File size: 12,661 Bytes
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
2a88bc9
b222eb5
 
2a88bc9
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed62109
 
 
 
 
 
 
 
 
b222eb5
 
 
 
ed62109
 
b222eb5
ed62109
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b222eb5
 
 
 
 
 
 
 
7d05566
 
 
 
 
 
 
b222eb5
 
7d05566
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7d05566
 
b222eb5
 
 
 
63502da
 
b222eb5
 
2a88bc9
 
 
 
 
 
b222eb5
 
 
 
2a88bc9
 
 
 
 
 
 
 
 
b222eb5
 
 
63502da
 
b222eb5
 
 
2a88bc9
 
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7d05566
b222eb5
 
 
02bdb20
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
02bdb20
 
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a88bc9
b222eb5
 
 
 
 
 
 
02bdb20
b222eb5
 
 
 
 
 
 
 
 
 
 
 
 
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
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
"""Evaluate Dream-7B-Instruct with baseline / CCD / CCD-DS on Trip Plan or HumanEval.

Hyperparameters follow the Dream authors' official eval scripts, which is what the
paper says it does ("we follow the base models' default settings without tuning"):

  Trip Plan  : steps=256,  max_new_tokens=256,  temperature=0,   top_p=1,   alg=entropy
               (DreamLM/Dream eval/eval_dream_gen_planning.sh)  -- 2-shot prompt
  HumanEval  : steps=768,  max_new_tokens=768,  temperature=0.1, top_p=0.9, alg=entropy
               (DreamLM/Dream eval_instruct/eval.sh)            -- 0-shot, chat template

Usage:
  python run_eval.py --task trip --method ccd_ds --limit 100 --out outputs/x.json
"""
import argparse, json, os, random, sys, time
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

import numpy as np
import torch
from transformers import AutoModel, AutoTokenizer

import ccd_decode
from trip_metric import parse_response, compute_example_score
from sanitize_utils import sanitize

MODEL = "Dream-org/Dream-v0-Instruct-7B"
HERE = os.path.dirname(os.path.abspath(__file__))
DATA = os.path.join(HERE, "..", "data")

TASK_DEFAULTS = {
    "trip":      dict(steps=256, max_new_tokens=256, temperature=0.0, top_p=1.0),
    "humaneval": dict(steps=768, max_new_tokens=768, temperature=0.1, top_p=0.9),
}


# ---------------------------------------------------------------- data / prompts

def load_trip(limit=None, num_cities=None, seed=0):
    """Load Trip Plan examples.

    IMPORTANT: the benchmark file is ordered by difficulty -- the first 200 of the
    1600 examples all have num_cities=3 (the easiest tier), then 200 with 4, etc.
    Taking a prefix therefore samples ONLY the easiest tier and inflates the score
    (measured: 55% on a prefix-60 vs the paper's 15.10% over the full set).
    When no explicit tier is requested we take a *stratified* sample: equal numbers
    from each num_cities tier, so the subset matches the full benchmark's mix.
    """
    with open(os.path.join(DATA, "trip_planning.json")) as f:
        data = json.load(f)
    items = list(data.values())
    if num_cities is not None:
        # single-tier request (e.g. the Claim 5 City=3 ablation): prefix is fine,
        # every example in the tier is equivalent for sampling purposes
        items = [i for i in items if i["num_cities"] == str(num_cities)]
        if limit:
            items = items[:limit]
    elif limit:
        tiers = sorted({i["num_cities"] for i in items}, key=int)
        per = limit // len(tiers)
        rng = random.Random(seed)
        picked = []
        for t in tiers:
            pool = [i for i in items if i["num_cities"] == t]
            picked.extend(rng.sample(pool, min(per, len(pool))))
        # top up any remainder deterministically from the unpicked pool
        if len(picked) < limit:
            rest = [i for i in items if i not in picked]
            picked.extend(rng.sample(rest, limit - len(picked)))
        items = picked
    # 2-shot prompt, exactly as DreamLM/Dream eval_planning.py::eval_trip
    prompts = []
    for i in items:
        splits = i["prompt_5shot"].split("TASK:")
        prompts.append("TASK:".join(splits[:3] + [splits[-1]]))
    return items, prompts


def load_humaneval(limit=None, offset=0):
    """HumanEval problems [offset, offset+limit).

