[KV Offload]: Tiered offloading + MoonCakeStore

[S1ro1]

Replace the flat inference.kv_cache_offload (single cpu_bytes) with a typed, backend-swappable, tiered config discriminated on type:

• native: vLLM OffloadingConnector (cpu) / TieringOffloadingSpec (cpu+disk, fs_python secondary). Fully self-contained.
• mooncake: per-node standalone-store (mooncake_master + mooncake_client, RDMA), cpu and/or disk tiers.

Composable cpu/disk tiers (cpu required, disk-only rejected); disk.num_bytes is mooncake-only (native fs tier is filesystem-bounded; warns if set on native).

Centralize the vLLM kv_transfer_config build in
InferenceConfig.build_kv_transfer_config (NIXL transfer + offload composed via MultiConnector); both SLURM templates stop hand-rolling the JSON and launch the node-local Mooncake store. New src/prime_rl/inference/mooncake.py launches the master+client and writes MOONCAKE_CONFIG_PATH; server.py brings it up for local runs.

Verified:

• Single-node (Qwen3-0.6B, H200): all four configs boot + serve; native-cpu and mooncake-rdma both round-trip KV (store ~1.1GB / reload ~205MB).
• Multi-node 3-node SLURM disaggregated (Qwen3-30B-A3B): NIXL P/D transfer serves coherent output; with mooncake offload, MultiConnector[Nixl, MooncakeStore] + per-node RDMA store serves coherent output.

---

Note

High Risk
Breaking inference config and changes to the KV transfer/offload path for multi-node and P/D SLURM jobs, plus new distributed Mooncake/RDMA dependencies.

Overview
Replaces the flat inference.kv_cache_offload.cpu_bytes setting with a type-discriminated offload config: required cpu.num_bytes, optional disk.path, and backends native (vLLM OffloadingConnector / TieringOffloadingSpec) or mooncake (shared RDMA store on SLURM). Old cpu_bytes is rejected — configs must migrate per CHANGELOG.

InferenceConfig.build_kv_transfer_config now assembles vLLM kv_transfer_config (NIXL for disaggregated P/D via auto-set use_pd_kv_transfer, plus offload; MultiConnector when both apply). to_vllm() uses that instead of inline offload JSON. SLURM inference / multi_node_rl templates drop hand-built kv_transfer_config in VLLM_EXTRA and, for Mooncake, include _mooncake_store.sh.j2 (master/client, MOONCAKE_CONFIG_PATH) with cleanup for mooncake_* processes. RL/inference entrypoints pass Mooncake template vars; validate_mooncake_offload_requires_slurm blocks Mooncake without SLURM. Adds mooncake-transfer-engine and docs/changelog for the new shape.

^{Reviewed by Cursor Bugbot for commit 44f951d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit cb7b4d4. Configure here.}