Bump sglang to v0.5.18

[yueming-yuan]

ci-megatron-pr: miles-main-20260819
ci-sglang-pr: sglang-miles-v0.5.18

Rebases the sglang-miles stack onto v0.5.18 and repoints the miles images at it. Targets #2673 so the Megatron and sglang bumps are validated together.

sglang side

sglang-miles-v0.5.18 = 35 commits on v0.5.18, replayed from sglang-miles-v0.5.17 rather than from production: production is still the v0.5.16 stack, so going through v0.5.17 halves the upstream drift (736 commits instead of 1317) and carries forward resolutions a green CI already validated. Production gained no commits in the meantime, so nothing is skipped. Rollback points: sglang-miles-v0.5.16-final (production) and sglang-miles-v0.5.17.

Only 9 of 35 picks conflicted. Where v0.5.18 has since removed machinery that the v0.5.17 resolutions leaned on, the newer shape wins and our behaviour was re-expressed on top of it:

• tokenizer_manager: v0.5.18 dropped the parallel-sampling child_rid_to_logical_rid / lifecycle_id / _remove_req_state machinery. Our abort-by-prefix gate, pre-dispatch flagging and LoRA lease release were rebuilt on v0.5.18's simpler form; the dispatch paths keep upstream's new CUDA-VMM prepare/cancel structure with our abort gate ahead of the reservation.
• forward_batch_info: v0.5.18 split mrope into decode/extend helpers and already forces the text-only path on rl_on_policy_target — the same intent as ours. Took its structure and added our contract switch at both decision points, since true_on_policy_contract is a separate switch from rl_on_policy_target.
• deepseek_v2: v0.5.18 moved the shared-expert fusion decision into the loader plus a shared_experts_fusion_disable_reason classmethod that carries our SBO/TBO and DeepEP reasons, so determine_num_fused_shared_experts is upstream's no-arg version.
• compile_utils: kept upstream's envs.SGLANG_DG_CACHE_DIR default and layered our SGLANG_DG_CACHE_DIR_PER_PROCESS suffix on it (colocated RL engines would otherwise race on one JIT cache).
• glm4_moe, dsa_indexer: each side adds a distinct method, both kept.

Post-pick repairs: restored _post_load_weights (the pick dropped the definition but kept its call sites) and get_exec in glm4_moe_nextn; a mangled determine_num_fused_shared_experts remnant was replaced with upstream's. ruff F821 is zero across every file the stack touches, and sys.exit(pytest.main(...)) now matches the repo convention in the parallelism-context test.

miles side

• docker/Dockerfile: SGLANG_IMAGE_TAG=v0.5.18, SGLANG_BRANCH=sglang-miles-v0.5.18 (switched to the stable sglang-miles name at merge time).
• docker/build.py: release variant pinned to v0.5.18-cu129.
• The cu12 pyproject.toml rewrite block was re-checked against v0.5.18's upstream Dockerfile — the three sed patterns and the three markers are unchanged, so it still applies as-is.

Rust files are untouched; local rustfmt/clippy hooks fail only for want of cargo on this machine.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}