Jeff Daily

@ProExpertProg fixed. I had aggressively removed that variable from vllm/model_executor/layers/quantization/utils/fp8_utils.py because it wasn't used elsewhere in that file. I failed to grep through all the other files for anyone that...

Looks like a ROCm version issue. `__hip_fp8_e4m3` and related types were added in ROCm 6.3. Not present in ROCm 6.2. I was testing in a ROCm 6.3 environment and assumed...

Working on a fix now. Can we forward-fix with a new PR or do you need to revert this one?

Fix shouldn't be complicated. Will need to isolate the use of `__hip_fp8_e4m3` behind a ROCM_VERSION macro limiting it to ROCm 6.3 or newer.

@ProExpertProg I'm not on slack yet. Here's the PR fix. https://github.com/vllm-project/vllm/pull/14709

I will make an exception for this. hipblaslt doesn't have a public rocblaslt interface or library. We really should be putting such extensions as "roc" APIs only. The HIP interfaces...

@pytorchbot merge

@pytorchbot merge

That feature is not currently supported.

It would be a great feature. Alas I have barely enough time to keep this project (both the C lib and the python bindings) maintained. Can you help design the...