adaptive flag for GSM eval
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
@@ -13,7 +13,8 @@ from pathlib import Path
|
|||||||
|
|
||||||
import torch
|
import torch
|
||||||
|
|
||||||
from loop_common import (DIRECT_SUFFIX, BandLooper, MergeAdapter, chat_prompt,
|
from loop_common import (DIRECT_SUFFIX, AdaptiveMergeAdapter, BandLooper,
|
||||||
|
MergeAdapter, chat_prompt,
|
||||||
last_number, num_eq)
|
last_number, num_eq)
|
||||||
import sys
|
import sys
|
||||||
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
||||||
@@ -96,6 +97,7 @@ def main():
|
|||||||
ap.add_argument("--prompt-only", action="store_true",
|
ap.add_argument("--prompt-only", action="store_true",
|
||||||
help="loop the prompt span only (unified regime, fast path)")
|
help="loop the prompt span only (unified regime, fast path)")
|
||||||
ap.add_argument("--no-spider", action="store_true")
|
ap.add_argument("--no-spider", action="store_true")
|
||||||
|
ap.add_argument("--adaptive", action="store_true")
|
||||||
ap.add_argument("--feedforward", action="store_true",
|
ap.add_argument("--feedforward", action="store_true",
|
||||||
help="no-recurrence control arm: adapter(e,e) once")
|
help="no-recurrence control arm: adapter(e,e) once")
|
||||||
args = ap.parse_args()
|
args = ap.parse_args()
|
||||||
@@ -104,7 +106,8 @@ def main():
|
|||||||
model, tok = load_model(dtype=torch.bfloat16)
|
model, tok = load_model(dtype=torch.bfloat16)
|
||||||
tok.padding_side = "left"
|
tok.padding_side = "left"
|
||||||
looper = BandLooper(model)
|
looper = BandLooper(model)
|
||||||
adapter = MergeAdapter(
|
cls = AdaptiveMergeAdapter if args.adaptive else MergeAdapter
|
||||||
|
adapter = cls(
|
||||||
d=model.config.get_text_config().hidden_size).cuda()
|
d=model.config.get_text_config().hidden_size).cuda()
|
||||||
if args.adapter:
|
if args.adapter:
|
||||||
adapter.load_state_dict(torch.load(args.adapter, map_location="cuda"))
|
adapter.load_state_dict(torch.load(args.adapter, map_location="cuda"))
|
||||||
|
|||||||
Reference in New Issue
Block a user