Files
gpu-turnstile/internal/comfy/client.go
T
mram 065c294a96 gpulock: GPU arbitration proxy for Ollama + ComfyUI
Implements SPEC.md: two listeners, one writer-preferring two-mode lock,
Ollama unload before image jobs, ComfyUI history polling + VRAM free,
optional model warm-up, healthz/metrics endpoints, streaming-safe
reverse proxies, Dockerfile and Gitea Actions CI.
2026-09-20 18:05:20 +02:00

139 lines
3.5 KiB
Go

// Package comfy is a minimal client for the ComfyUI endpoints gpulock needs
// after a prompt has been accepted: polling /history and freeing VRAM.
package comfy
import (
"bytes"
"context"
"encoding/json"
"fmt"
"io"
"log/slog"
"net/http"
"net/url"
"time"
)
// Client talks to a ComfyUI upstream.
type Client struct {
base string
hc *http.Client
log *slog.Logger
// PollInterval is how often /history/<id> is polled. Defaults to 1s;
// tests can shorten it.
PollInterval time.Duration
}
// New validates baseURL and returns a Client.
func New(baseURL string, log *slog.Logger) (*Client, error) {
u, err := url.Parse(baseURL)
if err != nil || u.Scheme == "" || u.Host == "" {
return nil, fmt.Errorf("invalid ComfyUI URL %q", baseURL)
}
return &Client{
base: u.String(),
hc: &http.Client{Timeout: 30 * time.Second},
log: log,
PollInterval: time.Second,
}, nil
}
// Probe checks that ComfyUI answers on /system_stats.
func (c *Client) Probe(ctx context.Context) error {
req, err := http.NewRequestWithContext(ctx, http.MethodGet, c.base+"/system_stats", nil)
if err != nil {
return err
}
resp, err := c.hc.Do(req)
if err != nil {
return err
}
defer resp.Body.Close()
io.Copy(io.Discard, resp.Body)
if resp.StatusCode/100 != 2 {
return fmt.Errorf("/system_stats returned %s", resp.Status)
}
return nil
}
type historyEntry struct {
Status struct {
Completed bool `json:"completed"`
StatusStr string `json:"status_str"`
} `json:"status"`
}
// jobDone reports whether the history entry for promptID shows a finished
// job (completed or errored).
func (c *Client) jobDone(ctx context.Context, promptID string) (bool, error) {
req, err := http.NewRequestWithContext(ctx, http.MethodGet, c.base+"/history/"+promptID, nil)
if err != nil {
return false, err
}
resp, err := c.hc.Do(req)
if err != nil {
return false, err
}
defer resp.Body.Close()
if resp.StatusCode/100 != 2 {
io.Copy(io.Discard, resp.Body)
return false, fmt.Errorf("/history/%s returned %s", promptID, resp.Status)
}
var history map[string]historyEntry
if err := json.NewDecoder(resp.Body).Decode(&history); err != nil {
return false, err
}
entry, ok := history[promptID]
if !ok {
return false, nil
}
return entry.Status.Completed || entry.Status.StatusStr == "error", nil
}
// WaitJob polls until the job for promptID is finished. The context should
// carry the job deadline; transient polling errors are logged and retried,
// so a returned error is always from the context.
func (c *Client) WaitJob(ctx context.Context, promptID string) error {
for {
done, err := c.jobDone(ctx, promptID)
if err != nil {
if ctx.Err() != nil {
return ctx.Err()
}
c.log.Warn("failed to poll job history", "prompt_id", promptID, "err", err)
}
if done {
return nil
}
select {
case <-ctx.Done():
return ctx.Err()
case <-time.After(c.PollInterval):
}
}
}
// Free asks ComfyUI to unload its models and free VRAM.
func (c *Client) Free(ctx context.Context) error {
buf, err := json.Marshal(map[string]bool{"unload_models": true, "free_memory": true})
if err != nil {
return err
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.base+"/free", bytes.NewReader(buf))
if err != nil {
return err
}
req.Header.Set("Content-Type", "application/json")
resp, err := c.hc.Do(req)
if err != nil {
return err
}
defer resp.Body.Close()
io.Copy(io.Discard, resp.Body)
if resp.StatusCode/100 != 2 {
return fmt.Errorf("/free returned %s", resp.Status)
}
return nil
}