1.2.19 (2026-10-02) Full Changelog: v1.2.18...v1.2.19 Features training-sdk: automatically select RDMA weight sync only for dedicated full-parameter trainers and inference replicas that all advertise support; older images, LoRA, and deployments using gateway shard fan-out retain save-weights followed by file hotload datasets: add fireworks.lib.dataset_upload.upload_dataset_shards to upload a .json…
1.2.18 (2026-10-01) Full Changelog: v1.2.17...v1.2.18 Features training-sdk: add forward-only reads for per-token projection outputs training-sdk: expose supports_router_replay on training clients, reported by the trainer at create_model, so callers can decide Router Replay without fetching the base model (None when the trainer does not report it) training-sdk: raise ModelDetailsUnavailableError (…
1.2.17 (2026-09-26) Full Changelog: v1.2.16...v1.2.17 Features training-sdk: add lora_init_method to create_training_client (init_method on create_lora_training_client) to pick each model's LoRA initialization at create time: "kaiming" (default) or "nora" (NoRA-init)
1.2.16 (2026-09-25) Full Changelog: v1.2.15...v1.2.16 Features training-sdk: add projection_head_dim to managed service and training-client creation for train-only projection heads, and output="projection" for forward_backward_custom gradients with respect to projection outputs
1.2.15 (2026-09-22) Full Changelog: v1.2.14...v1.2.15 Features training-sdk: retain per-token top-logprob token IDs and values in sampling results and TITO trajectory artifacts Bug Fixes training-sdk: preserve multidimensional custom-loss gradients as shaped tensor data during backward requests training-sdk: preserve failed trainer status when reporting provisioning errors to managed training jobs…
1.2.14 (2026-09-18) Full Changelog: v1.2.13...v1.2.14 Features training-sdk: select Parquet R3 using a small completion request; older inference servers retain inline routing without a capability endpoint training-sdk: allow sampling requests to set r3_ttl_seconds for shared Parquet routing files; the default retention is six hours
1.2.13 (2026-09-18) Full Changelog: v1.2.12...v1.2.13 Features deployments: add accept_shapeless_risk opt-out on deployments.create for intentionally creating a deployment without a pre-validated shape; document the deployment_shape="default" sentinel that asks the server to pick a default validated shape training-sdk: add accept_shapeless_risk to DeploymentConfig for the manual shapeless creation…
1.2.12 (2026-09-15) Full Changelog: v1.2.11...v1.2.12 Bug Fixes training-sdk: stop sending the obsolete managed-rollout deployment annotation training-sdk: wait for control-plane readiness when reattaching rollout deployments training-sdk: avoid copying dense tensor values when sizing requests for every client training-sdk: automatically negotiate optimized training requests for dedicated and serv…
1.2.11 (2026-08-31) Full Changelog: v1.2.10...v1.2.11 Features training-sdk: add environment-local TITO sidecars with full-history prompt construction by default and an experimental incremental renderer protocol training-sdk: preserve and attribute completion serving metrics when merging response headers with final performance metrics training: support explicit reservation targeting for dedicated…
1.2.10 (2026-08-30) Full Changelog: v1.2.9...v1.2.10 Features training-sdk: expose merged-base output precision with source-format export as the default Bug Fixes training-sdk: fail trainer readiness waits fast when the job stops instead of polling to the readiness timeout training-sdk: preserve canonical structured status details for managed recipe failures training-sdk: preserve Lifecycle, Tinke…
1.2.9 (2026-08-12) Full Changelog: v1.2.8...v1.2.9 Bug Fixes training-sdk: preserve per-request affinity for serverless training sampling
