File size: 11,203 Bytes
fa82a9d | 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 | #!/usr/bin/env python3
"""Batched, greedy, frozen-base/SFT ablations with isolated Scene loading and full denominators."""
from __future__ import annotations
import argparse
from contextlib import nullcontext
import hashlib
import json
import os
from pathlib import Path
import sys
import time
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / 'src'))
from vbench_prompts_compile.inference import AdapterRouter, SceneVerifier, parse_output, validate_direction_output
from vbench_prompts_compile.training import user_text
def sha(path):
digest = hashlib.sha256()
with Path(path).open('rb') as handle:
for block in iter(lambda: handle.read(8 * 1024 * 1024), b''):
digest.update(block)
return digest.hexdigest()
def jobs_for(rows):
jobs = []
scene_caps = sorted({r['input']['caption'] for r in rows if r['task'] == 'scene'})
for mode in ('base', 'sft'):
for row in rows:
caption_modes = ['intact']
if row['task'] == 'scene' and (row['variant'] == 'canonical' or row['source'] == 'snli_test_locations'):
caption_modes += ['missing', 'shuffled']
for caption_mode in caption_modes:
inp = dict(row['input'])
if caption_mode == 'missing':
inp['caption'] = ''
elif caption_mode == 'shuffled':
i = scene_caps.index(inp['caption'])
inp['caption'] = scene_caps[(i + 1) % len(scene_caps)]
jobs.append({'key': f'{row["id"]}/{mode}/{caption_mode}', 'id': row['id'],
'task': row['task'], 'mode': mode, 'caption_mode': caption_mode, 'input': inp})
return jobs
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument('--data', default='data/ablation-v1/eval.jsonl')
p.add_argument('--config', required=True)
p.add_argument('--out', required=True)
p.add_argument('--batch-size', type=int, default=8)
p.add_argument('--input-budget', type=int, default=1792)
args = p.parse_args()
import gc
import torch
torch.set_num_threads(4)
torch.manual_seed(20260919)
cfg = json.loads(Path(args.config).read_text())
rows = [json.loads(line) for line in Path(args.data).read_text().splitlines()]
out = Path(args.out)
out.parent.mkdir(parents=True, exist_ok=True)
weights = {task: {name: sha(Path(path) / name) for name in ('adapter_model.safetensors', 'adapter_config.json', 'training_config.json')}
for task, path in cfg['adapters'].items()}
base_weights = {p.name: sha(p) for p in sorted(Path(cfg['model']).glob('model*.safetensors'))}
if not base_weights:
raise SystemExit('no local safetensors base weights found')
plan = {'data_sha256': sha(args.data), 'config': cfg, 'runner_sha256': sha(__file__),
'inference_sha256': sha(ROOT / 'src/vbench_prompts_compile/inference.py'),
'training_sha256': sha(ROOT / 'src/vbench_prompts_compile/training.py'),
'adapter_sha256': weights, 'base_config_sha256': sha(Path(cfg['model']) / 'config.json'),
'base_weight_sha256': base_weights, 'lock_sha256': sha(ROOT / 'uv.lock'),
'tokenizer_sha256': sha(Path(cfg['model']) / 'tokenizer.json'),
'batch_size': args.batch_size, 'input_budget': args.input_budget,
'parse_max_new_tokens': 128, 'scene_max_new_tokens': 8, 'seed': 20260919,
'decoding': 'greedy, non-thinking, bf16, left padding; Scene newline boundary',
'torch_version': torch.__version__, 'gpu': torch.cuda.get_device_name(),
'cuda_visible_devices': os.environ.get('CUDA_VISIBLE_DEVICES'),
'gpu_total_memory_bytes': torch.cuda.get_device_properties(0).total_memory}
plan_path = out.with_suffix('.plan.json')
if plan_path.exists():
if json.loads(plan_path.read_text()) != plan:
raise SystemExit('frozen inference plan mismatch')
else:
if out.exists():
raise SystemExit('predictions without plan')
plan_path.write_text(json.dumps(plan, indent=2) + '\n')
complete = {}
if out.exists():
for line in out.read_text().splitlines():
r = json.loads(line)
if r['key'] in complete:
raise SystemExit('duplicate result key')
complete[r['key']] = r
all_jobs = jobs_for(rows)
resource_path = out.with_suffix('.resources.jsonl')
with out.open('a') as output, resource_path.open('a') as resources:
for system in ('parse', 'scene'):
tasks = ['spatial', 'action', 'objects'] if system == 'parse' else ['scene']
pending = [j for j in all_jobs if j['task'] in tasks and j['key'] not in complete]
if not pending:
continue
torch.cuda.reset_peak_memory_stats()
