Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20260928T225702Z-paired-ft-scsemi-delta/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scsemi/tools/make_queues.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+# Write the scSemiProfiler baseline queues.
+#   queues/server1_fewshot.txt : ji2025_chili, giroux2022, ye2026_sepsis, immicro (draws 0-4), run on server_1
+#   queues/server2_fewshot.txt : imperial_eskd (draws 0-4), run on server_2 (its data and outputs live there)
+#   queues/server1_cv.txt      : OPTIONAL 5-fold CV variant (representatives = all training donors of fold k);
+#                                expensive, see README "CV variant cost"; not started by default
+set -euo pipefail
+S=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi
+mkdir -p "$S/queues"
+{
+  echo "# scSemiProfiler paired few-pair baseline (5 representatives' donors), server_1; order fixed by the coordinator"
+  for ds in ji2025_chili giroux2022 ye2026_sepsis immicro; do
+    opt="|clean_sample_sc"   # server_1 disk: dense cell graphs 2-24 GB per draw, re-derivable (bit-identical), sha kept
+    for j in 0 1 2 3 4; do echo "fewshot|$ds|$j$opt"; done
+  done
+} > "$S/queues/server1_fewshot.txt"
+{
+  echo "# scSemiProfiler paired few-pair baseline, server_2 (Imperial data + outputs live on server_2)"
+  for j in 0 1 2 3 4; do echo "fewshot|imperial_eskd|$j"; done
+} > "$S/queues/server2_fewshot.txt"
+{
+  echo "# OPTIONAL scSemiProfiler 5-fold CV variant (all training donors of fold k are representatives). Expensive."
+  for ds in ji2025_chili giroux2022 ye2026_sepsis immicro; do
+    opt="|clean_sample_sc"
+    for k in 0 1 2 3 4; do echo "cv|$ds|$k$opt"; done
+  done
+} > "$S/queues/server1_cv.txt"
+wc -l "$S"/queues/*.txt

paper-releases/pbisc-diffusion/updates/20260928T225702Z-paired-ft-scsemi-delta/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scsemi/tools/prepare_draw_inputs.py

@@ -0,0 +1,226 @@
+#!/usr/bin/env python
+"""Build scSemiProfiler inputs for one paired split (CPU only).
+
+Representatives = every usable row of the split's training donors (paired sc + real bulk).
+Targets        = the split's test rows (real bulk only; their cells are never read here).
+
+  --mode fewshot --draw j : training donors = data/<ds>/fewshot_n5/draw<j>/train_donors.txt,
+                            targets = draw<j>/test_samples.txt, condition = draw<j>/test_raw.npz
+  --mode cv --fold k      : training donors = every usable donor not in fold k (inner-val donors included),
+                            targets = fold k's rows, condition = data/<ds>/fold<k>/test_raw.npz
+
+Modeled genes = final 6k panel genes measured in the bulk of every representative and target row AND in the
+single cells of every representative row (features.npz bulk_valid / sc_valid).  No target single-cell mask,
+depth or cell count enters.
