vllm 0.16 single node mi325 kimi k2.5 vllm tp8

[functionstackx]

@claude vllm 0.16 kimi k2.5 int4 tp8

similar to mi355 vllm int4 kimi recipe

remember that above the vllm.serve start command to have an comment saying it is following AMD andy's recipe https://x.com/linluo77/status/2017024513595301985

[Klaud-Cold]

Claude finished @functionstackx's task in 2m 38s —— View job • claude/issue-856-20260303-0249 • Create PR ➔

---

Task: Add Kimi K2.5 INT4 vLLM benchmark for MI325X (single node, TP8)

• Gather context - read existing MI355X Kimi K2.5 INT4 config and scripts
• Create MI325X Kimi K2.5 INT4 vLLM benchmark script
• Update AMD master config with new config key
• Update perf-changelog.yaml
• Commit and push changes

Summary

Added the kimik2.5-int4-mi325x-vllm benchmark config for Kimi K2.5 INT4 on MI325X single-node with TP8, based on the existing MI355X INT4 recipe:

• Benchmark script: benchmarks/single_node/kimik2.5_int4_mi325x.sh — includes AMD Andy Luo's recipe comment above vllm serve
• Config: kimik2.5-int4-mi325x-vllm in amd-master.yaml using vllm/vllm-openai-rocm:v0.16.0 image, model moonshotai/Kimi-K2.5, TP8 across all three seq-len configs (1k1k, 1k8k, 8k1k)
• perf-changelog.yaml updated