    The offset lets us EXTEND an existing run instead of repeating it: having
    already scored problems 0..31, we buy problems 32..63 and merge to n=64 for
    the price of the increment rather than re-running the whole prefix.
    """
    from datasets import load_dataset
    ds = load_dataset("openai/openai_humaneval", split="test")
    items = list(ds)[offset:]
    if limit:
        items = items[:limit]
    return items


def humaneval_prompt(tok, doc):
    """lm_eval humaneval_instruct: user doc_to_text + assistant gen_prefix, continued."""
    user = ("Write a solution to the following problem and make sure that it "
            "passes the tests:\n```" + doc["prompt"])
    prefix = "Here is the completed function:\n```python\n" + doc["prompt"] + "\n"
    chat = [{"role": "user", "content": user},
            {"role": "assistant", "content": prefix}]
    return tok.apply_chat_template(chat, tokenize=False,
                                   add_generation_prompt=False,
                                   continue_final_message=True)


# ---------------------------------------------------------------- scoring

def score_trip(items, responses):
    scores = []
    for item, r in zip(items, responses):
        r = r.split("<|endoftext|>")[0].split("\n\nTASK")[0]
        scores.append(compute_example_score(item["cities"], item["durations"],
                                            parse_response(r)))
    return 100.0 * sum(scores) / len(scores), scores


def score_humaneval(items, responses):
    """pass@1 via subprocess exec, mirroring lm_eval's code_eval + build_predictions_instruct."""
    import subprocess, tempfile
    scores = []
    for doc, r in zip(items, responses):
        for stop in ["\nclass", "\ndef", "\n#", "\nif", "\nprint"]:
            r = r.split(stop)[0]
        body = r.split("```python\n", 1)[-1].split("```")[0]
        code = sanitize(doc["prompt"] + "\n" + body, doc["entry_point"])
        program = code + "\n" + doc["test"] + "\n" + f"check({doc['entry_point']})\n"
        with tempfile.NamedTemporaryFile("w", suffix=".py", delete=False) as f:
            f.write(program); path = f.name
        try:
            p = subprocess.run([sys.executable, path], capture_output=True, timeout=15)
            scores.append(1.0 if p.returncode == 0 else 0.0)
        except subprocess.TimeoutExpired:
            scores.append(0.0)
        finally:
            os.unlink(path)
    return 100.0 * sum(scores) / len(scores), scores


# ---------------------------------------------------------------- main

class _NoMaskLogits(torch.nn.Module):
    """Dry-run only: forbid emitting <|mask|> as a clean-data prediction.

    A trained Dream never does; a randomly initialised tiny model does, which
    re-masks positions and hides real bugs behind noise.
    """
    def __init__(self, inner, mask_id):
        super().__init__()
        self.inner, self.mask_id = inner, mask_id
        self.config = inner.config

    def forward(self, *a, **kw):
        out = self.inner(*a, **kw)
        out.logits[..., self.mask_id] = -1e4
        return out


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--task", choices=["trip", "humaneval"], required=True)
    ap.add_argument("--method", choices=["baseline", "ccd", "ccd_ds"], required=True)
    ap.add_argument("--limit", type=int, default=None)
    ap.add_argument("--offset", type=int, default=0,
                    help="skip the first N problems (HumanEval; lets a run extend an earlier one)")
    ap.add_argument("--num-cities", type=int, default=None)
    ap.add_argument("--buffer-V", type=int, default=4)
    ap.add_argument("--history-d", type=int, default=3)
    ap.add_argument("--temperature", type=float, default=None)
    ap.add_argument("--top-p", type=float, default=None,
                    help="override top_p; 1.0 disables nucleus filtering")
    ap.add_argument("--steps", type=int, default=None)
    ap.add_argument("--out", required=True)
    ap.add_argument("--seed", type=int, default=0,
                    help="RNG seed for temperature sampling. Only bites when temperature>0: "
                         "at T=0 ccd_decode._pick_token takes the argmax, so those runs are "
                         "deterministic with or without it. Does NOT affect which examples are "
                         "evaluated -- load_trip's stratified sample is fixed at seed 0 so every "
                         "arm scores the same examples.")
    ap.add_argument("--tiny", action="store_true",
                    help="local CPU dry-run with tiny random weights")
    args = ap.parse_args()

