Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

trellis: run_parts (engine runs that survive a failed segment)

Niko1221b73bd26a5ccd
diff --git a/tools/trellis/run_parts.py b/tools/trellis/run_parts.py
new file mode 100644
index 000000000..c1014acbd
--- /dev/null
+++ b/tools/trellis/run_parts.py
@@ -0,0 +1,94 @@
+"""run_parts.py - tools/quant_quality.py run in parts that survive an engine failure.
+
+The long runs of this experiment met two kinds of "verify: timed out" (the engine's 20 s wait for the GPU) while the box
+was shared; one failed segment must not cost the whole run. The ids file is cut into --parts parts, each part is one engine
+run (retried up to --retries times, a failed try is discarded), and the part files are joined at the end into OUT.tsv +
+OUT.tsv.seg.json, which `quant_quality.py compare` and swap_compare.py read as usual. With --dump-dir every part writes its
+STRATA_ACT_DUMP files to <dump-dir>/p<i>/ (only the dumps of finished parts are kept).
+
+ python3 tools/trellis/run_parts.py --runner /data/dsv41-enc/run_qq.sh --cfg CFG --ids IDS --out OUT.tsv --parts 6 \
+ [--env K=V ...] [--dump-dir DIR]
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import pathlib
+import shutil
+import subprocess
+import sys
+import time
+
+
+def kill_engines():
+ out = subprocess.run(["pgrep", "-x", "strata"], capture_output=True, text=True).stdout.split()
+ for p in out:
+ subprocess.run(["kill", "-9", p])
+
+
+def main() -> int:
+ ap = argparse.ArgumentParser()
+ ap.add_argument("--runner", required=True)
+ ap.add_argument("--cfg", required=True)
+ ap.add_argument("--ids", required=True)
+ ap.add_argument("--out", required=True)
+ ap.add_argument("--parts", type=int, default=6)
+ ap.add_argument("--env", nargs="*", default=[])
+ ap.add_argument("--dump-dir", default=None)
+ ap.add_argument("--retries", type=int, default=3)
+ a = ap.parse_args()
+ out = pathlib.Path(a.out)
+ work = out.parent / (out.name + ".parts")
+ work.mkdir(parents=True, exist_ok=True)
+ d = json.load(open(a.ids))
+ S = d["segments"]
+ k = -(-len(S) // a.parts)
+ done = []
+ for i in range(a.parts):
+ seg = S[i * k:(i + 1) * k]
+ if not seg:
+ continue
+ ids = work / ("p%d.ids.json" % i)
+ json.dump(dict(d, segments=seg), open(ids, "w"))
+ tsv = work / ("p%d.tsv" % i)
+ ok = pathlib.Path(str(tsv) + ".seg.json").exists()
+ for attempt in range(a.retries):
+ if ok:
+ break
+ env = list(a.env)
+ if a.dump_dir:
+ dd = pathlib.Path(a.dump_dir) / ("p%d" % i)
+ shutil.rmtree(dd, ignore_errors=True)
+ dd.mkdir(parents=True)
+ env += ["STRATA_ACT_DUMP=%s/act" % dd]
+ t0 = time.time()
+ r = subprocess.run([a.runner, a.cfg, str(ids), str(tsv)] + env, capture_output=True, text=True)
+ ok = pathlib.Path(str(tsv) + ".seg.json").exists()
+ print("part %d try %d: %s (%.0f s)%s" % (i, attempt + 1, "ok" if ok else "FAILED", time.time() - t0,
+ "" if ok else " " + (r.stdout or "")[-200:].replace("\n", " ")), flush=True)
+ if not ok:
+ kill_engines()
+ if a.dump_dir:
+ shutil.rmtree(pathlib.Path(a.dump_dir) / ("p%d" % i), ignore_errors=True)
+ if not ok:
+ print("part %d gave up" % i, flush=True)
+ continue
+ done.append(i)
+ # join
+ lines, segs = [], []
+ for i in done:
+ tsv = work / ("p%d.tsv" % i)
+ part = open(tsv).read().splitlines()
+ sj = json.load(open(str(tsv) + ".seg.json"))
+ base = len(lines)
+ lines += part
+ for s in sj["segments"]:
+ segs.append({"name": s["name"], "first_line": base + s["first_line"], "end_line": base + s["end_line"]})
+ out.write_text("\n".join(lines) + "\n")
+ pathlib.Path(str(out) + ".seg.json").write_text(json.dumps({"segments": segs, "parts_done": done, "parts": a.parts}, indent=1))
+ print("joined %d parts, %d segments -> %s" % (len(done), len(segs), out))
+ return 0 if len(done) == a.parts else 1
+
+
+if __name__ == "__main__":
+ sys.exit(main())