Files
llmgw/internal/agent/engines/vllm.go
T
Ilker ManapandClaude Opus 5.5 2a1e0e6c4c LoRA adapters on vLLM, loaded on demand with replica affinity: max_loaded_adapters
A vLLM model can carry PEFT adapters, registered at start, or with
max_loaded_adapters held at most N per replica: the router loads an adapter
through vLLM's runtime LoRA API when a request first wants it, sends the
adapter's requests to a replica that holds it while that one is not
saturated, loads each cold adapter once, and unloads the least recently used
idle one to make room. llama.cpp keeps loading every adapter at start.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-05 01:14:38 +02:00

202 lines
6.7 KiB
Go

package engines
import (
"fmt"
"strconv"
"sync/atomic"
)
// VLLMConfig locates the vLLM installation and picks how it is launched.
type VLLMConfig struct {
// Bin is the vllm executable.
Bin string `koanf:"bin"`
// LaunchMode selects the multi-node mechanism: "mp", the default, or
// "ray".
//
// mp is what vLLM documents today — every node runs serve with its rank,
// the workers headless — and needs nothing but vLLM. ray is the older way,
// kept for a vLLM too old for mp's multi-node flags; it needs Ray on every
// node and has not been checked against a real cluster.
LaunchMode string `koanf:"launch_mode"`
// RayBin is used when LaunchMode is "ray".
RayBin string `koanf:"ray_bin"`
}
// VLLM builds vLLM launch commands.
type VLLM struct {
cfg VLLMConfig
installed atomic.Pointer[string]
}
// UseInstalled points the engine at the environment the agent installed and
// checked. It takes precedence over the configuration, as llama.cpp's does:
// it is the one seen to run. An operator who wants their own sets the
// control plane's engines.vllm to "".
func (v *VLLM) UseInstalled(path string) { v.installed.Store(&path) }
// bin is the vllm to run: the installed one, else the configured one, else
// whatever PATH has.
func (v *VLLM) bin() string {
if p := v.installed.Load(); p != nil {
return *p
}
if v.cfg.Bin != "" {
return v.cfg.Bin
}
return "vllm"
}
// Name identifies the engine in the catalog.
func (*VLLM) Name() string { return "vllm" }
// HealthPath is vLLM's readiness endpoint.
func (*VLLM) HealthPath() string { return "/health" }
// MetricsPath is vLLM's Prometheus endpoint.
func (*VLLM) MetricsPath() string { return "/metrics" }
// Validate checks the profile against the GPUs it would run on.
func (*VLLM) Validate(p Profile, gpus []GPUInfo) error {
if err := validateCommon(p, gpus); err != nil {
return err
}
// FP8 needs Hopper or Ada; on anything older vLLM either refuses or falls
// back silently to a slower path, and a silent fallback is worse.
if q := quantizationOf(p.Args); q == "fp8" {
if err := requireComputeCapability(p, gpus, 8.9, "fp8 quantization"); err != nil {
return err
}
}
return nil
}
// ServesHTTP reports whether this rank answers the API.
//
// Under Ray a worker runs nothing but the Ray join, and under multiprocessing
// only the driver binds the port. Either way it is the leader alone.
func (*VLLM) ServesHTTP(s ShardSpec) bool { return s.Leader() }
// Commands returns the processes for one shard, in start order.
//
// Single node, or multi-node under multiprocessing: one server process. Under
// Ray: the leader starts a head and then the server, a worker starts only the
// join. A Ray worker running no vLLM process at all is correct — the cluster
// gives the leader its GPUs.
func (v *VLLM) Commands(s ShardSpec) ([]Command, error) {
ray, useRay := v.rayCommand(s)
if useRay && !s.Leader() {
// A worker is the Ray join and nothing else. It blocks, so its exit is
// the shard's exit, which is what the supervisor needs.
return []Command{ray}, nil
}
server, err := v.serverCommand(s)
if err != nil {
return nil, err
}
if useRay {
return []Command{ray, server}, nil
}
return []Command{server}, nil
}
// serverCommand produces the vLLM server process.
