Two bugs found while verifying mTLS against real machines. session cancelled its own derived context and then asked that context whether the agent was shutting down. The answer was always yes, so every stream error read as a clean close: the log said err=<nil>, the backoff reset, and an agent whose certificate the control plane refused hammered it forever at the base interval with nothing to diagnose. It now keeps the parent context to answer that question, and a refusal reaches the operator naming both the certificate and the claim. Cancelling the context also did not stop the agent. HTTP/2 watches the context while the request body is being written, and a bidirectional stream spends its life blocked on the read side instead, so the read never returned. Measured on a deployed node: `systemctl stop llmgw-agent` took 90 seconds and ended in SIGKILL, on every restart. Closing the response when the context ends is what actually unblocks it. TestAgentStopsWhenContextEnds hangs reliably without the second fix, which is the only reason to trust that it tests it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
193 lines
5.6 KiB
Go
193 lines
5.6 KiB
Go
package cli
|
|
|
|
import (
|
|
"context"
|
|
"encoding/json"
|
|
"errors"
|
|
"fmt"
|
|
"os"
|
|
"os/signal"
|
|
"syscall"
|
|
"text/tabwriter"
|
|
"time"
|
|
|
|
"github.com/spf13/cobra"
|
|
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/agent"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/config"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/version"
|
|
)
|
|
|
|
func newAgentCmd() *cobra.Command {
|
|
var configPath string
|
|
|
|
var (
|
|
server string
|
|
nodeID string
|
|
mgmtAddr string
|
|
)
|
|
|
|
cmd := &cobra.Command{
|
|
Use: "agent",
|
|
Short: "Node agent: discovery and engine supervision",
|
|
Long: "Runs on every GPU node. It discovers the hardware, connects to the control\n" +
|
|
"plane, and starts, supervises and stops inference engines on demand.\n\n" +
|
|
"The agent dials out and keeps one stream open. Engines keep serving if the\n" +
|
|
"agent restarts or the control plane goes away: the router talks to them\n" +
|
|
"directly, and orphan adoption picks them up again.",
|
|
RunE: func(cmd *cobra.Command, _ []string) error {
|
|
log := logger()
|
|
|
|
cfg, err := config.LoadAgent(configPath)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
if server != "" {
|
|
cfg.Server = server
|
|
}
|
|
if nodeID != "" {
|
|
cfg.NodeID = nodeID
|
|
}
|
|
if mgmtAddr != "" {
|
|
cfg.MgmtAddr = mgmtAddr
|
|
}
|
|
if cfg.NodeID == "" {
|
|
host, err := os.Hostname()
|
|
if err != nil {
|
|
return fmt.Errorf("node_id is unset and the hostname is unavailable: %w", err)
|
|
}
|
|
cfg.NodeID = host
|
|
}
|
|
if cfg.Hostname == "" {
|
|
cfg.Hostname, _ = os.Hostname()
|
|
}
|
|
cfg.Version = version.Version
|
|
|
|
httpClient, err := agent.Client(cfg.TLS, 30*time.Second)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
a, err := agent.New(cfg, httpClient, log)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
ctx, stop := signal.NotifyContext(cmd.Context(), syscall.SIGINT, syscall.SIGTERM)
|
|
defer stop()
|
|
|
|
log.Info("agent starting",
|
|
"node", cfg.NodeID, "server", cfg.Server,
|
|
"fabrics", len(cfg.Fabrics), "nvml", discovery.NVMLSupported())
|
|
|
|
err = a.Run(ctx)
|
|
if errors.Is(err, context.Canceled) {
|
|
return nil
|
|
}
|
|
return err
|
|
},
|
|
}
|
|
cmd.Flags().StringVar(&configPath, "config", "", "configuration file")
|
|
cmd.Flags().StringVar(&server, "server", "", "control plane address (overrides the config)")
|
|
cmd.Flags().StringVar(&nodeID, "node-id", "", "node id (defaults to the hostname)")
|
|
cmd.Flags().StringVar(&mgmtAddr, "mgmt-addr", "", "address engines are reachable on")
|
|
cmd.AddCommand(newInventoryCmd())
|
|
return cmd
|
|
}
|
|
|
|
func newInventoryCmd() *cobra.Command {
|
|
var (
|
|
asJSON bool
|
|
smiBin string
|
|
ibRoot string
|
|
skipGPU bool
|
|
useNVML bool
|
|
)
|
|
|
|
cmd := &cobra.Command{
|
|
Use: "inventory",
|
|
Short: "Show this node's GPUs and InfiniBand adapters",
|
|
Long: "Reports what the scheduler would see on this node: GPU classes, health,\n" +
|
|
"and the fabric each adapter is attached to.\n\n" +
|
|
"Run it before adding a node to the cluster — it is the fastest way to find\n" +
|
|
"a card another process is already using, or a port with no subnet manager.",
|
|
RunE: func(cmd *cobra.Command, _ []string) error {
|
|
inv, err := discovery.Discover(cmd.Context(), discovery.Options{
|
|
NvidiaSMI: smiBin,
|
|
IBRoot: ibRoot,
|
|
PreferNVML: useNVML,
|
|
SkipGPUs: skipGPU,
|
|
}, logger())
|
|
if err != nil {
|
|
// Not fatal: a node whose GPUs cannot be read still has IB
|
|
// worth showing, and the error is itself the useful output.
