File size: 5,777 Bytes
24918f7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | #!/usr/bin/env python3
"""Score turns through the trainer's own forward and, optionally, through a pw-decide
launcher, so a head can be checked before the engine is blamed: the two paths agree
to the fourth decimal when the engine is right, and a head that scores nothing here
was never trained, whatever its training log said.
The head must record the encoding settings it trained with, and they must be the
ones the engine serves at. Exit 0 when they match and the engine agrees within
--tolerance, 1 on a mismatch or a larger gap, 2 when the head records no settings
or nothing was compared (no --engine, or no complete turns).
Usage: parity_probe.py --model ID --head FILE --examples FILE [--engine LAUNCHER] [--n 6]
[--corpus <decide dir>/corpus.json] [--tolerance 0.001]
"""
import argparse
import json
import os
import subprocess
import sys
os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
import torch # noqa: E402
import laya # noqa: E402
from safetensors.torch import load_file # noqa: E402
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import headfile # noqa: E402
import rows as rowfile # noqa: E402
from corpus import Corpus, decide_dir # noqa: E402
from train import forward_head, multi_item, precompute, state_text # noqa: E402
def torch_probs(agent, model, corpus, ex, device):
context = int(agent.cfg.get("max_len", 512))
head_max_len = min(int(corpus.state_spec.get("head_tokens", 512)), context - 64)
tools_q = corpus.multi_question("tool", ex["offered"]["loadable"])
item = multi_item(agent.tok, state_text(ex["state"]), tools_q, {k: 0 for k in tools_q["options"]}, context, head_max_len, "tools")
names = list(tools_q["options"].keys())[: len(item["markers"])]
f = precompute(agent, [item], device)[0]
with torch.no_grad():
logits = forward_head(model, f["h"][None].to(device), f["att"][None].to(device), f["marker_pos"][None].to(device),
f["marker_mask"][None].to(device), torch.tensor([item["qtype"]], device=device))[0].cpu()
return dict(zip(names, torch.sigmoid(logits[: len(names)]).tolist()))
def engine_probs(proc, corpus, ex):
req = {"method": "decide", "head": "turn-load", "state": ex["state"], "questions": corpus.row_questions(ex)}
proc.stdin.write(json.dumps(req) + "\n")
proc.stdin.flush()
return json.loads(proc.stdout.readline())["answers"]["tools"]["probabilities"]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--model", required=True)
ap.add_argument("--head", required=True)
ap.add_argument("--examples", required=True)
ap.add_argument("--engine", default=os.environ.get("LYCAON_DECIDE_ENGINE"))
ap.add_argument("--corpus", default=str(decide_dir() / "corpus.json"))
ap.add_argument("--n", type=int, default=6)
ap.add_argument("--tolerance", type=float, default=1e-3, help="the largest engine-trainer probability gap that passes")
args = ap.parse_args()
device = os.environ.get("LAYA_DEVICE") or ("mps" if torch.backends.mps.is_available() else "cpu")
agent = laya.load(args.model, device=device)
model = agent.model
tensors = load_file(args.head, device=device)
for module in ("head", "scorer", "type_emb"):
state = {k[len(module) + 1:]: v for k, v in tensors.items() if k.startswith(module + ".")}
if state:
getattr(model, module).load_state_dict(state)
model.eval()
# The probe encodes at the checkpoint's context, as the engine does; a head trained at
# another length would agree here and still have learned from different inputs.
meta = headfile.read_metadata(args.head)[0]
context = int(agent.cfg.get("max_len", 512))
budget = min(int(Corpus.load(args.corpus).state_spec.get("head_tokens", 512)), context - 64)
if "max_len" not in meta or "head_max_len" not in meta:
print("parity: unverified; the head records no encoding settings")
return 2
if (int(meta["max_len"]), int(meta["head_max_len"])) != (context, budget):
print("parity: the head was trained at max_len %s, head budget %s; the engine encodes at %d, %d" % (
meta["max_len"], meta["head_max_len"], context, budget))
return 1
corpus = Corpus.load(args.corpus)
proc = None
if args.engine:
proc = subprocess.Popen([args.engine], stdin=subprocess.PIPE, stdout=subprocess.PIPE, text=True, bufsize=1)
proc.stdin.write(json.dumps({"method": "hello"}) + "\n")
proc.stdin.flush()
proc.stdout.readline()
worst = 0.0
examples = [ex for ex in rowfile.load(args.examples) if not ex["partial"]][: args.n]
for ex in examples:
pt = torch_probs(agent, model, corpus, ex, device)
top = sorted(pt.items(), key=lambda kv: -kv[1])[:5]
truth = sorted(rowfile.truth_tools(ex))
print("truth", truth)
print(" trainer top", [(k, round(v, 3)) for k, v in top], "| truth p", [(t, round(pt.get(t, -1), 3)) for t in truth])
if proc is not None:
pe = engine_probs(proc, corpus, ex)
gap = max(abs(pe.get(k, 0.0) - v) for k, v in pt.items())
worst = max(worst, gap)
print(" engine top", [(k, round(pe.get(k, 0.0), 3)) for k, _ in top], "| max |engine - trainer| %.4f" % gap)
if proc is not None:
proc.stdin.close()
proc.wait()
print("parity: worst gap %.4f over %d turns" % (worst, len(examples)))
if worst > args.tolerance:
return 1
if proc is None or not examples:
print("parity: unverified; no engine answers were compared")
return 2
return 0
if __name__ == "__main__":
sys.exit(main())
|