Skip to content

armgit / kokon

Public Imported from GitHub
0

Edit photos by describing the change, on your own GPU. A free, open-source alternative to Nano Banana and ChatGPT image editing: photos stay on your computer, no daily limit. One installer for Windows, Mac and Linux sets up the best open model (Qwen-Image 2.1). MCP server for AI agents.

test: 8 GB emulation gives sd-server an explicit VRAM budget

Gabriele Lozupone63ffec707652
diff --git a/scripts/dev/enhancer-check.py b/scripts/dev/enhancer-check.py
index bee2087..29d3172 100644
--- a/scripts/dev/enhancer-check.py
+++ b/scripts/dev/enhancer-check.py
@@ -297,12 +297,7 @@ def start_isolated(card_mib, out):
for d in (cdir, ddir, odir):
d.mkdir(parents=True, exist_ok=True)
svc_port, srv_port = free_port(), free_port()
- cfg = dict(real)
- cfg.update({"servicePort": svc_port, "serverPort": srv_port, "outputDir": str(odir),
- "hardwarePreset": "8gb", "vramOverrideMib": card_mib, "setupComplete": True})
- cfg.setdefault("sdcppDir", str(data_dir() / "sdcpp"))
- (cdir / "config.json").write_text(json.dumps(cfg, indent=2), encoding="utf-8")
- sdcpp = cfg["sdcppDir"]
+ sdcpp = real.get("sdcppDir") or str(data_dir() / "sdcpp")
procs = []
blog = open(iso / "ballast.log", "w")
@@ -315,7 +310,19 @@ def start_isolated(card_mib, out):
break
if ballast.poll() is not None or "Holding" not in (iso / "ballast.log").read_text(errors="replace"):
raise SystemExit("VRAM ballast failed:\n" + (iso / "ballast.log").read_text(errors="replace"))
- print((iso / "ballast.log").read_text(errors="replace").splitlines()[0])
+ held = (iso / "ballast.log").read_text(errors="replace").splitlines()[0]
+ print(held)
+ # On Windows sd-server's own free-VRAM query does not see the ballast (it reported ~14.7 of
+ # 16 GiB free beside it), so its automatic budget (--max-vram -1) would overcommit and spill.
+ # Give it what a real card of this size detects: free VRAM now minus the same 1 GiB reserve.
+ free = int(re.search(r"(\d+) MiB free now", held).group(1))
+ budget_gib = round((free - 1024) / 1024, 2)
+ cfg = dict(real)
+ cfg.update({"servicePort": svc_port, "serverPort": srv_port, "outputDir": str(odir), "sdcppDir": sdcpp,
+ "hardwarePreset": "8gb", "vramOverrideMib": card_mib, "setupComplete": True,
+ "vramReserveGib": 0, "extraServerArgs": list(real.get("extraServerArgs") or []) + ["--max-vram", str(budget_gib)]})
+ (cdir / "config.json").write_text(json.dumps(cfg, indent=2), encoding="utf-8")
+ print("sd-server budget: --max-vram %s GiB (%d MiB free with the ballast, minus 1 GiB)" % (budget_gib, free))
env = dict(os.environ, QLS_DATA_DIR=str(ddir), QLS_CONFIG_DIR=str(cdir))
svc = subprocess.Popen([str(exe), "--port", str(svc_port), "--persistent"], env=env,
@@ -482,7 +489,7 @@ def write_report(out, a, sysinfo, release, markers, srv, checks, results):
L = ["# Prompt enhancer check %s" % out.name, ""]
L += ["- System: %s" % ", ".join("%s %s" % (k, v) for k, v in sysinfo.items())]
if a.card_mib:
- L += ["- Emulated card: %d MiB (VRAM ballast + 8gb preset, isolated service)" % a.card_mib]
+ L += ["- Emulated card: %d MiB (VRAM ballast + 8gb preset + explicit --max-vram, isolated service)" % a.card_mib]
L += ["- Engine: %s; markers: %s" % (release, "; ".join(markers) or "none"),
"- Server: model %s, text_generation %s" % (srv.get("model"), srv.get("text_generation")),
"- Quality: %s" % a.quality, ""]