Skip to content

v0.6.23

Choose a tag to compare

View all tags
@slrelease slrelease released this 24 Jul 04:44
· 31 commits to main since this release
v0.6.23
9d6ca6b
  • chore(main): release 0.6.23
  • fix(grammar): bound pointer index parsing
  • fix(grammar): harden schema normalization
  • fix(gateway): validate schema keywords by context
  • fix(sdk): accept nullable grammar metadata
  • fix(cloud): attest LightOnOCR image compatibility
  • fix(catalog): disable unsafe qwen speculative default
  • test(sdk): assert grammar validation failures precisely
  • fix(sdk): keep grammar types on supported module path
  • fix(generation): close structured grammar review gaps
  • feat(generation): type structured output contracts
  • test(server): assert visual muvera config removal
  • fix(server): hide unsupported visual muvera profiles
  • fix(server): pin launch guardrail policies
  • test(adapters): stdlib imports to file top + comment the skip-tail except (CodeRabbit/Copilot)
  • fix(server): defer LightOn OCR compatibility patch
  • fix(adapters): serialize output_hidden_states forwards to close the #2144 recorder-race class
  • fix(server): meter SigLIP tokens from attention masks
  • perf(vision): reuse siglip preprocessing
  • fix(vision): pad fused detection batches
  • test(sdk): harden Responses capacity contract
  • feat(sdk): add typed Responses client
  • test(api): harden multimodal contract fixtures
  • fix(api): close multimodal parity gaps
  • fix(api): address multimodal contract review
  • fix(server): preserve stream helper compatibility
  • feat(api): accept native multimodal generation
  • fix(cloud): align launch evidence contracts
  • fix(catalog): raise GTE Muvera repetitions
  • fix(catalog): route mxbai edge through native attention
  • fix(catalog): use pinned GTE query length
  • fix(server): serialize remaining tokenizer access
  • fix(server): preserve trusted code revision on fallback
  • fix(server): restore ColBERTv2 retrieval recipe
  • fix(server): apply trained Rotary ColBERT projection
  • fix(server): mask ColBERT expansion keys in flash attention
  • fix(server): restore Jina ColBERT retrieval recipe
  • style(test): D209 docstring closing quotes on own line + re-lock manifests
  • fix(eval): restore colbert recipe and cache integrity
  • docs(test): docstring the trust_remote_code pin invariant (CodeRabbit coverage)
  • fix(security): pin hf_revision on the 10 unpinned trust_remote_code models (#2109)
  • fix(catalog): preserve muvera quality similarity
  • test(ci): strengthen profile and install contracts
  • fix(eval): align governed benchmark identities
  • fix(vision): consolidate detection batch alignment
  • fix(ci): harden scoped quality and rust jobs
  • fix(vision): serialize Qwen processor tokenization
  • fix(sie_server): harden remote-code snapshot registration (review)
  • fix(vision): preserve fused request image alignment
  • fix(eval): preserve profile floor identity
  • fix(eval): scope quality floors to model profiles
  • fix(splade): enforce candle sequence cap
  • test(splade): cover inherited runtime caps
  • fix(retrieval): align GTE ColBERT serving recipe
  • fix(splade): preserve default retrieval recipe
  • test(encode): declare mocked output capabilities
  • style(encode): keep required imports global
  • fix(encode): align profile output contracts
  • fix(profiles): honor routed output capabilities
  • fix(vidore): per-thread processors in ImagePreprocessor instead of a pool-wide lock (#2098)
  • fix(vidore): close the image-path and qwen3_vl gaps in the tokenizer serialization (#2098 review)
  • fix(vidore): serialize Col-family processors and retry eval transport loss under load
  • fix(sie_server): register remote-code snapshot dir on sys.path for ST-Router models
  • fix(release): keep audio prep internal to builds
  • fix(cloud): make wallet balance own admission
  • feat(models): onboard naver/v-splade-quality (#2035)
  • fix(gateway): make sealed-standby schema drift a compile error, not a 5s log line
  • feat(1841): declare the sealed-lane metrics through the closed telemetry contract
  • fix(1841): close the post-merge gaps the real Postgres run found
  • fix(custom-models): a variant-named model's DPA erase no longer wedges forever
  • feat(custom-models): withdraw erased models from the catalog; quote only a real rate
  • feat(custom-models): gateway bundle-less sealed dispatch for org custom models (#1841 W4)
  • feat(gateway): resolve-seam org gate for the OpenAI body routes (#1895)
  • refactor(gateway): gate + list custom models on the RESOLVED id (#1895)
Assets 2
Loading