1.2.8 (2026-08-12) Full Changelog: v1.2.7...v1.2.8 Features training: default managed SFT/DPO and dedicated Training API trainers to reservation-first placement with an explicit opt-out
1.2.7 (2026-08-07) Full Changelog: v1.2.6...v1.2.7 Features training: support reservation-first placement for managed SFT/DPO and dedicated Training API trainers training-sdk: expose managed SFT and DPO renderer Hugging Face repo ID selection
1.2.6 (2026-08-03) Full Changelog: v1.2.5...v1.2.6 Features training: support CMEK-encrypted managed RFT sampler hot-loads and isolate reference trainers training-sdk: support repeated per-adapter model creation on one managed trainer Bug Fixes release: avoid false promotion-source failures after successful SDK releases
1.2.5 (2026-07-31) Full Changelog: v1.2.4...v1.2.5 Bug Fixes training-sdk: log backend training shape override warnings
1.2.4 (2026-07-27) Full Changelog: v1.2.3...v1.2.4 Features training-sdk: expose MoE model capability lookup for Router Replay training-sdk: expose structured Training API errors training-sdk: propagate training skill attribution headers Bug Fixes training-sdk: gate RLOR inactivity cleanup compatibility to Fireworks-owned jobs
1.2.3 (2026-07-21) Full Changelog: v1.2.2...v1.2.3 CI require generated promotion provenance before public SDK releases
1.2.2 (2026-07-21) Full Changelog: v1.2.1...v1.2.2
1.2.1 (2026-07-20) Full Changelog: v1.2.0...v1.2.1 Features serverless-rl: single retry owner + structured sampling error (SDK) (7113d1a) training-sdk: adjust rollout concurrency within RL steps (68595ff) training-sdk: expose routing matrices in sample responses (81d52c2) training: serverless cross-run resume — SDK config-from-checkpoint + cookbook unlock (PR3+PR5) (a284467) Bug Fixes training-sdk…
1.2.0 (2026-07-11) Full Changelog: v1.2.0-alpha.89...v1.2.0
1.2.0-alpha.89 (2026-07-10) Full Changelog: v1.2.0-alpha.88...v1.2.0-alpha.89 Bug Fixes training-sdk: sanitize public package contents (f0944c6) Chores training-sdk: raise async httpx connection limits (a562db0)
1.2.0-alpha.88 (2026-07-09) Full Changelog: v1.2.0-alpha.87...v1.2.0-alpha.88 Features training-sdk: default sampler concurrency to adaptive (53deb8e)
1.2.0-alpha.87 (2026-07-06) Full Changelog: v1.2.0-alpha.86...v1.2.0-alpha.87 Bug Fixes training-sdk: declare urllib3 runtime dependency (8c2c9d5) training: return 403 for training quota errors and surface the real cause (579b823)
1.2.0-alpha.86 (2026-07-05) Full Changelog: v1.2.0-alpha.85...v1.2.0-alpha.86 Bug Fixes managed: keep auto-shape selection when trainer.extra_args is set (a4b2a72) training-sdk: reject tombstone trainer reattach on create/poll (3c5ceee)
1.2.0-alpha.85 (2026-06-26) Full Changelog: v1.2.0-alpha.84...v1.2.0-alpha.85 Bug Fixes training: propagate auto-shape max context metadata (15d0a4d)
1.2.0-alpha.84 (2026-06-24) Full Changelog: v1.2.0-alpha.83...v1.2.0-alpha.84 Features training-sdk: surface serverless session (on the service) + run (on the training client) (ad77c33) Bug Fixes sdk: guard training deployment shapes (4b3ebbb) training: list and promote serverless session checkpoints by run (aefd3a8) training: require LoRA reference shapes (88a96e3)
1.2.0-alpha.83 (2026-06-18) Full Changelog: v1.2.0-alpha.82...v1.2.0-alpha.83 Chores public-repos: make SDK pyproject the version source (c349c92)
1.2.0-alpha.82 (2026-06-16) Full Changelog: v1.2.0-alpha.81...v1.2.0-alpha.82 Features cookbook/rl: expose lora_alpha in rl_loop and managed provisioning (5a67b14)
1.2.0-alpha.81 (2026-06-13) Full Changelog: v1.2.0-alpha.80...v1.2.0-alpha.81 Bug Fixes training-sdk: full-distribution RL rollouts + default LoRA alpha=32 (8e3df52)
1.2.0-alpha.80 (2026-06-12) Full Changelog: v1.2.0-alpha.79...v1.2.0-alpha.80 Features rlor: merged-base LoRA save + reject ineffective warm_start_from LoRA (#28107) (8b73705) Bug Fixes training-sdk: report probed deployments as ready (#29687) (d766124)