Repository navigation
Releases
sdk v0.6.0
Changed
SUPPORTED_SPEC_VERSIONSadmits runtimes1014015through1014020, and
SUPPORTED_SPEC_VERSION_RANGEends at1014020. They add the per-miner capacity reservations,
the model weights lock, the VRF committee seed and the D-0544 unsigned-open tariff. The wire
bytes the SDK signs are unchanged.
Fixed
-
Faucet POSTs are no longer automatically retried after ambiguous outcomes. Grant
confirmation no longer equates a sufficient balance with an attributable grant;
QuickStart stops on an unknown result instead of silently continuing. -
Preserve an unknown cooperative settlement outcome when finalized-event polling
reaches its deadline, preventing a competing force settlement or timeout refund.
Added
-
computeChannel.readMinerCapacityreads a miner's D-0542 capacity reservations at the
finalized head and returns its live accepted cap, reserved total, the rate the next open would
reserve, and whether it fits;undefinedon a runtime without the storage.minerCapacity,
reservationRateForSessionsandreservationRateUpdateare the pure rules behind it. -
fundFromFaucetDetaileddistinguishes refused from unknown requests. The current
faucet protocol lacks an authenticated request-to-transaction binding, so even a
successful response remains unknown.waitForAvailableBalanceexplicitly checks
balance; the older boolean helpers remain as deprecated compatibility aliases. -
SubmitResult.blockHashnames the finalized block when the submitter waits for finality;
settlementArgsandmapChainEventsare exported for callers that hold a signer lane. -
verifyMirroracceptsrevision, pinning a whole-repo Hugging Face check to that
commit instead of resolvingmain.
sdk v0.5.2
Fixed
- Resolve
flopSessionKeysandFlopSessionKeysin session-bound HTLC creation,
retaining legacy aliases, spending prechecks, signer-specific nonce and wrapper errors.
Added
packedImageChunks/packedImageStream/packedImageSizeemit a checkpoint
directory's canonical packed image as a stream, holding every file to the
(path, size, sha256)table as it is read, so a DA commitment can be taken over the
image without materializing it.normalizeRootis now exported from the package root.
python-sdk v0.3.0
Changed
- Session creation requires explicit keyword-only
per_tx_limitanddaily_cap
in both public helper families. This breaking correction supplies the runtime's
required fields without inventing financial defaults.
Fixed
- Address
FlopSessionKeyswith the runtime's snake-case create/revoke/execute
calls, query the delegate signer's live session nonce, and preserve wrapper,
dispatch and circuit-breaker failures. Session query/decode errors propagate;
only an absent option is reported as no session. Decoded booleans and bounded
integer fields are validated; stale-nonce errors propagate without retry.
Live execution is unverified.
pi v0.3.2
Fixed
- Bound the Pi AI peer to 1.4.x. Pi AI 1.5 removes the
Typeexport used by the Durable
and memory adapters, so the previous open-ended range could install a host that could
not load the published package. The packed-consumer smoke keeps resolving the declared
peer range and importing every public entry point.
Added
-
Bounded devnet credential renewal retains the aggregate parent allowance and
original conversation bindings, drains before rollover, and keeps cached task
routing and original-key recovery across Worker restarts. The existing-host
controller journals publication and has no payer funding operation.
Rotation preflights native FLOP estimates and the gateway's actual admission floor. -
An explicit bounded native Pi A2A text mode for ready text models, without tool
declarations, retaining the tool-capability requirement for tool-enabled users.
Execution mode and token bounds remain pinned to the original admission. -
A private operator-owned devnet A2A authority with hash-only bearer authentication, fixed
admission grants and concrete FLOP Labs deployment profiles, covered by real Workerd binding tests. -
An opt-in authenticated Pi Durable A2A application example with persistent task admission,
caller isolation, payload-bound retries and credential-bound lifecycle recovery. The private
Worker remains the default; deployment, genuine registration and verification remain operator gates.
node v0.2.0
Added
- First versioned
nodeimage forlinux/amd64, built with the default runtime.
miner v0.2.0
Added
- First versioned
minerimage forlinux/amd64, packagingzkv-relayand its runtime WASM.
miner-worker v0.2.0
Changed
-
On-demand model loads refuse a
max_model_lenthat differs from the decode policy's serving
profile, and a backend whose/healthdoes not report enforcing that profile. -
Standing spot offers bind fee profile 1 in v4 signatures when the finalized runtime exposes
ready inference credits. Uninitialized, initializing, faulted or unreadable credit state
refuses publication instead of downgrading the signed fee contract. -
Recover session evidence publication from durable, attempt-bound state after RPC loss. New paid
session admission waits for publication recovery; existing receipts and acknowledgments remain available.
