Optimize GLM-5.2-MXFP4 SGLang Agentic Performance on MI355X

[limou102]

Summary

• Update the GLM-5.2-MXFP4 MI355X AgentX SGLang image from v0.5.16 to v0.5.19.
• Pick up recent SGLang main-branch optimizations.
• Restore write_through as the default HiCache write policy.

Benchmark Results

Test configuration: MI355X, TP4/EP4, MTP, HiCache DRAM offload, 3600-second AgentX profiling.

Concurrency	P90 Interactivity	Throughput per Chip
10	66.92 tokens/s/user	13,002.49 tokens/s/GPU
12	57.87 tokens/s/user	12,340.00 tokens/s/GPU

Concurrency 10 provides the best balance of output interactivity and per-GPU throughput.

Validation

• Both benchmark runs completed successfully with no profiling request errors.
• Bash syntax, YAML parsing, and targeted matrix generation passed.

---

Note

Low Risk
Benchmark and config-only changes (container pin and HiCache default); no auth or core application logic.

Overview
Updates the glm5.2-fp4-mi355x-sglang-agentic-mtp recipe to SGLang ROCm v0.5.19 (20260907) from v0.5.16, picking up recent main-branch GLM-5.2-MXFP4 serving work.

In glm5.2_fp4_mi355x_sglang_mtp.sh, the default HiCache write policy switches from write_through_selective back to write_through as the validated baseline for AgentX interactivity and per-GPU throughput; HICACHE_WRITE_POLICY can still be set explicitly for selective-write experiments.

perf-changelog.yaml documents the image bump and write-policy change for this config key.

^{Reviewed by Cursor Bugbot for commit 5361b95. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[github-actions]

Dispatched trusted sweep run for approved external revision 5361b955f8451e45f23b23a9fdc4533c35898628.
New commits are not trusted automatically; remove and re-add the primary sweep label to approve a new SHA.

已为获批的外部提交 5361b955f8451e45f23b23a9fdc4533c35898628 调度trusted sweep run。
后续新提交不会自动获得信任；如需批准新的 SHA，请移除并重新添加主扫描标签。