Spaces:
Sleeping
Sleeping
File size: 9,676 Bytes
3afc977 | 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 | """Evaluate by execution and produce the accuracy-vs-latency curve (EVALUATION.md).
For each eval record the model predicts the hole; we reconstruct prefix + hole +
suffix and verify by execution against the held-out tests (pass@1). Latency is
wall-clock at batch 1. The diffusion model is swept over denoising steps N (the
test-time-compute knob); the AR baseline is a single sequential decode. Results
are grouped by difficulty so we can see where, at iso-latency, the diffusion
substrate wins.
Usage:
python -m ml.evaluate --diff runs/diff --ar runs/ar --eval data/eval.jsonl \
--steps 1,2,4,8,16,32 --out runs/curve.csv
"""
from __future__ import annotations
import argparse
import csv
import time
import torch
from . import ar, block_diffusion
from .ar import build_prompt
import numpy as np
from .config import ModelConfig, TaskConfig
from .data import (_ids_canvas, block_for_eval, encode_ar, encode_diffusion,
load_records, make_block, make_block_lua)
from .model import Transformer, amp_ctx
from .tokenizer import Tokenizer
from .verify import verify_batch
def load_model(path: str, device: str):
ckpt = torch.load(f"{path}/model.pt", map_location=device, weights_only=False)
mcfg = ModelConfig(**ckpt["model_cfg"])
model = Transformer(mcfg, causal=(ckpt["mode"] == "ar")).to(device)
model.load_state_dict(ckpt["model"])
model.eval()
tok = Tokenizer.load(f"{path}/tokenizer.json")
task = TaskConfig(**ckpt["task_cfg"])
return model, tok, task, ckpt["mode"]
def pick_device() -> str:
if torch.backends.mps.is_available():
return "mps"
if torch.cuda.is_available():
return "cuda"
return "cpu"
def rec_seed(rec) -> int:
"""Stable per-record seed for the eval block (independent of list position)."""
return rec.get("seed", 0) % 2147483647
def get_block(tok, source, frac, seed):
"""(pre, blk, suf) for the record — id lists in lua mode, strings in char."""
rng = np.random.RandomState(seed + 1)
return make_block_lua(source, frac, rng, tok) if tok.mode == "lua" else make_block(source, frac, rng)
def reconstruct(tok, blk, pred_ids):
"""Full program from prefix + predicted block + suffix."""
if tok.mode == "lua":
return tok.decode(list(blk[0]) + list(pred_ids) + list(blk[2]))
return blk[0] + tok.decode(pred_ids) + blk[2]
def eval_diffusion(model, tok, task, records, n_inner, device, frac):
"""Block-diffusion eval at one n_inner setting (latency knob)."""
cands, lats, kept = [], [], []
for i, rec in enumerate(records):
blk = get_block(tok, rec["source"], frac, rec_seed(rec))
if blk is None:
continue
enc = _ids_canvas(tok, blk[0], blk[1], blk[2], task, ar=False) if tok.mode == "lua" \
else encode_diffusion(tok, blk[0], blk[1], blk[2], task)
if enc is None:
continue
ids, region, _block_id, attn = enc
ids_row = torch.from_numpy(ids).to(device)
region_row = torch.from_numpy(region).to(device)
attn_row = torch.from_numpy(attn).to(device)
t0 = time.perf_counter()
with amp_ctx(device):
toks = block_diffusion.sample(model, ids_row, region_row, attn_row, tok, task, n_inner)
if device == "mps":
torch.mps.synchronize()
lat = (time.perf_counter() - t0) * 1000.0
cands.append({"source": reconstruct(tok, blk, toks), "tests": rec["tests"]})
lats.append(lat)
kept.append(i)
return cands, lats, kept
def eval_ar(model, tok, task, records, device, frac):
cands, lats, kept = [], [], []
for i, rec in enumerate(records):
blk = get_block(tok, rec["source"], frac, rec_seed(rec))
if blk is None:
continue
head = build_prompt(tok, blk[0], blk[2], task)
if head is None:
continue
head_ids = torch.tensor(head, device=device)
t0 = time.perf_counter()
with amp_ctx(device):
ids_out = model.generate(head_ids, max_new=task.max_decode, eos_id=tok.eos_id)
if device == "mps":
torch.mps.synchronize()
lat = (time.perf_counter() - t0) * 1000.0
cands.append({"source": reconstruct(tok, blk, ids_out), "tests": rec["tests"]})
lats.append(lat)
kept.append(i)
return cands, lats, kept
def summarize(rows, records, kept, passes, lats, model_name, steps, frac):
"""Group pass@1 and latency by difficulty, for one masking fraction."""