|
|
fmt.Fprintf(os.Stderr, "%v\n", err)
|
|
inv.Source = "unavailable"
|
|
}
|
|
|
|
if asJSON {
|
|
enc := json.NewEncoder(os.Stdout)
|
|
enc.SetIndent("", " ")
|
|
return enc.Encode(inv)
|
|
}
|
|
printInventory(inv)
|
|
return nil
|
|
},
|
|
}
|
|
|
|
f := cmd.Flags()
|
|
f.BoolVar(&asJSON, "json", false, "emit JSON")
|
|
f.StringVar(&smiBin, "nvidia-smi", "", "path to nvidia-smi")
|
|
f.StringVar(&ibRoot, "ib-root", discovery.SysfsRoot, "sysfs InfiniBand directory")
|
|
f.BoolVar(&skipGPU, "no-gpu", false, "skip GPU discovery")
|
|
f.BoolVar(&useNVML, "nvml", true, "prefer the NVML backend when this build has one (it is the only source of NVLink topology)")
|
|
|
|
return cmd
|
|
}
|
|
|
|
func printInventory(inv discovery.Inventory) {
|
|
w := tabwriter.NewWriter(os.Stdout, 0, 0, 2, ' ', 0)
|
|
|
|
fmt.Fprintf(w, "GPUs (%d, via %s)\n", len(inv.GPUs), inv.Source)
|
|
if len(inv.GPUs) == 0 {
|
|
fmt.Fprintln(w, " none")
|
|
} else {
|
|
fmt.Fprintln(w, " IDX\tCLASS\tNAME\tMEM\tCC\tHEALTH")
|
|
for _, g := range inv.GPUs {
|
|
health := "ok"
|
|
if !g.Health.OK() {
|
|
health = g.Health.Problems[0]
|
|
}
|
|
fmt.Fprintf(w, " %d\t%s\t%s\t%d MiB\t%s\t%s\n",
|
|
g.Index, g.Class(), g.Name, g.MemoryTotalMB, g.ComputeCapability, health)
|
|
}
|
|
fmt.Fprintf(w, " classes: %v\n", inv.Classes())
|
|
if !discovery.NVMLSupported() {
|
|
fmt.Fprintln(w, " NVLink topology unavailable: rebuild with -tags nvml.")
|
|
fmt.Fprintln(w, " Without it, placement cannot keep a tensor-parallel group inside")
|
|
fmt.Fprintln(w, " one NVLink domain (plan §6.2).")
|
|
}
|
|
}
|
|
|
|
fmt.Fprintf(w, "\nInfiniBand adapters (%d)\n", len(inv.HCAs))
|
|
if len(inv.HCAs) == 0 {
|
|
// Worth stating plainly: such a node is its own single-node island and
|
|
// can never host a multi-node replica.
|
|
fmt.Fprintln(w, " none — this node cannot host multi-node replicas")
|
|
} else {
|
|
fmt.Fprintln(w, " HCA\tPORT\tSTATE\tRATE\tLINK\tGID PREFIX\tNUMA")
|
|
for _, h := range inv.HCAs {
|
|
for _, p := range h.Ports {
|
|
fmt.Fprintf(w, " %s\t%d\t%s\t%d Gb/s\t%s\t%s\t%d\n",
|
|
h.Name, p.Num, p.State, p.RateGbps, p.LinkLayer, p.GIDPrefix, h.NUMANode)
|
|
}
|
|
}
|
|
fmt.Fprintln(w, "\n Fabric identity is not derived from the GID prefix by default:")
|
|
fmt.Fprintln(w, " OpenSM's default is fe80:: on every fabric, so a shared prefix proves")
|
|
fmt.Fprintln(w, " nothing. Label fabrics explicitly in agent.yaml (plan §5.2).")
|
|
}
|
|
|
|
_ = w.Flush()
|
|
}
|