File size: 10,862 Bytes
ed9153b | 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 | import json, os, subprocess, sys, hashlib, random, time, gc, collections
from pathlib import Path
CFG = json.loads(Path(sys.argv[1]).read_text("utf-8"))
ROOT = Path(CFG["OUT_ROOT"]); SHARD, NSH = CFG["SHARD_ID"], CFG["N_SHARDS"]
MODE, SEED, TIER = CFG["MODE"], CFG["SEED"], CFG["TIER"]
MODEL_ID = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
LU_SHA = "a98961956072224eaf244eb289d6c01700b63795"
N_PANEL, N_PROMPTS, RESERVE, MINWORK = 80, 5, 4.0, 20.0
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
q = subprocess.run(["nvidia-smi",
"--query-gpu=index,name,memory.total,memory.free,utilization.gpu",
"--format=csv,noheader,nounits"], capture_output=True, text=True, check=True).stdout
gpus = []
for line in q.strip().splitlines():
i, n, tot, free, u = [x.strip() for x in line.split(",")]
gpus.append({"index": int(i), "name": n, "total": int(tot)/1024,
"free": int(free)/1024, "util": int(u)})
print("devices: " + " ".join(f"[{g['index']}] {g['free']:.0f}/{g['total']:.0f}G {g['util']}%"
for g in gpus))
d = (max(gpus, key=lambda g: g["free"]) if CFG["DEVICE_ID"] is None
else next(g for g in gpus if g["index"] == CFG["DEVICE_ID"]))
os.environ["CUDA_VISIBLE_DEVICES"] = str(d["index"])
usable = d["free"] - RESERVE
if usable < MINWORK:
sys.exit(f"device {d['index']}: {usable:.1f} GiB usable, need {MINWORK}")
UTIL = round(max(0.0, min(0.92, usable / d["total"])), 3)
MAXLEN = 4096 if usable >= 26 else 2560
MAX_TOK = min(CFG["MAX_TOK"], MAXLEN - 640)
import torch
cap = torch.cuda.get_device_capability(0)
DTYPE = "bfloat16" if cap[0] >= 8 else "float16"
print(f"{MODE} shard {SHARD}/{NSH} device {d['index']} | usable {usable:.1f} GiB "
f"-> util={UTIL} maxlen={MAXLEN} max_tokens={MAX_TOK}")
vend = ROOT / "vendor" / "assistant-axis"
head = subprocess.run(["git", "-C", str(vend), "rev-parse", "HEAD"],
capture_output=True, text=True, check=True).stdout.strip()
assert head == LU_SHA, f"vendor at {head}, expected {LU_SHA}"
D = vend / "data"
role_list = json.loads((D/"roles"/"role_list.json").read_text("utf-8"))
questions = [json.loads(l) for l in
(D/"extraction_questions.jsonl").read_text("utf-8").splitlines() if l.strip()]
assert len(role_list) == 275 and len(questions) == 240
QT = {x["id"]: x["question"] for x in questions}
_ro = {}
def role_obj(n):
if n not in _ro:
_ro[n] = json.loads((D/"roles"/"instructions"/f"{n}.json").read_text("utf-8"))
return _ro[n]
pf = ROOT / "design" / "panel.json"
panel = json.loads(pf.read_text("utf-8")); PANEL = panel["question_ids"]
FP = hashlib.sha256(json.dumps([PANEL, panel["half_a_ids"],
panel["half_b_ids"]]).encode()).hexdigest()[:16]
if CFG["EXPECT_PANEL_FP"] and FP != CFG["EXPECT_PANEL_FP"]:
sys.exit(f"PANEL MISMATCH {FP} != {CFG['EXPECT_PANEL_FP']}")
print(f"panel {FP} ok")
roles = (sorted(role_list) if TIER == "A"
else sorted(random.Random(SEED).sample(sorted(role_list), 150)))
WRAP = ("Stay in character for your entire reply. Answer in the first person. "
"Never mention being an AI, a language model, or an assistant.")