Paid TOPLOC admission now requires the durable encrypted audit evidence store. Signer reconciliation
verifies finalized outcomes or complete mortal expiry before releasing a fence.
Finalized closed channels retire impossible publication admission retries while retaining audit
evidence, custody metadata and unresolved signer fences. -
On-demand model profiles classify a prepared checkpoint as dense GDN from its
config.json
model_type(qwen3_5,qwen3_5_text), not only from Qwen3.5/3.8 names, and refuse it at
startup unless it loads on one GPU without expert parallelism. A checkpoint config that exists
but cannot be read or parsed refuses the profile; an unprepared checkpoint is still checked by
its finalized architecture before loading. -
Session readiness, admission, force-ack capacity and offers read fresh miner-specific
calibration at one finalized block, without scanning unrelated calibration jobs.
Initial calibration and renewal discovery still run in background maintenance. -
The model gate and the qualification worker accept a registered root equal to the checkpoint's
content root (content-v1, D-0535) or its dm-verity root (verity-v1), and log which scheme
matched. Qualification checks the hub's published sha256 table against the governed manifest
root before any weight byte moves, and checks each file as it streams.
Added
-
Private qualification mode keeps the worker out of public discovery, pins its Ghost mailbox,
and stops new sessions at authorization expiry while allowing existing sessions to drain. -
Startup and hourly release-policy logs: warnings for newer worker builds or unavailable
version information, and errors for below-minimum or scheduled-for-deprecation versions.
Deprecation errors include the deadline and upgrade instructions; checks do not stop work
or install upgrades. -
Independent TOPLOC replay origin (
src/replay-origin.ts): the authenticatedPOST /prefilla
qualified checker serves for CLI positive replay. It shares the miner's turn translation and
backend call, checks the backend's measured identity, and refuses saturated lanes (#3994). -
Signed OCI cache preparation through the managed controller and qualified replay entrypoint,
with cancellation, preparation metrics, durable cache holds and offline readonly backend mounts. -
On a runtime with D-0542 capacity reservations, the worker sets its reservation rate to
floor(cap / MINER_MAX_CONCURRENT_SESSIONS)(default 4 sessions, at least 1) on startup and
when its accepted cap moves, if the stored rate is unset or too large.0disables it. -
First versioned
miner-workerimage forlinux/amd64. -
Operator-authenticated session diagnostics retain the latest 50 failed turns in memory,
with correlation IDs, stage and elapsed time. Abort logs distinguish the worker deadline,
a client disconnect and model/session cancellation without retaining request contents. -
Settlement repair for a transcript-less agent
force_settlethat defeated a sealed cooperative
close. The worker reposts the co-signed tariff checkpoint with the sealed transcript at most once
per channel. It checks fresh attestation and calibration, dispute and escalation state, and DA
retention through the extended dispute horizon. A write-once encrypted intent is stored before the
signer fence, and outcomes are reconciled from the finalized extrinsic./session/closeanswers
409 withrepair: "posted"(neversettled), and payout is recorded from the terminal finalize event. -
Best-effort timing diagnostics for direct-session retirement and DA publication,
with public identifiers and no evidence or signature contents. -
Read-only signer-fence collection for legacy workers and finalized chain
inspection; authenticated same-process recovery retries a proven failure's
archival without restarting or deleting pending-channel state. -
Native DA fixture evidence survives outer Node cancellation and is retained by
Compute-Channel CI without changing its failed result or existing deadlines. -
Standard versioned worker release packages include the shared classic/containerd
verifier and installer. Compose preparation changes only the worker image;
managed upgrades reuse the existing controller configuration. Release builds
produce qualification candidates, with immutable publication after review. -
Private candidate qualification worker uses bounded spare miner resources,
measures pinned checkpoint bytes and quantization metadata before loading,
and retains independent checker orchestration outside ordinary paid serving. -
Operator-only controller install/upgrade package stages one immutable hybrid
runtime with Docker-backed proof/DA wrappers, preserves host identity/state,
and refuses handoff while work, holds or signed recovery remain. Authenticated
cold monitoring reports version/capability and drain state. Its authenticated
preparation bridge consumes the query-only checkpoint helper, supports pinned
Hugging Face-compatible sources and manual checkpoints, and reports bounded
job progress without admitting inference. Source/CPU package only; publication
and fresh-GPU acceptance are pending. -
Local
prepare-modelverifies manual checkpoints or resumes a configured immutable
Hugging Face download before atomic publication, independently of paid sessions. -
Operator-enabled six-hour resident reservations keep a qualified model and its
occupied slot through idle gaps and Closed child channels. The fixed credential,
model, policy, miner and slot binding survives recovery; restarts require fresh
qualification, and existing child deadlines and proof gates stay unchanged. -
Reuse a qualified live model process across at most two independently bounded
channel leases, with at most one accepting new turns. Another process or a
controller restart requires fresh qualification. -
Authenticated retirement preparation seals the signed claim and produces proof
and DA evidence before the existing capacity-gated settlement call. -
One shared miner transaction lane retains nonce ownership through finality.
