v1.0.0-alpha.28
Pre-release
Pre-release
1.0.0-alpha.28 (2026-02-23)
Full Changelog: v1.0.0-alpha.27...v1.0.0-alpha.28
Features
- add R3 routing matrix side-channel injection for official tinker compat (f6b637b)
- default prompt_cache_max_len=0 to disable KV cache for verification (e49ecfb)
- default to greedy mode, add --stochastic flag, include dataset (049c4e9)
- port firetitan SDK and cookbook into fireworks.training (7134d64)
- port SDK unit tests from train-firetitan-py (dac3472)
- Regenerate API Reference / Docs (d97cb7a)
- Regenerate OpenAPI specs and SDK (v4.24.80) (ea622fd)
- Service mode LoRA documentation (32edbcd)
Bug Fixes
- correct stale "Server tokenization" log label to "Client tokenization" (5e6fdcc)
- exclude examples from mypy to avoid duplicate module conflict (c2ca3e5)
- harden client-side tokenization migration (7c586fd)
- internal: skip tests that depend on mock server (2aa4aca)
- move reportCallIssue ignore to weight_decay line (8146334)
- pass trust_remote_code=True to AutoTokenizer.from_pretrained (7ae2b5d)
- remove sdk/cookbook from training all to fix pyright error (d185f6f)
- resolve all ruff lint errors in ported SDK/cookbook code (5df914d)
- resolve e2e test issues found during test runs (ed2ed28)
- resolve pyright lint errors in training examples (13f97fd)
- suppress tinker SDK type stub errors in training examples (41ab1d6)
- use proper type annotation for all to satisfy pyright and ruff (71fb6bc)
Chores
- internal: add request options to SSE classes (91db2ea)
- internal: make
test_proxy_environment_variablesmore resilient (6b909f0) - internal: remove bad test (4ad6fcf)
- internal: remove mock server code (fbb2657)
- move verify_logprobs.py and watch_trainer_pods.sh to fireworks repo (ae3cf92)
- remove --stochastic flag, keep greedy as the only default (5502945)
- remove redundant tinker_patch.py (27b5521)
- sanitize training code for public release (36fc2ba)
- sanitize training docs/tests for public release (ee1fdb7)
- update mock server docs (db18242)
Documentation
- add compact training readmes and fix lint imports (12d7196)