fix(benchmark/sglang): keep the radix cache when the caller enables HiCache

[jiejingzhangamd]

Motivation

sglang_mi300x.sh / sglang_mi355x.sh add --disable-radix-cache to every server launch unless EXTRA_SGLANG_ARGS already contains it. SGLang refuses that flag next to --enable-hierarchical-cache (HiCache sits on the radix cache):

ValueError: The arguments enable-hierarchical-cache and disable-radix-cache are mutually exclusive and cannot be used at the same time. Please use only one of them.

So any recipe that enables HiCache — e.g. the SGLang cookbook GLM-5.2 MXFP4 MI355X high-throughput cell (--enable-hierarchical-cache --hicache-ratio 1.0) — cannot start through Magpie, and there is no caller-side workaround: passing --disable-radix-cache to suppress the default conflicts the same way. In one Hyperloom session this failed 8 times (baseline, roofline profiling ×3, explore candidates ×3, plus a second node).

Change

In both scripts' default-args loop, skip the --disable-radix-cache default when EXTRA_SGLANG_ARGS enables the hierarchical cache. Every other recipe keeps today's defaults (--mem-fraction-static=0.8, --disable-radix-cache).

Fixes #126. Related: AMD-AGI/Hyperloom#1601 (the same default as a prefix-cache regression on the AgentX path) and AMD-AGI/Hyperloom#1930.

Verification

• tests/test_sglang_default_args.py (new): runs each real script's server phase with a stub setsid that records the launch argv. HiCache recipe → --enable-hierarchical-cache present, --disable-radix-cache absent (fails before this change); plain recipe → both defaults still added. With tests/test_eval_concurrency_scripts.py: 58 passed. ruff check / ruff format --check clean on the new test (make lint reports the same 1555 pre-existing findings on main; this change touches no Python under Magpie/).
• Hardware, Docker, 8× MI355X (gfx950), image lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260923 (SGLang 0.5.20, torch 2.9.1+rocm7.2.0), model amd/GLM-5.2-MXFP4, TP4/EP4: patched sglang_mi355x.sh with MAGPIE_RUN_PHASE=server and the cookbook flags (--kv-cache-dtype fp8_e4m3 --dsa-*-backend triton, EAGLE 5/1/6, --enable-hierarchical-cache --hicache-ratio 1.0, --mem-fraction-static 0.85, --context-length 11264). Server args: enable_hierarchical_cache: True, disable_radix_cache: False; log Tree cache initialized: ... hicache_attached=True; /health 200 after ~180 s; /v1/completions "The capital of France is" → " Paris. It is one of the most". Before the change the same launch exits in argument parsing.
  (Without --context-length, that recipe OOMs at the model's default context in the DSA indexer's paged MQA logits — unrelated to this change; noted for anyone reproducing.)

🤖 Generated with Claude Code