TiedAlphaAdapter: learned per-dim alpha with tied B (anchored by construction); pre-registration item 14

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/train_merge_code.py

@@ -21,7 +21,8 @@ import torch
 import torch.nn.functional as F
 
 from loop_common import (AdaptiveMergeAdapter, BandLooper, MergeAdapter,
-                         ParcaeAdapter, PerDepthAdapter, RecurrentAdapter)
+                         ParcaeAdapter, PerDepthAdapter, RecurrentAdapter,
+                         TiedAlphaAdapter)
 from prep_mbpp import DIRECT_SUFFIX, mbpp_prompt
 
 sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
@@ -61,6 +62,9 @@ ap.add_argument("--recmax", type=int, default=16)
 ap.add_argument("--parcae", action="store_true",
                 help="rec regime with rho(A)<1 by construction "
                      "(diag-negative-exp ZOH parameterization)")
+ap.add_argument("--tiedalpha", action="store_true",
+                help="merge with learned per-dim constant alpha, B tied "
+                     "to (1-a); standard curriculum")
 ap.add_argument("--perdepth", type=int, default=0,
                 help="PerDepthAdapter: one merge per iteration depth "
                      "(Bae-style relaxation), standard curriculum")
@@ -80,7 +84,8 @@ SUFFIX = ((f"_s{SEED}" if SEED else "")
           + ("_warm" if ARGS.warm else "")
           + (("_parcae" if ARGS.parcae else "_rec") + str(ARGS.recmax)
              if ARGS.rec else "")
-          + (f"_pd{ARGS.perdepth}" if ARGS.perdepth else ""))
+          + (f"_pd{ARGS.perdepth}" if ARGS.perdepth else "")
+          + ("_ta" if ARGS.tiedalpha else ""))
 PAUSE_ID = 6  # <unused0>
 
 
@@ -139,7 +144,9 @@ def main():
         p.requires_grad_(False)
     looper = BandLooper(model)
     d = model.config.get_text_config().hidden_size
-    if ARGS.perdepth:
+    if ARGS.tiedalpha:
+        adapter = TiedAlphaAdapter(d=d, alpha=ARGS.alpha).cuda()
+    elif ARGS.perdepth:
         adapter = PerDepthAdapter(d=d, alpha=ARGS.alpha,
                                   n_depth=ARGS.perdepth).cuda()
     elif ARGS.parcae: