Any tutorial one how to use ROLL with ROCK and I-flow.

[shamanez]

I read the paper, let it flow, and it's ALE concept. Can you please provide how these three libraries can work together?

[PanAndy]

ROCK & ROLL 有一个快速开始指南：https://alibaba.github.io/ROCK/zh-Hans/docs/Getting%20Started/rockroll/
可以参考

[shamanez]

Thanks @PanAndy , I actually checked it , by any chance is there a tutorial where we connect ROCK+Iflow to a training instance ?

[PanAndy]

Rock ROLL Iflow is trained by chaining together the model-service and rock-agent provided by ROCK. The end-to-end training pipeline will be open-sourced in the future, but the training implementation currently needs to be refactored.

https://alibaba.github.io/ROCK/zh-Hans/docs/References/Python%20SDK%20References/model-service
https://alibaba.github.io/ROCK/zh-Hans/docs/References/Python%20SDK%20References/rock-agent

[shamanez]

Thanks. Well as a good starting point I will go through Rock’s iflow tutorial , then I need to change the model proxy to what is getting trained right ?

…

On Thu, 26 Feb 2026 at 8:46 pm, PanAndy ***@***.***> wrote: *PanAndy* left a comment (alibaba/ROLL#358) <#358 (comment)> Rock ROLL Iflow is trained by chaining together the model-service and rock-agent provided by ROCK. The end-to-end training pipeline will be open-sourced in the future, but the training implementation currently needs to be refactored. https://alibaba.github.io/ROCK/zh-Hans/docs/References/Python%20SDK%20References/model-service https://alibaba.github.io/ROCK/zh-Hans/docs/References/Python%20SDK%20References/rock-agent — Reply to this email directly, view it on GitHub <#358 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEA4FGW42ODDAVO7V5T3B734N26JHAVCNFSM6AAAAACV2RAMGKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSNRVGQYTOOBYG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[PanAndy]

Yes.
The model service has already implemented forwarding the agent’s requests to the LLM from within the sandbox.
On the ROLL side, the environment can capture these requests as observations and pass them to the inference engine (vLLM/SGLang) for generation.

[shamanez]

Excellent, thanks a lot

…

On Thu, 26 Feb 2026 at 10:34 pm, PanAndy ***@***.***> wrote: *PanAndy* left a comment (alibaba/ROLL#358) <#358 (comment)> Yes. The model service has already implemented forwarding the agent’s requests to the LLM from within the sandbox. On the ROLL side, the environment can capture these requests as observations and pass them to the inference engine (vLLM/SGLang) for generation. — Reply to this email directly, view it on GitHub <#358 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEA4FGQMH4UZJSHJFXJ64QD4N3K4NAVCNFSM6AAAAACV2RAMGKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSNRWGA2TAMRZG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[shamanez]

Got it to work with the IPC (inter process communication) protocol + Model Service, as sglang or VLLM cannot be accessed via http request during the training . I guess this is the correct way. This is more like we extract the iflow message after reading a log , then invoke the inference , send the response back to iflow etc.

@PanAndy

[shamanez]

@PanAndy

[PanAndy]

yes, and we plan to release a demo soon.

[shamanez]

Thanks a lot @PanAndy . Waiting for the release.