+
+Writes <work>/inputs/:
+  bulk.h5ad        raw bulk counts (float64) of reps + targets on the modeled genes, obs['sample_ids'],
+                   rows in features.npz order; upstream initsetup(normed=False) library-normalises + log1p's it
+  rep_sc.h5ad      raw UMI counts (float32 CSR) of every representative cell on the modeled genes,
+                   obs['sample_ids'] = the representative row's sample id; upstream scprocess(normed=False)
+  inputs_manifest.json   rows, genes, depth rule, source hashes
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+if os.environ.get("CUDA_VISIBLE_DEVICES") != "":
+    raise SystemExit("run with CUDA_VISIBLE_DEVICES=''")
+
+import numpy as np
+import pandas as pd
+import scipy.sparse as sp
+
+TASK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927")
+HERE = TASK / "scsemi"
+sys.path.insert(0, str(TASK / "tools"))
+import prepare_dataset as PD  # noqa: E402  (the benchmark's own loader: bulk_raw with unmeasured NaN -> 0, masks, use)
+
+SEED = 20260927
+
+
+def sha(path: Path) -> str:
+    h = hashlib.sha256()
+    with open(path, "rb") as f:
+        for block in iter(lambda: f.read(1 << 24), b""):
+            h.update(block)
+    return h.hexdigest()
+
+
+def split_rows(ds: str, mode: str, j: int, sample_ids: np.ndarray, folds: pd.DataFrame) -> dict:
+    data = TASK / "data" / ds
+    usable = folds["fold"].to_numpy() >= 0
+    donors = folds["donor_id"].astype(str).to_numpy()
+    if mode == "fewshot":
+        dd = data / "fewshot_n5" / f"draw{j}"
+        train_donors = [x.strip() for x in (dd / "train_donors.txt").read_text().splitlines() if x.strip()]
+        test = [x.strip() for x in (dd / "test_samples.txt").read_text().splitlines() if x.strip()]
+        cond = dd / "test_raw.npz"
+        sources = {"train_donors.txt": dd / "train_donors.txt", "test_samples.txt": dd / "test_samples.txt"}
+        rep = np.flatnonzero(np.isin(donors, train_donors) & usable)
+    elif mode == "cv":
+        fold = folds["fold"].to_numpy()
+        test = sample_ids[fold == j].tolist()
+        rep = np.flatnonzero(usable & (fold != j))
+        train_donors = sorted(set(donors[rep]))
+        cond = data / f"fold{j}" / "test_raw.npz"
+        sources = {}
+    else:
+        raise SystemExit(f"unknown mode {mode}")
+    pos = {s: i for i, s in enumerate(sample_ids)}
+    miss = [s for s in test if s not in pos]
+    if miss:
+        raise SystemExit(f"test samples not in features.npz: {miss[:5]}")
+    tgt = np.asarray([pos[s] for s in test], dtype=np.int64)
+    if not usable[tgt].all():
+        raise SystemExit("a test row is not usable (fold -1)")
+    if set(donors[rep]) & set(donors[tgt]):
+        raise SystemExit("a donor is both representative and target")
+    if len(set(test)) != len(test):
+        raise SystemExit("duplicate test samples")
+    return {"rep": rep, "tgt": tgt, "train_donors": train_donors, "test": test, "condition": cond, "sources": sources}
+
+
+def main(argv=None) -> int:
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--dataset", required=True)
+    ap.add_argument("--mode", choices=["fewshot", "cv"], default="fewshot")
+    ap.add_argument("--draw", type=int, help="fewshot draw index")
+    ap.add_argument("--fold", type=int, help="cv fold index")
+    ap.add_argument("--work", type=Path, required=True)
+    ap.add_argument("--smoke-reps", type=int, default=0, help="SMOKE ONLY: keep the first N representative rows")
+    ap.add_argument("--smoke-cells", type=int, default=0, help="SMOKE ONLY: at most N cells per representative")
+    ap.add_argument("--smoke-targets", default="", help="SMOKE ONLY: comma list of target sample ids")
+    ap.add_argument("--smoke-rep-ids", default="", help="SMOKE ONLY: comma list of representative sample ids to keep")
+    a = ap.parse_args(argv)
+    j = a.draw if a.mode == "fewshot" else a.fold
+    if j is None:
+        raise SystemExit("--draw (fewshot) or --fold (cv) required")
