File size: 14,091 Bytes
4968ea3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 | """Direct-ask cartridge eval for the trainable-KV length ablation.
Pure "inject cartridge -> ask the user's OWN question -> greedy answer -> judge".
NO retrieval, NO MS/ACT decision pipeline, NO SFT LoRA. This isolates the question
"how much did the trainable-KV cartridge actually memorize?" from the rest of the
MetaMem machinery, so accuracy reflects ONLY the cartridge at a given KV length p.
One (dataset, p) cell per invocation; ALL outputs land under ISOLATED paths keyed by
(dataset, p) — see run_kvlen_cell.sh. Reuses the FIR-probing generation path and the
production gpt-4o-mini LLMJudge so the answer/judge口径 matches the rest of the repo.
Usage:
python scripts/train/eval_direct_ask.py \
--dataset data/processed/longmemeval_s/dataset.json \
--users-list data/processed/longmemeval_s/splits/kvlen_sample200_seed42.json \
--cartridge-dir checkpoints/kvlen/lmes/p256 \
--output-path data/eval/kvlen/lmes/p256/preds.jsonl \
--judge-cache-path data/cache/kvlen/lmes/p256/llm_judge.jsonl \
--model-path /mnt/train-gui-agent/zhangzeyu/models/Qwen2.5-7B-Instruct
# generation only (no API), then judge separately:
python scripts/train/eval_direct_ask.py ... --no-judge --stage infer
python scripts/train/eval_direct_ask.py ... --stage eval
"""
import argparse
import json
import os
import sys
from collections import defaultdict
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
os.environ.setdefault("CARTRIDGES_DIR", os.path.join(PROJECT_ROOT, "cartridges-lib"))
os.environ.setdefault("CARTRIDGES_OUTPUT_DIR", os.path.join(PROJECT_ROOT, "checkpoints/cartridge"))
from src.utils import load_json, save_json, set_seed, setup_logger
logger = setup_logger(__name__)
DEFAULT_MODEL_PATH = "/mnt/train-gui-agent/zhangzeyu/models/Qwen2.5-7B-Instruct"
DEFAULT_MAX_NEW_TOKENS = 256
def _r(p):
return p if os.path.isabs(p) else os.path.join(PROJECT_ROOT, p)
# --------------------------------------------------------------------------------------
# Inference: inject cartridge, greedily answer the user's own questions
# --------------------------------------------------------------------------------------
def _generate_answer(model, tokenizer, cache, query, max_new_tokens):
"""Greedy (temp=0 -> argmax) answer with the cartridge as a CONSTANT KV prefix.
Mirrors src/data_construction/fir_probing.py::_generate_samples_cartridge but
single-sample at temperature 0 (deterministic). flex_generate already clears the
transient per-token KV at the end; we clear again as belt-and-suspenders so no KV
bleeds across queries.
"""
import torch
from cartridges.generation import flex_generate
from src.data_construction.fir_probing import format_probe_prompt
prompt = format_probe_prompt(query, tokenizer, mode="answer", enable_thinking=False)
input_ids = tokenizer.encode(prompt, add_special_tokens=False)
device = next(model.parameters()).device
input_ids = torch.tensor(input_ids, device=device)
seq_ids = torch.zeros_like(input_ids)
position_ids = torch.arange(len(input_ids), device=device)
out = flex_generate(
model=model, tokenizer=tokenizer,
input_ids=input_ids, seq_ids=seq_ids, position_ids=position_ids,
cache=cache, max_new_tokens=max_new_tokens, temperature=0.0, # GREEDY
) # Dict[seq_id, List[token_id]]
cache.clear()
tok_ids = list(out.values())[0] if out else []
return tokenizer.decode(tok_ids, skip_special_tokens=True).strip()
def run_inference(cfg):
import torch
from transformers import AutoTokenizer
from cartridges.cache import TrainableCache
from src.cartridge.model_factory import get_flex_model_cls
from src.evaluation.metrics.answer_metrics import judge_answer
from src.utils.cartridge_utils import find_cartridge_path
dataset = load_json(_r(cfg["dataset"]))
user_ids = load_json(_r(cfg["users_list"]))
cartridge_dir = _r(cfg["cartridge_dir"])
out_path = _r(cfg["output_path"])
max_new_tokens = cfg.get("max_new_tokens", DEFAULT_MAX_NEW_TOKENS)
model_path = cfg.get("model_path", DEFAULT_MODEL_PATH)
user_map = {ud["user_id"]: ud for ud in dataset}
# Fail loud if the sampled users aren't in this dataset (wrong --dataset path).
if sum(1 for u in user_ids if u in user_map) == 0:
raise ValueError(
f"None of the {len(user_ids)} sampled users exist in {cfg['dataset']}. "
f"Wrong --dataset? (lmes -> longmemeval_s, metamem5k -> metamem_5k)"
)
# Resume: skip query_ids already written.
done_qids = set()
if os.path.exists(out_path):
with open(out_path) as f:
for line in f:
line = line.strip()
if line:
try:
done_qids.add(json.loads(line)["query_id"])
except Exception:
pass
logger.info(f"Resuming: {len(done_qids)} queries already generated")
