v4.6.0

JEV Support (typed-decision / System-1)

Added the decision task type for JEV-style decision models: instead of generating text, the model returns calibrated probabilities over each question's fixed option set (choice / score / noul) in a single forward pass. (#10306)
Ready-to-use recipes for three released decision models — JEV-27B-VL, Clef and OmniJev — covering training, inference and LoRA merge/export. Any CausalLM can also be converted with swift sft --task_type decision.
Details: Decision (Best Practices) · examples

MiniCPM5 (1B / 2B) — LLM series with base and SFT variants, vLLM inference supported.
MiniCPM-V-4.7 (1B / 35B-A3B) — vision-language models with image and video support.
GLM-5.3-Flash — ZhipuAI GLM-5.3-Flash with vision and video support.
DeepSeek-V4.1-Flash — full-parameter SFT & GRPO on Megatron-SWIFT plus vLLM inference.
Xing4.0-29B-A4B — MLA + MoE + mHC; LoRA & full-parameter SFT (DDP / DeepSpeed ZeRO-3) with grouped_mm expert acceleration (~3.7x).
MonkeyOCRv2 (S / B) — OCR and document-understanding models.

Added a Moore Threads (MUSA) backend via torchada. (#10224)
Compatibility: transformers ≥ 5.17 kernel proxy, vLLM 0.28+ colocate weight sync / wake-up API, and TRL optional-dependency checks under Transformers 5. (#9770, #10311, #10315)
Updated qwen4_exp / glm5_next MTP examples with GLM-5.3 support. (#10219)

Hardened the WebUI and rollout: fixed unauthenticated RCE and arbitrary file read, a rollout-side vulnerability, and GitHub Actions workflow injection. (#10209, #10218, #10226, #10227)
Authenticate before request-body deserialization; weights_only=True for Megatron checkpoint loading; download security fixes. (#10062, #10066, #10234, #10235)

Loss & gradient alignment across SP, DPO, DAPO/CISPO/FIPO and GKD/GRPO (including Liger and off-policy paths); listwise reranker and partial gradient-accumulation fixes. (#9977, #10100, #10104, #10116, #10130, #10144, #10185, #10244, #10285)
Megatron & parallelism: contiguous CP token order, ring-padding handling in output unpacking / packed seq_acc / MoE aux loss, mcore checkpoint metadata. (#10071, #10113, #10181, #10182, #10184, #10242)
Training stability: ZeRO-3 hooks after generation errors, GaLore checkpointing, LISA init and layer selection, SP sampler options, CPU-offload activation views. (#10132, #10140, #10157, #10161, #10173, #10260)
RLHF & rollout: colocate offload / vLLM wake-up ordering, async rollout failure propagation, IPC bucket acknowledgment, vLLM LoRA metadata and local-teacher adapters. (#10092, #10107, #10109, #10207, #10231, #10232)
Embedding & reranker: InfoNCE negative sampling and gradient compensation under DDP / Megatron DP, DP metric gathering, WeMM duplicate forward. (#10059, #10068, #10069, #10081, #10088, #10091, #10150)

Fixed qwen-omni vLLM inference, video metadata handling after frame sampling, and video_second_per_grid tensor handling. (#10168, #10274, #10330)
infer: adapter-cache and async-request isolation, independent event-loop thread, streamed-request completion, worker error propagation, SSE heartbeats. (#10080, #10086, #10093, #10094, #10134, #10135, #10177)
Device-map memory resolution for MP+DDP; CLI list-value serialization; blank LOG_LEVEL / LOCAL_RANK handling; WebUI default port. (#10210, #10217, #10245, #10246, #10248, #10292)

Tool-call correctness across GLM / glm4_5 / Qwen3-Coder / kimi_k25 / ReAct / Seed-OSS, OpenAI & Anthropic tool-result alignment, and JSON-Schema type-union parsing. (#10169, #10174, #10270, #10271, #10284, #10293, #10294, #10299, #10300, #10304)
Template fixes for gemma4, marco_o1, Qwen3-VL video, Janus, rejected-response placement and tool-response names. (#10120, #10127, #10171, #10190, #10199, #10203, #10230, #10238)

Dataset pipeline: cross-rank cache races, Arrow tool-schema fidelity, streaming packing worker detection, versioned map cache, metadata cache invalidation, media sharding. (#10083, #10121, #10125, #10159, #10160, #10162, #10297, #10298, #10301)
WebUI: named datasets now list their hub IDs. (#10119)
Model loading: Idefics3 AutoModelForImageTextToText, MiniCPM position cache under DDP, MiniCPM-o padding offsets, seq-cls metadata round-trips. (#10187, #10202, #10228, #10253)

Vectorized ring-attention LSE extraction and packed zigzag CP reconstruction; faster dataloader for single-rank groups; batched rollout / teacher-API metric gathering. (#10064, #10082, #10239, #10240, #10241)

Added AGENTS.md and contribution guidelines (human-submitted PRs; first-time and AI-assisted contributors). (#10339, #10347)
Updated hardware and docker-image docs; fixed reward examples, fractional-loss-weight docs and in-page anchors. (#10067, #10118, #10170, #10172, #10188, #10233)

Full Changelog: v4.5.3...v4.6.0