An uncertain submission leaves an owner-only, fsynced fence across restarts
and requires operator reconciliation before further signing.
Fixed
-
Cooperative close can settle the latest acknowledged checkpoint after a rejected
turn. It verifies the separate full-escrow receipt, publishes only the accepted
prefix's evidence, and preserves retirement fences across retries. -
Prepare initial and upward recalibration claims inside the signer lane using refreshed
credentials, bond eligibility and a final chain clock, so discovery and queue delays do not
overstate sustained capacity. Initial stake top-ups remain separate finalized transactions. -
Offline image import on macOS stages under the resolved temp directory, so the symlinked
/varno longer makes every staged OCI member read as a symlink path. -
signer-fence-checkexplains a failedsettlecandidate.counterparty_force_settlenames
the earlierforceSettleby another signer that moved the channel out ofOpen. For each
Settlingchannel it reports whether a transcript is on chain, the challenge end, and whether
finalizewould forfeit. The release stays blocked. The inspection deadline is now 15 minutes,
because the full channel scan took about six on public devnet RPC. -
A turn that fails after its first streamed chunk now ends the stream with one
{"type":"error","error":"..."}line naming the failure, at most 300 characters, instead
of closing silently. Gateways saw only "missing terminal receipt" (304 in a day across 27
miners) and could not tell which check failed. The line is unsigned diagnostic text and
carries no sequence, so it is not part of the signed frame stream. -
Finalized dispatch failures retain their exact transaction, block and call
identity and are archived before the miner lane resumes. Unknown outcomes
remain fenced, and turn requests during a fence preserve the live transcript,
turn counter and encrypted checkpoint. -
Artifact-only controller handoffs verify the complete approved OCI closure and
bind Docker classic's config/rootfs identity without registry access. Pinned
proof/DA and optional backend artifact paths, parser-tested local examples and
candidate release manifests retain identity/state checks. Docker 28 classic and
intended-GPU installation qualification remain release gates. -
A separate artifact-import CLI supports an existing standing Compose miner's
worker-only upgrade on classic Docker, retaining its backend, model, keys and
state. Standing handoff manifests record that scope separately from a new
loading-controller or Qwen model qualification. -
Fresh controller installs discover absent systemd units from explicit manager state and use the configured RPC for historical recovery checks; manager failures and foreign supervisors still refuse handoff.
-
Controller installation can use an approved locally imported image without registry authentication only when its canonical repository digest, actual content ID and descriptor all match; all release, native-runtime and fenced-state checks still run.
-
Controller upgrades retain authenticated historical checkpo...