# Load model ONCE; swap the cartridge per user.
# 🔴 device_map={"": 0} forces the WHOLE model onto a single GPU — do NOT use "auto".
# The eval step runs with ALL 8 GPUs visible (unlike the per-GPU-pinned training
# steps), so "auto" would shard the 7B across cuda:0..7, and the cartridge attention
# path isn't model-parallel-safe (rotary cos/sin don't follow each layer's shard ->
# "tensors on cuda:1 and cuda:6" crash in apply_rotary_pos_emb). A 7B in bf16 (~15GB)
# fits on one card and serial greedy decode needs no sharding. The cartridge cache is
# loaded with device="cuda" (current device = GPU 0), so model + cache + inputs all
# land on the same device. Pin a specific GPU via CUDA_VISIBLE_DEVICES if GPU 0 is busy.
model_cls = get_flex_model_cls(model_path)
logger.info(f"Loading {model_cls.__name__} from {model_path} (single GPU)...")
model = model_cls.from_pretrained(
model_path, torch_dtype=torch.bfloat16, device_map={"": 0}
)
model.eval()
tokenizer = AutoTokenizer.from_pretrained(model_path)
os.makedirs(os.path.dirname(out_path), exist_ok=True)
f = open(out_path, "a", encoding="utf-8")
n_done = 0
n_no_cart = 0
n_stage1_only = 0
for ui, uid in enumerate(user_ids):
ud = user_map.get(uid)
if ud is None:
logger.warning(f"[{ui+1}/{len(user_ids)}] user {uid} not in dataset, skip")
continue
cpath = find_cartridge_path(cartridge_dir, uid)
if cpath is None:
n_no_cart += 1
logger.warning(f"[{ui+1}/{len(user_ids)}] user {uid}: no cartridge, skip")
continue
# REFUSE Stage-1-only fallback: with --skip-probing there's no done.flag gate, so
# find_cartridge_path may silently return the (weaker) ntp_recon_stage1 cartridge.
# Evaluating that would mix Stage-1-only results into this p-cell's accuracy.
# Require the Stage-2 CD warmstart cartridge; otherwise treat as "no cartridge".
if "warmstart" not in os.path.basename(os.path.dirname(cpath)):
n_stage1_only += 1
logger.warning(
f"[{ui+1}/{len(user_ids)}] user {uid}: only Stage-1 cartridge "
f"({cpath}) — refusing fallback, skip (Stage-2 warmstart missing)"
)
continue
cache = TrainableCache.from_pretrained(cpath, device="cuda").to("cuda")
for q in ud["queries"]:
qid = q["id"]
if qid in done_qids:
continue
ans = _generate_answer(model, tokenizer, cache, q["query"], max_new_tokens)
em, _partial, f1 = judge_answer(ans, q["answer"], [q["answer"]])
rec = {
"query_id": qid, "user_id": uid,
"query_type": q.get("query_type"),
"oracle_ms": q.get("ms_label"), # None for metamem_5k; present for lmes
"act": "DIRECT", # required key for downstream aggregators
"query_text": q["query"], "answer": ans, "gold": q["answer"],
"answer_em": bool(em), "answer_f1": float(f1),
"answer_judge": None, # filled in the judge stage
"cartridge_path": cpath,
}
f.write(json.dumps(rec, ensure_ascii=False) + "\n")
f.flush()
n_done += 1
if n_done % 50 == 0:
logger.info(f" generated {n_done} answers (user {ui+1}/{len(user_ids)})")
# free per-user VRAM before next cartridge
del cache
torch.cuda.empty_cache()
f.close()
logger.info(
f"Inference done: {n_done} new answers | {n_no_cart} users w/o cartridge | "
f"{n_stage1_only} users Stage-1-only (skipped)"
)
return out_path
# --------------------------------------------------------------------------------------
# Judge: batched gpt-4o-mini verdicts (correct/partial/wrong)
# --------------------------------------------------------------------------------------
def run_judge(cfg):
from src.evaluation.metrics.llm_judge import build_judge_from_cfg
out_path = _r(cfg["output_path"])
recs = []
with open(out_path) as fh:
for line in fh:
line = line.strip()
if line:
recs.append(json.loads(line))
if not cfg.get("no_judge", False):
judge = build_judge_from_cfg(
{"enabled": True, "model": cfg.get("judge_model", "gpt-4o-mini"),
"cache_path": cfg["judge_cache_path"]},
PROJECT_ROOT,
)
if judge is not None:
logger.info(f"Judging {len(recs)} answers with {judge.model} (batched)...")