loaded_at = time.monotonic()
if system == 'parse':
engine = AdapterRouter(base_model_path=cfg['model'], adapters={t: cfg['adapters'][t] for t in tasks})
else:
engine = SceneVerifier(base_model_path=cfg['model'], adapter_path=cfg['adapters']['scene'])
tokenizer, model = engine.tokenizer, engine.model
tokenizer.padding_side = 'left'
model.eval()
config = model.config.to_dict()
resources.write(json.dumps({'event': 'load', 'system': system, 'seconds': time.monotonic() - loaded_at,
'parameters_with_loaded_adapters': sum(p.numel() for p in model.parameters()),
'architecture': {k: config.get(k) for k in ('architectures', 'hidden_size', 'intermediate_size',
'num_hidden_layers', 'num_attention_heads', 'num_key_value_heads',
'head_dim', 'vocab_size', 'tie_word_embeddings')}}) + '\n')
resources.flush()
for task in tasks:
for mode in ('base', 'sft'):
if system == 'parse':
engine.route(task)
jobs = [j for j in pending if j['task'] == task and j['mode'] == mode]
for j in jobs:
user = user_text({'task': task, 'input': j['input']}, spatial_relation_space='four_directions')
j['text'] = tokenizer.apply_chat_template([{'role': 'user', 'content': user}], tokenize=False,
add_generation_prompt=True, enable_thinking=False)
j['input_tokens'] = len(tokenizer.encode(j['text']))
jobs.sort(key=lambda j: (j['input_tokens'], j['key']))
for start in range(0, len(jobs), args.batch_size):
batch = jobs[start:start + args.batch_size]
overflow = [j for j in batch if j['input_tokens'] > args.input_budget]
batch = [j for j in batch if j['input_tokens'] <= args.input_budget]
results = [{**{k: v for k, v in j.items() if k not in ('text', 'input')}, 'raw': '', 'value': None,
'errors': ['input_over_budget'], 'seconds_per_item_amortized': None} for j in overflow]
if batch:
inputs = tokenizer([j['text'] for j in batch], return_tensors='pt', padding=True)
inputs.pop('token_type_ids', None)
inputs = inputs.to(model.device)
torch.cuda.synchronize()
t = time.monotonic()
new_tokens = 8 if task == 'scene' else 128
kwargs = {'stop_strings': ['\n'], 'tokenizer': tokenizer} if task == 'scene' else {}
with torch.inference_mode(), (model.disable_adapter() if mode == 'base' else nullcontext()):
generated = model.generate(**inputs, max_new_tokens=new_tokens, do_sample=False, num_beams=1,
temperature=None, top_p=None, top_k=None,
pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id, **kwargs)
torch.cuda.synchronize()
elapsed = time.monotonic() - t
tokens = generated[:, inputs['input_ids'].shape[1]:]
raws = tokenizer.batch_decode(tokens, skip_special_tokens=True)
for j, raw, token_ids in zip(batch, raws, tokens.tolist()):
value, errors = parse_output(task, raw.strip())
ended = tokenizer.eos_token_id in token_ids or (task == 'scene' and '\n' in raw)
if not ended and len(token_ids) >= new_tokens:
value, errors = None, tuple(errors) + ('output_budget_exhausted',)
if task == 'spatial' and value is not None:
value, more = validate_direction_output(value)
errors = tuple(errors) + more
results.append({**{k: v for k, v in j.items() if k not in ('text', 'input')},
'raw': raw.strip(), 'value': value, 'errors': list(errors),
'generated_tokens_including_padding': len(token_ids),
'seconds_per_item_amortized': elapsed / len(batch)})
resources.write(json.dumps({'event': 'batch', 'task': task, 'mode': mode, 'items': len(batch),
'seconds': elapsed, 'peak_allocated_bytes': torch.cuda.max_memory_allocated()}) + '\n')
resources.flush()
for result in results:
output.write(json.dumps(result, ensure_ascii=False) + '\n')
complete[result['key']] = result
output.flush()
print(json.dumps({'system': system, 'task': task, 'mode': mode, 'complete': len(complete)}), flush=True)
del engine, model, tokenizer
gc.collect()
torch.cuda.empty_cache()
expected = {j['key'] for j in all_jobs}
missing, extra = expected - complete.keys(), complete.keys() - expected
report = {'planned': len(expected), 'completed': len(complete), 'missing': sorted(missing), 'extra': sorted(extra),
'predictions_sha256': sha(out), 'resources_sha256': sha(resource_path),
'plan_sha256': sha(plan_path), 'complete': not missing and not extra}
out.with_suffix('.report.json').write_text(json.dumps(report, indent=2) + '\n')
print(json.dumps(report), flush=True)
if __name__ == '__main__':
main()
|