miner-vllm-toploc v0.2.0
Added
-
serving_profile.py: a decode policy can bind a text serving profile (context window,
completion cap, tool-call parser, chat rendering profile) to the registered root through its
class_policy_hash. The window defaults to the checkpoint's nativeconfig.jsonlimit and can
only be lowered; the parser comes from its chat template (nonefor no template or one with no
tool handling; a template that readstoolsor names a tool call in a format no verified parser
covers fails derivation, so such a model has no profile until its format is supported). With such
a policy the backend re-derives the profile from the served checkpoint and its loaded tokenizer,
takesMAX_MODEL_LEN/MAX_NEW_TOKENSfrom it, refuses any other env value, and reports
serving_profilein/health. Policies without a profile serve as before. -
decode_witness.pyandscore_decode_witness.py: offline TOPLOC witness tooling (#4043). The
first captures, per prompt, the last prefill row and every decode row of native decode
(serial or batched; batched rows split by the engine's own request layout), and the same
positions from a prefill replay. The second scores those captures under last-token,
per-token and per-32-token commits, with signed-exponent, upstream and exact-bit metrics, and
hybrid serving's flattened top-k. Neither is copied into the image, and neither changes serving,
lanes or any threshold. -
Inactive private MI300X ROCm qualification scaffold, pinned provenance, CPU runtime
guards and isolated single-GPU pilot. The exact official vendor artifact has a
separate source/byte identity guard and immutable launch contract; the derived
image builds/imports on native x86; native-device acceptance remains pending. -
First versioned
miner-vllm-toplocimage forlinux/amd64, built from the default backend target. -
Calibration corpus v2 (
derive_calibration.py corpus --version v2 --tokenizer …, in
corpus_v2.py). It has 1,536 synthetic rows in eight length buckets, cut to exact
prefill lengths for one tokenizer, plus a token-count sidecar. v1 stays the default and
reproduces byte for byte. -
capture_calibration.pyprefills v2 random-id rows asTokensPrompt. On a v2 corpus it
refuses, naming the rows, when a row is under 2 tokens, misses itstarget_tokens, or
leaves no room under--max-model-len. It also refuses when--batch-pairswould not run
the planned pair slots. The generation probe caps a row's decode at the room its prompt
leaves under--max-model-len.derivereports per-class and per-bucket breakdowns and
refuses a v2 corpus that has a bucket missing from either holdout half.
Fixed
dense-gdntarget: prefill kernels run the first valid Triton config
(VLLM_TRITON_FORCE_FIRST_CONFIG=1), so captures are reproducible across processes.
Serving, calibration capture and replay refuse to start without it.- Dense-GDN serving refuses prompts shorter than 2 tokens, and calibration capture refuses
a corpus with any such row. vLLM 0.26 decodes them from an un-zeroed state slot. replay_chat_activations.pyno longer passes on equal lanes alone. It reports the
saturated-lane count and refuses a replay with any lane at the i16 clamp unless
unquantised rows agree with zero exponent errors (#2168, #3224).capture_calibration.pyno longer pairs rows with another run's metadata (#4297). A rerun
without--batch-pairsremoves a stale<config>.alt.f32. The previous<config>.jsonis
removed before any rows are written, and the new one is written last through a temp file and
os.replace. A crash mid-capture leaves no metadata, soderive_calibration.pyrefuses it.
da-validator v0.2.0
Added
- First versioned
da-validatorimage forlinux/amd64. submit --inflight <1-4>(default 2) uploads that many chunks at once and reads and encodes
the next one meanwhile. Each provider still gets one shard PUT at a time.submit --provider-fail-limit <K>(default 8) stops sending shards to a provider after K
refused PUTs in a row for the rest of the run, unless fewer than 3 providers would remain.
Its shards stay unconfirmed and the custody check after the push still decides.- Opt-in source pull (#4263).
serve --self-ingest
accepts a publisher-signed source hint atPUT /blob/{c}/source. It then builds this
validator's shards of an assigned blob from the hint's mirrors with HTTP Range reads. Each
chunk is checked against its manifest leaf before its shard is stored, and a refused chunk
tries the next mirror. Chunks no mirror serves are left to peer re-replication.
--self-ingest-hosts(defaulthuggingface.co,hf.co) bounds the hosts it may contact, and
--self-ingest-concurrency(default 2) the chunks fetched at once. submit --source-files <table> --source-mirror <url>offers that hint first, waits while
providers self-ingest (--source-stall-secs, default 600), then pushes only what they lack.- A custody worker submits
DaRegistry::confirm_custodyonce this validator's assigned index
of a custody-latched blob is custody-pinned. It is inert while the chain latches no blob.
Changed
-
submitlogs why a provider refuses a shard (its HTTP status or error) when it starts
refusing, and bounds each request (connect 10 s, 150 s in all, past the validator's own
60 s admission plus 60 s body deadline). -
A blob may have up to 65,536 chunks (4 TiB at the unchanged 64 MiB chunk size), up from
16,384 (1 TiB). The manifest body limit grows to 2 MiB to match. A validator on an older
image rejects these manifests, so roll every validator before such a blob is registered.
Fixed
retrieveno longer drops shards whose body takes longer than 30 s to arrive. The client bounds
connect (10 s) and each read (30 s) instead of the whole request, so--chunkspot checks pass
behind the public DA edge.
contracts v0.1.2
Added
- Generated exact-ID credential receipt, request ID and lookup error contracts,
preserving unknown values and exact terminal allocations in raw FLOP units.