def build(role, pi, qid, cond):
ins = role_obj(role)["instruction"][pi]["pos"]
qq = QT[qid]
if role == "default":
# NEVER wrap the default. Lu's five default conditions are identity
# statements ("You are an AI assistant.", "Respond as yourself.", and one
# empty), not characters. Appending a role-adoption instruction here would
# contradict the condition and corrupt mean(default), which is one half of
# the Axis.
ins = ins.replace("{model_name}", "DeepSeek-R1")
return f"{ins}\n\n{qq}" if ins else qq
if not ins: return qq
return f"{ins}\n\n{qq}" if cond == "USER_TRANSLATED_LU" else f"{ins}\n\n{WRAP}\n\n{qq}"
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained(MODEL_ID)
def render(u):
# add_special_tokens=False: the template emits BOS and add_bos_token is True.
return tok.apply_chat_template([{"role": "user", "content": u}],
tokenize=False, add_generation_prompt=True)
def uid(*p): return hashlib.sha256("|".join(map(str, p)).encode()).hexdigest()[:16]
conds = (["USER_TRANSLATED_LU", "USER_EXPLICIT"]
if CFG["CONDITION"] == "BOTH" else [CFG["CONDITION"]])
if MODE == "DEFAULT" and len(conds) > 1:
# The wrapper never applies to default, so both conditions render IDENTICAL
# prompts. uid includes the condition, so running both would silently
# double-weight the default vector. Collapse to one.
conds = conds[:1]
print("DEFAULT: conditions render identically here; using one to avoid "
"double-weighting mean(default)")
units = []
if MODE == "EXTRACTION":
for r in roles:
for i, qid in enumerate(PANEL):
if i % NSH != SHARD: continue # split by QUESTION index
for c in conds:
units.append((uid(r, i % N_PROMPTS, qid, c), r, i % N_PROMPTS, qid, c))
elif MODE == "DEFAULT":
for pi in range(N_PROMPTS): # 5 default conditions
for i, qid in enumerate(PANEL): # x 80 questions = 400
if i % NSH != SHARD: continue
for c in conds:
units.append((uid("default", pi, qid, c), "default", pi, qid, c))
else:
sys.exit(f"queue handles EXTRACTION/DEFAULT, not {MODE}")
pc = collections.Counter(u[2] for u in units)
print(f"{len(units)} units | {len({u[1] for u in units})} roles | "
f"prompt balance {dict(sorted(pc.items()))}")
od = ROOT / MODE.lower(); od.mkdir(parents=True, exist_ok=True)
sf = od / f"rollouts_shard{SHARD}.jsonl"
done = set()
if sf.exists():
for line in sf.read_text("utf-8").splitlines():
try: done.add(json.loads(line)["uid"])
except Exception: pass
todo = [u for u in units if u[0] not in done]
print(f"resume: {len(done)} done, {len(todo)} to go")
if todo:
print(f" ~{len(todo)/7000:.2f}-{len(todo)/5500:.2f} h at 5,500-7,000/h")
if todo:
from vllm import LLM, SamplingParams
llm = LLM(model=MODEL_ID, dtype=DTYPE, gpu_memory_utilization=UTIL,
max_model_len=MAXLEN, seed=SEED)
sp = SamplingParams(temperature=CFG["TEMP"], top_p=CFG["TOP_P"],
max_tokens=MAX_TOK, seed=SEED)
t0 = time.time()
for s in range(0, len(todo), CFG["CKPT_EVERY"]):
chunk = todo[s:s+CFG["CKPT_EVERY"]]
prompts = [render(build(r, pi, qid, c)) for _, r, pi, qid, c in chunk]
outs = llm.generate(prompts, sp)
nc = nt = 0
with sf.open("a", encoding="utf-8") as fh:
for (u, r, pi, qid, c), rp, o in zip(chunk, prompts, outs):
t = o.outputs[0].text
nc += "</think>" in t; nt += o.outputs[0].finish_reason == "length"
