SGLang and Miles add day-zero support for Inkling 2 Mini
SGLang lists single-GPU serving options, and Miles tested its training recipe on eight Nvidia B200 GPUs.
The SGLang and Miles teams added serving and reinforcement-learning support for Inkling 2 Mini on November 20, the day Thinking Machines Lab released the model. SGLang says Mini support is still outside its regular releases.
SGLang directs users to model-specific Docker images or the pull request that adds Mini support.
SGLang lists single-GPU recipes for Nvidia’s RTX PRO 6000 Blackwell, RTX 5090 and DGX Spark. Its cookbook also covers Nvidia B300/GB300, B200/GB200 and H200, AMD MI350X and MI355X, and Google TPU7x. On TPU7x, SGLang says it tested the recipe with DSpark, which uses a draft head to propose tokens. The cookbook lists the plain-decoding recipe as untested end to end.
Miles uses SGLang to generate groups of sample answers on the same GPUs that train the model. Miles scores those answers and uses them to update the model with a method called full-parameter GRPO. The recipe uses text prompts and keeps the vision encoder and draft head frozen.
Miles reports that in one 100-step run on the DAPO-Math-17k dataset, the mean raw reward rose from about 0.52 over the first 10 steps to about 0.85 over the last 10. The run used eight Nvidia B200 GPUs and capped each response at 4,096 tokens. Miles says the share of responses cut short fell from about 47% to about 9%, accounting for much of the reward gain. Miles reported no held-out comparison for the run.
Support for Mini follows the teams’ November 12 integration of Inkling 2 Small, the larger 594-billion-parameter model. vLLM also added day-zero serving support for Small on November 12.
More news

VAST Data announces per-job storage controls for VAST AI OS 5.6

AWS introduces cross-Region inference routing for SageMaker HyperPod

AWS adds managed user administration to HyperPod Slurm clusters
