fix: recover PR 1863 ingest via sweep reuse

perf-changelog.yaml

@@ -4086,3 +4086,16 @@
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260608-303757cc"
     - "6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP + wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1810
+
+- config-keys:
+    - minimaxm3-fp8-b300-dynamo-vllm
+  description:
+    - "Add MiniMax-M3 MXFP8 B300 disaggregated vLLM benchmarks via Dynamo for 1k1k and 8k1k STP."
+    - "Add local srt-slurm recipes under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8 and wire the B300 launcher to overlay them into srt-slurm."
+    - "Add right-sized TP4 decode variants with expert parallelism disabled and the Marlin MoE backend for selected low-concurrency 1k1k and 8k1k shapes."
+    - "Colocate the six-GPU TP4 prefill/decode pairs on one B300 node and enable CUDA IPC for NIXL KV transfer."
+    - "Patch the 0618 image's MiniMax M3 MSA prefill top-k slice to be contiguous before CSR construction."
+    - "Align 8k1k expert-parallel settings with the 1k1k recipes and correct the decode CUDA graph capture limit."
+    - "Backport NVIDIA/srt-slurm#38 to sanitize Slurm node-IP discovery output on the pinned submission branch."
+    - "Backport vllm-project/vllm#45879 so NIXL validates heterogeneous-TP KV block lengths using the GQA KV-head ratio."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1893