    # Seed every RNG that can touch decoding. NOTE: the temperature-sweep results
    # shipped in outputs/ (c6_*_t0.1/0.4/0.7/1.0) were produced BEFORE this was
    # added and therefore do not reproduce bit-exactly from this seed; the T=0
    # runs (Trip Plan, HumanEval, the ablations) are argmax and always did.
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    torch.cuda.manual_seed_all(args.seed)

    cfg = dict(TASK_DEFAULTS[args.task])
    if args.temperature is not None:
        cfg["temperature"] = args.temperature
    if args.top_p is not None:
        cfg["top_p"] = args.top_p
    if args.steps is not None:
        cfg["steps"] = args.steps; cfg["max_new_tokens"] = args.steps

    print(f"[cfg] task={args.task} method={args.method} V={args.buffer_V} d={args.history_d} "
          f"seed={args.seed} {cfg}", flush=True)

    tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
    if args.tiny:
        # Local CPU dry-run: real tokenizer/prompts/metrics, tiny random weights.
        # Exercises the whole pipeline end-to-end so no GPU minute is spent on a crash.
        from transformers import AutoConfig
        cfg_m = AutoConfig.from_pretrained(MODEL, trust_remote_code=True)
        cfg_m.num_hidden_layers, cfg_m.hidden_size = 2, 128
        cfg_m.intermediate_size, cfg_m.num_attention_heads = 256, 4
        cfg_m.num_key_value_heads, cfg_m.tie_word_embeddings = 2, True
        model = AutoModel.from_config(cfg_m, trust_remote_code=True).eval()
        model = _NoMaskLogits(model, cfg_m.mask_token_id)
        device = "cpu"
    else:
        model = AutoModel.from_pretrained(MODEL, torch_dtype=torch.bfloat16,
                                          trust_remote_code=True).to("cuda").eval()
        device = "cuda"
    mask_id = model.config.mask_token_id

    if args.task == "trip":
        items, prompts = load_trip(args.limit, args.num_cities)
    else:
        items = load_humaneval(args.limit, args.offset)
        prompts = [humaneval_prompt(tok, d) for d in items]

    responses, all_steps, all_budgets, n_fallback = [], [], [], 0
    all_ic, all_stable = [], []
    t0 = time.time()
    for n, prompt in enumerate(prompts):
        enc = tok(prompt, return_tensors="pt")
        input_ids = enc.input_ids.to(device)
        attn = enc.attention_mask.to(device)
        x, st = ccd_decode.generate(
            model, input_ids, attention_mask=attn,
            max_new_tokens=cfg["max_new_tokens"], steps=cfg["steps"],
            temperature=cfg["temperature"], top_p=cfg["top_p"],
            mask_token_id=mask_id, method=args.method,
            buffer_V=args.buffer_V, history_d=args.history_d,
        )
        gen = x[0, input_ids.shape[1]:]
        text = tok.decode(gen.tolist()).split(tok.eos_token)[0]
        responses.append(text)
        all_steps.append(st["steps"])
        all_budgets.append(st["budgets"])
        all_ic.append(st["ic_sizes"])
        all_stable.append(st["n_stable"])
        n_fallback += st["fallbacks"]
        if n == 0:
            print(f"--- example 0 response ---\n{text[:600]}\n---", flush=True)
        if (n + 1) % 10 == 0:
            el = time.time() - t0
            print(f"[{n+1}/{len(prompts)}] mean_steps={sum(all_steps)/len(all_steps):.2f} "
                  f"elapsed={el/60:.1f}min eta={el/(n+1)*(len(prompts)-n-1)/60:.1f}min", flush=True)

    if args.task == "trip":
        score, per_ex = score_trip(items, responses)
    else:
        score, per_ex = score_humaneval(items, responses)

    mean_steps = sum(all_steps) / len(all_steps)
    result = {
        "task": args.task, "method": args.method, "config": cfg,
        "buffer_V": args.buffer_V, "history_d": args.history_d,
        "n_examples": len(items), "score": score, "offset": args.offset, "seed": args.seed,
        "mean_steps": mean_steps,
        "speedup_vs_uniform": cfg["steps"] / mean_steps,
        "fallback_steps": n_fallback,
        "wall_clock_s": time.time() - t0,
        "per_example_score": per_ex, "per_example_steps": all_steps,
        "responses": responses,
        "budgets": all_budgets,
        "ic_sizes": all_ic, "n_stable": all_stable,
    }
    os.makedirs(os.path.dirname(args.out) or ".", exist_ok=True)
    with open(args.out, "w") as f:
        json.dump(result, f, indent=1)

    print(f"\n=== RESULT {args.task}/{args.method} V={args.buffer_V} d={args.history_d} ===")
    print(f"score={score:.2f}  mean_steps={mean_steps:.2f}  "
          f"speedup={cfg['steps']/mean_steps:.2f}x  n={len(items)}  "
          f"fallback_steps={n_fallback}  wall={result['wall_clock_s']/60:.1f}min")


if __name__ == "__main__":
    main()