CI / test (push) Successful in 1m39s
The stale-profile flag could not fire where it matters. The control plane keeps no record of a running replica across its own restart — it rebuilds them from what the agents report — and adoption left the fingerprint empty, so every replica read as unknown afterwards. That is precisely the moment the catalog is most likely to have been edited: change the configuration, restart the server, and nothing would ever say the fleet disagrees with it. The fingerprint now travels with StartShard, the agent keeps it beside the pool and fabric it already echoes back, and adoption takes the replica's version from the agent rather than from the catalog it is about to be compared against. An agent too old to report it sends nothing, which still reads as unknown rather than as matching. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
259 lines
8.3 KiB
Go
259 lines
8.3 KiB
Go
// Package agentsvc is the control plane's end of the agent stream (plan §5.6).
|
|
package agentsvc
|
|
|
|
import (
|
|
"fmt"
|
|
"time"
|
|
|
|
llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/agent/engines"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/agent/supervisor"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control/reconciler"
|
|
)
|
|
|
|
// NodeFromRegister converts a registration into an inventory node.
|
|
func NodeFromRegister(r *llmgwv1.Register) inventory.Node {
|
|
n := inventory.Node{
|
|
ID: r.GetNodeId(),
|
|
Hostname: r.GetHostname(),
|
|
MgmtAddr: r.GetMgmtAddr(),
|
|
Labels: r.GetLabels(),
|
|
State: inventory.Ready,
|
|
AgentVersion: r.GetAgentVersion(),
|
|
EngineVersions: r.GetEngineVersions(),
|
|
MemoryTotalMB: int(r.GetMemoryTotalMb()),
|
|
}
|
|
for _, w := range r.GetWeights() {
|
|
n.Weights = append(n.Weights, discovery.LocalWeights{
|
|
Path: w.GetPath(), SizeBytes: w.GetSizeBytes(), Shared: w.GetShared(),
|
|
})
|
|
}
|
|
for _, s := range r.GetStorage() {
|
|
n.Storage = append(n.Storage, discovery.Storage{
|
|
Path: s.GetPath(),
|
|
FreeBytes: s.GetFreeBytes(),
|
|
TotalBytes: s.GetTotalBytes(),
|
|
Shared: s.GetShared(),
|
|
})
|
|
}
|
|
for _, g := range r.GetGpus() {
|
|
n.GPUs = append(n.GPUs, discovery.GPU{
|
|
Index: int(g.GetIndex()),
|
|
UUID: g.GetUuid(),
|
|
Name: g.GetName(),
|
|
MemoryTotalMB: int(g.GetMemoryTotalMb()),
|
|
MemoryUsedMB: int(g.GetMemoryUsedMb()),
|
|
ComputeCapability: g.GetComputeCapability(),
|
|
NUMANode: int(g.GetNumaNode()),
|
|
PCIBusID: g.GetPciBusId(),
|
|
DriverVersion: g.GetDriverVersion(),
|
|
UtilizationPercent: int(g.GetUtilizationPercent()),
|
|
CUDAVersion: g.GetCudaVersion(),
|
|
NVLinkPeers: intSlice(g.GetNvlinkPeers()),
|
|
Health: discovery.GPUHealth{
|
|
ECCErrors: int(g.GetEccErrors()),
|
|
TemperatureC: int(g.GetTemperatureC()),
|
|
Problems: g.GetProblems(),
|
|
},
|
|
})
|
|
}
|
|
for _, f := range r.GetFabrics() {
|
|
n.Fabrics = append(n.Fabrics, inventory.FabricAttachment{
|
|
Fabric: f.GetFabric(),
|
|
HCAs: f.GetHcas(),
|
|
RateGbps: int(f.GetRateGbps()),
|
|
})
|
|
}
|
|
return n
|
|
}
|
|
|
|
// RegisterFromInventory builds the registration an agent sends.
