[NVIDIA] Add DSR1 MTP

[kedarpotdar-nv]

As per guidance we've recvd, wanted to add MTP ON configs for single node after GB200 is merged.
Only TRT MTP for now.

Changes:

1. Added benchmarks/dsr1_*_mtp_slurm.sh files for fp8 and fp4
2. modified dsr1-tmpl.yml to account for new runs.

[kedarpotdar-nv]

I'm not sure the mtp-mode: on actually triggers the dsr1_b200_fp4/fp8_trt_mtp.sh scripts.

[kedarpotdar-nv]

@kimbochen can you please guide how to add logic in runner/launch b200 to account for mtp flag?

[csahithi]

Can we add a variable called mtp-mode (default set to off) to benchmark-tmpl.yaml? We can pass this as an env variable which will be used by the runner script. If set as on, then we attach _mtp as a suffix to the filename so that the correct script gets picked up.

[kedarpotdar-nv]

Thanks @csahithi . i have made changes accordingly. ran several sample sweeps:

https://github.com/InferenceMAX/InferenceMAX/actions/runs/18050118322
https://github.com/InferenceMAX/InferenceMAX/actions/runs/18050505472
https://github.com/InferenceMAX/InferenceMAX/actions/runs/18051017127
https://github.com/InferenceMAX/InferenceMAX/actions/runs/18049715475

[functionstackx]

stale