chore: fill perf-changelog pr-link for #2448

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: functionstackx

perf-changelog.yaml

@@ -5367,4 +5367,4 @@
     - "Second delta: the checkpoint-resolution guard. runners/launch_b200-dgxc.sh rewrites MODEL to the cluster-local checkpoint path for single-node runs, so `hf download \"$MODEL\"` cannot work there; the script keeps the HF repo id in HF_MODEL_ID and downloads from it only when MODEL_PATH is missing or empty."
     - "Adds glm5.2/fp4 routing to runners/launch_b200-dgxc.sh, which previously hard-failed with 'Unsupported model prefix/precision' for this model. Probes /lustre/fsw/models/GLM-5.2-NVFP4, /scratch/fsw/models/GLM-5.2-NVFP4 and /lustre/fsw/gharunners/models/GLM-5.2-NVFP4 in that order (same shape as the dsv4 branch), falling back to the sa-shared-writable gharunners path and creating it, since --container-mounts needs the host path to exist and the bench script's download guard populates it on first use."
     - "Search space mirrors glm5.2-fp4-b300-sglang-agentic-mtp exactly so the two SKUs are comparable: one TP8 + HiCache arm at conc [1, 4, 8, 12, 16]. Steps of at least 2 because single-step sampling cannot separate configurations by more than run-to-run noise on the agentic corpus, and a hard stop at conc 16. TP8-only for memory as well: the ~433 GB NVFP4 checkpoint needs ~54 GB/GPU across 8 B200s and does not fit below 8. The DEP (attention-DP + EP8) throughput arm is not wired up for the same reason as on B300 -- its frontier peak sits well above the conc-16 cap -- but the branch, including the --speculative-moe-a2a-backend none / --speculative-moe-runner-backend triton pair that GLM-5.2-NVFP4's unquantized bf16 nextn layer needs under expert parallelism, is kept in the script."
-  pr-link: TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2448