[NV] update B300 disagg recipes

[biswapanda]

Note

Medium Risk
Large benchmark-only surface area, but the launcher mutates cloned srt-slurm and patches installed vLLM at runtime—failures would affect CI/cluster jobs rather than app users.

Overview
Adds MiniMax-M3 MXFP8 disaggregated Dynamo + vLLM fixed-sequence benchmarks on B300, including a new minimaxm3-fp8-b300-dynamo-vllm entry in nvidia-master.yaml with 1k/1k and 8k/1k search spaces (prefill DEP2, varied decode topologies and concurrencies).

Introduces local srt-slurm recipes under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/ and wires launch_b300-nv.sh to overlay them on sa-submission-q2-2026, set model paths, run minimax-m3-vllm-fixes.sh via srtctl --setup-script, apply the srt-slurm node-IP patch, optionally exclude b300-018, and verify #SBATCH --exclude in the generated script.

The setup script patches the installed vLLM image at job time: contiguous MSA prefill top-k for CSR, and NIXL KV block-length checks using GQA head ratios (heterogeneous TP). Recipes add TP4 + Marlin decode options, colocated 6-GPU pairs with CUDA IPC for NIXL on selected shapes, and 8k1k fp8 KV / attention settings aligned with 1k1k.

Documents the work in perf-changelog.yaml (PR #1863).

^{Reviewed by Cursor Bugbot for commit b2e71c8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d08cc43. Configure here.}