mirror of
https://github.com/onestardao/WFGY.git
synced 2026-04-28 11:40:07 +00:00
60 lines
2.2 KiB
Python
60 lines
2.2 KiB
Python
# example_07_flash_show.py
|
|
# Flashy showcase: 10 prompts, remote toggle (slow if True)
|
|
|
|
import pathlib, sys, numpy as np, torch, textwrap, time
|
|
sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1]))
|
|
|
|
import wfgy_sdk as w
|
|
from wfgy_sdk.evaluator import compare_logits
|
|
|
|
use_remote = False
|
|
MODEL_ID = "tiiuae/falcon-7b-instruct"
|
|
GAMMA = 1.0
|
|
NOISE = 0.12
|
|
|
|
PROMPTS = [
|
|
"Derive Maxwell's equations from first principles in 30 words.",
|
|
"Explain Gödel's incompleteness in terms of topological fixed points.",
|
|
"Predict 2120 climate using quantum chromodynamics metaphors.",
|
|
"Summarise category theory for a five-year-old using only emojis.",
|
|
"Describe consciousness as a phase transition in Hilbert space.",
|
|
"Translate the second law of thermodynamics into sushi-chef language.",
|
|
"Explain dark energy by quoting Shakespearean sonnets.",
|
|
"Model altruism as a non-convex optimization landscape.",
|
|
"Describe a black hole using only prime numbers.",
|
|
"Solve world peace with a single C++ template meta-program."
|
|
]
|
|
|
|
rng = np.random.default_rng(999)
|
|
eng = w.get_engine(reload=True); eng.gamma = GAMMA
|
|
|
|
if not use_remote:
|
|
from transformers import GPT2LMHeadModel, GPT2TokenizerFast
|
|
tok = GPT2TokenizerFast.from_pretrained("gpt2")
|
|
gpt2 = GPT2LMHeadModel.from_pretrained("gpt2").eval()
|
|
|
|
print("\n=== Example 07 · Flash-show ===")
|
|
records = []
|
|
for idx, prompt in enumerate(PROMPTS, 1):
|
|
if use_remote:
|
|
logits0 = w.call_remote_model(prompt, model_id=MODEL_ID)
|
|
else:
|
|
ids = tok(prompt, return_tensors="pt").input_ids
|
|
with torch.no_grad():
|
|
logits0 = gpt2(ids).logits[0, -1].cpu().numpy()
|
|
|
|
G = rng.normal(size=256); G /= np.linalg.norm(G)
|
|
I = G + rng.normal(scale=NOISE, size=256)
|
|
|
|
logits1 = eng.run(input_vec=I, ground_vec=G, logits=logits0)
|
|
m = compare_logits(logits0, logits1)
|
|
records.append(m)
|
|
|
|
print(f"[{idx:02d}] KL {m['kl_divergence']:.2f} | "
|
|
f"var↓ {(1-m['std_ratio'])*100:.0f}% | "
|
|
f"{textwrap.shorten(prompt, 45)}")
|
|
|
|
avg = {k: np.mean([r[k] for r in records]) for k in records[0]}
|
|
print("\n--- average over 10 prompts ---")
|
|
print(avg)
|
|
print("⚠ Larger LLM → stronger variance drop & higher KL.\n")
|