Retry Qwen edits that run out of VRAM with a 5 GB budget

Qwen-Image 2.1 edits with two reference sheets need about 10 GB of VRAM,
more than the 12 GB Arc leaves beside the desktop: prefix caching fails,
the retry without it fails too, and so does disabling caching up front.
Only an explicit --max-vram budget fits, by streaming weights (about
12 minutes per candidate), so retry once with vulkan0=5 when sd-cli
reports it cannot make enough memory available.

sd-cli logs [ERROR for the prefix caching failure it recovers from, so a
run whose images were all saved no longer counts as failed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 10 insertions, 1 deletions

diff --git a/fanfictioner b/fanfictioner
index 9917540..db801bc 100755
--- a/fanfictioner
+++ b/fanfictioner
@@ -436,7 +436,9 @@ def run_sd(args, log, preview=None):
             print()
             if viewer:
                 viewer.terminate()
-    if proc.returncode or "[ERROR" in Path(log).read_text(errors="replace"):
+    out = Path(log).read_text(errors="replace")
+    # sd-cli also logs [ERROR for steps it recovers from (Qwen prefix caching): every image saved wins.
+    if proc.returncode or ("[ERROR" in out and not re.search(r"\b(\d+)/\1 images saved", out)):
         return f"sd-cli failed (exit {proc.returncode})"
     return None
 
@@ -484,6 +486,11 @@ def generate(d, profile, prompt, size, refs=()):
         preview.unlink(missing_ok=True)
         args += ["--preview", "proj", "--preview-path", str(preview)]
     err = run_sd(args, log, preview)
+    if err and "cannot make enough memory" in log.read_text(errors="replace"):
+        # ponytail: fixed 5 GB budget, fits a Qwen edit beside the desktop but streams weights,
+        # ~12 min per candidate; closing VRAM-heavy apps avoids this path.
+        print("out of VRAM, retrying with a 5 GB budget (slow) ...", flush=True)
+        err = run_sd(args + ["--max-vram", "vulkan0=5"], log, preview)
     files = [d / f"{seed}_{i}.png" for i in (1, 2, 3)]
     if not err and not all(f.exists() for f in files):
         err = "sd-cli wrote no output"
@@ -901,6 +908,8 @@ def selftest():
         assert run_sd(["true"], t / "ok.log") is None
         assert (t / "ok.log").read_text().startswith("true")
         assert run_sd(["sh", "-c", "echo '[ERROR] boom'"], t / "e.log") == "sd-cli failed (exit 0)"
+        assert run_sd(["sh", "-c", "echo '[ERROR] boom'; echo '3/3 images saved'"], t / "r.log") is None
+        assert run_sd(["sh", "-c", "echo '[ERROR] boom'; echo '1/3 images saved'"], t / "p.log")
         assert run_sd(["false"], t / "f.log") == "sd-cli failed (exit 1)"
         assert run_sd(["/nonexistent-sd-cli"], t / "n.log").startswith("cannot run /nonexistent-sd-cli")
         global LOG