Hi,

https://docs.sglang.ai/start/install.html

The installation instructions for docker does not work. Running the stock docker image as recommended for ROCm yields:

triton.runtime.errors.OutOfResources: out of resource: shared memory, Required: 196608, Hardware limit: 65536. Reducing block sizes or num_stages may help.

Other folks have run into this via cuda and it was recommended to upgrade vllm to vllm>=8.5. I built vllm from source and picked up with sglang's docker/Dockerfile.rocm builds as well as attempting to build from source and have run into multiple version conflicts and issues.

System i was attempting to run this on had 16x MI300X AMD cards.

Please advise, this seems doable to patch on my end but wondering if someone with more intimate knowledge with sglang's builds can provide insight or advise around any 'gotchas' I may run into. Either that or upgrading the Image to work.

Best,

P.S.

vllm-project/vllm#17578 (comment)