items = [{"question": r["query_text"], "gold": r["gold"],
"answer": r["answer"], "qtype": r.get("query_type") or ""}
for r in recs]
for r, (v, ok) in zip(recs, judge.judge_batch(items)):
r["answer_judge"] = v if ok else None
# rewrite jsonl atomically with verdicts filled in
tmp = out_path + ".tmp"
with open(tmp, "w", encoding="utf-8") as g:
for r in recs:
g.write(json.dumps(r, ensure_ascii=False) + "\n")
os.replace(tmp, out_path)
stats = aggregate_direct(recs)
stats_path = out_path.replace(".jsonl", ".stats.json")
save_json(stats, stats_path)
logger.info(f"Aggregated {len(recs)} records -> {stats_path}")
print(json.dumps(stats, ensure_ascii=False, indent=2))
return stats
def aggregate_direct(recs):
"""Bespoke aggregator — only the metrics that MEAN something for direct-ask.
The stock src/evaluation/eval_metrics.aggregate() emits null/misleading retrieval +
strategy fields here (every record is act="DIRECT", metamem_5k has no ms_label), so
we report exactly: n / judge_acc / judge_correct_rate / em / f1, plus the same
broken out by_query_type. verdict->numeric mapping matches eval_metrics._judge_num.
"""
def jn(v):
if isinstance(v, str):
return {"correct": 1.0, "partial": 0.5, "wrong": 0.0}.get(v)
if isinstance(v, bool):
return 1.0 if v else 0.0
return None
def mean(xs):
xs = [x for x in xs if x is not None]
return round(sum(xs) / len(xs), 4) if xs else None
jc = [1.0 if r.get("answer_judge") == "correct" else 0.0
for r in recs if r.get("answer_judge") is not None]
out = {
"n": len(recs),
"n_judged": len(jc),
"judge_acc": mean([jn(r.get("answer_judge")) for r in recs]),
"judge_correct_rate": round(sum(jc) / len(jc), 4) if jc else None,
"em": mean([r["answer_em"] for r in recs]),
"f1": mean([r["answer_f1"] for r in recs]),
}
by = defaultdict(list)
for r in recs:
by[r.get("query_type")].append(r)
out["by_query_type"] = {
str(qt): {
"n": len(rs),
"judge_acc": mean([jn(r.get("answer_judge")) for r in rs]),
"em": mean([r["answer_em"] for r in rs]),
"f1": mean([r["answer_f1"] for r in rs]),
}
for qt, rs in sorted(by.items(), key=lambda kv: str(kv[0]))
}
return out
def main():
ap = argparse.ArgumentParser(description="Direct-ask cartridge eval (KV-length ablation)")
ap.add_argument("--dataset", required=True, help="Path to dataset.json")
ap.add_argument("--users-list", required=True, help="Seeded 200-user split JSON")
ap.add_argument("--cartridge-dir", required=True, help="ISOLATED cartridge_root for this cell")
ap.add_argument("--output-path", required=True, help="ISOLATED preds.jsonl for this cell")
ap.add_argument("--judge-cache-path", required=True, help="ISOLATED judge cache for this cell")
ap.add_argument("--model-path", default=DEFAULT_MODEL_PATH)
ap.add_argument("--max-new-tokens", type=int, default=DEFAULT_MAX_NEW_TOKENS)
ap.add_argument("--judge-model", default="gpt-4o-mini")
ap.add_argument("--no-judge", action="store_true",
help="Skip the LLM judge (no API). EM/F1 + label dist still produced.")
ap.add_argument("--stage", choices=["infer", "eval", "both"], default="both")
ap.add_argument("--seed", type=int, default=42)
args = ap.parse_args()
set_seed(args.seed)
cfg = {
"dataset": args.dataset, "users_list": args.users_list,
"cartridge_dir": args.cartridge_dir, "output_path": args.output_path,
"judge_cache_path": args.judge_cache_path, "model_path": args.model_path,
"max_new_tokens": args.max_new_tokens, "judge_model": args.judge_model,
"no_judge": args.no_judge,
}
if args.stage in ("infer", "both"):
run_inference(cfg)
if args.stage in ("eval", "both"):
run_judge(cfg)
if __name__ == "__main__":
main()
|