+    smoke = bool(a.smoke_reps or a.smoke_cells or a.smoke_targets or a.smoke_rep_ids)
+    work = a.work.resolve()
+    if smoke and "/scsemi/smoke/" not in str(work) + "/":
+        raise SystemExit("smoke subsetting only under scsemi/smoke/")
+    out = work / "inputs"
+    if (out / "inputs_manifest.json").exists():
+        raise SystemExit(f"{out}/inputs_manifest.json exists; inputs are immutable (preserve and inspect)")
+    out.mkdir(parents=True, exist_ok=True)
+    data = TASK / "data" / a.dataset
+    d = PD.load(data)
+    folds = pd.read_csv(data / "folds.csv")
+    if not np.array_equal(folds["sample_id"].astype(str).to_numpy(), d["sample_ids"]):
+        raise SystemExit("folds.csv rows differ from features.npz")
+    s = split_rows(a.dataset, a.mode, j, d["sample_ids"], folds)
+    rep, tgt = s["rep"], s["tgt"]
+    if a.smoke_rep_ids:
+        keep_r = set(a.smoke_rep_ids.split(","))
+        rep = np.asarray([r for r in rep if d["sample_ids"][r] in keep_r], dtype=np.int64)
+    if a.smoke_reps:
+        rep = rep[: a.smoke_reps]
+    if a.smoke_targets:
+        keep = set(a.smoke_targets.split(","))
+        tgt = np.asarray([t for t in tgt if d["sample_ids"][t] in keep], dtype=np.int64)
+        s["test"] = [x for x in s["test"] if x in keep]
+    if a.mode == "fewshot" and not smoke:
+        usable = np.flatnonzero(folds["fold"].to_numpy() >= 0)
+        if set(rep.tolist()) | set(tgt.tolist()) != set(usable.tolist()):
+            raise SystemExit("representatives + targets do not cover the usable rows")
+    # modeled genes
+    z = np.load(data / "cells_panel6k.npz", allow_pickle=False)
+    panel = z["panel_gene_ids"].astype(str)
+    gpos = {g: i for i, g in enumerate(d["gene_ids"])}
+    pcanon = np.asarray([gpos[g] for g in panel], dtype=np.int64)
+    rows = np.sort(np.concatenate([rep, tgt]))
+    bulk_ok = d["bulk_valid"][rows][:, pcanon].all(0)
+    sc_ok = d["sc_valid"][rep][:, pcanon].all(0)
+    model_pidx = np.flatnonzero(bulk_ok & sc_ok)
+    genes = panel[model_pidx]
+    if genes.size < 1000:
+        raise SystemExit(f"only {genes.size} modeled genes")
+    # scorer's scoring genes (all rows of the dataset) for the record
+    score_mask = d["bulk_valid"][:, pcanon].all(0) & d["sc_valid"][:, pcanon].all(0)
+    # bulk
+    import anndata as ad
+    B = d["bulk"][rows][:, pcanon[model_pidx]].astype(np.float64)
+    if not np.isfinite(B).all() or (B < 0).any():
+        raise SystemExit("non-finite or negative bulk on the modeled genes")
+    if (B.sum(1) <= 0).any():
+        raise SystemExit("a bulk row has no counts on the modeled genes")
+    sym = pd.read_csv(TASK.parents[1] / "artifacts/panel_ladder_floor01_1m_20260923/tiers/6k/selected_genes.csv")
+    sym = sym.set_index("gene_id").gene_symbol.astype(str)
+    var = pd.DataFrame({"gene_symbol": sym.reindex(genes).to_numpy()}, index=pd.Index(genes, name="gene_id"))
+    obs_b = pd.DataFrame({"sample_ids": d["sample_ids"][rows], "donor_id": d["donor_ids"][rows],
+                          "group": d["groups"][rows], "role": np.where(np.isin(rows, rep), "representative", "target")},
+                         index=pd.Index(d["sample_ids"][rows], name="obs"))
+    ad.AnnData(X=B, obs=obs_b, var=var).write_h5ad(out / "bulk.h5ad")
+    # representative cells
+    X = sp.csr_matrix((z["data"], z["indices"], z["indptr"]), shape=tuple(int(x) for x in z["shape"]))
+    rs = np.asarray(z["row_sample"], dtype=np.int64)
+    blocks, obs_rows, per_rep = [], [], []
+    rng = np.random.default_rng(SEED)
+    for r in rep:
+        cells = np.flatnonzero(rs == r)
+        if a.smoke_cells and cells.size > a.smoke_cells:
+            cells = np.sort(rng.choice(cells, a.smoke_cells, replace=False))
+        Xr = X[cells][:, model_pidx].tocsr().astype(np.float32)
+        lib = np.asarray(Xr.sum(1)).ravel()
+        if (lib <= 0).any():
+            raise SystemExit(f"{d['sample_ids'][r]}: {int((lib <= 0).sum())} cells without counts on the modeled genes")
