verify converted gate_up split adapter: clean load + 48-edit subset

verify_split.py

@@ -0,0 +1,93 @@
+"""Verify the gate_up-split adapter for ysharma/orbit-alpha-lora.
+
+1. Load orbit_alpha_lora_gate_up_split.safetensors into QwenImage21Pipeline
+   via load_lora_weights and CAPTURE all warnings/logs during the load.
+   PASS requires: no warning mentioning gate_up / gate_layer / proj / dropped keys.
+2. Run the fixed 48-edit selection subset (first 12 eval objects x 4) at 20 steps.
+   Compare against the ckpt2000 selection numbers (alpha_iou 0.836, lpips 0.087)
+   measured with the ORIGINAL ComfyUI-keyed file, whose gate_up keys were dropped
+   at load. The split adapter should be equal or better; a large drop would mean
+   the fused-order assumption (rows [0:12288]=gate_layer, rest=proj) is wrong.
+"""
+import argparse, io, json, logging, os, warnings
+import eval_edit as ee
+import select_ckpt as sc
+
+ee.STEPS = 20  # same subset budget as the selection run, for apples-to-apples numbers
+
+MODEL_REPO = 'ysharma/orbit-alpha-lora'
+SPLIT_FILE = 'checkpoints/steps2000res768/orbit_alpha_lora_gate_up_split.safetensors'
+DS_REPO = ee.DS_REPO
+OUT_REPO_PATH = 'eval/gate_up_split'
+REFERENCE = {'name': 'ckpt2000-selection-run', 'alpha_iou': 0.836, 'lpips': 0.087}
+
+
+def capture_load(pipe, path):
+    """Load adapter while recording every warning and diffusers/peft log line."""
+    buf = []
+
+    class Rec(logging.Handler):
+        def emit(self, record):
+            buf.append(record.getMessage())
+
+    h = Rec()
+    logging.getLogger().addHandler(h)
+    for n in ('diffusers', 'peft', 'diffusers.loaders', 'diffusers.models', 'root'):
+        lg = logging.getLogger(n)
+        old = lg.level
+        lg.setLevel(logging.DEBUG)
+        lg.addHandler(h)
+        setattr(h, '_saved_' + n.replace('.', '_'), (lg, old))
+    with warnings.catch_warnings(record=True) as wlist:
+        warnings.simplefilter('always')
+        pipe.load_lora_weights(path)
+    msgs = list(buf) + [str(w.message) for w in wlist]
+    for n in ('diffusers', 'peft', 'diffusers.loaders', 'diffusers.models', 'root'):
+        lg = logging.getLogger(n)
+        lg.removeHandler(h)
+    logging.getLogger().removeHandler(h)
+    return msgs
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--out', required=True)
+    args = ap.parse_args()
+    os.makedirs(args.out, exist_ok=True)
+
+    pairs = sc.pick_subset([json.loads(l) for l in open(ee.get_file('pairs_eval.jsonl'))])
+    print(f'subset: {len(pairs)} edits', flush=True)
+
+    pipe = ee.load_pipe()
+    m = ee.make_metrics()
+
+    local = ee.hf_hub_download(repo_id=MODEL_REPO, repo_type='model', filename=SPLIT_FILE)
+    msgs = capture_load(pipe, local)
+    print('LOAD MESSAGES:', json.dumps(msgs), flush=True)
+    bad = [x for x in msgs if any(t in x.lower() for t in
+          ('gate_up', 'not found', 'skipped', 'dropped', 'unexpected key', 'no module'))]
+    load_clean = not bad
+    print(f'LOAD CLEAN: {load_clean}', flush=True)
+    if not load_clean:
+        for b in bad:
+            print('  SUSPECT:', b, flush=True)
+
+    rows = ee.run_pairs(pipe, m, pairs, 'instruction', os.path.join(args.out, 'split48'))
+    r = sc.agg_rows(rows, 'split-adapter-48edits')
+    r['load_clean'] = load_clean
+    r['load_warnings'] = bad
+    r['reference_ckpt2000'] = REFERENCE
+    verdict = load_clean and r['alpha_iou'] >= REFERENCE['alpha_iou'] - 0.02 \
+        and r['lpips'] <= REFERENCE['lpips'] + 0.01
+    r['verdict'] = 'PASS' if verdict else 'FAIL'
+    print('VERDICT:', r['verdict'], flush=True)
+
+    with open(os.path.join(args.out, 'summary.json'), 'w') as f:
+        json.dump(r, f, indent=2)
+    sc.upload(os.path.join(args.out, 'summary.json'), f'{OUT_REPO_PATH}/summary.json')
+    sc.upload(os.path.join(args.out, 'split48', 'results.jsonl'), f'{OUT_REPO_PATH}/split48_results.jsonl')
+    print('DONE', flush=True)
+
+
+if __name__ == '__main__':
+    main()