|
|
func RegisterFromInventory(nodeID, hostname, mgmtAddr, agentVersion string,
|
|
labels map[string]string, inv discovery.Inventory, fabrics []inventory.FabricAttachment,
|
|
) *llmgwv1.Register {
|
|
r := &llmgwv1.Register{
|
|
NodeId: nodeID, Hostname: hostname, MgmtAddr: mgmtAddr,
|
|
Labels: labels, AgentVersion: agentVersion,
|
|
MemoryTotalMb: int64(inv.MemoryTotalMB),
|
|
}
|
|
for _, w := range inv.Weights {
|
|
r.Weights = append(r.Weights, &llmgwv1.LocalWeights{
|
|
Path: w.Path, SizeBytes: w.SizeBytes, Shared: w.Shared,
|
|
})
|
|
}
|
|
for _, s := range inv.Storage {
|
|
r.Storage = append(r.Storage, &llmgwv1.Storage{
|
|
Path: s.Path,
|
|
FreeBytes: s.FreeBytes,
|
|
TotalBytes: s.TotalBytes,
|
|
Shared: s.Shared,
|
|
})
|
|
}
|
|
for _, g := range inv.GPUs {
|
|
r.Gpus = append(r.Gpus, &llmgwv1.GPU{
|
|
Index: int32(g.Index),
|
|
Uuid: g.UUID,
|
|
Name: g.Name,
|
|
MemoryTotalMb: int32(g.MemoryTotalMB),
|
|
MemoryUsedMb: int32(g.MemoryUsedMB),
|
|
ComputeCapability: g.ComputeCapability,
|
|
NumaNode: int32(g.NUMANode),
|
|
PciBusId: g.PCIBusID,
|
|
DriverVersion: g.DriverVersion,
|
|
UtilizationPercent: int32(g.UtilizationPercent),
|
|
CudaVersion: g.CUDAVersion,
|
|
NvlinkPeers: int32Slice(g.NVLinkPeers),
|
|
Problems: g.Health.Problems,
|
|
TemperatureC: int32(g.Health.TemperatureC),
|
|
EccErrors: int32(g.Health.ECCErrors),
|
|
})
|
|
}
|
|
for _, f := range fabrics {
|
|
r.Fabrics = append(r.Fabrics, &llmgwv1.FabricAttachment{
|
|
Fabric: f.Fabric, Hcas: f.HCAs, RateGbps: int32(f.RateGbps),
|
|
})
|
|
}
|
|
return r
|
|
}
|
|
|
|
// AdoptedFromRegister converts an agent's running-shard report into the
|
|
// reconciler's view.
|
|
func AdoptedFromRegister(nodeID string, shards []*llmgwv1.RunningShard) []reconciler.AdoptedShard {
|
|
out := make([]reconciler.AdoptedShard, 0, len(shards))
|
|
for _, s := range shards {
|
|
a := reconciler.AdoptedShard{
|
|
NodeID: nodeID,
|
|
ReplicaID: s.GetReplicaId(),
|
|
ShardID: s.GetShardId(),
|
|
Rank: int(s.GetRank()),
|
|
GPUIndices: intSlice(s.GetGpuIndices()),
|
|
HCAs: s.GetHcas(),
|
|
HTTPPort: int(s.GetHttpPort()),
|
|
State: StateFromWire(s.GetState()),
|
|
Model: s.GetModel(),
|
|
Profile: s.GetProfile(),
|
|
Pool: s.GetPool(),
|
|
Fabric: s.GetFabric(),
|
|
ProfileFP: s.GetProfileFp(),
|
|
}
|
|
if a.Rank == 0 && s.GetLeaderAddr() != "" {
|
|
a.Endpoint = fmt.Sprintf("http://%s:%d", s.GetLeaderAddr(), s.GetHttpPort())
|
|
}
|
|
out = append(out, a)
|
|
}
|
|
return out
|
|
}
|
|
|
|
// StartShardFromCommand converts a reconciler command into a wire message.
|
|
func StartShardFromCommand(cmd reconciler.ShardCommand) *llmgwv1.StartShard {
|
|
p := cmd.Spec.Profile
|
|
return &llmgwv1.StartShard{
|
|
ReplicaId: cmd.ReplicaID,
|
|
ShardId: cmd.ShardID,
|
|
Rank: int32(cmd.Spec.Rank),
|
|
Nnodes: int32(cmd.Spec.NNodes),
|
|
LeaderAddr: cmd.Spec.LeaderAddr,
|
|
LeaderPort: int32(cmd.Spec.LeaderPort),
|
|
GpuIndices: int32Slice(cmd.Spec.GPUIndices),
|
|
Hcas: cmd.Spec.HCAs,
|
|
MgmtInterface: cmd.Spec.MgmtInterface,
|
|
HttpPort: int32(cmd.Spec.HTTPPort),
|
|
Profile: &llmgwv1.EngineProfile{
|
|
Name: p.Name, Engine: p.Engine, GpuClass: p.GPUClass,
|
|
Tp: int32(p.TP), Pp: int32(p.PP),
|
|
GpusPerNode: int32(p.GPUsPerNode), Nodes: int32(p.Nodes),
|
|
Args: p.Args, Env: p.Env,
|
|
ModelPath: p.ModelPath, ServedModelName: p.ServedModelName,
|
|
},
|
|
StartupTimeoutMs: cmd.StartupTimeout.Milliseconds(),
|
|
Pool: cmd.Pool,
|
|
Fabric: cmd.Fabric,
|
|
ProfileFp: cmd.ProfileFP,
|
|
}
|
|
}
|
|
|
|
// ShardSpecFromStart converts a wire message into an engine shard spec.