+        blocks.append(Xr)
+        obs_rows.append(pd.DataFrame({"sample_ids": d["sample_ids"][r], "donor_id": d["donor_ids"][r],
+                                      "source_cell_row": cells}))
+        per_rep.append({"sample_id": str(d["sample_ids"][r]), "donor_id": str(d["donor_ids"][r]),
+                        "group": str(d["groups"][r]), "n_cells": int(cells.size),
+                        "median_umi_modeled_genes": float(np.median(lib))})
+    Xs = sp.vstack(blocks, format="csr")
+    obs_s = pd.concat(obs_rows, ignore_index=True)
+    obs_s.index = pd.Index([f"{a_}|{b_}" for a_, b_ in zip(obs_s.sample_ids, obs_s.source_cell_row)], name="cell")
+    ad.AnnData(X=Xs, obs=obs_s, var=var.copy()).write_h5ad(out / "rep_sc.h5ad")
+    lib_all = np.asarray(Xs.sum(1)).ravel()
+    depth = float(np.median(lib_all))
+    cond = s["condition"]
+    man = {
+        "schema": "scsemi_paired.inputs.v1", "written_utc": datetime.now(timezone.utc).isoformat(),
+        "dataset": a.dataset, "mode": a.mode, "split": j, "smoke": smoke,
+        "smoke_settings": {"reps": a.smoke_reps, "cells": a.smoke_cells, "targets": a.smoke_targets,
+                           "rep_ids": a.smoke_rep_ids} if smoke else None,
+        "train_donors": s["train_donors"],
+        "representatives": per_rep, "targets": [str(x) for x in d["sample_ids"][tgt]],
+        "target_order_test_samples": s["test"],
+        "bulk_row_order": [str(x) for x in d["sample_ids"][rows]],
+        "n_representatives": int(rep.size), "n_targets": int(tgt.size), "n_rep_cells": int(Xs.shape[0]),
+        "modeled_genes": {"n": int(genes.size), "gene_ids": genes.tolist(), "panel_index": model_pidx.tolist(),
+                          "rule": "6k panel genes with bulk_valid in every rep+target row and sc_valid in every rep row",
+                          "equals_scorer_scoring_genes": bool(np.array_equal(model_pidx, np.flatnonzero(score_mask))),
+                          "n_scorer_scoring_genes": int(score_mask.sum())},
+        "depth": {"value": depth, "rule": "median over all representative cells (all rows of the training donors) of the "
+                  "per-cell UMI sum on the modeled genes; every generated cell is scaled to this expected library",
+                  "n_cells": int(lib_all.size)},
+        "condition_npz": str(cond), "condition_sha256": sha(cond),
+        "sources": {"features.npz": sha(data / "features.npz"), "folds.csv": sha(data / "folds.csv"),
+                    "cells_panel6k.npz": sha(data / "cells_panel6k.npz"),
+                    **{k: sha(v) for k, v in s["sources"].items()}},
+        "files": {"bulk.h5ad": sha(out / "bulk.h5ad"), "rep_sc.h5ad": sha(out / "rep_sc.h5ad")},
+        "builder_sha256": sha(Path(__file__)),
+    }
+    with np.load(cond, allow_pickle=False) as cz:
+        cond_ids = cz["sample_ids"].astype(str).tolist()
+    if not smoke and sorted(cond_ids) != sorted(man["targets"]):
+        raise SystemExit("condition NPZ sample_ids differ from the targets")
+    man["condition_sample_order"] = cond_ids
+    tmp = out / "inputs_manifest.json.tmp"
+    tmp.write_text(json.dumps(man, indent=1) + "\n")
+    tmp.replace(out / "inputs_manifest.json")
+    print(json.dumps({"dataset": a.dataset, "mode": a.mode, "split": j, "reps": int(rep.size), "targets": int(tgt.size),
+                      "rep_cells": int(Xs.shape[0]), "modeled_genes": int(genes.size),
+                      "equals_scorer_genes": man["modeled_genes"]["equals_scorer_scoring_genes"], "depth": depth}))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

paper-releases/pbisc-diffusion/updates/20260928T225702Z-paired-ft-scsemi-delta/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scsemi/tools/prepare_env_server2.sh

@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+# Build the server_2 scSemiProfiler venv equivalent to server_1's frozen runtime_env (CPU-side install only).
