feat: GPU interconnect multinode support (chain SDK pieces)

docs/proto/provider.md

@@ -299,6 +299,7 @@
  | `ephemeral_storage` | [ResourcePair](#akash.inventory.v1.ResourcePair) |  |  |
  | `volumes_attached` | [ResourcePair](#akash.inventory.v1.ResourcePair) |  |  |
  | `volumes_mounted` | [ResourcePair](#akash.inventory.v1.ResourcePair) |  |  |
+ | `gpu_interconnect` | [ResourcePair](#akash.inventory.v1.ResourcePair) |  | GPUInterconnect reports node GPU-interconnect HCA capacity. Capacity/Allocatable/Allocated are populated by the inventory operator from k8s allocatable for whichever rdma/rdma_shared_device_* extended resource the cluster's device plugin publishes (the `rdma/*` prefix is the device plugin's naming convention; see NodeCapabilities.interconnect_resource_name). |
 
 
 
@@ -347,6 +348,9 @@
  | Field | Type | Label | Description |
  | ----- | ---- | ----- | ----------- |
  | `storage_classes` | [string](#string) | repeated |  |
+ | `interconnect_resource_name` | [string](#string) |  | Kubernetes extended-resource name the cluster's device plugin publishes for GPU interconnect HCAs (e.g. rdma/rdma_shared_device_ib for an InfiniBand fabric, rdma/rdma_shared_device_eth for RoCE). The `rdma/*` prefix is the device-plugin's own convention (Mellanox/NVIDIA) and stays unchanged here. Empty when the node has no GPU interconnect capability. Discovered by the inventory operator from k8s allocatable. |
+ | `interconnect_fabric` | [string](#string) |  | GPU interconnect fabric type. "infiniband" or "roce". Internal / informational — the SDL surface is fabric-agnostic; tenants only declare `interconnect: []` or `interconnect: { group: <name> }`. Derived from /sys/class/infiniband/<dev>/ports/1/link_layer on the host node. |
+ | `nccl_hca_prefixes` | [string](#string) | repeated | NCCL HCA device-name prefixes present on this node, one per distinct family (e.g. ["mlx5"], or ["mlx5","bnxt_re"] on a mixed-vendor host). Same key for IB and RoCE since NCCL uses the IB verbs API for both. Joined with commas and injected as NCCL_IB_HCA when scheduling GPU interconnect workloads — NCCL accepts comma-separated device prefixes natively. Discovered from /sys/class/infiniband/<dev> on the host. |
 
 
 
@@ -598,6 +602,9 @@
  | `expose` | [ServiceExpose](#akash.manifest.v2beta3.ServiceExpose) | repeated |  |
  | `params` | [ServiceParams](#akash.manifest.v2beta3.ServiceParams) |  |  |
  | `credentials` | [ImageCredentials](#akash.manifest.v2beta3.ImageCredentials) |  |  |
+ | `interconnect_group` | [string](#string) |  | InterconnectGroup carries the SDL gpu.attributes.interconnect_group peer-group label. Lifted from Resources.GPU.Attributes by the manifest builder so the off-chain workload builder can label pods for per-group anti-affinity. Services sharing the same value form one NCCL peer group; the provider schedules them on distinct nodes. Empty when the service is not part of any GPU interconnect group.
+
+JSON / YAML tags carry `omitempty`: the on-chain manifest `version` is a SHA hash of the JSON-serialized off-chain manifest, so any field that always serializes (even at zero value) would shift the hash for every non-interconnect SDL and break send-manifest validation on existing leases. |