|
|
func ShardSpecFromStart(m *llmgwv1.StartShard) (engines.ShardSpec, time.Duration) {
|
|
p := m.GetProfile()
|
|
spec := engines.ShardSpec{
|
|
ReplicaID: m.GetReplicaId(),
|
|
ShardID: m.GetShardId(),
|
|
Rank: int(m.GetRank()),
|
|
NNodes: int(m.GetNnodes()),
|
|
LeaderAddr: m.GetLeaderAddr(),
|
|
LeaderPort: int(m.GetLeaderPort()),
|
|
GPUIndices: intSlice(m.GetGpuIndices()),
|
|
HCAs: m.GetHcas(),
|
|
MgmtInterface: m.GetMgmtInterface(),
|
|
HTTPPort: int(m.GetHttpPort()),
|
|
Pool: m.GetPool(),
|
|
Fabric: m.GetFabric(),
|
|
ProfileFP: m.GetProfileFp(),
|
|
Profile: engines.Profile{
|
|
Name: p.GetName(), Engine: p.GetEngine(), GPUClass: p.GetGpuClass(),
|
|
TP: int(p.GetTp()), PP: int(p.GetPp()),
|
|
GPUsPerNode: int(p.GetGpusPerNode()), Nodes: int(p.GetNodes()),
|
|
Args: p.GetArgs(), Env: p.GetEnv(),
|
|
ModelPath: p.GetModelPath(), ServedModelName: p.GetServedModelName(),
|
|
},
|
|
}
|
|
return spec, time.Duration(m.GetStartupTimeoutMs()) * time.Millisecond
|
|
}
|
|
|
|
// stateWire maps supervisor states onto the wire enum.
|
|
var stateWire = map[supervisor.State]llmgwv1.ShardStatus_State{
|
|
supervisor.Pending: llmgwv1.ShardStatus_STATE_PENDING,
|
|
supervisor.Starting: llmgwv1.ShardStatus_STATE_STARTING,
|
|
supervisor.Loading: llmgwv1.ShardStatus_STATE_LOADING,
|
|
supervisor.Ready: llmgwv1.ShardStatus_STATE_READY,
|
|
supervisor.Draining: llmgwv1.ShardStatus_STATE_DRAINING,
|
|
supervisor.Stopped: llmgwv1.ShardStatus_STATE_STOPPED,
|
|
supervisor.Failed: llmgwv1.ShardStatus_STATE_FAILED,
|
|
}
|
|
|
|
// StateToWire converts a supervisor state for transmission.
|
|
func StateToWire(s supervisor.State) llmgwv1.ShardStatus_State {
|
|
if w, ok := stateWire[s]; ok {
|
|
return w
|
|
}
|
|
return llmgwv1.ShardStatus_STATE_UNSPECIFIED
|
|
}
|
|
|
|
// StateFromWire converts a received state.
|
|
//
|
|
// An unrecognised state is treated as Failed rather than guessed at: a newer
|
|
// agent reporting something this control plane does not understand must not be
|
|
// mistaken for a healthy shard.
|
|
func StateFromWire(w llmgwv1.ShardStatus_State) supervisor.State {
|
|
for s, v := range stateWire {
|
|
if v == w {
|
|
return s
|
|
}
|
|
}
|
|
return supervisor.Failed
|
|
}
|
|
|
|
func intSlice(v []int32) []int {
|
|
if len(v) == 0 {
|
|
return nil
|
|
}
|
|
out := make([]int, len(v))
|
|
for i, x := range v {
|
|
out[i] = int(x)
|
|
}
|
|
return out
|
|
}
|
|
|
|
func int32Slice(v []int) []int32 {
|
|
if len(v) == 0 {
|
|
return nil
|
|
}
|
|
out := make([]int32, len(v))
|
|
for i, x := range v {
|
|
out[i] = int32(x)
|
|
}
|
|
return out
|
|
}
|