+#   base: /home/sj_server_1/miniforge3/envs/scanvi (same path and versions on both hosts)
+#   venv: scsemi/runtime_env (--system-site-packages), + the frozen env's pip set with the same pins; certifi and
+#   charset-normalizer (+ the optional regex, defusedxml) are pinned INTO the venv (on server_1 they come from ~/.local), so server_2 runs with
+#   PYTHONNOUSERSITE=1 and nothing from the user site.  Records scsemi_environment.json.  Run as a CPU unit.
+set -euo pipefail
+S=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi
+BASE=/home/sj_server_1/miniforge3/envs/scanvi
+ENV=$S/runtime_env
+export PYTHONNOUSERSITE=1 CUDA_VISIBLE_DEVICES=
+[[ -x $ENV/bin/python ]] || "$BASE/bin/python" -m venv --system-site-packages "$ENV"
+"$ENV/bin/python" -m pip install --no-deps --no-input \
+  faiss-cpu==1.12.0 gseapy==1.3.0 igraph==0.11.9 texttable==1.7.0 \
+  requests==2.32.5 urllib3==2.8.0 certifi==2026.1.4 charset-normalizer==3.4.4 regex==2026.1.15 defusedxml==0.7.1
+"$ENV/bin/python" - <<'PY'
+import importlib.metadata as m, json, os, sys
+from pathlib import Path
+names = ["torch", "anndata", "scanpy", "scvi-tools", "lightning", "pytorch-lightning", "numpy", "scipy", "scikit-learn",
+         "pandas", "h5py", "faiss-cpu", "gseapy", "igraph", "texttable", "requests", "urllib3", "certifi",
+         "charset-normalizer", "idna", "regex", "defusedxml"]
+sys.path.insert(0, os.environ.get("VENDOR", "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/vendor/scSemiProfiler"))
+import scSemiProfiler, scvi, torch  # noqa: F401  (CUDA hidden: import only)
+user = [n for n, mod in list(sys.modules.items()) if getattr(mod, "__file__", None) and "/.local/" in str(mod.__file__)]
+rec = {"python": sys.executable, "versions": {n: m.version(n) for n in names}, "user_site_modules_loaded": user,
+       "cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES")}
+Path(sys.prefix, "scsemi_environment.json").write_text(json.dumps(rec, indent=2) + "\n")
+print(json.dumps(rec, indent=2))
+if user:
+    raise SystemExit(f"modules imported from the user site: {user}")
+PY

paper-releases/pbisc-diffusion/updates/20260928T225702Z-paired-ft-scsemi-delta/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scsemi/tools/run_queue_scsemi.sh

@@ -0,0 +1,149 @@
+#!/usr/bin/env bash
+# scSemiProfiler paired-baseline queue for this host, one split at a time (style of ../tools/run_queue.sh).
+#   queue line:  MODE|DATASET|SPLIT[|OPTIONS]      MODE fewshot (SPLIT = draw) or cv (SPLIT = fold)
+#                OPTIONS (comma list): clean_sample_sc = after the split's NPZ is exported AND validated, delete
+#                scsemi_project/sample_sc/*.h5ad (dense upstream cell graphs, ~1.5 GB per Giroux representative;
+#                their sha256 stay in prepared.json)
+# Per split:
+#   1. done check: generated NPZ + receipt + passing <npz>.validation.json -> skipped
+#   2. CPU units (tools/cpu_unit.sh, ml.slice, <= 6 threads, MemoryMax <= 16G): input build, upstream prepare
+#   3. GPU steps, each after the thermal gate and gpu_usable.sh, each through $HOME/.local/bin/ml-run (ml.slice,
+#      shared GPU lock): pretrain CHUNK_REPS representatives per call, then CHUNK_TARGETS targets per call, until
+#      status.json says complete.  Per-target runtime is in work/.../shards/<target>.json and events.jsonl.
+#   4. CPU units: export (adapter) + independent validation; optional scoring (SCORE=1, benchmark scorer)
+# A failure is logged, the split's partial outputs are preserved (no retry, no cleanup) and the queue moves on.