by_diff = {}
for k, p, lat in zip(kept, passes, lats):
d = records[k]["difficulty"]
by_diff.setdefault(d, []).append((p, lat))
for d in sorted(by_diff):
ps = [p for p, _ in by_diff[d]]
ls = [lat for _, lat in by_diff[d]]
rows.append({
"model": model_name, "frac": frac, "steps": steps, "difficulty": d,
"n": len(ps), "pass@1": round(sum(ps) / len(ps), 4),
"mean_latency_ms": round(sum(ls) / len(ls), 2),
})
rows.append({
"model": model_name, "frac": frac, "steps": steps,
"difficulty": "all", "n": len(passes),
"pass@1": round(sum(passes) / max(1, len(passes)), 4),
"mean_latency_ms": round(sum(lats) / max(1, len(lats)), 2),
})
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--diff", help="diffusion run dir")
ap.add_argument("--ar", help="AR run dir")
ap.add_argument("--eval", default="data/eval.jsonl")
ap.add_argument("--steps", default="1,2,4,8", help="block-diffusion n_inner values to sweep")
ap.add_argument("--fracs", default="0.15,0.3,0.45", help="block fractions to sweep")
ap.add_argument("--tile_size", type=int, default=0, help="override remask tile size (0=keep checkpoint)")
ap.add_argument("--limit", type=int, default=0, help="cap eval records (0=all)")
ap.add_argument("--out", default="runs/curve.csv")
args = ap.parse_args()
device = pick_device()
all_records = load_records(args.eval)
if args.limit:
all_records = all_records[: args.limit]
print(f"eval records={len(all_records)} device={device}")
def warmup(model, task):
ids = torch.zeros(1, task.seq_len, dtype=torch.long, device=device)
keep = torch.ones(1, task.seq_len, dtype=torch.bool, device=device)
with amp_ctx(device):
model(ids, keep)
if device == "mps":
torch.mps.synchronize()
rows = []
step_list = [int(s) for s in args.steps.split(",") if s]
frac_list = [float(x) for x in args.fracs.split(",") if x]
# Load both models once; reuse across fractions.
def warmup_block(model, tok, task):
# Warm the cached block-decode kernels so the first timed record is clean.
ids = torch.zeros(task.seq_len, dtype=torch.long, device=device)
region = torch.zeros(task.seq_len, dtype=torch.bool, device=device)
region[4 : 4 + 2 * task.block_len] = True
attn = torch.ones(task.seq_len, dtype=torch.bool, device=device)
with amp_ctx(device):
block_diffusion.sample(model, ids, region, attn, tok, task, 2)
if device == "mps":
torch.mps.synchronize()
dmodel = dtok = dtask = None
if args.diff:
dmodel, dtok, dtask, m = load_model(args.diff, device)
assert m == "diffusion"
if args.tile_size:
dtask.tile_size = args.tile_size
warmup(dmodel, dtask)
warmup_block(dmodel, dtok, dtask)
amodel = atok = atask = None
if args.ar:
amodel, atok, atask, m = load_model(args.ar, device)
assert m == "ar"
warmup(amodel, atask)
tok0 = dtok if dtok is not None else atok
task0 = dtask if dtask is not None else atask
for frac in frac_list:
# Common-fit at THIS fraction: both encodings must fit the same records.
def fits_both(rec):
blk = get_block(tok0, rec["source"], frac, rec_seed(rec))
if blk is None:
return False
if tok0.mode == "lua":
d = _ids_canvas(tok0, blk[0], blk[1], blk[2], task0, ar=False)
a = _ids_canvas(tok0, blk[0], blk[1], blk[2], task0, ar=True)
else:
d = encode_diffusion(tok0, blk[0], blk[1], blk[2], task0)
a = encode_ar(tok0, blk[0], blk[1], blk[2], task0)
return d is not None and a is not None
records = [r for r in all_records if fits_both(r)] if (args.diff and args.ar) else all_records
print(f"\n--- frac={frac} common-fit={len(records)} ---")
if args.diff:
for N in step_list:
cands, lats, kept = eval_diffusion(dmodel, dtok, dtask, records, N, device, frac)
passes = verify_batch(cands)
summarize(rows, records, kept, passes, lats, "diffusion", N, frac)
o = rows[-1]
print(f" diffusion n_inner={N:>2} pass@1={o['pass@1']:.3f} lat={o['mean_latency_ms']:.1f}ms")
if args.ar:
cands, lats, kept = eval_ar(amodel, atok, atask, records, device, frac)
passes = verify_batch(cands)
summarize(rows, records, kept, passes, lats, "ar", "NA", frac)
o = rows[-1]
print(f" ar pass@1={o['pass@1']:.3f} lat={o['mean_latency_ms']:.1f}ms")
with open(args.out, "w", newline="") as f:
w = csv.DictWriter(f, fieldnames=["model", "frac", "steps", "difficulty", "n", "pass@1", "mean_latency_ms"])
w.writeheader()
w.writerows(rows)
print(f"\nwrote {args.out}")
if __name__ == "__main__":
main()
|