[HiCache] Fall back to host memory when no cgroup fs is mounted

[raghotham]

Motivation

HiCache sizes its host pool from available_host_memory_bytes(), which bounds psutil's available memory by the process's cgroup headroom. If the process belongs to a cgroup but no cgroup/cgroup2 filesystem is mounted (for example, a container that shares the host cgroup namespace without mounting cgroupfs), discovery raises

RuntimeError: Cannot locate the process memory cgroup in mounted cgroup filesystems

and the server fails to start whenever a HiCache host pool is created. There was also no way to give the budget directly.

Modifications

• mem_cache/host_memory.py
  • No cgroup filesystem mounted at all: no limit is visible, so log a one-time warning and size against host available memory instead of raising.
  • A cgroup filesystem is mounted but doesn't contain the process: still raise, since that may hide a real limit. The message now points to the override below.
  • SGLANG_HICACHE_HOST_MEMORY_BYTES: when set, available_host_memory_bytes() returns it and skips both cgroup and psutil discovery. The existing reserve and per-local-rank split in pool_host/base.py still apply. Values <= 0 raise ValueError.
• environ.py: registers SGLANG_HICACHE_HOST_MEMORY_BYTES (EnvInt, default unset).

Behavior is unchanged when cgroup discovery succeeds, and when no cgroup membership is present.

Tests

test/registered/unit/mem_cache/test_host_memory.py drives available_host_memory_bytes() through the real discovery code over a synthetic procfs:

• cgroup membership with no cgroupfs mounted (v1 and v2) returns host available memory;
• the explicit budget is returned even in a layout where discovery would raise, and 0 is rejected;
• a mounted but unmatched cgroup still raises, and the error names the variable.

These replace the old test that expected an empty mountinfo to raise. All three fail without the change.

Ran locally:

• pytest test/registered/unit/mem_cache/test_host_memory.py: 9 passed (13 subtests).
• ruff 0.15.1 check/format, isort 7.0.0 and codespell 2.4.1 on the changed files: clean.

GPU and e2e tests were not run locally.

---

CI States

Latest PR Test (Base): ✅ Run #36918037160
Latest PR Test (Extra): ❌ Run #36920089666
Latest PR Test (AMD ROCm 10): ❌ Run #36918036853

[metamergebot]

This port clears our bar on 489a69413d2b6ca9311d14ffdc1f455787af5e42: every base-a and base-b job is green.
base-c and the accelerator suites are reported but not gated, because they run
on scarce hardware and upstream merges through them routinely.

gate 20✓✗6 lint✓ build✓ A✓ B✓ C 0/10⏳ other✓ · accel ✗amd,mlx,musa

@merrymercy @raghotham — could one of you take a look and merge it if you are happy? Nothing else is waiting on our side.