+#   env: QUEUE (required), MAX_WAIT_H (24), CHUNK_REPS (2), CHUNK_TARGETS (12), SCORE (0), SCSEMI_PY (host default)
+#        SMOKE_CPU=1: queue-logic test on "smoke|DATASET|DRAW|OPTIONS|T1,T2,.." lines only (2 reps x 300 cells, 1-2 epochs,
+#        CPU units instead of ml-run, outputs under scsemi/smoke/queue/); never touches the GPU
+#        PREP_ONLY=1: CPU steps only (input build + upstream prepare), no GPU, no export.  Safe to run while the GPU
+#        queue runs: every split is processed under an exclusive lock (work/.../.split.lock, waits up to MAX_WAIT_H).
+#        PREP_MEM (16): MemoryMax GiB of the upstream-prepare CPU unit (cpu_unit.sh caps it; see README memory note)
+set -uo pipefail
+export XDG_RUNTIME_DIR="${XDG_RUNTIME_DIR:-/run/user/$(id -u)}"
+export DBUS_SESSION_BUS_ADDRESS="${DBUS_SESSION_BUS_ADDRESS:-unix:path=$XDG_RUNTIME_DIR/bus}"
+M8=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion
+T=$M8/plans/paired_finetune_intradataset_20260927
+S=$T/scsemi
+case "$(hostname -s)" in
+  spark-fa6b) PYD=$M8/plans/immicro_scsemi_comparison_20260927/runtime_env/bin/python; USERSITE=1 ;;   # server_1: frozen env
+  spark-0110) PYD=$S/runtime_env/bin/python; USERSITE=0 ;;                                            # server_2: equivalent venv
+  *) echo "unknown host $(hostname -s)" >&2; exit 2 ;;
+esac
+PY=${SCSEMI_PY:-$PYD}
+SPY=/home/sj_server_1/miniforge3/envs/cxg_corpus/bin/python     # benchmark scorer env (both hosts)
+QUEUE="${QUEUE:?QUEUE}"; MAX_WAIT_H="${MAX_WAIT_H:-24}"; CHUNK_REPS="${CHUNK_REPS:-2}"; CHUNK_TARGETS="${CHUNK_TARGETS:-12}"
+SCORE="${SCORE:-0}"; SMOKE_CPU="${SMOKE_CPU:-0}"; PREP_ONLY="${PREP_ONLY:-0}"; PREP_MEM="${PREP_MEM:-16}"
+mkdir -p "$S/logs"
+LOG=$S/logs/queue_$(basename "$QUEUE" .txt)_$(hostname -s)_$(date +%Y%m%dT%H%M%S).log
+log() { echo "$(date +%FT%T) $*" | tee -a "$LOG"; }
+start=$(date +%s)
+gate() {
+  local verdict usable
+  if [[ $SMOKE_CPU == 1 ]]; then echo "smoke: CPU only, no GPU gate"; return; fi
+  until verdict=$(bash "$M8/plans/floor01_campaign_20260923/tools/thermal_guard.sh") \
+        && usable=$(bash "$M8/plans/arch_ablation_6k_20260913/tools/gpu_usable.sh" 2>/dev/null); do
+    (( $(date +%s) - start > MAX_WAIT_H * 3600 )) && { log "ABORT: no usable GPU within ${MAX_WAIT_H} h"; exit 1; }
+    sleep 120
+  done
+  echo "$verdict; GPU $usable"
+}
+cpu() {   # cpu NAME THREADS MEMMAX CMD...   (waits for the memory floor instead of failing on it)
+  local mem=$3
+  until (( $(awk '/MemAvailable/ {print int($2/1048576)}' /proc/meminfo) >= 32 + mem )); do
+    (( $(date +%s) - start > MAX_WAIT_H * 3600 )) && { log "ABORT: memory floor not met within ${MAX_WAIT_H} h"; exit 1; }
+    sleep 120
+  done
+  ALLOW_USER_SITE=${US:-$USERSITE} bash "$S/tools/cpu_unit.sh" "$@" >> "$LOG" 2>&1 < /dev/null
+}
+st() { python3 -c "import json,sys;print(json.load(open(sys.argv[1]))[sys.argv[2]])" "$1/status.json" "$2" 2>/dev/null; }
+gpu() {   # gpu WORK PHASE EXTRA...