fh.write(json.dumps({
"uid": u, "shard": SHARD, "role": r, "prompt_idx": pi,
"question_id": qid, "condition": c, "rendered_prompt": rp,
"completion": t, "n_out_tokens": len(o.outputs[0].token_ids),
"finish_reason": o.outputs[0].finish_reason,
"closed_think": "</think>" in t,
"gen_max_tokens": MAX_TOK, "gen_temperature": CFG["TEMP"],
"gen_top_p": CFG["TOP_P"], "gen_seed": SEED,
}, ensure_ascii=False) + "\n")
n, el = min(s+len(chunk), len(todo)), time.time()-t0
print(f" {n}/{len(todo)} {n/max(el,1)*3600:.0f}/h "
f"eta {(len(todo)-n)/max(n,1)*el/3600:.2f}h "
f"[closed {nc/len(chunk):.1%} trunc {nt/len(chunk):.1%}]", flush=True)
if CFG["HF_PUSH"]:
try:
from huggingface_hub import HfApi
HfApi().upload_file(path_or_fileobj=str(sf), repo_id=CFG["HF_REPO"],
repo_type="dataset",
path_in_repo=f"{CFG['RUN_ID']}/{MODE.lower()}/{sf.name}")
except Exception as e:
print(f" (checkpoint push failed: {type(e).__name__}; disk ok)")
del llm; gc.collect(); torch.cuda.empty_cache()
rows = [json.loads(l) for l in sf.read_text("utf-8").splitlines() if l.strip()]
N = max(len(rows), 1)
cfgs = collections.Counter((r.get("gen_max_tokens"), r.get("gen_temperature")) for r in rows)
per_role = collections.Counter(r["role"] for r in rows)
rep = {"mode": MODE, "shard": SHARD, "n_shards": NSH, "tier": TIER,
"condition": CFG["CONDITION"], "conditions_used": conds,
"n_outputs": len(rows), "n_expected": len(units),
"complete": len(rows) >= len(units), "panel_fingerprint": FP,
"lu_commit": LU_SHA, "model": MODEL_ID, "device_index": d["index"],
"dtype": DTYPE, "max_model_len": MAXLEN, "gpu_memory_utilization": UTIL,
"generation_configs_in_data": {str(k): v for k, v in cfgs.items()},
"config_homogeneous": len(cfgs) == 1,
"closed_think_rate": round(sum(r["closed_think"] for r in rows)/N, 4),
"truncated_rate": round(sum(r["finish_reason"] == "length" for r in rows)/N, 4),
"mean_out_tokens": round(sum(r["n_out_tokens"] for r in rows)/N, 1),
"roles_present": len(per_role),
"min_outputs_per_role": min(per_role.values()) if per_role else 0,
"default_wrapper_applied": False if MODE == "DEFAULT" else None}
rf = od / f"run_report_shard{SHARD}.json"
rf.write_text(json.dumps(rep, indent=2), "utf-8")
print("\n" + json.dumps(rep, indent=2))
if CFG["HF_PUSH"]:
try:
from huggingface_hub import HfApi
api = HfApi(); R, base = CFG["HF_REPO"], CFG["RUN_ID"]
for local, dest in ((sf, f"{MODE.lower()}/{sf.name}"),
(rf, f"{MODE.lower()}/{rf.name}"),
(pf, "design/panel.json")):
api.upload_file(path_or_fileobj=str(local), path_in_repo=f"{base}/{dest}",
repo_id=R, repo_type="dataset")
print(f" pushed {dest}")
api.upload_file(path_or_fileobj=Path(sys.argv[0]).read_text("utf-8").encode(),
path_in_repo=f"{base}/code/worker_{MODE.lower()}_shard{SHARD}.py",
repo_id=R, repo_type="dataset")
api.upload_file(path_or_fileobj=json.dumps(CFG, indent=2).encode(),
path_in_repo=f"{base}/code/config_{MODE.lower()}_shard{SHARD}.json",
repo_id=R, repo_type="dataset")
print(" pushed code + config")
except Exception as e:
print(f"WARNING HF push failed: {type(e).__name__}: {e}")
print(f"Outputs safe on disk at {sf}")
print(f"JOBDONE {MODE} shard {SHARD}: {rep['n_outputs']}/{rep['n_expected']} "
f"complete={rep['complete']}")
|