v5.0.6.rc1

qwen3-vl supported.

v5.0.5

qwen3-4b with tp>1 was supportted.
RLHF:
Skipping loading weights from local files when starting up vllm engines.
LLM was replaced by AsyncLLM. So manual batching was replaced by auto-continuous batching. Which means training waitting time has be shorten as the first response time.
max_num_seqs of training was no longer needed. Adjusting max_num_seqs and num_vllm_engines of vllm to balance training speed and inferrence speed of one batch samples will benefite a highest performance.

v5.0.4.post1

RLHF:
More vllm engines corresponsed to one training duplicate can be started, max_num_seqs*num_vllm_engines repsponces could be generatted at every inferrence request.
Adjusting arguments max_num_seqs and num_vllm_engines to make training speed of max_num_seqs*num_vllm_engines*num_generations samples approximately equals to inferrence speed of one request will benefite a highest performance.

v5.0.4

Enjoy an efficient RL.

v5.0.4.rc2

Support GRPO on deepseek-685b.

v5.0.4.rc1

Cyclically make vLLM inference and jllm training asynchronous during RL.
Support zero loss of RL.
Support training a micro batch of the generated samples belong to one group at a micro step during RL.

v5.0.3

MoE:
Experts in tp-ep group will be recombined at several steps for load balance if setting swap_experts_per_steps>0.
RLHF:
Multiple vllm engines can be started.
Generated completions of one group can be distributed on diffrent data parallel when sequence parallel is on.

v5.0.3.rc1

qwen3/qwen3-moe supportted.

5.0.2.post1

qwen2-vl/qwen2.5-vl supportted on NPUs.

v5.0.2

More efficient training on MoE with big EP.