+  local W=$1 phase=$2; shift 2
+  if [[ $SMOKE_CPU == 1 ]]; then   # queue-logic test: same wrapper, CPU, reduced epochs, never ml-run
+    cpu "smk-$phase" 4 8 "$PY" -u "$S/tools/run_scsemi_paired.py" --work "$W" --phase "$phase" --device cpu --smoke --threads 4 "$@"
+    return
+  fi
+  local ns=
+  [[ $USERSITE == 1 ]] || ns=PYTHONNOUSERSITE=1
+  "$HOME/.local/bin/ml-run" /usr/bin/env CUDA_VISIBLE_DEVICES=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
+    OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 PYTHONUNBUFFERED=1 PYTHONDONTWRITEBYTECODE=1 $ns \
+    "$PY" -u "$S/tools/run_scsemi_paired.py" --work "$W" --phase "$phase" --device cuda:0 --threads 4 "$@" \
+    >> "$W/gpu.log" 2>&1 < /dev/null
+}
+n_fail=0
+fail() { n_fail=$((n_fail+1)); log "WARN: $*"; }
+log "queue $QUEUE on $(hostname -s); python $PY; chunks reps=$CHUNK_REPS targets=$CHUNK_TARGETS; score=$SCORE; smoke_cpu=$SMOKE_CPU"
+while IFS= read -r line; do
+  [[ -z "${line// }" || "$line" == \#* ]] && continue
+  IFS='|' read -r mode ds j opts extra <<< "$line"
+  SMK=()
+  case $mode in
+    fewshot) W=$S/work/$ds/fewshot_n5/draw$j; OUT=$T/generated/$ds/fewshot_n5/draw$j/scsemi5__raw.npz; ARM=scsemi5
+             SC=$T/scores/$ds/fewshot_n5/draw${j}_scsemi5; DON=$T/data/$ds/fewshot_n5/draw$j/test_samples.txt
+             imode=fewshot; sflag=(--draw "$j") ;;
+    cv)      W=$S/work/$ds/cv/fold$j; OUT=$T/generated/$ds/fold$j/scsemi__raw.npz; ARM=scsemi
+             SC=$T/scores/$ds/fold${j}_scsemi; DON=; imode=cv; sflag=(--fold "$j") ;;
+    smoke)   [[ $SMOKE_CPU == 1 ]] || { fail "smoke line needs SMOKE_CPU=1: $line"; continue; }
+             W=$S/smoke/queue/${ds}_draw$j; OUT=$W/generated/scsemi5__raw.npz; ARM=scsemi5; SC=$W/score
+             DON=$W/targets.txt; imode=fewshot; SMK=(--smoke)
+             sflag=(--draw "$j" --smoke-reps 2 --smoke-cells 300 --smoke-targets "$extra")
+             mkdir -p "$W"; tr ',' '\n' <<< "$extra" > "$DON" ;;
+    *) fail "unknown line: $line"; continue ;;
+  esac
+  [[ $SMOKE_CPU == 1 && $mode != smoke ]] && { fail "SMOKE_CPU=1 runs smoke lines only: $line"; continue; }
+  tag="$ds/$mode$j"
+  mkdir -p "$W"
+  exec 9>"$W/.split.lock"          # re-opening fd 9 releases the previous split's lock
+  flock -w $((MAX_WAIT_H * 3600)) 9 || { fail "$tag: split lock held by another queue for ${MAX_WAIT_H} h"; continue; }
+  VAL=${OUT%.npz}.validation.json
+  if [[ -s $VAL ]] && python3 -c "import json,sys;sys.exit(0 if json.load(open(sys.argv[1]))['passed'] else 1)" "$VAL"; then
+    log "$tag: exported + validated already; skipped"
+  else
+    mkdir -p "$W"
+    if [[ ! -s $W/inputs/inputs_manifest.json ]]; then
+      log "$tag: building inputs (CPU unit)"
+      cpu "in-${ds:0:3}-$mode$j" 4 12 "$PY" -u "$S/tools/prepare_draw_inputs.py" --dataset "$ds" --mode "$imode" "${sflag[@]}" --work "$W" \
+        || { fail "$tag: input build failed"; continue; }
