Files
llmgw/internal/agent/engines/vllm.go
T
ilkerandClaude Opus 5 6a13e3dddf
CI / test (push) Successful in 45s
Phase 2 (1/2): node discovery, engine adapters, process supervisor
Everything here works without GPUs and is covered by tests. The control plane
connection is not built yet, so llmgw agent itself is still a stub and the phase
exit criterion is not met.

- discovery: IB from /sys/class/infiniband, parsing the formats sysfs actually
  uses and telling native InfiniBand from RoCE; GPUs via nvidia-smi, the
  cgo-free path, with health flags for uncorrected ECC and memory already held
  by another process. A node without IB is not an error, it is a single-node
  island.
- engines: vLLM and SGLang command and environment construction, kept free of
  process handling so it stays testable. The NCCL environment is the part the
  plan marks critical: NCCL_IB_HCA lists only the chosen fabric's adapters,
  because leaving it open lets NCCL pick an adapter on a fabric where the peer
  is unreachable, and the job then hangs instead of failing.
- validation before launch: no mixed GPU types in a tensor-parallel group, the
  node/GPU/TP arithmetic has to add up, and fp8 is refused below compute
  capability 8.9 rather than silently falling back to a slower path.
- supervisor: process groups, because engines fork workers that hold GPU memory
  and outlive a signal sent only to the leader; a bounded log tail, since engine
  output is the only diagnosis when loading fails; Starting separate from
  Loading, because a 70B model legitimately takes minutes; and orphan adoption
  through pid files fingerprinted with the command line, since pids get reused
  and adopting an unrelated process is worse than leaking the original.
- readiness is two questions: /health says the process is up, /v1/models says
  the weights loaded. Routing on /health alone sends traffic into a loading model.
- llmgw agent inventory reports what the scheduler would see on a node.

One plan correction: GPU class sizes come from the driver in GiB, which does not
always match the marketing name. An L40S is sold as 48GB and reports 44.99 GiB,
so its class is l40s-45g, not l40s-48g. Consistency matters more than matching
the box, because that number decides whether a model fits. Plan §2 and §5.1
updated.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-16 02:01:59 +02:00

118 lines
3.6 KiB
Go

package engines
import (
"fmt"
"strconv"
)
// 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: "ray" or "mp".
//
// Which one works depends on the vLLM version, and the flags have changed
// more than once (plan §14). Both paths exist so a cluster can pick the one
// its version supports instead of waiting for a code change.
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 }
// 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
}
// BuildCommand produces the launch command for one shard.
func (v *VLLM) BuildCommand(s ShardSpec) (Command, error) {
bin := v.cfg.Bin
if bin == "" {
bin = "vllm"
}
p := s.Profile
args := []string{"serve", p.ModelPath,
"--host", "0.0.0.0",
"--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":
env["VLLM_HOST_IP"] = s.LeaderAddr
env["VLLM_PORT"] = strconv.Itoa(s.LeaderPort)
args = append(args, "--distributed-executor-backend", "mp")
default:
return Command{}, fmt.Errorf("vllm: unknown launch_mode %q (want \"ray\" or \"mp\")", v.cfg.LaunchMode)
}
}
// Profile arguments last, so an operator can override anything above.
args = append(args, p.Args...)
return Command{Path: bin, Args: args, Env: materializeEnv(env)}, nil
}
// RayCommand returns the command that joins this node to the Ray cluster, or
// ok=false when the launch mode does not use Ray.
//
// It is separate from BuildCommand because it has to run first and finish
// before vLLM starts: the supervisor sequences the two.
func (v *VLLM) RayCommand(s ShardSpec) (Command, bool) {
if s.NNodes <= 1 || (v.cfg.LaunchMode != "" && v.cfg.LaunchMode != "ray") {
return Command{}, false
}
bin := v.cfg.RayBin
if bin == "" {
bin = "ray"
}
args := []string{"start", "--block"}
if s.Leader() {
args = append(args, "--head", "--port", strconv.Itoa(s.LeaderPort))
} 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{Path: bin, Args: args, Env: materializeEnv(baseEnv(s))}, true
}