v0.22.3

The th#767 completion ladder: 0.19.1 arm_compile · 0.20.0 allow_lora eviction + provider/ref alias retirement · 0.21.0 dynamic slot materialization (instance-per-pick residency; hub-connected workers never self-fetch raw upstream) · 0.22.0 ctx hierarchy + ResolvedCompute trim · 0.22.1 discovery lazy-import shim (module-top torch) · 0.22.2 selective component download · 0.22.3 lora repo-metadata composition. Plus the pb2 regen hotfix.

v0.18.1

Re-release of 0.18.0 (whose publish failed on the schema-export golden under msgspec 0.21) + the inspect.cleandoc description normalization (PR #231). Carries the full pgw#520 Slot/FamilyDefaults surface.

v0.18.0

pgw#520 (th#767 SDK half): FamilyDefaults vocabulary + JSON-schema export CLI; Slot(pipeline_cls, selected_by=, default=, fallback=); discovery emits functions[].slots; ctx.slots[...] typed resolution chain (repo metadata over code fallback, whole-object precedence); hub-less default path. ModelBinding.inference_defaults wire field (parity with tensorhub #305).

v0.17.4

pgw#515: de-fork cli/run.py onto the executor's load+place core (models/provision.py). Net −774 LOC; single fp8 fit factor (0.55); CLI gains the TRT compile lane; GEN_WORKER_LOCAL_DEVICE env channel and --variant deleted.

v0.17.2

pgw#519: stamp Compile(family=) onto model.choices[].binding — unblocks builder-path deploys of ModelChoice endpoints (th#586 gate). Build-time error when an allow_lora choice binding has no declared family.

v0.16.0

BREAKING (pgw#514): dead-surface sweep — deletes dead exports (PositivePrompt/NegativePrompt, TokenStreamSignal, InputTooLargeError, 3 models exports), 6 dead RequestContext init params, dead Settings fields; bare ValueError now maps to FATAL (retryable) not INVALID; FATAL messages redact absolute paths; net −241 LOC. Endpoints pin <0.16 and are unaffected.

v0.15.0

BREAKING (pgw#509): variants= and route= deleted from @endpoint.

• Checkpoint selection is a typed model= payload arg: ModelChoice[D] enum of Model(id, ref, defaults, hot=, price=) rows
• BYOM = field type (model: X | ModelRef); no decorator
• Discovery emits the placement-key manifest block {field, byom, slot, choices[]} (th#761 contract)
• Endpoints pinned <0.15 are unaffected until they migrate (ie#469)

v0.14.8

• Ship py.typed (PEP 561) — SDK types now reach downstream endpoint checkers (was silently Any fleet-wide)
• Typing: @endpoint overloads, StringEnum=StrEnum alias, typed ctx.models/loras/save_*
• ie#463: diffusers_step_callback gains window=(start,end) for multi-stage progress + per-step cancel