Fix SGLang disagg on ionic/RoCE fabrics (AINIC transport env)

[i-kosarev]

Summary

Minimal fix (analogous to #1) enabling the SGLang disaggregated-serving workload
to run on ionic-based RoCE clusters such as mia1.

$MOONCAKE_COOKBOOK_PATH/set_env_vars.sh assumes a Mellanox/mlx5 fabric. On
ionic clusters it corrupts the transport environment three ways:

1. forces NCCL_IB_DISABLE=1 (kills RDMA entirely),
2. hardcodes mlx5 IB device names (wrong devices → no IB),
3. mangles NCCL_SOCKET_IFNAME via an ip route awk parse.

The fix saves the launcher-provided NCCL_SOCKET_IFNAME before sourcing, then
re-asserts the ionic-correct values afterward: NCCL_IB_DISABLE=0, IBDEVICES
from IB_DEVICES, and NCCL/GLOO socket ifnames. host_ip is also derived from
the actual transport interface instead of the first hostname -I address (which
on multi-homed nodes may be the wrong NIC).

Single file, single hunk: scripts/sglang_disagg/sglang_disagg_server.sh.

Validation

Validated on mia1 (SLURM job 11211):

• DeepSeek-R1-0528, 1 prefill + 1 decode + proxy disagg, TP=8, Mooncake KV
• RCCL 78e8ba0 overlay + AINIC (ionic RDMA) transport confirmed in NCCL logs
• Both servers fired up, sglang_router bound on port 2322
• Full benchmark sweep, 100% request success:

ISL/OSL	con8	con16	con32	(total_token_throughput tok/s)
1024/1024	154.47	309.46	630.84
8192/1024	704.67	1361.27	2712.20

Notes / prerequisites

On gfx950 (MI355X) this workload also requires the kernel + runtime work from the
private stack #341 (gfx950 kernel overlay) and #342 (runtime kernel guards +
Mooncake P/D default). This PR is orthogonal — it is the fabric/transport fix
needed on ionic RoCE clusters regardless of GPU arch.

Test plan

• SGLang R1 disagg comes up and serves on mia1 (ionic RoCE)
• NCCL logs show IB/ionic transport (RDMA engaged, not disabled)
• Full benchmark sweep completes with perf CSV captured

[mkuznet1]

Reuse the IP that's already resolved upstream instead of re-deriving it here.

run.sh already resolves this node's IP (same iface → hostname -I logic) and hands us the rank-ordered, post-rendezvous node list via IPADDRS, and NODE_RANK is known here. So this node's IP is simply IPADDRS[NODE_RANK] — no second ip addr / ip route / hostname -I pass is needed.

More important than the dedup: host_ip feeds --host and socket_barrier --local-ip, which must match what peers registered for this node in the router/barrier (both are built from IPADDRS). An independent re-derivation can resolve to a different NIC on a multi-homed node and desync the two — the server binds/advertises one address while the router looks for another. Sourcing it from IPADDRS[NODE_RANK] keeps them consistent by construction.

The RDMA/socket env re-assert above (NCCL_IB_DISABLE / IBDEVICES / *_SOCKET_IFNAME) is the actual ionic fix and stays as-is.

Suggested change

	# Derive host_ip from the actual transport interface rather than the first
	# `hostname -I` address (which on multi-homed nodes may be the wrong NIC).
	_XPORT_IFACE="${_SAVED_NCCL_SOCKET_IFNAME%%,*}"
	host_ip=""
	if [[ -n "$_XPORT_IFACE" ]]; then
	host_ip=$(ip -4 -o addr show "$_XPORT_IFACE" 2>/dev/null | awk '{print $4}' | cut -d/ -f1 | head -n1)
	fi
	[[ -z "$host_ip" ]] && host_ip=$(ip route get 1.1.1.1 2>/dev/null | awk '{for(i=1;i<=NF;i++) if($i=="src"){print $(i+1); exit}}')
	[[ -z "$host_ip" ]] && host_ip=$(hostname -I | awk '{print $1}')
	host_name=$(hostname)
	unset _SAVED_NCCL_SOCKET_IFNAME _XPORT_IFACE
	# This node's IP is already resolved by run.sh and handed to us (rank-ordered,
	# post-rendezvous) via IPADDRS. Reuse IPADDRS[NODE_RANK] so the address we bind
	# and advertise (--host, socket_barrier --local-ip) matches exactly what peers
	# registered for us in the router/barrier; re-deriving it here can pick a
	# different NIC on multi-homed nodes and desync the two.
	IFS=',' read -ra IP_ARRAY <<< "$IPADDRS"
	host_ip="${IP_ARRAY[NODE_RANK]:-$(hostname -I | awk '{print $1}')}"
	host_name=$(hostname)
	unset _SAVED_NCCL_SOCKET_IFNAME

Note: the Cluster Topology section further down re-parses IP_ARRAY from IPADDRS; with the read moved up that later read becomes redundant (harmless) and can be dropped in a follow-up.

Validated the equivalent change on an AINIC/ionic node.

[mkuznet1]

To fully match the validated test commit, one more spot: after the suggestion above lands, the Cluster Topology section still parses the array again:

-IFS=',' read -ra IP_ARRAY <<< "$IPADDRS"
+# IP_ARRAY was already parsed from IPADDRS above (where host_ip is derived).

The suggestion above introduces an earlier IFS=',' read -ra IP_ARRAY <<< "$IPADDRS" next to host_ip, so this later read becomes redundant (harmless, but dead). The test commit replaces it with the one-line breadcrumb above.

I couldn't post this as a one-click suggestion: that line isn't part of this PR's diff (it's unchanged, and only turns redundant once the host_ip change is applied), so GitHub rejects a review comment anchored there.