File size: 2,991 Bytes
4737431 | 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 | #!/bin/bash
# Isolate the all-EOS HumanEval bug: same prompt through (A) Dream's OFFICIAL
# diffusion_generate and (B) our decoder. If A is also EOS -> prompt/format issue.
# If A gives code and B does not -> our decoder has a bug.
set -uo pipefail
pip install -q "transformers==4.46.2" "huggingface_hub<1.0" "datasets<4" "accelerate" 2>&1 | tail -1
python -c "
from huggingface_hub import snapshot_download
snapshot_download('ashishk1331/ccd-repro-code', repo_type='dataset', local_dir='/work')"
cd /work
python - <<'EOF'
import sys, torch
sys.path.insert(0, "scripts")
from transformers import AutoModel, AutoTokenizer
from datasets import load_dataset
import ccd_decode
from run_eval import humaneval_prompt
M = "Dream-org/Dream-v0-Instruct-7B"
tok = AutoTokenizer.from_pretrained(M, trust_remote_code=True)
model = AutoModel.from_pretrained(M, torch_dtype=torch.bfloat16,
trust_remote_code=True).to("cuda").eval()
doc = load_dataset("openai/openai_humaneval", split="test")[0]
p = humaneval_prompt(tok, doc)
enc = tok(p, return_tensors="pt")
ids = enc.input_ids.to("cuda")
attn = enc.attention_mask.to("cuda")
def show(tag, gen_ids):
txt = tok.decode(gen_ids)
n_eos = sum(1 for i in gen_ids if i == tok.eos_token_id)
print(f"\n########## {tag}")
print(f" eos {n_eos}/{len(gen_ids)} first ids {gen_ids[:10]}")
print(" text:", repr(txt.split(tok.eos_token)[0][:300]))
# ---- A: Dream's OFFICIAL generate, exactly the authors' eval settings
out = model.diffusion_generate(
ids, attention_mask=attn, max_new_tokens=256, output_history=False,
return_dict_in_generate=True, steps=256, temperature=0.1, top_p=0.9,
top_k=None, alg="entropy", alg_temp=0.0)
show("A: OFFICIAL diffusion_generate (temp=0.1 top_p=0.9 alg=entropy)",
out.sequences[0, ids.shape[1]:].tolist())
# ---- B: our decoder, same settings
x, st = ccd_decode.generate(model, ids, attention_mask=attn, max_new_tokens=256,
steps=256, temperature=0.1, top_p=0.9,
mask_token_id=model.config.mask_token_id, method="baseline")
show("B: OUR decoder baseline (temp=0.1 top_p=0.9)", x[0, ids.shape[1]:].tolist())
# ---- C: our decoder at temp 0 (isolates the temperature/top_p path)
x, st = ccd_decode.generate(model, ids, attention_mask=attn, max_new_tokens=256,
steps=256, temperature=0.0, top_p=1.0,
mask_token_id=model.config.mask_token_id, method="baseline")
show("C: OUR decoder baseline (temp=0 top_p=1)", x[0, ids.shape[1]:].tolist())
# ---- D: official generate at 768 tokens (the paper/authors' real length)
out = model.diffusion_generate(
ids, attention_mask=attn, max_new_tokens=768, output_history=False,
return_dict_in_generate=True, steps=768, temperature=0.1, top_p=0.9,
top_k=None, alg="entropy", alg_temp=0.0)
show("D: OFFICIAL diffusion_generate, 768 tokens/steps", out.sequences[0, ids.shape[1]:].tolist())
EOF
|