func (v *VLLM) serverCommand(s ShardSpec) (Command, error) {
bin := v.bin()
p := s.Profile
args := []string{"serve", p.ModelPath,
"--host", s.Host(),
"--port", strconv.Itoa(s.HTTPPort),
"--tensor-parallel-size", strconv.Itoa(p.TP),
}
if p.PP > 1 {
args = append(args, "--pipeline-parallel-size", strconv.Itoa(p.PP))
}
if p.ServedModelName != "" {
args = append(args, "--served-model-name", p.ServedModelName)
}
env := baseEnv(s)
if s.NNodes > 1 {
switch v.cfg.LaunchMode {
case "ray":
// Ray is started separately by the supervisor; vLLM then finds the
// cluster through the address in the environment.
env["VLLM_HOST_IP"] = s.LeaderAddr
env["RAY_ADDRESS"] = fmt.Sprintf("%s:%d", s.LeaderAddr, s.LeaderPort)
args = append(args, "--distributed-executor-backend", "ray")
case "", "mp":
// vLLM's own multi-node launch: every node runs the same serve
// with its rank and the leader's address, and the workers run
// headless — no API server, only their GPUs. VLLM_HOST_IP is
// this node's address for vLLM to find itself, so it is left
// to vLLM; the leader's, which workers used to be given, made
// every worker announce itself as the leader.
args = append(args, "--distributed-executor-backend", "mp",
"--nnodes", strconv.Itoa(s.NNodes),
"--node-rank", strconv.Itoa(s.Rank),
"--master-addr", s.LeaderAddr,
"--master-port", strconv.Itoa(s.LeaderPort))
if !s.Leader() {
args = append(args, "--headless")
}
default:
return Command{}, fmt.Errorf("vllm: unknown launch_mode %q (want \"ray\" or \"mp\")", v.cfg.LaunchMode)
}
}
// Adapters loaded at start are registered under their names, and a
// request names one as its model. Adapters loaded on demand start
// unregistered: the router loads each when a request first wants it
// here, by the link the agent made (RuntimeLoRADir). --max-loras is
// vLLM's default of one at start — raise it in the profile's arguments
// to mix more adapters in a batch — and the cap on demand, which
// WithRuntimeLoRA put in the arguments.
if len(p.Adapters) > 0 {
args = append(args, "--enable-lora")
if !p.RuntimeLoRA() {
args = append(args, loraModules(p.Adapters)...)
}
}
// Profile arguments last, so an operator can override anything above.
args = append(args, p.Args...)
if s.APIKey != "" {
env["VLLM_API_KEY"] = s.APIKey
}
return Command{Name: "vllm", Path: bin, Args: args, Env: materializeEnv(env)}, nil
}
// rayCommand returns the command that starts or joins the Ray cluster, or
// ok=false when the launch mode does not use Ray.
func (v *VLLM) rayCommand(s ShardSpec) (Command, bool) {
if s.NNodes <= 1 || v.cfg.LaunchMode != "ray" {
return Command{}, false
}
bin := v.cfg.RayBin
if bin == "" {
bin = "ray"
}
args := []string{"start", "--block"}
if s.Leader() {
// Ray's head has no authentication: whoever reaches its port can run
// code on the node. Its dashboard, and the job API behind it, is the
// easiest way in, and nothing here uses it.
args = append(args, "--head", "--port", strconv.Itoa(s.LeaderPort), "--include-dashboard=false")
} else {
args = append(args, "--address", fmt.Sprintf("%s:%d", s.LeaderAddr, s.LeaderPort))
}
args = append(args, "--num-gpus", strconv.Itoa(len(s.GPUIndices)))
return Command{
Name: "ray", Path: bin, Args: args, Env: materializeEnv(baseEnv(s)),
// The head must be listening before anything else happens: a worker
// cannot join, and the leader's server cannot claim GPUs, until it is.
WaitForLeaderPort: true,
}, true
}