ParcaeAdapter: rho(A)<1 by construction (ZOH negative-diag), rho logging in rec arms, pre-registration item 12
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
@@ -16,7 +16,7 @@ from pathlib import Path
|
||||
import torch
|
||||
|
||||
from loop_common import (AdaptiveMergeAdapter, BandLooper, MergeAdapter,
|
||||
RecurrentAdapter)
|
||||
ParcaeAdapter, RecurrentAdapter)
|
||||
from prep_mbpp import DIRECT_SUFFIX, extract_code, mbpp_prompt, run_tests
|
||||
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
||||
@@ -79,13 +79,16 @@ def main():
|
||||
ap.add_argument("--adaptive", action="store_true")
|
||||
ap.add_argument("--rec", action="store_true",
|
||||
help="RecurrentAdapter (noise h0, learned A/B)")
|
||||
ap.add_argument("--parcae", action="store_true",
|
||||
help="ParcaeAdapter (rho(A)<1 by construction)")
|
||||
args = ap.parse_args()
|
||||
ks = [int(x) for x in args.ks.split(",")]
|
||||
|
||||
model, tok = load_model(dtype=torch.bfloat16)
|
||||
tok.padding_side = "left"
|
||||
looper = BandLooper(model)
|
||||
cls = (RecurrentAdapter if args.rec
|
||||
cls = (ParcaeAdapter if args.parcae
|
||||
else RecurrentAdapter if args.rec
|
||||
else AdaptiveMergeAdapter if args.adaptive else MergeAdapter)
|
||||
kw = ({"alpha0": args.alpha} if args.adaptive else {"alpha": args.alpha})
|
||||
adapter = cls(d=model.config.get_text_config().hidden_size, **kw).cuda()
|
||||
|
||||
Reference in New Issue
Block a user