functionstackx
functionstackx
Estimate: 2-3 dev hours Change the data type from single selector to dropdown multi-selector for precision field (i.e. can select fp8 and fp4 in 1 graph). Acceptance Criteria: The user...
### Suggestion Description on every commit/release that changes the enroot/pyxis code, there should be open source CI that validates that pyxis/enroot works for multinode mpi rccl-tests, single node simple pytorch...
# Overview [SLURM Pyxis Container Plugin](https://github.com/NVIDIA/pyxis) has a very clean UX & first class support for using container with `SLURM` with just an additional arg `--container-image`passed into `srun`. Many nvidia...
### Suggestion Description There is lots of room for improvement on ROCm docker UX. On NVIDIA GPUs, end users just do `--gpus all` but on AMD, i need to type...
dcgm only has an global tensor core active metric and only has 3 dtype specific ones (IMMA for int8, HMMA for fp16/bf16, DMMA for tf32/fp32), missing is fp8 and fp4/6....