+    fi
+    if [[ ! -s $W/prepared.json ]]; then
+      log "$tag: upstream prepare (initsetup + assignment + scprocess, CPU unit)"
+      cpu "prep-${ds:0:3}-$mode$j" 6 "$PREP_MEM" "$PY" -u "$S/tools/run_scsemi_paired.py" --work "$W" --phase prepare --device cpu --threads 6 "${SMK[@]}" \
+        || { fail "$tag: prepare failed (see $W/events.jsonl)"; continue; }
+    fi
+    if [[ $PREP_ONLY == 1 ]]; then log "$tag: CPU-prepared (PREP_ONLY; GPU steps left to the GPU queue)"; continue; fi
+    ok=1
+    while [[ "$(st "$W" pretrain_complete)" != True ]]; do
+      before=$(st "$W" reps_pretrained)
+      g=$(gate) || { log "ABORT: no usable GPU within ${MAX_WAIT_H} h"; exit 1; }
+      log "$tag: pretrain ${before}/$(st "$W" n_reps) reps; next $CHUNK_REPS ($g)"
+      gpu "$W" pretrain --max-reps "$CHUNK_REPS" || { fail "$tag: pretrain step failed (see $W/gpu.log, $W/events.jsonl)"; ok=0; break; }
+      [[ "$(st "$W" reps_pretrained)" != "$before" ]] || { fail "$tag: pretrain made no progress"; ok=0; break; }
+    done
+    (( ok )) || continue
+    while [[ "$(st "$W" infer_complete)" != True ]]; do
+      before=$(st "$W" targets_done)
+      g=$(gate) || { log "ABORT: no usable GPU within ${MAX_WAIT_H} h"; exit 1; }
+      log "$tag: infer ${before}/$(st "$W" n_targets) targets; next $CHUNK_TARGETS ($g)"
+      gpu "$W" infer --max-targets "$CHUNK_TARGETS" || { fail "$tag: infer step failed (see $W/gpu.log, $W/events.jsonl)"; ok=0; break; }
+      [[ "$(st "$W" targets_done)" != "$before" ]] || { fail "$tag: infer made no progress"; ok=0; break; }
+    done
+    (( ok )) || continue
+    if [[ ! -s $OUT ]]; then
+      cpu "exp-${ds:0:3}-$mode$j" 2 16 "$PY" -u "$S/tools/export_npz.py" --work "$W" --out "$OUT" || { fail "$tag: export failed"; continue; }
+    fi
+    cpu "val-${ds:0:3}-$mode$j" 2 16 "$PY" -u "$S/tools/validate_export.py" --work "$W" --npz "$OUT" || { fail "$tag: validation failed ($VAL)"; continue; }
+    log "$tag: exported + validated: $(python3 -c "import json,sys;d=json.load(open(sys.argv[1]));print('targets',d['bulks'],'median target s %.1f'%d['median_target_seconds'],'total target s %.0f'%d['total_target_seconds'])" "${OUT%.npz}.log")"
+    if [[ ",$opts," == *",clean_sample_sc,"* ]]; then
+      rm -f "$W"/scsemi_project/sample_sc/*.h5ad && log "$tag: removed sample_sc/*.h5ad (sha256 kept in prepared.json)"
+    fi
+  fi
+  if [[ $SCORE == 1 && $PREP_ONLY != 1 && $mode != cv && -s $VAL && ! -s $SC/summary.json ]]; then
+    read -r -a cargs <<< "$(bash "$T/tools/contrast_args.sh" "$ds")"
+    log "$tag: scoring arm $ARM -> $SC"
+    mkdir -p "$SC"
+    US=0 cpu "sc-${ds:0:3}-$mode$j" 6 16 "$SPY" -u "$T/tools/score_ft.py" --dataset "$ds" --out "$SC" --threads 6 \
+      --donors "$DON" "${cargs[@]}" --arm "$ARM=$OUT" || fail "$tag: scoring failed"
+  fi
+done < "$QUEUE"
+exec 9>&-
+log "queue finished: failed=$n_fail"
+(( n_fail == 0 ))