna: keep draft projections on fixed overlay (732be598)
• xdna: parallelize DSpark CPU reductions (3b830cab)
• engine: measure GPU DSpark preprojection (54270615)
• engine: stage DSpark main projection on GPU (f9f9d141)
• xdna: reuse weights across draft context (4fc7f551)
• xdna: measure task-blocked Q8 projections (a449b3ec)
• xdna: add compensated Q8 draft shared expert (efad17c8)
• xdna: mask routed rows in gen8 (5c97cd24)
• xdna: add five-row ROCMFP4 expert kernel (6d3b7acb)
• engine: measure fused verifier MoE budget (e888ef1a)
• xdna: batch shared experts in gen6 (de55413a)
• xdna: measure concurrent GPU load (f6e91bd7)
• xdna: validate zero-copy GPU buffers (76b9d3cf)
• xdna: batch gen5 draft blocks (8729f21b)
• xdna: fuse expert pipeline in gen5 (252b813f)
• xdna: expose provider phase timings (aaa0cf49)
• xdna: vectorize ROCMFP2 decode in gen4 (e811dfde)
• xdna: preserve projection outputs in fp32 (2de48a4f)
• xdna: prototype second-generation expert kernel (4d3c55a6)

Changed

• engine: prune unsupported backends and prototypes (eade92dd)

Build and CI

• xdna: refresh runtime for queued drafts (452c2e7d)

Documentation

• xdna: pin gen43 benchmark candidate (8d0c90fd)
• xdna: record whole-draft contention results (8a05387a)
• xdna: record AIE-RT queue rejection (571d5d8c)
• xdna: record gen5 validation scope (b50e00bb)
• ci: record the self-hosted builder and its fork-PR constraint (8b9c5f99)

Testing

• json: cover UTF-8 output boundary (48e0ca53)
• xdna: enforce resident speculation gate (37f7f6bf)
• xdna: reject bfp16 draft projections (9a8b0647)
• xdna: validate trained draft experts (41454f7f)

Maintenance

• ci: move checkout actions to Node 24 (7987558a)

Ember 2026.8.10

• Prepared the first source release.
• Added a one-command Docker Compose workflow with resumable model download.
• Pinned the default model revision and added expected-size, free-space, and
  SHA-256 checks before startup.
• Added host/container preflight diagnostics, Compose health reporting, and
  read-only plus generation smoke tests.
• Added OCI image metadata and operations/upstream-release guidance.
• Split the container into a full-toolchain dev target and a minimal
  dependency-closure release target used by Compose.
• Added licensing, vendor provenance, security, support, and contribution policy.
• Synced the latest agent-progress work: repeated-call diagnostics, a
  result-based progress lease, empty/degenerate-turn telemetry, and detection
  of tool markup delivered as visible text.
• Added opt-in stateless automatic loop recovery and Compose environment
  controls for progress reporting and recovery thresholds.
• Hardened request parsing against embedded NULs, duplicate or mistyped fields,
  unsupported HTTP transfer framing, and excess headers.
• Made strict tool-schema number equality exact beyond IEEE-754's integer range
  and replaced direct POSIX interpretation of ECMA-262 regex patterns with a
  translated, fail-closed subset.
• Removed one-off development probes and made static analysis, exact-commit
  tests, and gfx1151 certification blocking image-release gates.
• Added a history-free, policy-gated three-way promotion workflow for bringing
  reviewed implementation changes over from ember-dev without overwriting
  release-owned files.
• Promoted the gfx1151 HIP GEMM batch sweep and recorded the successful
  two-session hardware acceptance result.
• Changed the default Compose path to pull an immutable GHCR image while
  retaining an explicit local source-build override.
• Added project and third-party license notices to the runtime image and made
  the vulnerability-reporting fallback actionable.
• Added GitHub-hosted CPU and large-disk container workflows, direct GHCR
  publishing, and a manual environment-gated gfx1151 certification workflow.
• Fixed the ROCm 7.14 runtime image assembly to package both the gfx1151
  rocBLAS KPACK and its architecture-specific Tensile metadata/code-object
  tree from the pinned toolchain.
• Normalized loader-reported /lib paths into merged /usr locations so the
  collected runtime can be copied into the Ubuntu 24.04 release stage.
• Hardened gfx1151 certification to verify the immutable model pair,
  separately require exercised DSpark and disk-cache paths, use locked memory,
  and verify live speculative decoding outside resident batching.
• Repointed first-run acquisition to the current published quant and DSpark
  drafter, made both artifacts mandatory and digest-locked, and removed support
  for substituting unverified model files.

Container image

ghcr.io/otheru-ai/ember:2026.8.10

Digest sha256:ac3be24624021320ccf9eaa1c42a15435031d746b1b567f06404079469ba1838,
built from f16b4ae28fb071851007c509b9b7beef79c5e6a3.

A CycloneDX SBOM for the image is attached to this release. This image was
published manually because the tagged CI publish could not schedule a runner;
see docs/ci.md for the runner requirements now recorded in the workflow.